v2.17.0 (#822)

Co-authored-by: Clément Sirieix <clement.sirieix@mistral.ai>
Co-authored-by: Guillaume LE GOFF <guillaume.lgf@gmail.com>
Co-authored-by: Hdandria <henri.dandria@mistral.ai>
Co-authored-by: Ivana Dunisijevic <ivana.dunisijevic@mistral.ai>
Co-authored-by: Jean Burellier <sheplu@users.noreply.github.com>
Co-authored-by: Mathias Gesbert <mathias.gesbert@mistral.ai>
Co-authored-by: Mert Unsal <mert.unsal@mistral.ai>
Co-authored-by: Michel Thomazo <51709227+michelTho@users.noreply.github.com>
Co-authored-by: Paul VEZIA <166131032+le-codeur-rapide@users.noreply.github.com>
Co-authored-by: Pierre Rossinès <pierre.rossines@mistral.ai>
Co-authored-by: Val <102326092+vdeva@users.noreply.github.com>
Co-authored-by: Vincent G <10739306+VinceOPS@users.noreply.github.com>
Co-authored-by: renovate-mistral[bot] <253709520+renovate-mistral[bot]@users.noreply.github.com>
Co-authored-by: Mistral Vibe <vibe@mistral.ai>

CHANGELOG.md

@@ -5,6 +5,30 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [2.17.0] - 2026-06-19
+
+### Added
+
+- `/mcp login`, `/mcp logout`, and `/mcp status` commands to authenticate OAuth-backed MCP servers from the TUI
+- `vibe --check-upgrade` to force an immediate update check and exit
+- `--yolo` as an alias for `--auto-approve`
+- ACP now accepts inline image content blocks
+
+### Changed
+
+- API keys are now stored in the OS keyring instead of plain text
+- Edit diff view now shows all replaced occurrences instead of just the first
+- Completion popup now uses a two-column layout
+- Chat messages now have right padding so text no longer collapses into the scrollbar
+- Faster CLI shutdown by deferring resource cleanup on exit
+
+### Fixed
+
+- Skill autocomplete popup now dismisses after Tab completion
+- Stdio MCP connections now persist across tool calls
+- Retryable 5xx responses from the Mistral backend are now retried instead of failing
+
+
 ## [2.16.1] - 2026-06-16
 
 ### Added

README.md

@@ -137,8 +137,8 @@ custom agent file in `~/.vibe/agents/` or the project's `.vibe/agents/`
 directory. Subagents such as `explore` are not accepted.
 
 > Note: `default_agent` applies in both interactive and programmatic
-> (`-p` / `--prompt`) sessions. Pass `--auto-approve` when a run should
-> approve all tool calls without prompting.
+> (`-p` / `--prompt`) sessions. Pass `--auto-approve` or `--yolo` when
+> a run should approve all tool calls without prompting.
 
 ### Subagents and Task Delegation
 
@@ -262,8 +262,8 @@ vibe --prompt "Refactor the main function in cli/main.py to be more modular."
 ```
 
 By default, it uses your configured `default_agent` (`default` unless changed).
-To approve all tool calls without prompting, pass `--auto-approve` (also
-available for interactive sessions):
+To approve all tool calls without prompting, pass `--auto-approve` or `--yolo`
+(also available for interactive sessions):
 
 ```bash
 vibe --prompt "Refactor the main function in cli/main.py to be more modular." --auto-approve
@@ -277,7 +277,7 @@ When using `--prompt`, you can specify additional options:
 - **`--max-price DOLLARS`**: Set a maximum cost limit in dollars. The session will be interrupted if the cost exceeds this limit.
 - **`--max-tokens N`**: Set a maximum cumulative LLM token budget for the session, counting both prompt and completion tokens. The session will be interrupted if usage exceeds this limit.
 - **`--agent NAME`**: Select the agent profile for this run.
-- **`--auto-approve`**: Shortcut for `--agent auto-approve`. Approves all tool calls without prompting, including in interactive sessions.
+- **`--auto-approve`, `--yolo`**: Shortcut for `--agent auto-approve`. Approves all tool calls without prompting, including in interactive sessions.
 - **`--enabled-tools TOOL`**: Enable specific tools. In programmatic mode, this disables all other tools. Can be specified multiple times. Supports exact names, glob patterns (e.g., `bash*`), or regex with `re:` prefix (e.g., `re:^serena_.*$`).
 - **`--output FORMAT`**: Set the output format. Options:
   - `text` (default): Human-readable text output
@@ -730,6 +730,8 @@ Each path is implicitly trusted (no trust prompt) and contributes its `AGENTS.md
 
 Vibe checks PyPI at most once per day during a session. When a newer version is found, the next launch shows an update prompt before opening the chat, offering to either update immediately (via `uv tool upgrade mistral-vibe` or `brew upgrade mistral-vibe`) or continue with the current version.
 
+Run `vibe --check-upgrade` to check PyPI immediately, prompt to install a newer version if one exists, and exit.
+
 To disable the daily check entirely, add this to your `config.toml`:
 
 ```toml

distribution/zed/extension.toml

@@ -1,7 +1,7 @@
 id = "mistral-vibe"
 name = "Mistral Vibe"
 description = "Mistral's open-source coding assistant"
-version = "2.16.1"
+version = "2.17.0"
 schema_version = 1
 authors = ["Mistral AI"]
 repository = "https://github.com/mistralai/mistral-vibe"
@@ -11,21 +11,21 @@ name = "Mistral Vibe"
 icon = "./icons/mistral_vibe.svg"
 
 [agent_servers.mistral-vibe.targets.darwin-aarch64]
-archive = "https://github.com/mistralai/mistral-vibe/releases/download/v2.16.1/vibe-acp-darwin-aarch64-2.16.1.zip"
+archive = "https://github.com/mistralai/mistral-vibe/releases/download/v2.17.0/vibe-acp-darwin-aarch64-2.17.0.zip"
 cmd = "./vibe-acp"
 
 [agent_servers.mistral-vibe.targets.darwin-x86_64]
-archive = "https://github.com/mistralai/mistral-vibe/releases/download/v2.16.1/vibe-acp-darwin-x86_64-2.16.1.zip"
+archive = "https://github.com/mistralai/mistral-vibe/releases/download/v2.17.0/vibe-acp-darwin-x86_64-2.17.0.zip"
 cmd = "./vibe-acp"
 
 [agent_servers.mistral-vibe.targets.linux-aarch64]
-archive = "https://github.com/mistralai/mistral-vibe/releases/download/v2.16.1/vibe-acp-linux-aarch64-2.16.1.zip"
+archive = "https://github.com/mistralai/mistral-vibe/releases/download/v2.17.0/vibe-acp-linux-aarch64-2.17.0.zip"
 cmd = "./vibe-acp"
 
 [agent_servers.mistral-vibe.targets.linux-x86_64]
-archive = "https://github.com/mistralai/mistral-vibe/releases/download/v2.16.1/vibe-acp-linux-x86_64-2.16.1.zip"
+archive = "https://github.com/mistralai/mistral-vibe/releases/download/v2.17.0/vibe-acp-linux-x86_64-2.17.0.zip"
 cmd = "./vibe-acp"
 
 [agent_servers.mistral-vibe.targets.windows-x86_64]
-archive = "https://github.com/mistralai/mistral-vibe/releases/download/v2.16.1/vibe-acp-windows-x86_64-2.16.1.zip"
+archive = "https://github.com/mistralai/mistral-vibe/releases/download/v2.17.0/vibe-acp-windows-x86_64-2.17.0.zip"
 cmd = "./vibe-acp.exe"

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "mistral-vibe"
-version = "2.16.1"
+version = "2.17.0"
 description = "Minimal CLI coding agent by Mistral"
 readme = "README.md"
 requires-python = ">=3.12"
@@ -38,7 +38,7 @@ dependencies = [
     "charset-normalizer==3.4.7",
     "click==8.3.3 ; sys_platform != 'emscripten'",
     "colorama==0.4.6 ; sys_platform == 'win32'",
-    "cryptography==47.0.0",
+    "cryptography==48.0.1",
     "eval-type-backport==0.3.1",
     "gitdb==4.0.12",
     "gitpython==3.1.47",
@@ -65,10 +65,10 @@ dependencies = [
     "linkify-it-py==2.1.0",
     "markdown-it-py==4.0.0",
     "markdownify==1.2.2",
-    "mcp==1.27.1",
+    "mcp==1.27.2",
     "mdit-py-plugins==0.5.0",
     "mdurl==0.1.2",
-    "mistralai==2.4.4",
+    "mistralai==2.4.9",
     "more-itertools==11.0.2",
     "opentelemetry-api==1.39.1",
     "opentelemetry-exporter-otlp-proto-common==1.39.1",
@@ -88,11 +88,11 @@ dependencies = [
     "pydantic-core==2.46.3",
     "pydantic-settings==2.14.0",
     "pygments==2.20.0",
-    "pyjwt==2.12.1",
+    "pyjwt==2.13.0",
     "pyperclip==1.11.0",
     "python-dateutil==2.9.0.post0",
     "python-dotenv==1.2.2",
-    "python-multipart==0.0.27",
+    "python-multipart==0.0.32",
     "pywin32==311 ; sys_platform == 'win32'",
     "pywin32-ctypes==0.2.3 ; sys_platform == 'win32'",
     "pyyaml==6.0.3",
@@ -116,7 +116,7 @@ dependencies = [
     "typing-extensions==4.15.0",
     "typing-inspection==0.4.2",
     "uc-micro-py==2.0.0",
-    "urllib3==2.6.3",
+    "urllib3==2.7.0",
     "uvicorn==0.46.0 ; sys_platform != 'emscripten'",
     "watchfiles==1.1.1",
     "websockets==16.0",
@@ -147,7 +147,7 @@ vibe-acp = "vibe.acp.entrypoint:main"
 
 
 [tool.uv]
-exclude-newer = "7 days"
+exclude-newer = "2 days"
 
 package = true
 required-version = ">=0.8.0"

tests/acp/test_acp.py

@@ -553,6 +553,7 @@ class TestSessionUpdates:
 
             assert tool_call.params.update.session_update == "tool_call"
             assert tool_call.params.update.kind == "search"
+            assert tool_call.params.update.status == "pending"
             assert tool_call.params.update.title == "Grepping 'auth'"
             assert (
                 tool_call.params.update.raw_input

tests/acp/test_auth_status.py

@@ -4,10 +4,12 @@ import os
 from pathlib import Path
 
 from dotenv import dotenv_values
+import keyring
+from keyring.errors import KeyringError
 import pytest
 
 from vibe.acp.acp_agent_loop import VibeAcpAgentLoop
-from vibe.acp.exceptions import InvalidRequestError
+from vibe.acp.exceptions import InternalError, InvalidRequestError
 from vibe.core.config import (
     DEFAULT_MISTRAL_API_ENV_KEY,
     ProviderConfig,
@@ -18,6 +20,15 @@ from vibe.setup.auth import AuthStateKind
 from vibe.setup.onboarding.context import OnboardingContext
 
 
+@pytest.fixture(autouse=True)
+def disable_keyring(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setattr(keyring, "get_password", lambda service, username: None)
+    monkeypatch.setattr(
+        keyring, "set_password", lambda service, username, password: None
+    )
+    monkeypatch.setattr(keyring, "delete_password", lambda service, username: None)
+
+
 def build_mistral_provider(
     *, api_key_env_var: str = DEFAULT_MISTRAL_API_ENV_KEY
 ) -> ProviderConfig:
@@ -127,7 +138,7 @@ class TestACPAuthStatus:
         }
 
     @pytest.mark.asyncio
-    async def test_returns_combined_source_when_process_env_existed_before_dotenv(
+    async def test_returns_process_env_when_process_env_existed_before_dotenv(
         self, config_dir: Path, monkeypatch: pytest.MonkeyPatch
     ) -> None:
         monkeypatch.setenv(DEFAULT_MISTRAL_API_ENV_KEY, "process-key")
@@ -138,7 +149,25 @@ class TestACPAuthStatus:
 
         assert response == {
             "authenticated": True,
-            "authState": AuthStateKind.VIBE_HOME_ENV_FILE_OVERRIDES_PROCESS_ENV.value,
+            "authState": AuthStateKind.PROCESS_ENV.value,
+            "signOutAvailable": False,
+        }
+
+    @pytest.mark.asyncio
+    async def test_returns_keyring_when_key_only_exists_in_keyring(
+        self, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        monkeypatch.delenv(DEFAULT_MISTRAL_API_ENV_KEY, raising=False)
+        monkeypatch.setattr(
+            keyring, "get_password", lambda service, username: "keyring-key"
+        )
+        acp_agent_loop = build_acp_agent_loop(build_mistral_provider())
+
+        response = await acp_agent_loop.ext_method("auth/status", {})
+
+        assert response == {
+            "authenticated": True,
+            "authState": AuthStateKind.OS_KEYRING.value,
             "signOutAvailable": True,
         }
 
@@ -213,23 +242,58 @@ class TestACPAuthSignOut:
         assert os.environ[DEFAULT_MISTRAL_API_ENV_KEY] == "process-key"
 
     @pytest.mark.asyncio
-    async def test_removes_dotenv_key_and_restores_process_env_key(
+    async def test_refuses_dotenv_key_when_process_env_key_exists(
         self, config_dir: Path, monkeypatch: pytest.MonkeyPatch
     ) -> None:
         monkeypatch.setenv(DEFAULT_MISTRAL_API_ENV_KEY, "process-key")
         write_env_file(config_dir, f"{DEFAULT_MISTRAL_API_ENV_KEY}=file-key\n")
         acp_agent_loop = build_acp_agent_loop(build_mistral_provider())
 
+        with pytest.raises(InvalidRequestError, match=AuthStateKind.PROCESS_ENV.value):
+            await acp_agent_loop.ext_method("auth/signOut", {})
+
+        assert (
+            dotenv_values(config_dir / ".env")[DEFAULT_MISTRAL_API_ENV_KEY]
+            == "file-key"
+        )
+        assert os.environ[DEFAULT_MISTRAL_API_ENV_KEY] == "process-key"
+
+    @pytest.mark.asyncio
+    async def test_removes_keyring_key(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        deleted: list[str] = []
+        monkeypatch.delenv(DEFAULT_MISTRAL_API_ENV_KEY, raising=False)
+        monkeypatch.setattr(
+            keyring, "get_password", lambda service, username: "keyring-key"
+        )
+        monkeypatch.setattr(
+            keyring,
+            "delete_password",
+            lambda service, username: deleted.append(username),
+        )
+        acp_agent_loop = build_acp_agent_loop(build_mistral_provider())
+
         response = await acp_agent_loop.ext_method("auth/signOut", {})
 
         assert response == {}
-        assert DEFAULT_MISTRAL_API_ENV_KEY not in dotenv_values(config_dir / ".env")
-        assert os.environ[DEFAULT_MISTRAL_API_ENV_KEY] == "process-key"
-        assert await acp_agent_loop.ext_method("auth/status", {}) == {
-            "authenticated": True,
-            "authState": AuthStateKind.PROCESS_ENV.value,
-            "signOutAvailable": False,
-        }
+        assert deleted == [DEFAULT_MISTRAL_API_ENV_KEY]
+
+    @pytest.mark.asyncio
+    async def test_surfaces_internal_error_when_keyring_delete_fails(
+        self, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        monkeypatch.delenv(DEFAULT_MISTRAL_API_ENV_KEY, raising=False)
+        monkeypatch.setattr(
+            keyring, "get_password", lambda service, username: "keyring-key"
+        )
+
+        def _failed(service: str, username: str) -> None:
+            raise KeyringError("delete failed")
+
+        monkeypatch.setattr(keyring, "delete_password", _failed)
+        acp_agent_loop = build_acp_agent_loop(build_mistral_provider())
+
+        with pytest.raises(InternalError, match="Failed to sign out"):
+            await acp_agent_loop.ext_method("auth/signOut", {})
 
     @pytest.mark.asyncio
     async def test_refuses_unsupported_provider_key(

tests/acp/test_compact_session_updates.py

@@ -101,6 +101,7 @@ class TestCompactEventHandling:
             cwd=str(Path.cwd()), mcp_servers=[]
         )
         session = acp_agent_loop.sessions[session_response.session_id]
+        await session.agent_loop.wait_until_ready()
         # Need >1 message so /compact does not early-return.
         session.agent_loop.messages.append(LLMMessage(role=Role.user, content="hello"))
 

tests/acp/test_image_blocks.py

@@ -0,0 +1,113 @@
+from __future__ import annotations
+
+import base64
+import hashlib
+from pathlib import Path
+
+from acp.helpers import ImageContentBlock, TextContentBlock
+import pytest
+
+from vibe.acp.exceptions import INVALID_IMAGE_ATTACHMENT, InvalidImageAttachmentError
+from vibe.acp.image_blocks import extract_image_attachments
+from vibe.core.types import (
+    MAX_IMAGE_BYTES,
+    MAX_IMAGES_PER_MESSAGE,
+    FileImageSource,
+    InlineImageSource,
+)
+
+PNG_BYTES = b"\x89PNG\r\n\x1a\n" + b"\x00" * 16
+
+
+def _image_block(uri: str | None = None) -> ImageContentBlock:
+    return ImageContentBlock(
+        type="image",
+        data=base64.b64encode(PNG_BYTES).decode("ascii"),
+        mime_type="image/png",
+        uri=uri,
+    )
+
+
+def test_inlines_bytes_when_session_dir_is_none() -> None:
+    [att] = extract_image_attachments([_image_block()], session_dir=None)
+
+    assert isinstance(att.source, InlineImageSource)
+    assert att.source.data == base64.b64encode(PNG_BYTES).decode("ascii")
+    assert att.mime_type == "image/png"
+
+
+def test_writes_attachment_file_when_session_dir_is_set(tmp_path: Path) -> None:
+    [att] = extract_image_attachments([_image_block()], session_dir=tmp_path)
+
+    digest = hashlib.sha1(PNG_BYTES, usedforsecurity=False).hexdigest()
+    assert isinstance(att.source, FileImageSource)
+    assert att.source.path == (tmp_path / "attachments" / f"{digest}.png").resolve()
+    assert att.source.path.read_bytes() == PNG_BYTES
+
+
+def test_derives_alias_from_uri_basename() -> None:
+    [att] = extract_image_attachments([_image_block(uri="cat.png")], session_dir=None)
+
+    assert att.alias == "cat.png"
+
+
+def test_falls_back_to_default_alias_without_uri() -> None:
+    [att] = extract_image_attachments([_image_block()], session_dir=None)
+
+    assert att.alias == "pasted-image.png"
+
+
+def test_ignores_non_image_blocks() -> None:
+    block = TextContentBlock(type="text", text="hello")
+
+    assert extract_image_attachments([block], session_dir=None) == []
+
+
+def test_rejects_unsupported_mime() -> None:
+    block = ImageContentBlock(
+        type="image",
+        data=base64.b64encode(PNG_BYTES).decode("ascii"),
+        mime_type="image/tiff",
+    )
+
+    with pytest.raises(InvalidImageAttachmentError):
+        extract_image_attachments([block], session_dir=None)
+
+
+def test_image_block_error_is_structured_acp_error() -> None:
+    block = ImageContentBlock(
+        type="image",
+        data=base64.b64encode(PNG_BYTES).decode("ascii"),
+        mime_type="image/tiff",
+    )
+
+    with pytest.raises(InvalidImageAttachmentError) as exc_info:
+        extract_image_attachments([block], session_dir=None)
+
+    assert exc_info.value.code == INVALID_IMAGE_ATTACHMENT
+    assert exc_info.value.data == {"reason": "wrong_type"}
+
+
+def test_rejects_invalid_base64() -> None:
+    block = ImageContentBlock(type="image", data="not base64!!!", mime_type="image/png")
+
+    with pytest.raises(InvalidImageAttachmentError):
+        extract_image_attachments([block], session_dir=None)
+
+
+def test_rejects_too_many_images() -> None:
+    blocks = [_image_block() for _ in range(MAX_IMAGES_PER_MESSAGE + 1)]
+
+    with pytest.raises(InvalidImageAttachmentError):
+        extract_image_attachments(blocks, session_dir=None)
+
+
+def test_rejects_oversized_image() -> None:
+    block = ImageContentBlock(
+        type="image",
+        data=base64.b64encode(b"x" * (MAX_IMAGE_BYTES + 1)).decode("ascii"),
+        mime_type="image/png",
+    )
+
+    with pytest.raises(InvalidImageAttachmentError):
+        extract_image_attachments([block], session_dir=None)

tests/acp/test_initialize.py

@@ -63,7 +63,7 @@ class TestACPInitialize:
         assert response.agent_capabilities == AgentCapabilities(
             load_session=True,
             prompt_capabilities=PromptCapabilities(
-                audio=False, embedded_context=True, image=False
+                audio=False, embedded_context=True, image=True
             ),
             session_capabilities=SessionCapabilities(
                 close=SessionCloseCapabilities(),
@@ -72,7 +72,7 @@ class TestACPInitialize:
             ),
         )
         assert response.agent_info == Implementation(
-            name="@mistralai/mistral-vibe", title="Mistral Vibe", version="2.16.1"
+            name="@mistralai/mistral-vibe", title="Mistral Vibe", version="2.17.0"
         )
 
         assert response.auth_methods is not None
@@ -163,7 +163,7 @@ class TestACPInitialize:
         assert response.agent_capabilities == AgentCapabilities(
             load_session=True,
             prompt_capabilities=PromptCapabilities(
-                audio=False, embedded_context=True, image=False
+                audio=False, embedded_context=True, image=True
             ),
             session_capabilities=SessionCapabilities(
                 close=SessionCloseCapabilities(),
@@ -172,7 +172,7 @@ class TestACPInitialize:
             ),
         )
         assert response.agent_info == Implementation(
-            name="@mistralai/mistral-vibe", title="Mistral Vibe", version="2.16.1"
+            name="@mistralai/mistral-vibe", title="Mistral Vibe", version="2.17.0"
         )
 
         assert response.auth_methods is not None

tests/acp/test_load_session.py

@@ -1,13 +1,11 @@
 from __future__ import annotations
 
 from pathlib import Path
-from unittest.mock import AsyncMock
 
 from acp import RequestError
 from acp.schema import (
     AgentMessageChunk,
     AgentThoughtChunk,
-    ClientCapabilities,
     ToolCallProgress,
     ToolCallStart,
     UserMessageChunk,
@@ -17,7 +15,8 @@ import pytest
 from tests.conftest import build_test_vibe_config
 from tests.stubs.fake_backend import FakeBackend
 from tests.stubs.fake_client import FakeClient
-from vibe.acp.acp_agent_loop import WORKSPACE_TRUST_CAPABILITY, VibeAcpAgentLoop
+from vibe.acp.acp_agent_loop import VibeAcpAgentLoop
+from vibe.acp.user_display_content import USER_DISPLAY_CONTENT_META_KEY
 from vibe.core.agent_loop import AgentLoop
 from vibe.core.agents.models import BuiltinAgentName
 from vibe.core.config import ModelConfig, SessionLoggingConfig
@@ -126,35 +125,42 @@ class TestLoadSession:
         assert response.config_options[2].current_value == "off"
 
     @pytest.mark.asyncio
-    async def test_load_session_resolves_workspace_trust_before_loading_config(
+    async def test_load_session_returns_trust_details_without_loading_project_docs(
         self,
         acp_agent_with_session_config: tuple[VibeAcpAgentLoop, FakeClient],
         temp_session_dir: Path,
         create_test_session,
         tmp_working_directory: Path,
-        monkeypatch: pytest.MonkeyPatch,
     ) -> None:
-        acp_agent, client = acp_agent_with_session_config
-        acp_agent.client_capabilities = ClientCapabilities(
-            field_meta={WORKSPACE_TRUST_CAPABILITY: True}
-        )
-        request_trust = AsyncMock(return_value={"decision": "trust_cwd"})
-        monkeypatch.setattr(client, "ext_method", request_trust)
+        acp_agent, _client = acp_agent_with_session_config
         (tmp_working_directory / "AGENTS.md").write_text(
             "Loaded session project instructions", encoding="utf-8"
         )
         session_id = "test-sess-trust"
         create_test_session(temp_session_dir, session_id, str(tmp_working_directory))
 
-        await acp_agent.load_session(
+        response = await acp_agent.load_session(
             cwd=str(tmp_working_directory), mcp_servers=[], session_id=session_id
         )
 
-        request_trust.assert_awaited_once()
-        assert trusted_folders_manager.is_trusted(tmp_working_directory) is True
+        assert response is not None
+        payload = response.model_dump(mode="json", by_alias=True)
+        assert payload.get("_meta") == {
+            "workspace_trust": {
+                "status": "untrusted",
+                "details": {
+                    "cwd": str(tmp_working_directory.resolve()),
+                    "repoRoot": None,
+                    "ignoredFiles": ["AGENTS.md"],
+                    "availableDecisions": ["trust_cwd", "trust_session", "decline"],
+                },
+            }
+        }
+        assert trusted_folders_manager.is_trusted(tmp_working_directory) is None
+        await acp_agent.sessions[session_id].agent_loop.wait_until_ready()
         system_prompt = acp_agent.sessions[session_id].agent_loop.messages[0].content
         assert system_prompt is not None
-        assert "Loaded session project instructions" in system_prompt
+        assert "Loaded session project instructions" not in system_prompt
 
     @pytest.mark.asyncio
     async def test_load_session_registers_session_with_original_id(
@@ -223,6 +229,51 @@ class TestLoadSession:
         ]
         assert len(user_updates) == 1
         assert user_updates[0].update.content.text == "Hello world"
+        assert user_updates[0].update.field_meta is None
+
+    @pytest.mark.asyncio
+    async def test_load_session_replays_user_display_content(
+        self,
+        acp_agent_with_session_config: tuple[VibeAcpAgentLoop, FakeClient],
+        temp_session_dir: Path,
+        create_test_session,
+    ) -> None:
+        acp_agent, client = acp_agent_with_session_config
+
+        session_id = "replay-dsp-123456"
+        cwd = str(Path.cwd())
+        user_display_content = {
+            "version": "1.0.0",
+            "host": "mistral-vscode",
+            "content": [
+                {"type": "text", "text": "Look at "},
+                {
+                    "type": "workspace_mention",
+                    "kind": "file",
+                    "uri": "file:///repo/src/app.ts",
+                    "name": "app.ts",
+                },
+            ],
+        }
+        messages = [
+            {
+                "role": "user",
+                "content": "Look at app.ts",
+                "user_display_content": user_display_content,
+            }
+        ]
+        create_test_session(temp_session_dir, session_id, cwd, messages=messages)
+
+        await acp_agent.load_session(cwd=cwd, mcp_servers=[], session_id=session_id)
+
+        user_updates = [
+            u for u in client._session_updates if isinstance(u.update, UserMessageChunk)
+        ]
+        assert len(user_updates) == 1
+        assert user_updates[0].update.content.text == "Look at app.ts"
+        assert user_updates[0].update.field_meta == {
+            USER_DISPLAY_CONTENT_META_KEY: user_display_content
+        }
 
     @pytest.mark.asyncio
     async def test_load_session_replays_assistant_messages(
@@ -278,7 +329,12 @@ class TestLoadSession:
                     }
                 ],
             },
-            {"role": "tool", "tool_call_id": "call_123", "content": "file contents"},
+            {
+                "role": "tool",
+                "tool_call_id": "call_123",
+                "name": "read",
+                "content": "file contents",
+            },
         ]
         create_test_session(temp_session_dir, session_id, cwd, messages=messages)
 
@@ -288,15 +344,58 @@ class TestLoadSession:
             u for u in client._session_updates if isinstance(u.update, ToolCallStart)
         ]
         assert len(tool_call_starts) == 1
-        assert tool_call_starts[0].update.title == "read"
-        assert tool_call_starts[0].update.tool_call_id == "call_123"
+        start = tool_call_starts[0].update
+        assert start.tool_call_id == "call_123"
+        assert start.kind == "read"
+        # The host drops created events without a status; replayed calls are
+        # historical, so they must carry one.
+        assert start.status == "completed"
+        assert start.field_meta is not None
+        assert start.field_meta["tool_name"] == "read"
 
         tool_results = [
             u for u in client._session_updates if isinstance(u.update, ToolCallProgress)
         ]
         assert len(tool_results) == 1
-        assert tool_results[0].update.tool_call_id == "call_123"
-        assert tool_results[0].update.status == "completed"
+        result = tool_results[0].update
+        assert result.tool_call_id == "call_123"
+        assert result.status == "completed"
+        # The host drops tool_call_update events without a kind.
+        assert result.kind == "read"
+        assert result.field_meta is not None
+        assert result.field_meta["tool_name"] == "read"
+
+    @pytest.mark.asyncio
+    async def test_load_session_skips_result_whose_call_was_not_replayed(
+        self,
+        acp_agent_with_session_config: tuple[VibeAcpAgentLoop, FakeClient],
+        temp_session_dir: Path,
+        create_test_session,
+    ) -> None:
+        acp_agent, client = acp_agent_with_session_config
+
+        session_id = "replay-orphan-12345"
+        cwd = str(Path.cwd())
+        # A tool result whose call was not replayed (here no matching tool call
+        # at all, mirroring a hidden tool whose call replay returns None).
+        # Emitting it would orphan a tool_call_update with no preceding call.
+        messages = [
+            {"role": "user", "content": "Do something"},
+            {
+                "role": "tool",
+                "tool_call_id": "call_orphan",
+                "name": "read",
+                "content": "file contents",
+            },
+        ]
+        create_test_session(temp_session_dir, session_id, cwd, messages=messages)
+
+        await acp_agent.load_session(cwd=cwd, mcp_servers=[], session_id=session_id)
+
+        tool_results = [
+            u for u in client._session_updates if isinstance(u.update, ToolCallProgress)
+        ]
+        assert tool_results == []
 
     @pytest.mark.asyncio
     async def test_load_session_replays_reasoning_content(

tests/acp/test_new_session.py

@@ -3,29 +3,32 @@ from __future__ import annotations
 from pathlib import Path
 from unittest.mock import AsyncMock, patch
 
-from acp import RequestError
-from acp.schema import ClientCapabilities
+from acp import NewSessionResponse
 import pytest
 
 from tests.acp.conftest import _create_acp_agent
 from tests.conftest import build_test_vibe_config
-from vibe.acp.acp_agent_loop import WORKSPACE_TRUST_CAPABILITY, VibeAcpAgentLoop
-from vibe.acp.exceptions import InvalidRequestError
+from vibe.acp.acp_agent_loop import VibeAcpAgentLoop
 from vibe.core.agent_loop import AgentLoop
 from vibe.core.agents.models import BuiltinAgentName
 from vibe.core.config import ModelConfig
+from vibe.core.feedback import _CACHE_SECTION, _LAST_SHOWN_KEY, record_feedback_asked
 from vibe.core.trusted_folders import trusted_folders_manager
 
 
-def _system_prompt(acp_agent_loop: VibeAcpAgentLoop, session_id: str) -> str:
+async def _system_prompt(acp_agent_loop: VibeAcpAgentLoop, session_id: str) -> str:
     session = acp_agent_loop.sessions[session_id]
+    await session.agent_loop.wait_until_ready()
     return session.agent_loop.messages[0].content or ""
 
 
-def _enable_workspace_trust(acp_agent_loop: VibeAcpAgentLoop) -> None:
-    acp_agent_loop.client_capabilities = ClientCapabilities(
-        field_meta={WORKSPACE_TRUST_CAPABILITY: True}
-    )
+def _workspace_trust_meta(session_response: NewSessionResponse) -> dict:
+    payload = session_response.model_dump(mode="json", by_alias=True)
+    meta = payload.get("_meta")
+    assert isinstance(meta, dict)
+    workspace_trust = meta.get("workspace_trust")
+    assert isinstance(workspace_trust, dict)
+    return workspace_trust
 
 
 @pytest.fixture
@@ -153,7 +156,27 @@ class TestACPNewSession:
         assert len(thinking_config.options) == 5
 
     @pytest.mark.asyncio
-    async def test_new_session_loads_root_agents_md_from_workspace_cwd(
+    async def test_new_session_uses_file_backed_feedback_cache(
+        self, acp_agent_loop: VibeAcpAgentLoop, config_dir: Path
+    ) -> None:
+        session_response = await acp_agent_loop.new_session(
+            cwd=str(Path.cwd()), mcp_servers=[]
+        )
+        acp_session = acp_agent_loop.sessions[session_response.session_id]
+
+        record_feedback_asked(acp_session.agent_loop.cache_store)
+
+        cache_path = config_dir / "cache.toml"
+        assert cache_path.exists()
+        assert (
+            acp_session.agent_loop.cache_store.read_section(_CACHE_SECTION)[
+                _LAST_SHOWN_KEY
+            ]
+            > 0
+        )
+
+    @pytest.mark.asyncio
+    async def test_new_session_returns_actionable_trust_details_without_prompting(
         self,
         acp_agent_loop: VibeAcpAgentLoop,
         tmp_working_directory: Path,
@@ -162,7 +185,6 @@ class TestACPNewSession:
         (tmp_working_directory / "AGENTS.md").write_text(
             "Root project instructions", encoding="utf-8"
         )
-        _enable_workspace_trust(acp_agent_loop)
         request_trust = AsyncMock(return_value={"decision": "trust_cwd"})
         monkeypatch.setattr(acp_agent_loop.client, "ext_method", request_trust)
 
@@ -171,26 +193,24 @@ class TestACPNewSession:
         )
         assert session_response.session_id is not None
 
-        request_trust.assert_awaited_once()
-        await_args = request_trust.await_args
-        assert await_args is not None
-        method, params = await_args.args
-        assert method == "trust/request"
-        assert params["cwd"] == str(tmp_working_directory.resolve())
-        assert params["detectedFiles"] == ["AGENTS.md"]
-        assert params["repoDetectedFiles"] == []
-        assert params["availableDecisions"] == ["trust_cwd", "trust_session", "decline"]
-        assert trusted_folders_manager.is_trusted(tmp_working_directory) is True
-        assert "Root project instructions" in _system_prompt(
+        request_trust.assert_not_awaited()
+        assert _workspace_trust_meta(session_response) == {
+            "status": "untrusted",
+            "details": {
+                "cwd": str(tmp_working_directory.resolve()),
+                "repoRoot": None,
+                "ignoredFiles": ["AGENTS.md"],
+                "availableDecisions": ["trust_cwd", "trust_session", "decline"],
+            },
+        }
+        assert trusted_folders_manager.is_trusted(tmp_working_directory) is None
+        assert "Root project instructions" not in await _system_prompt(
             acp_agent_loop, session_response.session_id
         )
 
     @pytest.mark.asyncio
-    async def test_new_session_can_trust_full_repo_from_subdirectory(
-        self,
-        acp_agent_loop: VibeAcpAgentLoop,
-        tmp_path: Path,
-        monkeypatch: pytest.MonkeyPatch,
+    async def test_new_session_returns_repo_trust_details_from_subdirectory(
+        self, acp_agent_loop: VibeAcpAgentLoop, tmp_path: Path
     ) -> None:
         repo = tmp_path / "repo"
         cwd = repo / "src" / "pkg"
@@ -198,79 +218,68 @@ class TestACPNewSession:
         (repo / ".git" / "HEAD").write_text("ref: refs/heads/main\n", encoding="utf-8")
         cwd.mkdir(parents=True)
         (repo / "AGENTS.md").write_text("Repo instructions", encoding="utf-8")
-        _enable_workspace_trust(acp_agent_loop)
-        request_trust = AsyncMock(return_value={"decision": "trust_repo"})
-        monkeypatch.setattr(acp_agent_loop.client, "ext_method", request_trust)
 
         session_response = await acp_agent_loop.new_session(
             cwd=str(cwd), mcp_servers=[]
         )
         assert session_response.session_id is not None
 
-        request_trust.assert_awaited_once()
-        await_args = request_trust.await_args
-        assert await_args is not None
-        method, params = await_args.args
-        assert method == "trust/request"
-        assert params["cwd"] == str(cwd.resolve())
-        assert params["repoRoot"] == str(repo.resolve())
-        assert params["detectedFiles"] == []
-        assert params["repoDetectedFiles"] == ["AGENTS.md"]
-        assert params["availableDecisions"] == [
-            "trust_repo",
-            "trust_cwd",
-            "trust_session",
-            "decline",
-        ]
-        assert trusted_folders_manager.is_trusted(repo) is True
-        assert trusted_folders_manager.is_trusted(cwd) is True
-        assert "Repo instructions" in _system_prompt(
+        assert _workspace_trust_meta(session_response) == {
+            "status": "untrusted",
+            "details": {
+                "cwd": str(cwd.resolve()),
+                "repoRoot": str(repo.resolve()),
+                "ignoredFiles": ["AGENTS.md"],
+                "availableDecisions": [
+                    "trust_repo",
+                    "trust_cwd",
+                    "trust_session",
+                    "decline",
+                ],
+            },
+        }
+        assert trusted_folders_manager.is_trusted(repo) is None
+        assert trusted_folders_manager.is_trusted(cwd) is None
+        assert "Repo instructions" not in await _system_prompt(
             acp_agent_loop, session_response.session_id
         )
 
     @pytest.mark.asyncio
-    async def test_new_session_decline_skips_project_docs(
-        self,
-        acp_agent_loop: VibeAcpAgentLoop,
-        tmp_working_directory: Path,
-        monkeypatch: pytest.MonkeyPatch,
+    async def test_new_session_returns_details_after_explicit_decline(
+        self, acp_agent_loop: VibeAcpAgentLoop, tmp_working_directory: Path
     ) -> None:
         (tmp_working_directory / "AGENTS.md").write_text(
             "Do not load this", encoding="utf-8"
         )
-        _enable_workspace_trust(acp_agent_loop)
-        monkeypatch.setattr(
-            acp_agent_loop.client,
-            "ext_method",
-            AsyncMock(return_value={"decision": "decline"}),
-        )
+        trusted_folders_manager.add_untrusted(tmp_working_directory)
 
         session_response = await acp_agent_loop.new_session(
             cwd=str(tmp_working_directory), mcp_servers=[]
         )
         assert session_response.session_id is not None
 
+        assert _workspace_trust_meta(session_response) == {
+            "status": "untrusted",
+            "details": {
+                "cwd": str(tmp_working_directory.resolve()),
+                "repoRoot": None,
+                "ignoredFiles": ["AGENTS.md"],
+                "availableDecisions": ["trust_cwd", "trust_session", "decline"],
+            },
+        }
         assert trusted_folders_manager.is_trusted(tmp_working_directory) is False
-        assert "Do not load this" not in _system_prompt(
+        assert "Do not load this" not in await _system_prompt(
             acp_agent_loop, session_response.session_id
         )
 
     @pytest.mark.asyncio
     async def test_new_session_session_trust_loads_docs_without_persisting(
-        self,
-        acp_agent_loop: VibeAcpAgentLoop,
-        tmp_working_directory: Path,
-        monkeypatch: pytest.MonkeyPatch,
+        self, acp_agent_loop: VibeAcpAgentLoop, tmp_working_directory: Path
     ) -> None:
         (tmp_working_directory / "AGENTS.md").write_text(
             "Session-only instructions", encoding="utf-8"
         )
-        _enable_workspace_trust(acp_agent_loop)
-        monkeypatch.setattr(
-            acp_agent_loop.client,
-            "ext_method",
-            AsyncMock(return_value={"decision": "trust_session"}),
-        )
+        trusted_folders_manager.trust_for_session(tmp_working_directory)
 
         session_response = await acp_agent_loop.new_session(
             cwd=str(tmp_working_directory), mcp_servers=[]
@@ -278,10 +287,14 @@ class TestACPNewSession:
         assert session_response.session_id is not None
 
         normalized = str(tmp_working_directory.resolve())
+        assert _workspace_trust_meta(session_response) == {
+            "status": "session",
+            "details": None,
+        }
         assert trusted_folders_manager.is_trusted(tmp_working_directory) is True
         assert normalized in trusted_folders_manager._session_trusted
         assert normalized not in trusted_folders_manager._trusted
-        assert "Session-only instructions" in _system_prompt(
+        assert "Session-only instructions" in await _system_prompt(
             acp_agent_loop, session_response.session_id
         )
 
@@ -292,90 +305,45 @@ class TestACPNewSession:
         tmp_working_directory: Path,
         monkeypatch: pytest.MonkeyPatch,
     ) -> None:
-        _enable_workspace_trust(acp_agent_loop)
-        request_trust = AsyncMock(return_value={"decision": "trust_cwd"})
-        monkeypatch.setattr(acp_agent_loop.client, "ext_method", request_trust)
-
-        await acp_agent_loop.new_session(cwd=str(tmp_working_directory), mcp_servers=[])
-
-        request_trust.assert_not_awaited()
-        assert trusted_folders_manager.is_trusted(tmp_working_directory) is None
-
-    @pytest.mark.asyncio
-    async def test_new_session_direct_client_fallback_skips_project_docs(
-        self,
-        acp_agent_loop: VibeAcpAgentLoop,
-        tmp_working_directory: Path,
-        monkeypatch: pytest.MonkeyPatch,
-    ) -> None:
-        (tmp_working_directory / "AGENTS.md").write_text(
-            "Direct client should skip this", encoding="utf-8"
-        )
-        _enable_workspace_trust(acp_agent_loop)
-        monkeypatch.setattr(
-            acp_agent_loop.client,
-            "ext_method",
-            AsyncMock(side_effect=RequestError.method_not_found("trust/request")),
-        )
-
-        session_response = await acp_agent_loop.new_session(
-            cwd=str(tmp_working_directory), mcp_servers=[]
-        )
-        assert session_response.session_id is not None
-
-        assert trusted_folders_manager.is_trusted(tmp_working_directory) is None
-        assert "Direct client should skip this" not in _system_prompt(
-            acp_agent_loop, session_response.session_id
-        )
-
-    @pytest.mark.asyncio
-    async def test_new_session_without_workspace_trust_capability_skips_prompt(
-        self,
-        acp_agent_loop: VibeAcpAgentLoop,
-        tmp_working_directory: Path,
-        monkeypatch: pytest.MonkeyPatch,
-    ) -> None:
-        (tmp_working_directory / "AGENTS.md").write_text(
-            "Unsupported client should skip this", encoding="utf-8"
-        )
         request_trust = AsyncMock(return_value={"decision": "trust_cwd"})
         monkeypatch.setattr(acp_agent_loop.client, "ext_method", request_trust)
 
         session_response = await acp_agent_loop.new_session(
             cwd=str(tmp_working_directory), mcp_servers=[]
         )
-        assert session_response.session_id is not None
 
         request_trust.assert_not_awaited()
-        assert trusted_folders_manager.is_trusted(tmp_working_directory) is None
-        assert "Unsupported client should skip this" not in _system_prompt(
-            acp_agent_loop, session_response.session_id
+        assert _workspace_trust_meta(session_response) == {
+            "status": "untrusted",
+            "details": None,
+        }
+        assert (
+            trusted_folders_manager.trust_status(tmp_working_directory) == "untrusted"
         )
+        assert trusted_folders_manager.is_trusted(tmp_working_directory) is None
 
     @pytest.mark.asyncio
-    async def test_new_session_cancelled_trust_prompt_cancels_session_creation(
-        self,
-        acp_agent_loop: VibeAcpAgentLoop,
-        tmp_working_directory: Path,
-        monkeypatch: pytest.MonkeyPatch,
+    async def test_new_session_trusted_folder_loads_docs_with_no_details(
+        self, acp_agent_loop: VibeAcpAgentLoop, tmp_working_directory: Path
     ) -> None:
         (tmp_working_directory / "AGENTS.md").write_text(
-            "Cancelled prompt", encoding="utf-8"
+            "Trusted folder instructions", encoding="utf-8"
         )
-        _enable_workspace_trust(acp_agent_loop)
-        monkeypatch.setattr(
-            acp_agent_loop.client,
-            "ext_method",
-            AsyncMock(return_value={"decision": "cancelled"}),
+        trusted_folders_manager.add_trusted(tmp_working_directory)
+
+        session_response = await acp_agent_loop.new_session(
+            cwd=str(tmp_working_directory), mcp_servers=[]
         )
+        assert session_response.session_id is not None
 
-        with pytest.raises(InvalidRequestError):
-            await acp_agent_loop.new_session(
-                cwd=str(tmp_working_directory), mcp_servers=[]
-            )
-
-        assert trusted_folders_manager.is_trusted(tmp_working_directory) is None
-        assert acp_agent_loop.sessions == {}
+        assert trusted_folders_manager.is_trusted(tmp_working_directory) is True
+        assert _workspace_trust_meta(session_response) == {
+            "status": "trusted",
+            "details": None,
+        }
+        assert "Trusted folder instructions" in await _system_prompt(
+            acp_agent_loop, session_response.session_id
+        )
 
     @pytest.mark.skip(reason="TODO: Fix this test")
     @pytest.mark.asyncio

tests/acp/test_set_model.py

@@ -235,6 +235,7 @@ class TestACPSetModel:
             (s for s in acp_agent_loop.sessions.values() if s.id == session_id), None
         )
         assert acp_session is not None
+        await acp_session.agent_loop.wait_until_ready()
 
         user_msg = LLMMessage(role=Role.user, content="Hello")
         assistant_msg = LLMMessage(role=Role.assistant, content="Hi there!")

tests/acp/test_tool_call_session_update.py

@@ -40,3 +40,32 @@ class TestToolCallSessionUpdate:
         assert update.tool_call_id == "test_call_123"
         assert update.kind == "read"
         assert update.raw_input is None
+
+    def test_whole_file_read_emits_plain_file_location(self) -> None:
+        event = self._create_event()
+
+        update = tool_call_session_update(event)
+
+        assert isinstance(update, ToolCallStart)
+        assert update.locations is not None
+        location = update.locations[0]
+        assert location.field_meta == {"type": "file"}
+        assert location.line is None
+
+    def test_bounded_read_emits_file_range_location(self) -> None:
+        event = ToolCallEvent(
+            tool_name="read",
+            tool_call_id="test_call_123",
+            args=ReadArgs(file_path="/tmp/test.txt", offset=10, limit=20),
+            tool_class=Read,
+        )
+
+        update = tool_call_session_update(event)
+
+        assert isinstance(update, ToolCallStart)
+        assert update.locations is not None
+        location = update.locations[0]
+        assert location.field_meta is not None
+        assert location.field_meta["type"] == "file_range"
+        assert location.field_meta["offset"] == 10
+        assert location.field_meta["limit"] == 20

tests/acp/test_tool_field_meta.py

@@ -112,14 +112,25 @@ class TestReadFieldMeta:
         assert loc.field_meta == {"type": "file_range", "offset": 10, "limit": 50}
         assert update.field_meta == {"tool_name": "read"}
 
-    def test_call_defaults_offset_none_limit_default(self) -> None:
+    def test_whole_file_read_emits_plain_file_location(self) -> None:
         event = _call_event("read", Read, ReadArgs(file_path="/tmp/f.txt"))
         update = tool_call_session_update(event)
 
         assert isinstance(update, ToolCallStart)
         assert update.locations is not None
         loc = update.locations[0]
-        assert loc.field_meta == {"type": "file_range", "offset": None, "limit": 2000}
+        assert loc.field_meta == {"type": "file"}
+        assert loc.line is None
+
+    def test_read_from_offset_to_end_emits_file_location_with_line(self) -> None:
+        event = _call_event("read", Read, ReadArgs(file_path="/tmp/f.txt", offset=42))
+        update = tool_call_session_update(event)
+
+        assert isinstance(update, ToolCallStart)
+        assert update.locations is not None
+        loc = update.locations[0]
+        assert loc.field_meta == {"type": "file"}
+        assert loc.line == 42
 
     def test_result_location_has_start_line_and_num_lines(self) -> None:
         result = ReadResult(
@@ -128,6 +139,7 @@ class TestReadFieldMeta:
             num_lines=3,
             start_line=10,
             total_lines=20,
+            requested_limit=50,
         )
         event = _result_event("read", Read, result)
         update = tool_result_session_update(event)
@@ -137,6 +149,77 @@ class TestReadFieldMeta:
         loc = update.locations[0]
         assert loc.field_meta == {"type": "file_range", "offset": 10, "limit": 3}
 
+    def test_whole_file_result_emits_plain_file_location(self) -> None:
+        result = ReadResult(
+            file_path="/tmp/f.txt",
+            content="     1→line1\n     2→line2",
+            num_lines=2,
+            start_line=1,
+            total_lines=2,
+        )
+        event = _result_event("read", Read, result)
+        update = tool_result_session_update(event)
+
+        assert isinstance(update, ToolCallProgress)
+        assert update.locations is not None
+        loc = update.locations[0]
+        assert loc.field_meta == {"type": "file"}
+        assert loc.line is None
+
+    def test_truncated_default_limit_result_emits_range(self) -> None:
+        # No limit given, but the file exceeded the default limit: the read was
+        # partial, so the chip must show the range, not imply the whole file.
+        result = ReadResult(
+            file_path="/tmp/f.txt",
+            content="     1→line1",
+            num_lines=2000,
+            start_line=1,
+            total_lines=None,
+            was_truncated=True,
+        )
+        event = _result_event("read", Read, result)
+        update = tool_result_session_update(event)
+
+        assert isinstance(update, ToolCallProgress)
+        assert update.locations is not None
+        loc = update.locations[0]
+        assert loc.field_meta == {"type": "file_range", "offset": 1, "limit": 2000}
+
+    def test_offset_to_end_result_emits_file_location_with_line(self) -> None:
+        result = ReadResult(
+            file_path="/tmp/f.txt",
+            content="    42→line42",
+            num_lines=1,
+            start_line=42,
+            total_lines=42,
+            requested_offset=42,
+        )
+        event = _result_event("read", Read, result)
+        update = tool_result_session_update(event)
+
+        assert isinstance(update, ToolCallProgress)
+        assert update.locations is not None
+        loc = update.locations[0]
+        assert loc.field_meta == {"type": "file"}
+        assert loc.line == 42
+
+    def test_bounded_result_clamps_limit_to_lines_read(self) -> None:
+        result = ReadResult(
+            file_path="/tmp/f.txt",
+            content="     1→line1",
+            num_lines=1,
+            start_line=1,
+            total_lines=1,
+            requested_limit=100,
+        )
+        event = _result_event("read", Read, result)
+        update = tool_result_session_update(event)
+
+        assert isinstance(update, ToolCallProgress)
+        assert update.locations is not None
+        loc = update.locations[0]
+        assert loc.field_meta == {"type": "file_range", "offset": 1, "limit": 1}
+
 
 class TestWebSearchFieldMeta:
     def test_call_meta_contains_query(self) -> None:

tests/acp/test_user_display_content.py

@@ -0,0 +1,144 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any
+
+from acp.schema import TextContentBlock
+from pydantic import ValidationError
+import pytest
+
+from tests.stubs.fake_backend import FakeBackend
+from tests.stubs.fake_client import FakeClient
+from vibe.acp.acp_agent_loop import VibeAcpAgentLoop
+from vibe.acp.exceptions import InvalidRequestError
+from vibe.acp.user_display_content import (
+    USER_DISPLAY_CONTENT_META_KEY,
+    parse_user_display_content_metadata,
+)
+from vibe.core.session.session_loader import SessionLoader
+from vibe.core.types import Role, UserDisplayContentMetadata
+
+
+def _metadata_payload() -> dict[str, object]:
+    return {
+        "version": "1.0.0",
+        "host": "mistral-vscode",
+        "content": [
+            {"type": "text", "text": "Look at "},
+            {
+                "type": "workspace_mention",
+                "kind": "file",
+                "uri": "file:///repo/src/app.ts",
+                "name": "app.ts",
+            },
+        ],
+    }
+
+
+def _metadata_kwargs(value: object) -> dict[str, Any]:
+    return {USER_DISPLAY_CONTENT_META_KEY: value}
+
+
+def test_user_display_content_meta_key_is_snake_case() -> None:
+    assert USER_DISPLAY_CONTENT_META_KEY == "user_display_content"
+
+
+def test_parse_returns_none_when_metadata_is_missing() -> None:
+    assert parse_user_display_content_metadata(None) is None
+
+
+def test_parse_validates_present_metadata() -> None:
+    metadata = parse_user_display_content_metadata({
+        "version": "1.0.0",
+        "host": "mistral-vscode",
+        "content": [],
+    })
+
+    assert metadata == UserDisplayContentMetadata(
+        version="1.0.0", host="mistral-vscode", content=[]
+    )
+
+
+def test_parse_rejects_invalid_metadata() -> None:
+    with pytest.raises(ValidationError):
+        parse_user_display_content_metadata({
+            "version": 2,
+            "host": "mistral-vscode",
+            "content": [],
+        })
+
+
+@pytest.mark.asyncio
+async def test_prompt_attaches_user_display_content_to_user_message(
+    acp_agent_loop: VibeAcpAgentLoop, backend: FakeBackend
+) -> None:
+    session_response = await acp_agent_loop.new_session(
+        cwd=str(Path.cwd()), mcp_servers=[]
+    )
+    payload = _metadata_payload()
+
+    response = await acp_agent_loop.prompt(
+        prompt=[TextContentBlock(type="text", text="Look at app.ts")],
+        session_id=session_response.session_id,
+        **_metadata_kwargs(payload),
+    )
+
+    assert response.stop_reason == "end_turn"
+    user_message = next(
+        (msg for msg in backend.requests_messages[0] if msg.role == Role.user), None
+    )
+    assert user_message is not None
+    assert user_message.content == "Look at app.ts"
+    assert (
+        user_message.user_display_content
+        == UserDisplayContentMetadata.model_validate(payload)
+    )
+
+
+@pytest.mark.asyncio
+async def test_prompt_rejects_invalid_user_display_content(
+    acp_agent_loop: VibeAcpAgentLoop,
+) -> None:
+    session_response = await acp_agent_loop.new_session(
+        cwd=str(Path.cwd()), mcp_servers=[]
+    )
+
+    with pytest.raises(InvalidRequestError, match="Invalid user display content"):
+        await acp_agent_loop.prompt(
+            prompt=[TextContentBlock(type="text", text="Look at app.ts")],
+            session_id=session_response.session_id,
+            **_metadata_kwargs({"version": 2, "host": "mistral-vscode", "content": []}),
+        )
+
+
+@pytest.mark.asyncio
+async def test_prompt_persists_user_display_content(
+    acp_agent_with_session_config: tuple[VibeAcpAgentLoop, FakeClient],
+    temp_session_dir: Path,
+) -> None:
+    acp_agent, _client = acp_agent_with_session_config
+    session_response = await acp_agent.new_session(cwd=str(Path.cwd()), mcp_servers=[])
+    payload = _metadata_payload()
+
+    await acp_agent.prompt(
+        prompt=[TextContentBlock(type="text", text="Look at app.ts")],
+        session_id=session_response.session_id,
+        **_metadata_kwargs(payload),
+    )
+
+    session_dir = next(temp_session_dir.glob("session_*"))
+    messages_file = session_dir / "messages.jsonl"
+    messages = [
+        json.loads(line)
+        for line in messages_file.read_text(encoding="utf-8").splitlines()
+    ]
+    user_message = next(msg for msg in messages if msg["role"] == "user")
+
+    assert user_message["user_display_content"] == payload
+
+    loaded_messages, _metadata = SessionLoader.load_session(session_dir)
+    loaded_user_message = next(msg for msg in loaded_messages if msg.role == Role.user)
+    assert loaded_user_message.user_display_content == (
+        UserDisplayContentMetadata.model_validate(payload)
+    )

tests/acp/test_utils.py

@@ -1,14 +1,32 @@
 from __future__ import annotations
 
+from acp.schema import ToolCallStart
+
+from vibe.acp.tools.builtins.task import Task as AcpTask
+from vibe.acp.tools.builtins.todo import Todo as AcpTodo, TodoArgs
+from vibe.acp.user_display_content import USER_DISPLAY_CONTENT_META_KEY
 from vibe.acp.utils import (
     TOOL_OPTIONS,
     ToolOption,
     build_permission_options,
+    create_tool_call_replay,
+    create_tool_result_replay,
+    create_user_message_replay,
     get_proxy_help_text,
+    tool_call_replay_update,
 )
+from vibe.core.llm.format import ResolvedToolCall
 from vibe.core.paths import GLOBAL_ENV_FILE
 from vibe.core.proxy_setup import SUPPORTED_PROXY_VARS
+from vibe.core.tools.builtins.task import TaskArgs
 from vibe.core.tools.permissions import PermissionScope, RequiredPermission
+from vibe.core.types import (
+    FunctionCall,
+    LLMMessage,
+    Role,
+    ToolCall,
+    UserDisplayContentMetadata,
+)
 
 
 def _write_env_file(content: str) -> None:
@@ -124,3 +142,126 @@ class TestBuildPermissionOptions:
         assert reject_once.name == "Deny"
         assert allow_once.field_meta is None
         assert reject_once.field_meta is None
+
+
+class TestCreateUserMessageReplay:
+    def test_replays_plain_text_without_field_meta(self) -> None:
+        replay = create_user_message_replay(
+            LLMMessage(role=Role.user, content="Hello", message_id="msg-1")
+        )
+
+        assert replay.content.text == "Hello"
+        assert replay.message_id == "msg-1"
+        assert replay.field_meta is None
+
+    def test_replays_user_display_content_as_field_meta(self) -> None:
+        user_display_content = UserDisplayContentMetadata(
+            version="1.0.0",
+            host="mistral-vscode",
+            content=[
+                {"type": "text", "text": "Look at "},
+                {
+                    "type": "workspace_mention",
+                    "kind": "file",
+                    "uri": "file:///repo/src/app.ts",
+                    "name": "app.ts",
+                },
+            ],
+        )
+
+        replay = create_user_message_replay(
+            LLMMessage(
+                role=Role.user,
+                content="Look at app.ts",
+                message_id="msg-1",
+                user_display_content=user_display_content,
+            )
+        )
+
+        assert replay.content.text == "Look at app.ts"
+        assert replay.field_meta == {
+            USER_DISPLAY_CONTENT_META_KEY: user_display_content.model_dump(mode="json")
+        }
+
+
+class TestCreateToolCallReplay:
+    def test_carries_status_and_resolved_kind(self) -> None:
+        update = create_tool_call_replay("call_1", "grep", '{"pattern": "foo"}')
+
+        assert update.status == "completed"
+        assert update.kind == "search"
+        assert update.field_meta == {"tool_name": "grep"}
+        assert update.raw_input == '{"pattern": "foo"}'
+
+    def test_unknown_tool_defaults_kind_to_other(self) -> None:
+        update = create_tool_call_replay("call_2", "mystery_tool", None)
+
+        assert update.status == "completed"
+        assert update.kind == "other"
+        assert update.field_meta == {"tool_name": "mystery_tool"}
+
+
+class TestCreateToolResultReplay:
+    def test_carries_kind_and_tool_name_meta(self) -> None:
+        msg = LLMMessage(
+            role=Role.tool, tool_call_id="call_1", name="bash", content="exit 0"
+        )
+
+        update = create_tool_result_replay(msg)
+
+        assert update is not None
+        assert update.kind == "execute"
+        assert update.status == "completed"
+        assert update.field_meta == {"tool_name": "bash"}
+
+    def test_returns_none_without_tool_call_id(self) -> None:
+        msg = LLMMessage(role=Role.tool, name="bash", content="exit 0")
+
+        assert create_tool_result_replay(msg) is None
+
+
+def _tool_call(call_id: str, name: str, arguments: str | None) -> ToolCall:
+    return ToolCall(id=call_id, function=FunctionCall(name=name, arguments=arguments))
+
+
+class TestToolCallReplayUpdate:
+    def test_resolved_task_carries_agent_and_task_meta(self) -> None:
+        tool_call = _tool_call(
+            "call_1", "task", '{"agent": "explore", "task": "find the bug"}'
+        )
+        resolved = ResolvedToolCall(
+            tool_name="task",
+            tool_class=AcpTask,
+            validated_args=TaskArgs(agent="explore", task="find the bug"),
+            call_id="call_1",
+        )
+
+        update = tool_call_replay_update(resolved, tool_call)
+
+        assert isinstance(update, ToolCallStart)
+        assert update.status == "completed"
+        assert update.field_meta is not None
+        assert update.field_meta["tool_name"] == "task"
+        assert update.field_meta["agent"] == "explore"
+        assert update.field_meta["task"] == "find the bug"
+
+    def test_unresolved_call_falls_back_to_generic_replay(self) -> None:
+        tool_call = _tool_call("call_2", "grep", '{"pattern": "foo"}')
+
+        update = tool_call_replay_update(None, tool_call)
+
+        assert isinstance(update, ToolCallStart)
+        assert update.status == "completed"
+        assert update.kind == "search"
+        assert update.field_meta == {"tool_name": "grep"}
+
+    def test_hidden_tool_call_is_skipped(self) -> None:
+        tool_call = _tool_call("call_3", "todo", "{}")
+        resolved = ResolvedToolCall(
+            tool_name="todo",
+            tool_class=AcpTodo,
+            validated_args=TodoArgs(action="read"),
+            call_id="call_3",
+        )
+
+        assert tool_call_replay_update(resolved, tool_call) is None

tests/acp/test_workspace_trust.py

@@ -0,0 +1,140 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+from tests.acp.conftest import _create_acp_agent
+from tests.conftest import build_test_vibe_config
+from tests.stubs.fake_backend import FakeBackend
+from vibe.acp.acp_agent_loop import VibeAcpAgentLoop
+from vibe.acp.exceptions import InvalidRequestError, SessionNotFoundError
+from vibe.core.agent_loop import AgentLoop
+from vibe.core.config import ModelConfig
+from vibe.core.trusted_folders import trusted_folders_manager
+
+
+async def _system_prompt(acp_agent_loop: VibeAcpAgentLoop, session_id: str) -> str:
+    session = acp_agent_loop.sessions[session_id]
+    await session.agent_loop.wait_until_ready()
+    return session.agent_loop.messages[0].content or ""
+
+
+@pytest.fixture
+def acp_agent_loop(
+    backend: FakeBackend, monkeypatch: pytest.MonkeyPatch
+) -> VibeAcpAgentLoop:
+    config = build_test_vibe_config(
+        active_model="devstral-latest",
+        models=[
+            ModelConfig(
+                name="devstral-latest", provider="mistral", alias="devstral-latest"
+            ),
+            ModelConfig(
+                name="devstral-small", provider="mistral", alias="devstral-small"
+            ),
+        ],
+    )
+
+    class PatchedAgentLoop(AgentLoop):
+        def __init__(self, *args, **kwargs) -> None:
+            super().__init__(*args, **{**kwargs, "backend": backend})
+            self._base_config = config
+            self.agent_manager.invalidate_config()
+
+    monkeypatch.setattr("vibe.acp.acp_agent_loop.AgentLoop", PatchedAgentLoop)
+    return _create_acp_agent()
+
+
+class TestWorkspaceTrustExtMethods:
+    @pytest.mark.asyncio
+    async def test_workspace_trust_status_returns_details_even_after_decline(
+        self, acp_agent_loop: VibeAcpAgentLoop, tmp_working_directory: Path
+    ) -> None:
+        (tmp_working_directory / "AGENTS.md").write_text(
+            "Trust me later", encoding="utf-8"
+        )
+        trusted_folders_manager.add_untrusted(tmp_working_directory)
+
+        response = await acp_agent_loop.ext_method(
+            "trust/status", {"cwd": str(tmp_working_directory)}
+        )
+
+        assert response == {
+            "trust_status": "untrusted",
+            "details": {
+                "cwd": str(tmp_working_directory.resolve()),
+                "repoRoot": None,
+                "ignoredFiles": ["AGENTS.md"],
+                "availableDecisions": ["trust_cwd", "trust_session", "decline"],
+            },
+        }
+
+    @pytest.mark.asyncio
+    async def test_workspace_trust_decision_trusts_session_and_reloads_session(
+        self, acp_agent_loop: VibeAcpAgentLoop, tmp_working_directory: Path
+    ) -> None:
+        (tmp_working_directory / "AGENTS.md").write_text(
+            "Reloaded session instructions", encoding="utf-8"
+        )
+
+        session_response = await acp_agent_loop.new_session(
+            cwd=str(tmp_working_directory), mcp_servers=[]
+        )
+        assert session_response.session_id is not None
+        assert "Reloaded session instructions" not in await _system_prompt(
+            acp_agent_loop, session_response.session_id
+        )
+
+        response = await acp_agent_loop.ext_method(
+            "trust/decision",
+            {
+                "cwd": str(tmp_working_directory),
+                "decision": "trust_session",
+                "session_id": session_response.session_id,
+            },
+        )
+
+        assert response == {"trust_status": "session", "details": None}
+        normalized = str(tmp_working_directory.resolve())
+        assert normalized in trusted_folders_manager._session_trusted
+        assert normalized not in trusted_folders_manager._trusted
+        assert "Reloaded session instructions" in await _system_prompt(
+            acp_agent_loop, session_response.session_id
+        )
+
+    @pytest.mark.asyncio
+    async def test_workspace_trust_decision_rejects_unavailable_decision(
+        self, acp_agent_loop: VibeAcpAgentLoop, tmp_working_directory: Path
+    ) -> None:
+        (tmp_working_directory / "AGENTS.md").write_text(
+            "No repo decision here", encoding="utf-8"
+        )
+
+        with pytest.raises(InvalidRequestError):
+            await acp_agent_loop.ext_method(
+                "trust/decision",
+                {"cwd": str(tmp_working_directory), "decision": "trust_repo"},
+            )
+
+        assert trusted_folders_manager.is_trusted(tmp_working_directory) is None
+
+    @pytest.mark.asyncio
+    async def test_workspace_trust_decision_rejects_unknown_session_id(
+        self, acp_agent_loop: VibeAcpAgentLoop, tmp_working_directory: Path
+    ) -> None:
+        (tmp_working_directory / "AGENTS.md").write_text(
+            "Unknown session", encoding="utf-8"
+        )
+
+        with pytest.raises(SessionNotFoundError):
+            await acp_agent_loop.ext_method(
+                "trust/decision",
+                {
+                    "cwd": str(tmp_working_directory),
+                    "decision": "trust_session",
+                    "session_id": "missing-session",
+                },
+            )
+
+        assert trusted_folders_manager.is_trusted(tmp_working_directory) is None

tests/agent_loop/e2e/conftest.py

@@ -0,0 +1,181 @@
+from __future__ import annotations
+
+from collections.abc import Iterator
+import json
+from typing import Any
+
+import httpx
+import pytest
+import respx
+
+from tests.conftest import build_test_agent_loop, build_test_vibe_config
+from tests.constants import (
+    ANTHROPIC_BASE_URL,
+    ANTHROPIC_MESSAGES_PATH,
+    CHAT_COMPLETIONS_PATH,
+    CONNECTORS_BOOTSTRAP_PATH,
+    MISTRAL_BASE_URL,
+    OPENAI_BASE_URL,
+    OPENAI_RESPONSES_PATH,
+)
+from vibe.core.agent_loop import AgentLoop
+from vibe.core.agents.models import BuiltinAgentName
+from vibe.core.config import ModelConfig, ProviderConfig, VibeConfig
+from vibe.core.llm.backend.factory import BACKEND_FACTORY
+from vibe.core.types import AssistantEvent, Backend, BaseEvent
+
+
+def assistant_text(events: list[BaseEvent]) -> str:
+    return "".join(
+        e.content for e in events if isinstance(e, AssistantEvent) and e.content
+    ).strip()
+
+
+@pytest.fixture(autouse=True)
+def _git_offline(monkeypatch: pytest.MonkeyPatch) -> None:
+    # Restrict git to the local file protocol so no test can fetch/push over the
+    # network, even if a remote is misconfigured to a real URL.
+    monkeypatch.setenv("GIT_ALLOW_PROTOCOL", "file")
+
+
+def e2e_config(**overrides: Any) -> VibeConfig:
+    # Defaults give the mistral provider/model (api.mistral.ai, the respx-mocked
+    # base). The default model reasons, but that only adds reasoning_effort to the
+    # request, which the mocked responses ignore.
+    # The e2e harness mocks the connector bootstrap endpoint via respx, so
+    # connector discovery is enabled (and fast) here unlike the unit-test default.
+    return build_test_vibe_config(
+        enabled_tools=overrides.pop("enabled_tools", []),
+        enable_connectors=overrides.pop("enable_connectors", True),
+        system_prompt_id="tests",
+        include_project_context=False,
+        include_prompt_detail=False,
+        **overrides,
+    )
+
+
+def _generic_e2e_config(
+    *, name: str, api_base: str, api_style: str, model: str, **overrides: Any
+) -> VibeConfig:
+    provider = ProviderConfig(
+        name=name,
+        api_base=api_base,
+        api_key_env_var=f"{name.upper()}_API_KEY",
+        api_style=api_style,
+        backend=Backend.GENERIC,
+    )
+    models = [ModelConfig(name=model, provider=name, alias=name)]
+    return e2e_config(
+        active_model=name, models=models, providers=[provider], **overrides
+    )
+
+
+def anthropic_e2e_config(**overrides: Any) -> VibeConfig:
+    return _generic_e2e_config(
+        name="anthropic",
+        api_base=ANTHROPIC_BASE_URL,
+        api_style="anthropic",
+        model="claude-test",
+        **overrides,
+    )
+
+
+def openai_responses_e2e_config(**overrides: Any) -> VibeConfig:
+    return _generic_e2e_config(
+        name="openai",
+        api_base=f"{OPENAI_BASE_URL}/v1",
+        api_style="openai-responses",
+        model="gpt-test",
+        **overrides,
+    )
+
+
+@pytest.fixture
+def mock_mistral() -> Iterator[respx.MockRouter]:
+    with respx.mock(base_url=MISTRAL_BASE_URL, assert_all_called=False) as router:
+        router.get(CONNECTORS_BOOTSTRAP_PATH).mock(
+            return_value=httpx.Response(200, json={"connectors": []})
+        )
+        yield router
+
+
+def build_e2e_agent_loop(
+    *, config: VibeConfig | None = None, enable_streaming: bool = False, **kwargs: Any
+) -> AgentLoop:
+    resolved_config = config or e2e_config()
+    provider = resolved_config.providers[0]
+    # todo now we use build_test_agent_loop for the fake mcp registry.
+    # Maybe mock http tool calls and instantiate a real registry later for more coverage
+    return build_test_agent_loop(
+        config=resolved_config,
+        agent_name=kwargs.pop("agent_name", BuiltinAgentName.AUTO_APPROVE),
+        backend=BACKEND_FACTORY[provider.backend](provider=provider),
+        enable_streaming=enable_streaming,
+        **kwargs,
+    )
+
+
+class ProviderAPI:
+    """Stubs a provider's completion wire; the AgentLoop stays the subject."""
+
+    def __init__(self, router: respx.MockRouter, path: str) -> None:
+        self.route = router.post(path)
+
+    def reply(self, *completions: dict[str, Any]) -> None:
+        responses = [httpx.Response(200, json=c) for c in completions]
+        if len(responses) == 1:
+            self.route.mock(return_value=responses[0])
+        else:
+            self.route.mock(side_effect=responses)
+
+    def reply_stream(self, chunks: list[bytes]) -> None:
+        self.route.mock(return_value=self._stream_response(chunks))
+
+    def reply_streams(self, *chunk_lists: list[bytes]) -> None:
+        self.route.mock(
+            side_effect=[self._stream_response(chunks) for chunks in chunk_lists]
+        )
+
+    @staticmethod
+    def _stream_response(chunks: list[bytes]) -> httpx.Response:
+        return httpx.Response(
+            200,
+            stream=httpx.ByteStream(stream=b"\n\n".join(chunks)),
+            headers={"Content-Type": "text/event-stream"},
+        )
+
+    @property
+    def request_json(self) -> dict[str, Any]:
+        return json.loads(self.route.calls.last.request.content)
+
+
+class MistralAPI(ProviderAPI):
+    def __init__(self, router: respx.MockRouter) -> None:
+        super().__init__(router, CHAT_COMPLETIONS_PATH)
+
+
+@pytest.fixture
+def mistral_api(mock_mistral: respx.MockRouter) -> MistralAPI:
+    return MistralAPI(mock_mistral)
+
+
+@pytest.fixture
+def mock_anthropic() -> Iterator[respx.MockRouter]:
+    with respx.mock(base_url=ANTHROPIC_BASE_URL, assert_all_called=False) as router:
+        yield router
+
+
+@pytest.fixture
+def anthropic_api(mock_anthropic: respx.MockRouter) -> ProviderAPI:
+    return ProviderAPI(mock_anthropic, ANTHROPIC_MESSAGES_PATH)
+
+
+@pytest.fixture
+def mock_openai() -> Iterator[respx.MockRouter]:
+    with respx.mock(base_url=OPENAI_BASE_URL, assert_all_called=False) as router:
+        yield router
+
+
+@pytest.fixture
+def openai_responses_api(mock_openai: respx.MockRouter) -> ProviderAPI:
+    return ProviderAPI(mock_openai, OPENAI_RESPONSES_PATH)

tests/agent_loop/e2e/test_e2e_agent_loop.py

@@ -0,0 +1,111 @@
+from __future__ import annotations
+
+from typing import Any
+
+import pytest
+
+from tests.agent_loop.e2e.conftest import (
+    MistralAPI,
+    assistant_text,
+    build_e2e_agent_loop,
+    e2e_config,
+)
+from tests.backend.data.mistral import (
+    STREAMED_SIMPLE_CONVERSATION_PARAMS,
+    mistral_completion,
+)
+from vibe.core.types import (
+    AssistantEvent,
+    ToolCallEvent,
+    ToolResultEvent,
+    UserMessageEvent,
+)
+
+TODO_TOOL_CALL = [
+    {
+        "id": "call_todo_1",
+        "function": {"name": "todo", "arguments": '{"action": "read"}'},
+        "index": 0,
+    }
+]
+
+
+def _chunked_completion(reasoning: str, answer: str) -> dict[str, Any]:
+    completion = mistral_completion(answer)
+    completion["choices"][0]["message"]["content"] = [
+        {"type": "thinking", "thinking": [{"type": "text", "text": reasoning}]},
+        {"type": "text", "text": answer},
+    ]
+    return completion
+
+
+@pytest.mark.asyncio
+async def test_act_completes_against_real_backend(mistral_api: MistralAPI) -> None:
+    # A plain prompt yields a user event then the backend's assistant reply, and
+    # the loop accumulates prompt+completion tokens from the usage block.
+    mistral_api.reply(
+        mistral_completion("Hi there!", prompt_tokens=120, completion_tokens=30)
+    )
+    agent = build_e2e_agent_loop()
+
+    events = [event async for event in agent.act("Hello")]
+
+    assert [type(e) for e in events] == [UserMessageEvent, AssistantEvent]
+    assert isinstance(events[1], AssistantEvent)
+    assert events[1].content == "Hi there!"
+    assert agent.stats.context_tokens == 150
+
+
+@pytest.mark.asyncio
+async def test_act_streaming(mistral_api: MistralAPI) -> None:
+    # Streamed SSE chunks are reassembled into the final assistant content.
+    _, chunks, _ = STREAMED_SIMPLE_CONVERSATION_PARAMS[0]
+    mistral_api.reply_stream(chunks)
+    agent = build_e2e_agent_loop(enable_streaming=True)
+
+    events = [event async for event in agent.act("Hi")]
+
+    assert assistant_text(events).endswith("Some content")
+
+
+@pytest.mark.asyncio
+async def test_act_tool_call_round_trip(mistral_api: MistralAPI) -> None:
+    # A tool call is parsed, executed, and its result fed back for a final reply.
+    mistral_api.reply(
+        mistral_completion("", tool_calls=TODO_TOOL_CALL),
+        mistral_completion("All done"),
+    )
+    agent = build_e2e_agent_loop(config=e2e_config(enabled_tools=["todo"]))
+
+    events = [event async for event in agent.act("Show my todos")]
+
+    assert any(isinstance(e, ToolCallEvent) for e in events)
+    assert any(isinstance(e, ToolResultEvent) for e in events)
+    final = [e for e in events if isinstance(e, AssistantEvent)]
+    assert final and final[-1].content == "All done"
+
+
+@pytest.mark.asyncio
+async def test_act_serializes_tools_in_request_payload(mistral_api: MistralAPI) -> None:
+    # Enabled tools are serialized into the outgoing chat-completions request.
+    mistral_api.reply(mistral_completion("ok"))
+    agent = build_e2e_agent_loop(config=e2e_config(enabled_tools=["todo"]))
+
+    _ = [event async for event in agent.act("Hello")]
+
+    names = {t["function"]["name"] for t in mistral_api.request_json.get("tools", [])}
+    assert "todo" in names
+
+
+@pytest.mark.asyncio
+async def test_act_extracts_text_from_chunked_content_array(
+    mistral_api: MistralAPI,
+) -> None:
+    # A thinking+text content array has its text part extracted as the reply.
+    mistral_api.reply(_chunked_completion("thinking hard", "the answer"))
+    agent = build_e2e_agent_loop()
+
+    events = [event async for event in agent.act("Why?")]
+
+    assistant = [e for e in events if isinstance(e, AssistantEvent)]
+    assert assistant and assistant[-1].content == "the answer"

tests/agent_loop/e2e/test_e2e_bash.py

@@ -0,0 +1,170 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any, cast
+
+from pydantic import BaseModel
+import pytest
+
+from tests.agent_loop.e2e.conftest import MistralAPI, build_e2e_agent_loop, e2e_config
+from tests.backend.data.mistral import mistral_completion
+from vibe.core.agents.models import BuiltinAgentName
+from vibe.core.tools.builtins.bash import BashResult
+from vibe.core.tools.permissions import RequiredPermission
+from vibe.core.types import ApprovalResponse, BaseEvent, ToolResultEvent
+
+
+def _bash_call(command: str, timeout: int | None = None) -> dict[str, Any]:
+    arguments: dict[str, Any] = {"command": command}
+    if timeout is not None:
+        arguments["timeout"] = timeout
+    return mistral_completion(
+        "",
+        tool_calls=[
+            {
+                "id": "call_bash",
+                "function": {"name": "bash", "arguments": json.dumps(arguments)},
+                "index": 0,
+            }
+        ],
+    )
+
+
+async def _run_bash(
+    mistral_api: MistralAPI,
+    command: str,
+    *,
+    timeout: int | None = None,
+    agent_name: str = BuiltinAgentName.AUTO_APPROVE,
+    approval: ApprovalResponse | None = None,
+) -> ToolResultEvent:
+    mistral_api.reply(_bash_call(command, timeout), mistral_completion("done"))
+    agent = build_e2e_agent_loop(
+        config=e2e_config(enabled_tools=["bash"]), agent_name=agent_name
+    )
+    if approval is not None:
+
+        async def approval_callback(
+            _tool_name: str,
+            _args: BaseModel,
+            _tool_call_id: str,
+            _rp: list[RequiredPermission] | None = None,
+        ) -> tuple[ApprovalResponse, str | None]:
+            return (approval, None)
+
+        agent.set_approval_callback(approval_callback)
+
+    events: list[BaseEvent] = [event async for event in agent.act("go")]
+    return next(e for e in events if isinstance(e, ToolResultEvent))
+
+
+@pytest.mark.asyncio
+async def test_bash_captures_stdout(mistral_api: MistralAPI) -> None:
+    result = await _run_bash(mistral_api, "echo hello")
+
+    bash_result = cast(BashResult, result.result)
+    assert bash_result.returncode == 0
+    assert "hello" in bash_result.stdout
+
+
+@pytest.mark.asyncio
+async def test_bash_captures_stderr(mistral_api: MistralAPI) -> None:
+    result = await _run_bash(mistral_api, "echo oops >&2")
+
+    bash_result = cast(BashResult, result.result)
+    assert "oops" in bash_result.stderr
+
+
+@pytest.mark.asyncio
+async def test_bash_nonzero_exit_surfaces_as_error(mistral_api: MistralAPI) -> None:
+    result = await _run_bash(mistral_api, "exit 3")
+
+    assert result.error is not None
+    assert "Return code: 3" in result.error
+
+
+@pytest.mark.asyncio
+async def test_bash_timeout_surfaces_as_error(mistral_api: MistralAPI) -> None:
+    result = await _run_bash(mistral_api, "sleep 5", timeout=1)
+
+    assert result.error is not None
+    assert "timed out" in result.error.lower()
+
+
+@pytest.mark.asyncio
+async def test_bash_output_truncated_to_max_bytes(mistral_api: MistralAPI) -> None:
+    result = await _run_bash(mistral_api, "yes x | head -c 100000")
+
+    bash_result = cast(BashResult, result.result)
+    assert len(bash_result.stdout) <= 16_000
+
+
+@pytest.mark.asyncio
+async def test_bash_denylisted_command_is_skipped(mistral_api: MistralAPI) -> None:
+    result = await _run_bash(
+        mistral_api, "vim file.txt", agent_name=BuiltinAgentName.DEFAULT
+    )
+
+    assert result.skipped is True
+    assert result.skip_reason is not None
+    assert "denied" in result.skip_reason.lower()
+
+
+@pytest.mark.asyncio
+async def test_bash_allowlisted_command_runs_without_approval(
+    mistral_api: MistralAPI,
+) -> None:
+    # No approval callback registered; an allowlisted command must run anyway.
+    result = await _run_bash(
+        mistral_api, "echo allowed", agent_name=BuiltinAgentName.DEFAULT
+    )
+
+    assert result.skipped is False
+    assert "allowed" in cast(BashResult, result.result).stdout
+
+
+@pytest.mark.asyncio
+async def test_bash_non_allowlisted_command_requires_approval(
+    mistral_api: MistralAPI,
+) -> None:
+    result = await _run_bash(
+        mistral_api,
+        "touch newfile.txt",
+        agent_name=BuiltinAgentName.DEFAULT,
+        approval=ApprovalResponse.YES,
+    )
+
+    assert result.skipped is False
+    assert (Path.cwd() / "newfile.txt").exists()
+
+
+@pytest.mark.asyncio
+async def test_bash_non_allowlisted_command_denied_at_prompt_is_skipped(
+    mistral_api: MistralAPI,
+) -> None:
+    result = await _run_bash(
+        mistral_api,
+        "touch denied.txt",
+        agent_name=BuiltinAgentName.DEFAULT,
+        approval=ApprovalResponse.NO,
+    )
+
+    assert result.skipped is True
+    assert not (Path.cwd() / "denied.txt").exists()
+
+
+@pytest.mark.asyncio
+async def test_bash_command_touching_outside_workdir_requires_approval(
+    mistral_api: MistralAPI, tmp_path: Path
+) -> None:
+    outside = tmp_path / "outside.txt"
+    result = await _run_bash(
+        mistral_api,
+        f"touch {outside}",
+        agent_name=BuiltinAgentName.DEFAULT,
+        approval=ApprovalResponse.NO,
+    )
+
+    assert result.skipped is True
+    assert not outside.exists()

tests/agent_loop/e2e/test_e2e_connectors.py

@@ -0,0 +1,136 @@
+from __future__ import annotations
+
+from typing import Any
+
+import httpx
+import pytest
+import respx
+
+from tests.agent_loop.e2e.conftest import MistralAPI, build_e2e_agent_loop, e2e_config
+from tests.backend.data.mistral import mistral_completion
+from tests.constants import CONNECTORS_BOOTSTRAP_PATH
+from vibe.core.config import ConnectorConfig
+
+
+def _connector(
+    *,
+    connector_id: str = "conn-1",
+    name: str = "wiki",
+    is_ready: bool = True,
+    tools: list[dict[str, Any]] | None = None,
+    auth_action: dict[str, Any] | None = None,
+) -> dict[str, Any]:
+    return {
+        "id": connector_id,
+        "name": name,
+        "status": {"is_ready": is_ready},
+        "tools": tools or [],
+        "auth_action": auth_action,
+    }
+
+
+def _tool(name: str = "search", description: str = "Search docs") -> dict[str, Any]:
+    return {
+        "name": name,
+        "description": description,
+        "inputSchema": {"type": "object", "properties": {"query": {"type": "string"}}},
+    }
+
+
+def _set_bootstrap(router: respx.MockRouter, connectors: list[dict[str, Any]]) -> None:
+    router.get(CONNECTORS_BOOTSTRAP_PATH).mock(
+        return_value=httpx.Response(200, json={"connectors": connectors})
+    )
+
+
+def _connectors_config(*aliases: str) -> Any:
+    return e2e_config(
+        connectors=[ConnectorConfig(name=alias, disabled=False) for alias in aliases]
+    )
+
+
+async def _offered_tool_names(
+    mistral_api: MistralAPI, *enabled_connectors: str
+) -> set[str]:
+    # Drive one turn and read which tools the AgentLoop serialized into the
+    # outgoing chat-completions request — the only connector-visible surface.
+    mistral_api.reply(mistral_completion("ok"))
+    agent = build_e2e_agent_loop(config=_connectors_config(*enabled_connectors))
+    _ = [event async for event in agent.act("hello")]
+    return {t["function"]["name"] for t in mistral_api.request_json.get("tools", [])}
+
+
+@pytest.mark.asyncio
+async def test_connector_tools_are_offered_to_the_model(
+    mistral_api: MistralAPI, mock_mistral: respx.MockRouter
+) -> None:
+    _set_bootstrap(
+        mock_mistral, [_connector(name="wiki", tools=[_tool("search"), _tool("read")])]
+    )
+
+    names = await _offered_tool_names(mistral_api, "wiki")
+
+    assert "connector_wiki_search" in names
+    assert "connector_wiki_read" in names
+
+
+@pytest.mark.asyncio
+async def test_colliding_connector_aliases_are_disambiguated(
+    mistral_api: MistralAPI, mock_mistral: respx.MockRouter
+) -> None:
+    _set_bootstrap(
+        mock_mistral,
+        [
+            _connector(connector_id="c-1", name="mcp", tools=[_tool("a")]),
+            _connector(connector_id="c-2", name="mcp", tools=[_tool("b")]),
+        ],
+    )
+
+    names = await _offered_tool_names(mistral_api, "mcp", "mcp_2")
+
+    assert "connector_mcp_a" in names
+    assert "connector_mcp_2_b" in names
+
+
+@pytest.mark.asyncio
+async def test_not_ready_connector_offers_no_tools(
+    mistral_api: MistralAPI, mock_mistral: respx.MockRouter
+) -> None:
+    _set_bootstrap(
+        mock_mistral,
+        [
+            _connector(
+                name="linear",
+                is_ready=False,
+                tools=[_tool("search")],
+                auth_action={"type": "oauth"},
+            )
+        ],
+    )
+
+    names = await _offered_tool_names(mistral_api, "linear")
+
+    assert not any(name.startswith("connector_linear") for name in names)
+
+
+@pytest.mark.asyncio
+async def test_disabled_connector_tools_are_withheld(
+    mistral_api: MistralAPI, mock_mistral: respx.MockRouter
+) -> None:
+    _set_bootstrap(mock_mistral, [_connector(name="wiki", tools=[_tool("search")])])
+
+    # No ConnectorConfig entry → the connector stays disabled by default.
+    names = await _offered_tool_names(mistral_api)
+
+    assert "connector_wiki_search" not in names
+
+
+@pytest.mark.asyncio
+async def test_bootstrap_failure_still_lets_the_agent_run(
+    mistral_api: MistralAPI, mock_mistral: respx.MockRouter
+) -> None:
+    mock_mistral.get(CONNECTORS_BOOTSTRAP_PATH).mock(return_value=httpx.Response(500))
+
+    names = await _offered_tool_names(mistral_api, "wiki")
+
+    assert not any(name.startswith("connector_") for name in names)

tests/agent_loop/e2e/test_e2e_fork_and_plan.py

@@ -0,0 +1,144 @@
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+import pytest
+
+from tests.agent_loop.e2e.conftest import MistralAPI, build_e2e_agent_loop, e2e_config
+from tests.backend.data.mistral import mistral_completion
+from vibe.core.agents.models import BuiltinAgentName
+from vibe.core.tools.builtins.ask_user_question import (
+    AskUserQuestionArgs,
+    AskUserQuestionResult,
+)
+from vibe.core.types import PlanReviewEndedEvent, PlanReviewRequestedEvent, Role
+from vibe.core.utils.tags import VIBE_WARNING_TAG
+
+EXIT_PLAN_TOOL_CALL = [
+    {
+        "id": "call_exit_plan_1",
+        "function": {"name": "exit_plan_mode", "arguments": "{}"},
+        "index": 0,
+    }
+]
+
+
+def _user_message_ids(agent: Any) -> list[str]:
+    return [m.message_id for m in agent.messages if m.role == Role.user]
+
+
+def _assistant_message_id(agent: Any) -> str:
+    return next(m.message_id for m in agent.messages if m.role == Role.assistant)
+
+
+@pytest.mark.asyncio
+async def test_fork_copies_all_non_system_messages(mistral_api: MistralAPI) -> None:
+    # fork() with no anchor clones every non-system message into the child loop.
+    mistral_api.reply(mistral_completion("Hi there!"))
+    agent = build_e2e_agent_loop()
+    _ = [event async for event in agent.act("Hello")]
+
+    forked = await agent.fork()
+
+    non_system = [m for m in forked.messages if m.role != Role.system]
+    assert [m.role for m in non_system] == [Role.user, Role.assistant]
+    assert forked.parent_session_id == agent.session_id
+    assert forked.session_id != agent.session_id
+
+
+@pytest.mark.asyncio
+async def test_fork_from_message_id_truncates_at_next_user_turn(
+    mistral_api: MistralAPI,
+) -> None:
+    # Forking from a user message keeps that turn but drops everything from the next one.
+    mistral_api.reply(mistral_completion("First"), mistral_completion("Second"))
+    agent = build_e2e_agent_loop()
+    _ = [event async for event in agent.act("Turn one")]
+    _ = [event async for event in agent.act("Turn two")]
+
+    first_user_id = _user_message_ids(agent)[0]
+    forked = await agent.fork(first_user_id)
+
+    contents = [m.content for m in forked.messages if m.role != Role.system]
+    assert contents == ["Turn one", "First"]
+
+
+@pytest.mark.asyncio
+async def test_fork_from_unknown_message_id_raises(mistral_api: MistralAPI) -> None:
+    # An unknown anchor id is rejected rather than silently forking everything.
+    mistral_api.reply(mistral_completion("Hi"))
+    agent = build_e2e_agent_loop()
+    _ = [event async for event in agent.act("Hello")]
+
+    with pytest.raises(ValueError, match="unknown message_id"):
+        await agent.fork("does-not-exist")
+
+
+@pytest.mark.asyncio
+async def test_fork_from_assistant_message_id_raises(mistral_api: MistralAPI) -> None:
+    # Forking is only allowed from user turns; an assistant anchor is rejected.
+    mistral_api.reply(mistral_completion("Hi"))
+    agent = build_e2e_agent_loop()
+    _ = [event async for event in agent.act("Hello")]
+
+    assistant_id = _assistant_message_id(agent)
+    with pytest.raises(ValueError, match="only supported for user messages"):
+        await agent.fork(assistant_id)
+
+
+def _plan_agent(mistral_api: MistralAPI) -> Any:
+    mistral_api.reply(
+        mistral_completion("", tool_calls=EXIT_PLAN_TOOL_CALL),
+        mistral_completion("Staying in plan mode."),
+    )
+    return build_e2e_agent_loop(
+        config=e2e_config(enabled_tools=["exit_plan_mode"]),
+        agent_name=BuiltinAgentName.PLAN,
+    )
+
+
+@pytest.mark.asyncio
+async def test_plan_mode_emits_review_requested_and_ended_events(
+    mistral_api: MistralAPI,
+) -> None:
+    # An exit_plan_mode round trip brackets the tool with review-requested/ended events.
+    agent = _plan_agent(mistral_api)
+
+    async def stay_in_plan(_: AskUserQuestionArgs) -> AskUserQuestionResult:
+        return AskUserQuestionResult(cancelled=True, answers=[])
+
+    agent.set_user_input_callback(stay_in_plan)
+
+    events = [event async for event in agent.act("Make a plan")]
+
+    assert any(isinstance(e, PlanReviewRequestedEvent) for e in events)
+    assert any(isinstance(e, PlanReviewEndedEvent) for e in events)
+
+
+@pytest.mark.asyncio
+async def test_plan_mode_injects_updated_plan_when_file_changed(
+    mistral_api: MistralAPI,
+) -> None:
+    # If the plan file changes during review, its new content is injected back as context.
+    agent = _plan_agent(mistral_api)
+
+    plan_path: Path | None = None
+
+    async def edit_plan_then_decline(_: AskUserQuestionArgs) -> AskUserQuestionResult:
+        assert plan_path is not None
+        plan_path.parent.mkdir(parents=True, exist_ok=True)
+        plan_path.write_text("# Updated plan\nStep 1")
+        return AskUserQuestionResult(cancelled=True, answers=[])
+
+    agent.set_user_input_callback(edit_plan_then_decline)
+
+    async for event in agent.act("Make a plan"):
+        if isinstance(event, PlanReviewRequestedEvent):
+            plan_path = event.file_path
+
+    injected = [m for m in agent.messages if getattr(m, "injected", False)]
+    assert any(
+        m.content and VIBE_WARNING_TAG in m.content and "# Updated plan" in m.content
+        for m in injected
+    )

tests/agent_loop/e2e/test_e2e_provider_apis.py

@@ -0,0 +1,178 @@
+from __future__ import annotations
+
+import pytest
+
+from tests.agent_loop.e2e.conftest import (
+    ProviderAPI,
+    anthropic_e2e_config,
+    assistant_text,
+    build_e2e_agent_loop,
+    openai_responses_e2e_config,
+)
+from tests.backend.data.anthropic import (
+    anthropic_message,
+    anthropic_reasoning_tool_use_stream,
+    anthropic_request_content_blocks,
+    anthropic_text_stream,
+    anthropic_tool_use,
+)
+from tests.backend.data.openai_responses import (
+    openai_function_call_item,
+    openai_message_item,
+    openai_reasoning_tool_call_stream,
+    openai_response,
+    openai_text_stream,
+)
+from vibe.core.types import ToolResultEvent
+
+
+class TestAnthropic:
+    @pytest.mark.asyncio
+    async def test_agent_answers(self, anthropic_api: ProviderAPI) -> None:
+        # A plain prompt is answered through the Anthropic messages wire.
+        anthropic_api.reply(anthropic_message("pong"))
+        agent = build_e2e_agent_loop(config=anthropic_e2e_config())
+
+        events = [event async for event in agent.act("Reply with exactly: pong")]
+
+        assert assistant_text(events) == "pong"
+        assert agent.stats.context_tokens == 15
+
+    @pytest.mark.asyncio
+    async def test_agent_streams(self, anthropic_api: ProviderAPI) -> None:
+        # Anthropic SSE deltas are reassembled into the final assistant content.
+        anthropic_api.reply_stream(anthropic_text_stream("pong"))
+        agent = build_e2e_agent_loop(
+            config=anthropic_e2e_config(), enable_streaming=True
+        )
+
+        events = [event async for event in agent.act("Reply with exactly: pong")]
+
+        assert assistant_text(events) == "pong"
+
+    @pytest.mark.asyncio
+    async def test_agent_executes_tool_call(self, anthropic_api: ProviderAPI) -> None:
+        # A tool_use turn runs the tool, then Anthropic returns the final answer.
+        anthropic_api.reply(
+            anthropic_tool_use("todo", {"action": "read"}),
+            anthropic_message("Your list is empty."),
+        )
+        agent = build_e2e_agent_loop(
+            config=anthropic_e2e_config(enabled_tools=["todo"])
+        )
+
+        events = [event async for event in agent.act("What's on my todo list?")]
+
+        assert any(isinstance(e, ToolResultEvent) for e in events)
+        assert "Your list is empty." in assistant_text(events)
+
+    @pytest.mark.asyncio
+    async def test_agent_streams_reasoning_and_tool_call(
+        self, anthropic_api: ProviderAPI
+    ) -> None:
+        # A streamed thinking + tool_use turn runs the tool, then replays that
+        # reasoning/tool history into the follow-up Anthropic request.
+        anthropic_api.reply_streams(
+            anthropic_reasoning_tool_use_stream("todo", '{"action": "read"}'),
+            anthropic_text_stream("Your list is empty."),
+        )
+        agent = build_e2e_agent_loop(
+            config=anthropic_e2e_config(enabled_tools=["todo"]), enable_streaming=True
+        )
+
+        events = [event async for event in agent.act("What's on my todo list?")]
+
+        assert "Your list is empty." in assistant_text(events)
+        blocks = anthropic_request_content_blocks(anthropic_api.request_json)
+        thinking = [b["thinking"] for b in blocks if b.get("type") == "thinking"]
+        tool_uses = [b for b in blocks if b.get("type") == "tool_use"]
+
+        assert any("thinking..." in text for text in thinking)
+        assert tool_uses
+
+
+class TestOpenAIResponses:
+    @pytest.mark.asyncio
+    async def test_agent_answers(self, openai_responses_api: ProviderAPI) -> None:
+        # A plain prompt is answered through the OpenAI Responses wire.
+        openai_responses_api.reply(openai_response([openai_message_item("pong")]))
+        agent = build_e2e_agent_loop(config=openai_responses_e2e_config())
+
+        events = [event async for event in agent.act("Reply with exactly: pong")]
+
+        assert assistant_text(events) == "pong"
+        assert agent.stats.context_tokens == 12
+
+    @pytest.mark.asyncio
+    async def test_agent_streams(self, openai_responses_api: ProviderAPI) -> None:
+        # OpenAI Responses SSE deltas are reassembled into the final content.
+        openai_responses_api.reply_stream(openai_text_stream("pong"))
+        agent = build_e2e_agent_loop(
+            config=openai_responses_e2e_config(), enable_streaming=True
+        )
+
+        events = [event async for event in agent.act("Reply with exactly: pong")]
+
+        assert assistant_text(events) == "pong"
+
+    @pytest.mark.asyncio
+    async def test_agent_captures_reasoning(
+        self, openai_responses_api: ProviderAPI
+    ) -> None:
+        # Commentary phase output is captured as reasoning on the assistant message.
+        openai_responses_api.reply(
+            openai_response(
+                [
+                    openai_message_item("Let me think.", phase="commentary"),
+                    openai_message_item("pong", phase="final_answer"),
+                ],
+                output_tokens=5,
+            )
+        )
+        agent = build_e2e_agent_loop(config=openai_responses_e2e_config())
+
+        events = [event async for event in agent.act("Reply with exactly: pong")]
+
+        assert assistant_text(events) == "pong"
+        assert any(
+            m.reasoning_content and "Let me think." in m.reasoning_content
+            for m in agent.messages
+        )
+
+    @pytest.mark.asyncio
+    async def test_agent_executes_tool_call(
+        self, openai_responses_api: ProviderAPI
+    ) -> None:
+        # A function_call item runs the tool, then OpenAI returns the final answer.
+        openai_responses_api.reply(
+            openai_response([openai_function_call_item("todo", '{"action": "read"}')]),
+            openai_response([openai_message_item("Your list is empty.")]),
+        )
+        agent = build_e2e_agent_loop(
+            config=openai_responses_e2e_config(enabled_tools=["todo"])
+        )
+
+        events = [event async for event in agent.act("What's on my todo list?")]
+
+        assert any(isinstance(e, ToolResultEvent) for e in events)
+        assert "Your list is empty." in assistant_text(events)
+
+    @pytest.mark.asyncio
+    async def test_agent_streams_reasoning_and_tool_call(
+        self, openai_responses_api: ProviderAPI
+    ) -> None:
+        # A streamed commentary + function_call turn runs the tool, then OpenAI
+        # streams the final answer on the follow-up request.
+        openai_responses_api.reply_streams(
+            openai_reasoning_tool_call_stream("todo", '{"action": "read"}'),
+            openai_text_stream("Your list is empty."),
+        )
+        agent = build_e2e_agent_loop(
+            config=openai_responses_e2e_config(enabled_tools=["todo"]),
+            enable_streaming=True,
+        )
+
+        events = [event async for event in agent.act("What's on my todo list?")]
+
+        assert any(isinstance(e, ToolResultEvent) for e in events)
+        assert "Your list is empty." in assistant_text(events)

tests/agent_loop/e2e/test_e2e_teleport.py

@@ -0,0 +1,224 @@
+from __future__ import annotations
+
+from collections.abc import Iterator
+from pathlib import Path
+
+from git import Repo
+import httpx
+import pytest
+import respx
+
+from tests.agent_loop.e2e.conftest import build_e2e_agent_loop
+from tests.constants import (
+    CONNECTORS_BOOTSTRAP_PATH,
+    MISTRAL_BASE_URL,
+    TELEPORT_COMPLETE_URL,
+    TELEPORT_SESSIONS_PATH,
+)
+from vibe.core.agent_loop import AgentLoop, TeleportError
+from vibe.core.teleport.types import (
+    TeleportCheckingGitEvent,
+    TeleportCompleteEvent,
+    TeleportPushingEvent,
+    TeleportPushRequiredEvent,
+    TeleportPushResponseEvent,
+    TeleportStartingWorkflowEvent,
+)
+
+# vibe_code_sessions_base_url is an exclude=True field, so it is dropped when the
+# agent manager re-derives config via model_dump(); the default url is unavoidable.
+SESSIONS_BASE_URL = "https://chat.mistral.ai"
+SESSIONS_URL = f"{SESSIONS_BASE_URL}{TELEPORT_SESSIONS_PATH}"
+
+
+def _sessions_ok() -> httpx.Response:
+    return httpx.Response(
+        200,
+        json={
+            "sessionId": "controller-session-id",
+            "webSessionId": "web-session-id",
+            "projectId": "project-id",
+            "status": "running",
+            "url": TELEPORT_COMPLETE_URL,
+        },
+    )
+
+
+def _commit(repo: Repo, message: str) -> str:
+    (Path(repo.working_dir) / "file.txt").write_text(f"{message}\n")
+    repo.index.add(["file.txt"])
+    repo.index.commit(message)
+    return repo.head.commit.hexsha
+
+
+def _init_repo(workdir: Path) -> Repo:
+    # origin is a local bare repo so fetch/push stay offline and instant; a
+    # separate github-url remote satisfies Teleport's GitHub-only detection.
+    bare = Repo.init(workdir.with_name(f"{workdir.name}_origin.git"), bare=True)
+    repo = Repo.init(workdir, initial_branch="work")
+    repo.config_writer().set_value("user", "name", "Tester").release()
+    repo.config_writer().set_value("user", "email", "t@example.com").release()
+    repo.create_remote("origin", str(bare.git_dir))
+    repo.create_remote("hub", "https://github.com/owner/repo.git")
+    return repo
+
+
+def _repo_with_pushed_branch(workdir: Path) -> Repo:
+    repo = _init_repo(workdir)
+    _commit(repo, "initial")
+    repo.git.push("origin", "work")
+    return repo
+
+
+@pytest.fixture
+def mock_sessions() -> Iterator[respx.MockRouter]:
+    # One router stubs both hosts: the connector bootstrap (api.mistral.ai, hit on
+    # agent init) and the teleport sessions endpoint (chat.mistral.ai).
+    with respx.mock(assert_all_called=False) as router:
+        router.get(f"{MISTRAL_BASE_URL}{CONNECTORS_BOOTSTRAP_PATH}").mock(
+            return_value=httpx.Response(200, json={"connectors": []})
+        )
+        router.post(SESSIONS_URL).mock(return_value=_sessions_ok())
+        yield router
+
+
+async def _drain(
+    agent: AgentLoop, prompt: str | None, *, approve: bool | None = None
+) -> list[object]:
+    gen = agent.teleport_to_vibe_code(prompt)
+    events: list[object] = []
+    response: TeleportPushResponseEvent | None = None
+    while True:
+        try:
+            event = await gen.asend(response)
+        except StopAsyncIteration:
+            break
+        events.append(event)
+        response = None
+        if isinstance(event, TeleportPushRequiredEvent) and approve is not None:
+            response = TeleportPushResponseEvent(approved=approve)
+    return events
+
+
+@pytest.mark.asyncio
+async def test_teleport_completes_when_branch_already_pushed(
+    tmp_working_directory: Path, mock_sessions: respx.MockRouter
+) -> None:
+    _repo_with_pushed_branch(tmp_working_directory)
+
+    events = await _drain(build_e2e_agent_loop(), "do the thing")
+
+    assert [type(e) for e in events] == [
+        TeleportCheckingGitEvent,
+        TeleportStartingWorkflowEvent,
+        TeleportCompleteEvent,
+    ]
+    assert isinstance(events[-1], TeleportCompleteEvent)
+    assert events[-1].url == TELEPORT_COMPLETE_URL
+
+
+@pytest.mark.asyncio
+async def test_teleport_sends_repo_metadata_and_diff(
+    tmp_working_directory: Path, mock_sessions: respx.MockRouter
+) -> None:
+    repo = _repo_with_pushed_branch(tmp_working_directory)
+    commit = repo.head.commit.hexsha
+    (tmp_working_directory / "file.txt").write_text("uncommitted change\n")
+
+    await _drain(build_e2e_agent_loop(), "ship it")
+
+    payload = mock_sessions.calls.last.request.read().decode()
+    assert "https://github.com/owner/repo.git" in payload
+    assert "work" in payload
+    assert commit in payload
+    assert "zstd" in payload
+
+
+@pytest.mark.asyncio
+async def test_teleport_pushes_then_completes_when_approved(
+    tmp_working_directory: Path, mock_sessions: respx.MockRouter
+) -> None:
+    repo = _repo_with_pushed_branch(tmp_working_directory)
+    head = _commit(repo, "second")
+
+    events = await _drain(build_e2e_agent_loop(), "ship it", approve=True)
+
+    assert [type(e) for e in events] == [
+        TeleportCheckingGitEvent,
+        TeleportPushRequiredEvent,
+        TeleportPushingEvent,
+        TeleportStartingWorkflowEvent,
+        TeleportCompleteEvent,
+    ]
+    assert repo.remote("origin").refs["work"].commit.hexsha == head
+
+
+@pytest.mark.asyncio
+async def test_teleport_aborts_when_push_declined(
+    tmp_working_directory: Path, mock_sessions: respx.MockRouter
+) -> None:
+    repo = _repo_with_pushed_branch(tmp_working_directory)
+    _commit(repo, "second")
+
+    with pytest.raises(TeleportError, match="not pushed"):
+        await _drain(build_e2e_agent_loop(), "ship it", approve=False)
+
+    assert not mock_sessions.post(SESSIONS_URL).called
+
+
+@pytest.mark.asyncio
+async def test_teleport_fails_when_push_fails(
+    tmp_working_directory: Path, mock_sessions: respx.MockRouter
+) -> None:
+    repo = _repo_with_pushed_branch(tmp_working_directory)
+    _commit(repo, "second")
+    repo.git.remote("set-url", "--push", "origin", "/nonexistent/repo.git")
+
+    with pytest.raises(TeleportError, match="Failed to push"):
+        await _drain(build_e2e_agent_loop(), "ship it", approve=True)
+
+
+@pytest.mark.asyncio
+async def test_teleport_requires_a_branch(
+    tmp_working_directory: Path, mock_sessions: respx.MockRouter
+) -> None:
+    repo = _repo_with_pushed_branch(tmp_working_directory)
+    repo.git.checkout(repo.head.commit.hexsha)  # detach HEAD
+
+    with pytest.raises(TeleportError, match="checked-out branch"):
+        await _drain(build_e2e_agent_loop(), "ship it")
+
+
+@pytest.mark.asyncio
+async def test_teleport_rejects_empty_prompt(
+    tmp_working_directory: Path, mock_sessions: respx.MockRouter
+) -> None:
+    repo = _init_repo(tmp_working_directory)
+    _commit(repo, "initial")
+
+    with pytest.raises(TeleportError, match="non-empty prompt"):
+        await _drain(build_e2e_agent_loop(), None)
+
+
+@pytest.mark.asyncio
+async def test_teleport_surfaces_http_error(
+    tmp_working_directory: Path, mock_sessions: respx.MockRouter
+) -> None:
+    _repo_with_pushed_branch(tmp_working_directory)
+    mock_sessions.post(SESSIONS_URL).mock(return_value=httpx.Response(500, text="boom"))
+
+    with pytest.raises(TeleportError, match="start failed"):
+        await _drain(build_e2e_agent_loop(), "ship it")
+
+
+@pytest.mark.asyncio
+async def test_teleport_unsupported_without_github_remote(
+    tmp_working_directory: Path, mock_sessions: respx.MockRouter
+) -> None:
+    repo = Repo.init(tmp_working_directory, initial_branch="work")
+    repo.config_writer().set_value("user", "name", "Tester").release()
+    repo.config_writer().set_value("user", "email", "t@example.com").release()
+    _commit(repo, "initial")
+
+    with pytest.raises(TeleportError, match="GitHub"):
+        await _drain(build_e2e_agent_loop(), "ship it")

tests/agent_loop/e2e/test_e2e_tools.py

@@ -0,0 +1,93 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any, cast
+
+import pytest
+
+from tests.agent_loop.e2e.conftest import MistralAPI, build_e2e_agent_loop, e2e_config
+from tests.backend.data.mistral import mistral_completion
+from vibe.core.tools.builtins.grep import GrepResult
+from vibe.core.tools.builtins.read import ReadResult
+from vibe.core.tools.builtins.todo import TodoResult
+from vibe.core.types import BaseEvent, ToolResultEvent
+
+
+def _tool_call(name: str, arguments: dict[str, Any]) -> dict[str, Any]:
+    return mistral_completion(
+        "",
+        tool_calls=[
+            {
+                "id": f"call_{name}",
+                "function": {"name": name, "arguments": json.dumps(arguments)},
+                "index": 0,
+            }
+        ],
+    )
+
+
+async def _run_tool(
+    mistral_api: MistralAPI, name: str, arguments: dict[str, Any]
+) -> ToolResultEvent:
+    # Drive one tool call against the real cwd, then a final reply, and return
+    # the single tool result for the test to assert on.
+    mistral_api.reply(_tool_call(name, arguments), mistral_completion("done"))
+    agent = build_e2e_agent_loop(config=e2e_config(enabled_tools=[name]))
+    events: list[BaseEvent] = [event async for event in agent.act("go")]
+    result = next(e for e in events if isinstance(e, ToolResultEvent))
+    assert result.error is None
+    return result
+
+
+@pytest.mark.asyncio
+async def test_write_file_tool_creates_file(mistral_api: MistralAPI) -> None:
+    target = Path.cwd() / "note.txt"
+
+    await _run_tool(mistral_api, "write_file", {"path": str(target), "content": "hi\n"})
+
+    assert target.read_text() == "hi\n"
+
+
+@pytest.mark.asyncio
+async def test_read_tool_returns_file_content(mistral_api: MistralAPI) -> None:
+    target = Path.cwd() / "note.txt"
+    target.write_text("hello world\n")
+
+    result = await _run_tool(mistral_api, "read", {"file_path": str(target)})
+
+    assert "hello world" in cast(ReadResult, result.result).content
+
+
+@pytest.mark.asyncio
+async def test_edit_tool_replaces_text(mistral_api: MistralAPI) -> None:
+    target = Path.cwd() / "note.txt"
+    target.write_text("hello world\n")
+
+    await _run_tool(
+        mistral_api,
+        "edit",
+        {"file_path": str(target), "old_string": "hello", "new_string": "goodbye"},
+    )
+
+    assert target.read_text() == "goodbye world\n"
+
+
+@pytest.mark.asyncio
+async def test_grep_tool_finds_matches(mistral_api: MistralAPI) -> None:
+    (Path.cwd() / "note.txt").write_text("needle here\n")
+
+    result = await _run_tool(mistral_api, "grep", {"pattern": "needle", "path": "."})
+
+    assert cast(GrepResult, result.result).match_count >= 1
+
+
+@pytest.mark.asyncio
+async def test_todo_tool_writes_items(mistral_api: MistralAPI) -> None:
+    result = await _run_tool(
+        mistral_api,
+        "todo",
+        {"action": "write", "todos": [{"id": "1", "content": "ship it"}]},
+    )
+
+    assert cast(TodoResult, result.result).total_count == 1

tests/agent_loop/test_agent_loop_images.py

@@ -9,7 +9,7 @@ from tests.mock.utils import mock_llm_chunk
 from tests.stubs.fake_backend import FakeBackend
 from vibe.core.agent_loop import ImagesNotSupportedError
 from vibe.core.config import ModelConfig, ProviderConfig, VibeConfig
-from vibe.core.types import Backend, ImageAttachment, LLMMessage, Role
+from vibe.core.types import Backend, FileImageSource, ImageAttachment, LLMMessage, Role
 
 PNG_BYTES = b"\x89PNG\r\n\x1a\n" + b"\x00" * 16
 
@@ -68,7 +68,9 @@ def _config_with_both_models() -> VibeConfig:
 def png_attachment(tmp_path: Path) -> ImageAttachment:
     p = tmp_path / "x.png"
     p.write_bytes(PNG_BYTES)
-    return ImageAttachment(path=p, alias="x.png", mime_type="image/png")
+    return ImageAttachment(
+        source=FileImageSource(path=p), alias="x.png", mime_type="image/png"
+    )
 
 
 @pytest.mark.asyncio

tests/agent_loop/test_deferred_init.py

@@ -18,6 +18,7 @@ from vibe.core.agent_loop import AgentLoop
 from vibe.core.config import MCPStdio
 from vibe.core.telemetry.types import TerminalEmulator
 from vibe.core.tools.manager import ToolManager
+from vibe.core.tools.mcp import AuthStatus
 from vibe.core.tools.mcp.tools import RemoteTool
 
 
@@ -157,6 +158,23 @@ class TestIntegrateMcpIdempotency:
         # so a future call with servers would still run discovery.
         assert not manager._mcp_integrated
 
+    def test_no_servers_syncs_shared_registry_status(self) -> None:
+        config = build_test_vibe_config(
+            mcp_servers=[MCPStdio(name="srv", transport="stdio", command="echo")]
+        )
+        registry = FakeMCPRegistry()
+        manager = ToolManager(lambda: config, mcp_registry=registry, defer_mcp=True)
+
+        manager.integrate_mcp()
+        assert registry.status() == {"srv": AuthStatus.STDIO}
+
+        config = build_test_vibe_config(mcp_servers=[])
+        manager = ToolManager(lambda: config, mcp_registry=registry, defer_mcp=True)
+        manager.integrate_mcp()
+
+        assert registry.status() == {}
+        assert not manager._mcp_integrated
+
 
 class TestRefreshRemoteTools:
     @pytest.mark.asyncio

tests/autocompletion/test_path_completion_controller.py

@@ -24,7 +24,9 @@ class StubView(CompletionView):
     def clear_completion_suggestions(self) -> None:
         self.clears += 1
 
-    def replace_completion_range(self, start: int, end: int, replacement: str) -> None:
+    def replace_completion_range(
+        self, start: int, end: int, replacement: str, *, suppress_update: bool = False
+    ) -> None:
         self.replacements.append((start, end, replacement))
 
 

tests/autocompletion/test_slash_command_controller.py

@@ -40,7 +40,9 @@ class StubView(CompletionView):
     def clear_completion_suggestions(self) -> None:
         self.reset_count += 1
 
-    def replace_completion_range(self, start: int, end: int, replacement: str) -> None:
+    def replace_completion_range(
+        self, start: int, end: int, replacement: str, *, suppress_update: bool = False
+    ) -> None:
         self.replacements.append(Replacement(start, end, replacement))
 
 

tests/autocompletion/test_ui_chat_autocompletion.py

@@ -3,14 +3,13 @@ from __future__ import annotations
 from pathlib import Path
 
 import pytest
-from textual.content import Content
-from textual.style import Style
 from textual.widgets import Markdown
 
 from vibe.cli.textual_ui.app import VibeApp
 from vibe.cli.textual_ui.widgets.chat_input.completion_popup import (
     CompletionPopup,
     _CompletionItem,
+    _CompletionRow,
 )
 from vibe.cli.textual_ui.widgets.chat_input.container import ChatInputContainer
 
@@ -42,7 +41,7 @@ async def test_popup_hides_when_input_cleared(vibe_app: VibeApp) -> None:
 
 
 @pytest.mark.asyncio
-async def test_pressing_tab_writes_selected_command_and_leaves_popup_visible(
+async def test_pressing_tab_completes_command_and_hides_popup_when_exact_match(
     vibe_app: VibeApp,
 ) -> None:
     async with vibe_app.run_test() as pilot:
@@ -53,24 +52,18 @@ async def test_pressing_tab_writes_selected_command_and_leaves_popup_visible(
         await pilot.press("tab")
 
         assert chat_input.value == "/config"
-        assert popup.styles.display == "block"
+        assert popup.styles.display == "none"
 
 
 def ensure_selected_command(popup: CompletionPopup, expected_alias: str) -> None:
-    selected_aliases: list[str] = []
-    for item in popup.query(_CompletionItem):
-        renderable = item.render()
-        assert isinstance(renderable, Content)
-        content = str(renderable)
-        for span in renderable.spans:
-            style = span.style
-            if isinstance(style, Style) and style.reverse:
-                alias_text = content[span.start : span.end].strip()
-                alias = alias_text.split()[0] if alias_text else ""
-                selected_aliases.append(alias)
-
-    assert len(selected_aliases) == 1
-    assert selected_aliases[0] == expected_alias
+    selected_rows = [
+        row
+        for row in popup.query(_CompletionRow)
+        if row.has_class("completion-selected")
+    ]
+    assert len(selected_rows) == 1
+    command = selected_rows[0].query_one(".completion-command", _CompletionItem)
+    assert str(command.render()).strip() == expected_alias
 
 
 @pytest.mark.asyncio
@@ -297,20 +290,6 @@ async def test_finds_files_recursively_with_partial_path(
         assert popup.styles.display == "block"
 
 
-@pytest.mark.asyncio
-async def test_popup_is_positioned_near_cursor(vibe_app: VibeApp) -> None:
-    async with vibe_app.run_test() as pilot:
-        popup = vibe_app.query_one(CompletionPopup)
-
-        await pilot.press(*"/com")
-
-        assert popup.styles.display == "block"
-        offset = popup.styles.offset
-        # The popup should have an explicit offset set by _position_popup
-        assert offset.x is not None
-        assert offset.y is not None
-
-
 @pytest.mark.asyncio
 async def test_does_not_trigger_completion_when_navigating_history(
     file_tree: Path, vibe_app: VibeApp

tests/backend/data/anthropic.py

@@ -0,0 +1,132 @@
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from tests.backend.data import Chunk, JsonResponse
+
+
+def _sse_event(data: dict[str, Any]) -> Chunk:
+    return f"data: {json.dumps(data, separators=(',', ':'))}".encode()
+
+
+def anthropic_request_content_blocks(payload: dict[str, Any]) -> list[dict[str, Any]]:
+    # Flatten every structured content block across a request's messages
+    # (text/thinking/tool_use/tool_result blocks).
+    return [
+        block
+        for message in payload["messages"]
+        if isinstance(message["content"], list)
+        for block in message["content"]
+    ]
+
+
+def anthropic_message(
+    text: str,
+    *,
+    input_tokens: int = 12,
+    output_tokens: int = 3,
+    stop_reason: str = "end_turn",
+) -> JsonResponse:
+    return {
+        "content": [{"type": "text", "text": text}],
+        "usage": {"input_tokens": input_tokens, "output_tokens": output_tokens},
+        "stop_reason": stop_reason,
+    }
+
+
+def anthropic_tool_use(
+    name: str,
+    tool_input: dict[str, Any],
+    *,
+    tool_id: str = "toolu_1",
+    input_tokens: int = 20,
+    output_tokens: int = 5,
+) -> JsonResponse:
+    return {
+        "content": [
+            {"type": "tool_use", "id": tool_id, "name": name, "input": tool_input}
+        ],
+        "usage": {"input_tokens": input_tokens, "output_tokens": output_tokens},
+        "stop_reason": "tool_use",
+    }
+
+
+def anthropic_reasoning_tool_use_stream(
+    name: str,
+    arguments: str,
+    *,
+    reasoning: str = "thinking...",
+    signature: str = "sig",
+    tool_id: str = "toolu_1",
+    input_tokens: int = 20,
+    output_tokens: int = 5,
+) -> list[Chunk]:
+    return [
+        _sse_event(e)
+        for e in (
+            {
+                "type": "message_start",
+                "message": {"usage": {"input_tokens": input_tokens}},
+            },
+            {
+                "type": "content_block_start",
+                "index": 0,
+                "content_block": {"type": "thinking", "thinking": reasoning},
+            },
+            {
+                "type": "content_block_delta",
+                "index": 0,
+                "delta": {"type": "signature_delta", "signature": signature},
+            },
+            {"type": "content_block_stop", "index": 0},
+            {
+                "type": "content_block_start",
+                "index": 1,
+                "content_block": {"type": "tool_use", "id": tool_id, "name": name},
+            },
+            {
+                "type": "content_block_delta",
+                "index": 1,
+                "delta": {"type": "input_json_delta", "partial_json": arguments},
+            },
+            {"type": "content_block_stop", "index": 1},
+            {
+                "type": "message_delta",
+                "delta": {"stop_reason": "tool_use"},
+                "usage": {"output_tokens": output_tokens},
+            },
+            {"type": "message_stop"},
+        )
+    ]
+
+
+def anthropic_text_stream(
+    text: str, *, input_tokens: int = 12, output_tokens: int = 3
+) -> list[Chunk]:
+    return [
+        _sse_event(e)
+        for e in (
+            {
+                "type": "message_start",
+                "message": {"usage": {"input_tokens": input_tokens}},
+            },
+            {
+                "type": "content_block_start",
+                "index": 0,
+                "content_block": {"type": "text", "text": ""},
+            },
+            {
+                "type": "content_block_delta",
+                "index": 0,
+                "delta": {"type": "text_delta", "text": text},
+            },
+            {"type": "content_block_stop", "index": 0},
+            {
+                "type": "message_delta",
+                "delta": {"stop_reason": "end_turn"},
+                "usage": {"output_tokens": output_tokens},
+            },
+            {"type": "message_stop"},
+        )
+    ]

tests/backend/data/mistral.py

@@ -1,7 +1,41 @@
 from __future__ import annotations
 
+from typing import Any
+
 from tests.backend.data import Chunk, JsonResponse, ResultData, Url
 
+
+def mistral_completion(
+    content: str,
+    *,
+    prompt_tokens: int = 100,
+    completion_tokens: int = 50,
+    tool_calls: list[dict[str, Any]] | None = None,
+) -> JsonResponse:
+    return {
+        "id": "cmpl_test",
+        "created": 1234567890,
+        "model": "devstral-latest",
+        "object": "chat.completion",
+        "usage": {
+            "prompt_tokens": prompt_tokens,
+            "completion_tokens": completion_tokens,
+            "total_tokens": prompt_tokens + completion_tokens,
+        },
+        "choices": [
+            {
+                "index": 0,
+                "finish_reason": "tool_calls" if tool_calls else "stop",
+                "message": {
+                    "role": "assistant",
+                    "content": content,
+                    "tool_calls": tool_calls,
+                },
+            }
+        ],
+    }
+
+
 SIMPLE_CONVERSATION_PARAMS: list[tuple[Url, JsonResponse, ResultData]] = [
     (
         "https://api.mistral.ai",

tests/backend/data/openai_responses.py

@@ -4,8 +4,7 @@ import json
 from typing import Any
 
 from tests.backend.data import Chunk, JsonResponse, ResultData, Url
-
-OPENAI_RESPONSES_TEST_BASE_URL = "https://api.openai.com"
+from tests.constants import OPENAI_BASE_URL
 
 
 def _sse_event(data: dict[str, Any] | str) -> Chunk:
@@ -90,9 +89,134 @@ def _function_call_item(
     }
 
 
+def openai_message_item(
+    text: str, *, phase: str | None = None, message_id: str = "msg_1"
+) -> dict[str, Any]:
+    return _message_output_item(message_id, text, phase=phase)
+
+
+def openai_function_call_item(
+    name: str, arguments: str, *, item_id: str = "fc_1", call_id: str = "call_1"
+) -> dict[str, Any]:
+    return _function_call_item(item_id, call_id, name, arguments)
+
+
+def openai_response(
+    output: list[dict[str, Any]],
+    *,
+    input_tokens: int = 10,
+    output_tokens: int = 2,
+    response_id: str = "resp_1",
+    model: str = "gpt-test",
+) -> JsonResponse:
+    return {
+        "id": response_id,
+        "object": "response",
+        "created_at": 1234567890,
+        "model": model,
+        "output": output,
+        "usage": {
+            "input_tokens": input_tokens,
+            "output_tokens": output_tokens,
+            "total_tokens": input_tokens + output_tokens,
+        },
+    }
+
+
+def openai_sse(*events: dict[str, Any] | str) -> list[Chunk]:
+    return [_sse_event(event) for event in events]
+
+
+def openai_reasoning_tool_call_stream(
+    name: str,
+    arguments: str,
+    *,
+    reasoning: str = "thinking...",
+    call_id: str = "call_1",
+    item_id: str = "fc_1",
+    input_tokens: int = 20,
+    output_tokens: int = 5,
+) -> list[Chunk]:
+    return openai_sse(
+        {"type": "response.created", "response": {"id": "resp_1", "output": []}},
+        {
+            "type": "response.output_item.added",
+            "output_index": 0,
+            "item": _stream_message_item("msg_r", phase="commentary"),
+        },
+        {
+            "type": "response.reasoning_summary_text.delta",
+            "output_index": 0,
+            "delta": reasoning,
+        },
+        {
+            "type": "response.output_item.added",
+            "output_index": 1,
+            "item": _function_call_item(
+                item_id, call_id, name, "", status="in_progress"
+            ),
+        },
+        {
+            "type": "response.function_call_arguments.delta",
+            "output_index": 1,
+            "call_id": call_id,
+            "delta": arguments,
+        },
+        {
+            "type": "response.function_call_arguments.done",
+            "output_index": 1,
+            "call_id": call_id,
+            "name": name,
+            "arguments": arguments,
+        },
+        {
+            "type": "response.output_item.done",
+            "output_index": 1,
+            "item": _function_call_item(item_id, call_id, name, arguments),
+        },
+        {
+            "type": "response.completed",
+            "response": openai_response(
+                [_function_call_item(item_id, call_id, name, arguments)],
+                input_tokens=input_tokens,
+                output_tokens=output_tokens,
+            ),
+        },
+        "[DONE]",
+    )
+
+
+def openai_text_stream(
+    text: str, *, input_tokens: int = 10, output_tokens: int = 2
+) -> list[Chunk]:
+    return openai_sse(
+        {"type": "response.created", "response": {"id": "resp_1", "output": []}},
+        {
+            "type": "response.output_item.added",
+            "output_index": 0,
+            "item": _stream_message_item("msg_1"),
+        },
+        {
+            "type": "response.output_text.delta",
+            "output_index": 0,
+            "content_index": 0,
+            "delta": text,
+        },
+        {
+            "type": "response.completed",
+            "response": openai_response(
+                [openai_message_item(text)],
+                input_tokens=input_tokens,
+                output_tokens=output_tokens,
+            ),
+        },
+        "[DONE]",
+    )
+
+
 SIMPLE_CONVERSATION_PARAMS: list[tuple[Url, JsonResponse, ResultData]] = [
     (
-        OPENAI_RESPONSES_TEST_BASE_URL,
+        OPENAI_BASE_URL,
         {
             "id": "resp_fake_id_1234",
             "object": "response",
@@ -113,7 +237,7 @@ SIMPLE_CONVERSATION_PARAMS: list[tuple[Url, JsonResponse, ResultData]] = [
 
 TOOL_CONVERSATION_PARAMS: list[tuple[Url, JsonResponse, ResultData]] = [
     (
-        OPENAI_RESPONSES_TEST_BASE_URL,
+        OPENAI_BASE_URL,
         {
             "id": "resp_fake_id_9012",
             "object": "response",
@@ -144,7 +268,7 @@ TOOL_CONVERSATION_PARAMS: list[tuple[Url, JsonResponse, ResultData]] = [
 
 STREAMED_SIMPLE_CONVERSATION_PARAMS: list[tuple[Url, list[Chunk], list[ResultData]]] = [
     (
-        OPENAI_RESPONSES_TEST_BASE_URL,
+        OPENAI_BASE_URL,
         [
             _sse_event({
                 "type": "response.created",
@@ -235,7 +359,7 @@ STREAMED_SIMPLE_CONVERSATION_PARAMS: list[tuple[Url, list[Chunk], list[ResultDat
 
 COMMENTARY_CONVERSATION_PARAMS: list[tuple[Url, JsonResponse, ResultData]] = [
     (
-        OPENAI_RESPONSES_TEST_BASE_URL,
+        OPENAI_BASE_URL,
         {
             "id": "resp_thinking_1234",
             "object": "response",
@@ -268,7 +392,7 @@ STREAMED_COMMENTARY_CONVERSATION_PARAMS: list[
     tuple[Url, list[Chunk], list[ResultData]]
 ] = [
     (
-        OPENAI_RESPONSES_TEST_BASE_URL,
+        OPENAI_BASE_URL,
         [
             _sse_event({
                 "type": "response.created",
@@ -419,7 +543,7 @@ STREAMED_COMMENTARY_CONVERSATION_PARAMS: list[
 
 STREAMED_TOOL_CONVERSATION_PARAMS: list[tuple[Url, list[Chunk], list[ResultData]]] = [
     (
-        OPENAI_RESPONSES_TEST_BASE_URL,
+        OPENAI_BASE_URL,
         [
             _sse_event({
                 "type": "response.created",

tests/backend/test_anthropic_adapter.py

@@ -4,6 +4,7 @@ import json
 
 import pytest
 
+from tests.constants import ANTHROPIC_BASE_URL, ANTHROPIC_MESSAGES_PATH
 from vibe.core.config import ProviderConfig
 from vibe.core.llm.backend.anthropic import AnthropicAdapter, AnthropicMapper
 from vibe.core.types import (
@@ -30,7 +31,7 @@ def adapter():
 def provider():
     return ProviderConfig(
         name="anthropic",
-        api_base="https://api.anthropic.com",
+        api_base=ANTHROPIC_BASE_URL,
         api_key_env_var="ANTHROPIC_API_KEY",
         api_style="anthropic",
     )
@@ -229,75 +230,6 @@ class TestMapperParseResponse:
         assert chunk.usage.completion_tokens == 7
 
 
-class TestMapperStreamingEvents:
-    def test_text_delta(self, mapper):
-        chunk, idx = mapper.parse_streaming_event(
-            "content_block_delta",
-            {"delta": {"type": "text_delta", "text": "hi"}, "index": 0},
-            0,
-        )
-        assert chunk.message.content == "hi"
-
-    def test_thinking_delta(self, mapper):
-        chunk, _ = mapper.parse_streaming_event(
-            "content_block_delta",
-            {"delta": {"type": "thinking_delta", "thinking": "hmm"}, "index": 0},
-            0,
-        )
-        assert chunk.message.reasoning_content == "hmm"
-
-    def test_tool_use_start(self, mapper):
-        chunk, idx = mapper.parse_streaming_event(
-            "content_block_start",
-            {
-                "content_block": {"type": "tool_use", "id": "t1", "name": "search"},
-                "index": 2,
-            },
-            0,
-        )
-        assert chunk.message.tool_calls[0].id == "t1"
-        assert idx == 2
-
-    def test_input_json_delta(self, mapper):
-        chunk, _ = mapper.parse_streaming_event(
-            "content_block_delta",
-            {
-                "delta": {"type": "input_json_delta", "partial_json": '{"q":'},
-                "index": 1,
-            },
-            0,
-        )
-        assert chunk.message.tool_calls[0].function.arguments == '{"q":'
-
-    def test_message_start_usage(self, mapper):
-        chunk, _ = mapper.parse_streaming_event(
-            "message_start",
-            {"message": {"usage": {"input_tokens": 50, "cache_read_input_tokens": 10}}},
-            0,
-        )
-        assert chunk.usage.prompt_tokens == 60
-
-    def test_message_delta_usage(self, mapper):
-        chunk, _ = mapper.parse_streaming_event(
-            "message_delta", {"usage": {"output_tokens": 42}}, 0
-        )
-        assert chunk.usage.completion_tokens == 42
-
-    def test_unknown_event(self, mapper):
-        chunk, idx = mapper.parse_streaming_event("ping", {}, 5)
-        assert chunk is None
-        assert idx == 5
-
-    def test_signature_delta(self, mapper):
-        chunk, _ = mapper.parse_streaming_event(
-            "content_block_delta",
-            {"delta": {"type": "signature_delta", "signature": "sig"}, "index": 0},
-            0,
-        )
-        assert chunk is not None
-        assert chunk.message.reasoning_signature == "sig"
-
-
 class TestAdapterPrepareRequest:
     def test_basic(self, adapter, provider):
         messages = [LLMMessage(role=Role.user, content="Hello")]
@@ -316,7 +248,7 @@ class TestAdapterPrepareRequest:
         assert payload["model"] == "claude-sonnet-4-20250514"
         assert payload["max_tokens"] == 1024
         assert "temperature" not in payload
-        assert req.endpoint == "/v1/messages"
+        assert req.endpoint == ANTHROPIC_MESSAGES_PATH
         assert req.headers["anthropic-version"] == "2023-06-01"
 
     def test_beta_features(self, adapter, provider):

tests/backend/test_backend.py

@@ -36,8 +36,9 @@ from tests.backend.data.mistral import (
     STREAMED_TOOL_CONVERSATION_PARAMS as MISTRAL_STREAMED_TOOL_CONVERSATION_PARAMS,
     TOOL_CONVERSATION_PARAMS as MISTRAL_TOOL_CONVERSATION_PARAMS,
 )
+from tests.constants import CHAT_COMPLETIONS_PATH
 from vibe.core.config import ModelConfig, ProviderConfig
-from vibe.core.llm.backend.factory import BACKEND_FACTORY
+from vibe.core.llm.backend.factory import BACKEND_FACTORY, create_backend
 from vibe.core.llm.backend.generic import GenericBackend
 from vibe.core.llm.backend.mistral import MistralBackend, MistralMapper
 from vibe.core.llm.exceptions import BackendError, BackendErrorBuilder
@@ -71,7 +72,7 @@ class TestBackend:
         self, base_url: Url, json_response: JsonResponse, result_data: ResultData
     ):
         with respx.mock(base_url=base_url) as mock_api:
-            mock_api.post("/v1/chat/completions").mock(
+            mock_api.post(CHAT_COMPLETIONS_PATH).mock(
                 return_value=httpx.Response(status_code=200, json=json_response)
             )
             provider = ProviderConfig(
@@ -139,7 +140,7 @@ class TestBackend:
         self, base_url: Url, chunks: list[Chunk], result_data: list[ResultData]
     ):
         with respx.mock(base_url=base_url) as mock_api:
-            mock_api.post("/v1/chat/completions").mock(
+            mock_api.post(CHAT_COMPLETIONS_PATH).mock(
                 return_value=httpx.Response(
                     status_code=200,
                     stream=httpx.ByteStream(stream=b"\n\n".join(chunks)),
@@ -291,7 +292,7 @@ class TestBackend:
         response: httpx.Response,
     ):
         with respx.mock(base_url=base_url) as mock_api:
-            mock_api.post("/v1/chat/completions").mock(return_value=response)
+            mock_api.post(CHAT_COMPLETIONS_PATH).mock(return_value=response)
             provider = ProviderConfig(
                 name="provider_name",
                 api_base=f"{base_url}/v1",
@@ -335,7 +336,7 @@ class TestBackend:
         self, base_url: Url, provider_name: str, expected_stream_options: dict
     ):
         with respx.mock(base_url=base_url) as mock_api:
-            route = mock_api.post("/v1/chat/completions").mock(
+            route = mock_api.post(CHAT_COMPLETIONS_PATH).mock(
                 return_value=httpx.Response(
                     status_code=200,
                     stream=httpx.ByteStream(
@@ -399,7 +400,7 @@ class TestBackend:
             ],
         }
         with respx.mock(base_url=base_url) as mock_api:
-            mock_api.post("/v1/chat/completions").mock(
+            mock_api.post(CHAT_COMPLETIONS_PATH).mock(
                 return_value=httpx.Response(status_code=200, json=json_response)
             )
 
@@ -441,7 +442,7 @@ class TestBackend:
                 stream=httpx.ByteStream(stream=b"\n\n".join(chunks)),
                 headers={"Content-Type": "text/event-stream"},
             )
-            mock_api.post("/v1/chat/completions").mock(return_value=mock_response)
+            mock_api.post(CHAT_COMPLETIONS_PATH).mock(return_value=mock_response)
 
             provider = ProviderConfig(
                 name="provider_name",
@@ -470,13 +471,29 @@ class TestBackend:
 
 class TestMistralRetry:
     @staticmethod
-    def _create_test_backend() -> MistralBackend:
+    def _create_test_backend(
+        timeout: float = 720.0, retry_max_elapsed_time: float = 300.0
+    ) -> MistralBackend:
         provider = ProviderConfig(
             name="test_provider",
             api_base="https://api.mistral.ai/v1",
             api_key_env_var="API_KEY",
         )
-        return MistralBackend(provider=provider)
+        return MistralBackend(
+            provider=provider,
+            timeout=timeout,
+            retry_max_elapsed_time=retry_max_elapsed_time,
+        )
+
+    @staticmethod
+    def _build_fast_http_retry_config() -> RetryConfig:
+        return RetryConfig(
+            strategy="backoff",
+            backoff=BackoffStrategy(
+                initial_interval=1, max_interval=1, exponent=1, max_elapsed_time=10000
+            ),
+            retry_connection_errors=True,
+        )
 
     @pytest.mark.asyncio
     async def test_client_creation_includes_timeout_and_retry_config(self):
@@ -492,6 +509,61 @@ class TestMistralRetry:
             assert call_kwargs["retry_config"] is backend._retry_config
             assert "async_client" in call_kwargs
 
+    def test_retry_budget_uses_explicit_config(self):
+        backend = self._create_test_backend(
+            timeout=7200.0, retry_max_elapsed_time=1234.0
+        )
+
+        assert backend._timeout == 7200.0
+        assert backend._retry_config.backoff.max_elapsed_time == 1234000
+
+    def test_create_backend_passes_retry_budget(self):
+        provider = ProviderConfig(
+            name="test_provider",
+            api_base="https://api.mistral.ai/v1",
+            api_key_env_var="API_KEY",
+            backend=Backend.MISTRAL,
+        )
+
+        backend = create_backend(
+            provider=provider, timeout=7200.0, retry_max_elapsed_time=1234.0
+        )
+
+        assert isinstance(backend, MistralBackend)
+        assert backend._timeout == 7200.0
+        assert backend._retry_config.backoff.max_elapsed_time == 1234000
+
+    @pytest.mark.asyncio
+    async def test_complete_retries_retryable_http_error(self):
+        with respx.mock(base_url="https://api.mistral.ai") as mock_api:
+            route = mock_api.post("/v1/chat/completions").mock(
+                side_effect=[
+                    httpx.Response(status_code=502, text="Bad Gateway"),
+                    httpx.Response(
+                        status_code=200, json=MISTRAL_SIMPLE_CONVERSATION_PARAMS[0][1]
+                    ),
+                ]
+            )
+            backend = self._create_test_backend()
+            backend._retry_config = self._build_fast_http_retry_config()
+            model = ModelConfig(
+                name="model_name", provider="test_provider", alias="model_alias"
+            )
+            messages = [LLMMessage(role=Role.user, content="Just say hi")]
+
+            result = await backend.complete(
+                model=model,
+                messages=messages,
+                temperature=0.2,
+                tools=None,
+                max_tokens=None,
+                tool_choice=None,
+                extra_headers=None,
+            )
+
+            assert result.message.content == "Some content"
+            assert route.call_count == 2
+
 
 class TestMistralMapperPrepareMessage:
     """Tests for MistralMapper.prepare_message thinking-block handling.

tests/backend/test_image_encoding_cache.py

@@ -11,7 +11,7 @@ from vibe.core.llm.backend._image import (
     to_base64,
     to_data_uri,
 )
-from vibe.core.types import ImageAttachment
+from vibe.core.types import FileImageSource, ImageAttachment, InlineImageSource
 
 PNG_BYTES = b"\x89PNG\r\n\x1a\n" + b"\x00" * 16
 
@@ -24,7 +24,9 @@ def _clear_cache() -> None:
 def _att(tmp_path: Path, name: str = "shot.png") -> ImageAttachment:
     p = tmp_path / name
     p.write_bytes(PNG_BYTES)
-    return ImageAttachment(path=p, alias=name, mime_type="image/png")
+    return ImageAttachment(
+        source=FileImageSource(path=p), alias=name, mime_type="image/png"
+    )
 
 
 def test_repeated_calls_hit_cache_and_skip_disk(tmp_path: Path) -> None:
@@ -46,13 +48,14 @@ def test_cache_invalidates_when_file_mtime_changes(tmp_path: Path) -> None:
     to_base64(att)
     assert _encode_cached.cache_info().misses == 1
 
+    assert isinstance(att.source, FileImageSource)
     new_bytes = PNG_BYTES + b"\x01"
-    att.path.write_bytes(new_bytes)
+    att.source.path.write_bytes(new_bytes)
     # Force a distinct mtime even on coarse-resolution filesystems.
     import os
 
-    stat = att.path.stat()
-    os.utime(att.path, ns=(stat.st_atime_ns, stat.st_mtime_ns + 1_000_000))
+    stat = att.source.path.stat()
+    os.utime(att.source.path, ns=(stat.st_atime_ns, stat.st_mtime_ns + 1_000_000))
 
     refreshed = to_base64(att)
     assert refreshed == base64.b64encode(new_bytes).decode("ascii")
@@ -61,8 +64,23 @@ def test_cache_invalidates_when_file_mtime_changes(tmp_path: Path) -> None:
 
 def test_missing_file_raises_image_read_error(tmp_path: Path) -> None:
     att = ImageAttachment(
-        path=tmp_path / "nope.png", alias="nope.png", mime_type="image/png"
+        source=FileImageSource(path=tmp_path / "nope.png"),
+        alias="nope.png",
+        mime_type="image/png",
     )
 
     with pytest.raises(ImageReadError):
         to_data_uri(att)
+
+
+def test_inline_data_is_returned_without_touching_disk() -> None:
+    encoded = base64.b64encode(PNG_BYTES).decode("ascii")
+    att = ImageAttachment(
+        source=InlineImageSource(data=encoded),
+        alias="pasted.png",
+        mime_type="image/png",
+    )
+
+    assert to_base64(att) == encoded
+    assert to_data_uri(att) == f"data:image/png;base64,{encoded}"
+    assert _encode_cached.cache_info().misses == 0

tests/backend/test_image_mapping.py

@@ -14,7 +14,7 @@ from vibe.core.llm.backend.generic import OpenAIAdapter
 from vibe.core.llm.backend.mistral import MistralMapper
 from vibe.core.llm.backend.openai_responses import OpenAIResponsesAdapter
 from vibe.core.llm.backend.reasoning_adapter import ReasoningAdapter
-from vibe.core.types import ImageAttachment, LLMMessage, Role
+from vibe.core.types import FileImageSource, ImageAttachment, LLMMessage, Role
 
 PNG_BYTES = b"\x89PNG\r\n\x1a\n" + b"\x00" * 16
 EXPECTED_B64 = base64.b64encode(PNG_BYTES).decode("ascii")
@@ -25,7 +25,9 @@ EXPECTED_DATA_URI = f"data:image/png;base64,{EXPECTED_B64}"
 def image_attachment(tmp_path: Path) -> ImageAttachment:
     path = tmp_path / "shot.png"
     path.write_bytes(PNG_BYTES)
-    return ImageAttachment(path=path, alias="shot.png", mime_type="image/png")
+    return ImageAttachment(
+        source=FileImageSource(path=path), alias="shot.png", mime_type="image/png"
+    )
 
 
 def _user_message(image: ImageAttachment) -> LLMMessage:

tests/backend/test_openai_responses_adapter.py

@@ -19,13 +19,13 @@ import respx
 from tests.backend.data import Chunk, JsonResponse, ResultData, Url
 from tests.backend.data.openai_responses import (
     COMMENTARY_CONVERSATION_PARAMS,
-    OPENAI_RESPONSES_TEST_BASE_URL,
     SIMPLE_CONVERSATION_PARAMS,
     STREAMED_COMMENTARY_CONVERSATION_PARAMS,
     STREAMED_SIMPLE_CONVERSATION_PARAMS,
     STREAMED_TOOL_CONVERSATION_PARAMS,
     TOOL_CONVERSATION_PARAMS,
 )
+from tests.constants import OPENAI_BASE_URL, OPENAI_RESPONSES_PATH
 from vibe.core.config import ModelConfig, ProviderConfig
 from vibe.core.llm.backend.generic import GenericBackend
 from vibe.core.llm.backend.openai_responses import OpenAIResponsesAdapter
@@ -55,7 +55,7 @@ def model():
     return _make_model()
 
 
-def _make_provider(base_url: Url = OPENAI_RESPONSES_TEST_BASE_URL) -> ProviderConfig:
+def _make_provider(base_url: Url = OPENAI_BASE_URL) -> ProviderConfig:
     return ProviderConfig(
         name="openai",
         api_base=f"{base_url}/v1",
@@ -68,7 +68,7 @@ def _make_model() -> ModelConfig:
     return ModelConfig(name="gpt-4o", provider="openai", alias="gpt-4o")
 
 
-def _make_backend(base_url: Url = OPENAI_RESPONSES_TEST_BASE_URL) -> GenericBackend:
+def _make_backend(base_url: Url = OPENAI_BASE_URL) -> GenericBackend:
     return GenericBackend(provider=_make_provider(base_url))
 
 
@@ -1229,7 +1229,7 @@ class TestGenericBackendIntegration:
         self, base_url: Url, json_response: JsonResponse, result_data: ResultData
     ):
         with respx.mock(base_url=base_url) as mock_api:
-            mock_api.post("/v1/responses").mock(
+            mock_api.post(OPENAI_RESPONSES_PATH).mock(
                 return_value=httpx.Response(status_code=200, json=json_response)
             )
             backend = _make_backend(base_url)
@@ -1261,7 +1261,7 @@ class TestGenericBackendIntegration:
         self, base_url: Url, chunks: list[Chunk], result_data: list[ResultData]
     ):
         with respx.mock(base_url=base_url) as mock_api:
-            mock_api.post("/v1/responses").mock(
+            mock_api.post(OPENAI_RESPONSES_PATH).mock(
                 return_value=httpx.Response(
                     status_code=200,
                     stream=httpx.ByteStream(stream=b"\n\n".join(chunks)),
@@ -1289,9 +1289,9 @@ class TestGenericBackendIntegration:
 
     @pytest.mark.asyncio
     async def test_streaming_payload_includes_stream_flag(self):
-        base_url = OPENAI_RESPONSES_TEST_BASE_URL
+        base_url = OPENAI_BASE_URL
         with respx.mock(base_url=base_url) as mock_api:
-            route = mock_api.post("/v1/responses").mock(
+            route = mock_api.post(OPENAI_RESPONSES_PATH).mock(
                 return_value=httpx.Response(
                     status_code=200,
                     stream=httpx.ByteStream(

tests/cli/plan_offer/test_decide_plan_offer.py

@@ -176,6 +176,27 @@ def test_resolve_api_key_for_plan_with_missing_env_var() -> None:
         environ["MISTRAL_API_KEY"] = previous_api_key
 
 
+def test_resolve_api_key_for_plan_falls_back_to_keyring(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("MISTRAL_API_KEY", raising=False)
+    keyring_api_key = "keyring_mistral_api_key"
+    monkeypatch.setattr(
+        "vibe.core.config._settings.keyring.get_password",
+        lambda service, username: keyring_api_key,
+    )
+
+    provider = ProviderConfig(
+        name="test_mistral",
+        api_base="https://api.mistral.ai",
+        backend=Backend.MISTRAL,
+        api_key_env_var="MISTRAL_API_KEY",
+    )
+
+    result = resolve_api_key_for_plan(provider)
+    assert result == keyring_api_key
+
+
 @pytest.mark.parametrize(
     ("plan_info", "expected_cta"),
     [

tests/cli/test_cache.py

@@ -1,73 +0,0 @@
-from __future__ import annotations
-
-from pathlib import Path
-import tomllib
-
-from vibe.cli.cache import read_cache, write_cache
-
-
-class TestReadCache:
-    def test_reads_valid_toml(self, tmp_path: Path) -> None:
-        cache_path = tmp_path / "cache.toml"
-        cache_path.write_text('[update_cache]\nlatest_version = "1.0.0"\n')
-
-        result = read_cache(cache_path)
-
-        assert result["update_cache"]["latest_version"] == "1.0.0"
-
-    def test_returns_empty_dict_when_missing(self, tmp_path: Path) -> None:
-        assert read_cache(tmp_path / "missing.toml") == {}
-
-    def test_returns_empty_dict_when_corrupted(self, tmp_path: Path) -> None:
-        cache_path = tmp_path / "cache.toml"
-        cache_path.write_text("{bad toml")
-
-        assert read_cache(cache_path) == {}
-
-
-class TestWriteCache:
-    def test_writes_new_file(self, tmp_path: Path) -> None:
-        cache_path = tmp_path / "cache.toml"
-
-        write_cache(cache_path, "feedback", {"last_shown_at": 100.0})
-
-        with cache_path.open("rb") as f:
-            data = tomllib.load(f)
-        assert data["feedback"]["last_shown_at"] == 100.0
-
-    def test_merges_with_existing(self, tmp_path: Path) -> None:
-        cache_path = tmp_path / "cache.toml"
-        cache_path.write_text('[update_cache]\nlatest_version = "1.0.0"\n')
-
-        write_cache(cache_path, "feedback", {"last_shown_at": 200.0})
-
-        with cache_path.open("rb") as f:
-            data = tomllib.load(f)
-        assert data["update_cache"]["latest_version"] == "1.0.0"
-        assert data["feedback"]["last_shown_at"] == 200.0
-
-    def test_merges_within_section_and_leaves_other_sections_alone(
-        self, tmp_path: Path
-    ) -> None:
-        cache_path = tmp_path / "cache.toml"
-        cache_path.write_text(
-            "[update_cache]\n"
-            'latest_version = "1.0.0"\n'
-            "stored_at_timestamp = 1\n"
-            'seen_whats_new_version = "1.0.0"\n\n'
-            "[feedback]\n"
-            "last_shown_at = 100.0\n"
-        )
-
-        write_cache(
-            cache_path,
-            "update_cache",
-            {"latest_version": "2.0.0", "stored_at_timestamp": 2},
-        )
-
-        with cache_path.open("rb") as f:
-            data = tomllib.load(f)
-        assert data["update_cache"]["latest_version"] == "2.0.0"
-        assert data["update_cache"]["stored_at_timestamp"] == 2
-        assert data["update_cache"]["seen_whats_new_version"] == "1.0.0"
-        assert data["feedback"]["last_shown_at"] == 100.0

tests/cli/test_commands.py

@@ -128,6 +128,14 @@ class TestCommandRegistry:
         result = registry.parse_command("/connectors filesystem")
         assert result == ("mcp", registry.commands["mcp"], "filesystem")
 
+    def test_mcp_command_description_surfaces_auth_subcommands(self) -> None:
+        registry = CommandRegistry()
+        command = registry.commands["mcp"]
+
+        assert "status" in command.description
+        assert "login <alias>" in command.description
+        assert "logout <alias>" in command.description
+
     def test_data_retention_command_registration(self) -> None:
         registry = CommandRegistry()
         result = registry.parse_command("/data-retention")

tests/cli/test_feedback_bar_manager.py

@@ -6,6 +6,7 @@ import tomllib
 from unittest.mock import MagicMock, patch
 
 from vibe.cli.textual_ui.widgets.feedback_bar_manager import FeedbackBarManager
+from vibe.core.cache_store import FileSystemVibeCodeCacheStore
 from vibe.core.feedback import (
     _CACHE_SECTION,
     _LAST_SHOWN_KEY,
@@ -15,24 +16,18 @@ from vibe.core.feedback import (
 from vibe.core.types import LLMMessage, Role
 
 
-def _patch_cache_file(tmp_path: Path):
-    from vibe.core.paths._vibe_home import GlobalPath
-
-    return patch(
-        "vibe.core.feedback.CACHE_FILE", GlobalPath(lambda: tmp_path / "cache.toml")
-    )
-
-
 def _patch_probability(value: float):
     return patch("vibe.core.feedback.FEEDBACK_PROBABILITY", value)
 
 
 def _make_agent_loop(
+    cache_path: Path,
     user_message_count: int = MIN_USER_MESSAGES_FOR_FEEDBACK,
     telemetry_active: bool = True,
 ) -> MagicMock:
     loop = MagicMock()
     loop.telemetry_client.is_active.return_value = telemetry_active
+    loop.cache_store = FileSystemVibeCodeCacheStore(cache_path)
     messages = [
         LLMMessage(role=Role.user, content=f"msg {i}")
         for i in range(user_message_count)
@@ -45,20 +40,22 @@ class TestShouldShow:
     def test_shows_when_conditions_met(self, tmp_path: Path) -> None:
         manager = FeedbackBarManager()
         with (
-            _patch_cache_file(tmp_path),
             _patch_probability(0.2),
             patch("vibe.core.feedback.random.random", return_value=0.0),
         ):
-            assert manager.should_show(_make_agent_loop()) is True
+            assert (
+                manager.should_show(_make_agent_loop(tmp_path / "cache.toml")) is True
+            )
 
     def test_does_not_show_when_random_misses(self, tmp_path: Path) -> None:
         manager = FeedbackBarManager()
         with (
-            _patch_cache_file(tmp_path),
             _patch_probability(0.2),
             patch("vibe.core.feedback.random.random", return_value=1.0),
         ):
-            assert manager.should_show(_make_agent_loop()) is False
+            assert (
+                manager.should_show(_make_agent_loop(tmp_path / "cache.toml")) is False
+            )
 
     def test_does_not_show_within_cooldown(self, tmp_path: Path) -> None:
         (tmp_path / "cache.toml").write_text(
@@ -66,11 +63,12 @@ class TestShouldShow:
         )
         manager = FeedbackBarManager()
         with (
-            _patch_cache_file(tmp_path),
             _patch_probability(0.2),
             patch("vibe.core.feedback.random.random", return_value=0.0),
         ):
-            assert manager.should_show(_make_agent_loop()) is False
+            assert (
+                manager.should_show(_make_agent_loop(tmp_path / "cache.toml")) is False
+            )
 
     def test_shows_after_cooldown_expires(self, tmp_path: Path) -> None:
         (tmp_path / "cache.toml").write_text(
@@ -78,30 +76,38 @@ class TestShouldShow:
         )
         manager = FeedbackBarManager()
         with (
-            _patch_cache_file(tmp_path),
             _patch_probability(0.2),
             patch("vibe.core.feedback.random.random", return_value=0.0),
         ):
-            assert manager.should_show(_make_agent_loop()) is True
+            assert (
+                manager.should_show(_make_agent_loop(tmp_path / "cache.toml")) is True
+            )
 
     def test_does_not_show_when_telemetry_inactive(self, tmp_path: Path) -> None:
         manager = FeedbackBarManager()
-        with _patch_cache_file(tmp_path), _patch_probability(0.2):
+        with _patch_probability(0.2):
             assert (
-                manager.should_show(_make_agent_loop(telemetry_active=False)) is False
+                manager.should_show(
+                    _make_agent_loop(tmp_path / "cache.toml", telemetry_active=False)
+                )
+                is False
             )
 
     def test_does_not_show_when_too_few_user_messages(self, tmp_path: Path) -> None:
         manager = FeedbackBarManager()
         with (
-            _patch_cache_file(tmp_path),
             _patch_probability(0.2),
             patch("vibe.core.feedback.random.random", return_value=0.0),
         ):
-            assert manager.should_show(_make_agent_loop(user_message_count=1)) is False
+            assert (
+                manager.should_show(
+                    _make_agent_loop(tmp_path / "cache.toml", user_message_count=1)
+                )
+                is False
+            )
 
     def test_skips_injected_messages_in_count(self, tmp_path: Path) -> None:
-        loop = _make_agent_loop(user_message_count=0)
+        loop = _make_agent_loop(tmp_path / "cache.toml", user_message_count=0)
         loop.messages = [
             LLMMessage(role=Role.user, content="real"),
             LLMMessage(role=Role.user, content="injected", injected=True),
@@ -109,7 +115,6 @@ class TestShouldShow:
         ]
         manager = FeedbackBarManager()
         with (
-            _patch_cache_file(tmp_path),
             _patch_probability(0.2),
             patch("vibe.core.feedback.random.random", return_value=0.0),
         ):
@@ -121,8 +126,7 @@ class TestRecordFeedbackAsked:
     def test_writes_timestamp_to_cache(self, tmp_path: Path) -> None:
         manager = FeedbackBarManager()
         before = int(time.time())
-        with _patch_cache_file(tmp_path):
-            manager.record_feedback_asked()
-            with (tmp_path / "cache.toml").open("rb") as f:
-                data = tomllib.load(f)
+        manager.record_feedback_asked(_make_agent_loop(tmp_path / "cache.toml"))
+        with (tmp_path / "cache.toml").open("rb") as f:
+            data = tomllib.load(f)
         assert data[_CACHE_SECTION][_LAST_SHOWN_KEY] >= before

tests/cli/test_help.py

@@ -16,13 +16,41 @@ def test_help_shows_auto_approve_flag(
     assert exc_info.value.code == 0
     output = capsys.readouterr().out
     assert "--auto-approve" in output
+    assert "--yolo" in output
     assert "Shortcut for --agent auto-approve" in output
 
 
-def test_auto_approve_conflicts_with_agent(
+def test_help_shows_check_upgrade_flag(
     monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
 ) -> None:
-    monkeypatch.setattr("sys.argv", ["vibe", "--agent", "plan", "--auto-approve"])
+    monkeypatch.setattr("sys.argv", ["vibe", "--help"])
+
+    with pytest.raises(SystemExit) as exc_info:
+        parse_arguments()
+
+    assert exc_info.value.code == 0
+    output = capsys.readouterr().out
+    assert "--check-upgrade" in output
+    assert "Check for a Vibe update now" in output
+
+
+# def test_auto_approve_conflicts_with_agent(
+#     monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+
+
+def test_yolo_alias_selects_auto_approve(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setattr("sys.argv", ["vibe", "--yolo"])
+
+    args = parse_arguments()
+
+    assert args.auto_approve is True
+
+
+@pytest.mark.parametrize("flag", ["--auto-approve", "--yolo"])
+def test_auto_approve_aliases_conflict_with_agent(
+    flag: str, monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+) -> None:
+    monkeypatch.setattr("sys.argv", ["vibe", "--agent", "plan", flag])
 
     with pytest.raises(SystemExit) as exc_info:
         parse_arguments()

tests/cli/test_programmatic_setup.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import argparse
+from collections.abc import Callable
 from pathlib import Path
 
 import pytest
@@ -21,6 +22,7 @@ def _make_args(**overrides: object) -> argparse.Namespace:
         "enabled_tools": None,
         "output": "text",
         "agent": "default",
+        "check_upgrade": False,
         "setup": False,
         "workdir": None,
         "add_dir": [],
@@ -134,16 +136,15 @@ def test_trust_flag_trusts_cwd_for_session_only(
 
     args = _make_args(trust=True, prompt=None)
     monkeypatch.setattr(entrypoint_mod, "parse_arguments", lambda: args)
-    monkeypatch.setattr(
-        entrypoint_mod, "check_and_resolve_trusted_folder", lambda _cwd: None
-    )
     monkeypatch.setattr(
         entrypoint_mod, "init_harness_files_manager", lambda *a, **k: None
     )
+
     # Stop main() before it runs the actual CLI.
-    monkeypatch.setattr(
-        "vibe.cli.cli.run_cli", lambda _args: (_ for _ in ()).throw(SystemExit(0))
-    )
+    def fake_run_cli(_args: argparse.Namespace, **_kwargs: object) -> None:
+        raise SystemExit(0)
+
+    monkeypatch.setattr("vibe.cli.cli.run_cli", fake_run_cli)
 
     with pytest.raises(SystemExit) as exc_info:
         entrypoint_mod.main()
@@ -172,9 +173,11 @@ def test_trust_flag_works_in_programmatic_mode(
     monkeypatch.setattr(
         entrypoint_mod, "init_harness_files_manager", lambda *a, **k: None
     )
-    monkeypatch.setattr(
-        "vibe.cli.cli.run_cli", lambda _args: (_ for _ in ()).throw(SystemExit(0))
-    )
+
+    def fake_run_cli(_args: argparse.Namespace, **_kwargs: object) -> None:
+        raise SystemExit(0)
+
+    monkeypatch.setattr("vibe.cli.cli.run_cli", fake_run_cli)
 
     with pytest.raises(SystemExit):
         entrypoint_mod.main()
@@ -183,6 +186,78 @@ def test_trust_flag_works_in_programmatic_mode(
     assert trusted_folders_manager._trusted == []
 
 
+def test_check_upgrade_does_not_pass_trust_resolver(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    project = tmp_path / "proj"
+    project.mkdir()
+    (project / "AGENTS.md").write_text("hello", encoding="utf-8")
+    monkeypatch.chdir(project)
+
+    args = _make_args(prompt=None, check_upgrade=True)
+    monkeypatch.setattr(entrypoint_mod, "parse_arguments", lambda: args)
+    monkeypatch.setattr(
+        entrypoint_mod,
+        "check_and_resolve_trusted_folder",
+        lambda _cwd: pytest.fail("check-upgrade must not prompt for trust"),
+    )
+    monkeypatch.setattr(
+        entrypoint_mod, "init_harness_files_manager", lambda *a, **k: None
+    )
+
+    def fake_run_cli(
+        _args: argparse.Namespace,
+        *,
+        resolve_trusted_folder: Callable[[], None] | None = None,
+    ) -> None:
+        assert resolve_trusted_folder is None
+        raise SystemExit(0)
+
+    monkeypatch.setattr("vibe.cli.cli.run_cli", fake_run_cli)
+
+    with pytest.raises(SystemExit) as exc_info:
+        entrypoint_mod.main()
+
+    assert exc_info.value.code == 0
+
+
+def test_interactive_start_passes_trust_resolver_to_cli(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    project = tmp_path / "proj"
+    project.mkdir()
+    monkeypatch.chdir(project)
+
+    args = _make_args(prompt=None)
+    calls: list[str] = []
+    monkeypatch.setattr(entrypoint_mod, "parse_arguments", lambda: args)
+    monkeypatch.setattr(
+        entrypoint_mod,
+        "check_and_resolve_trusted_folder",
+        lambda _cwd: calls.append("trust"),
+    )
+    monkeypatch.setattr(
+        entrypoint_mod, "init_harness_files_manager", lambda *a, **k: None
+    )
+
+    def fake_run_cli(
+        _args: argparse.Namespace,
+        *,
+        resolve_trusted_folder: Callable[[], None] | None = None,
+    ) -> None:
+        assert callable(resolve_trusted_folder)
+        resolve_trusted_folder()
+        raise SystemExit(0)
+
+    monkeypatch.setattr("vibe.cli.cli.run_cli", fake_run_cli)
+
+    with pytest.raises(SystemExit) as exc_info:
+        entrypoint_mod.main()
+
+    assert exc_info.value.code == 0
+    assert calls == ["trust"]
+
+
 def test_session_trust_does_not_write_to_disk(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ) -> None:
@@ -224,3 +299,60 @@ def test_run_cli_passes_max_tokens_to_run_programmatic(
 
     assert exc_info.value.code == 0
     assert call["max_session_tokens"] == 123
+
+
+def test_run_cli_runs_update_prompt_before_trust_resolver(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    args = _make_args(prompt=None)
+    config = build_test_vibe_config()
+    calls: list[str] = []
+
+    monkeypatch.setattr(cli_mod, "bootstrap_config_files", lambda: None)
+    monkeypatch.setattr(cli_mod, "load_config_or_exit", lambda interactive: config)
+    monkeypatch.setattr(
+        cli_mod,
+        "_maybe_run_startup_update_prompt",
+        lambda _config, _repository: calls.append("update"),
+    )
+
+    def resolve_trusted_folder() -> None:
+        calls.append("trust")
+        raise SystemExit(0)
+
+    with pytest.raises(SystemExit) as exc_info:
+        cli_mod.run_cli(args, resolve_trusted_folder=resolve_trusted_folder)
+
+    assert exc_info.value.code == 0
+    assert calls == ["update", "trust"]
+
+
+def test_run_cli_check_upgrade_exits_before_loading_config(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    args = _make_args(prompt=None, check_upgrade=True)
+    call: dict[str, object] = {}
+
+    monkeypatch.setattr(cli_mod, "bootstrap_config_files", lambda: None)
+    monkeypatch.setattr(
+        cli_mod,
+        "load_config_or_exit",
+        lambda interactive: pytest.fail("check-upgrade should not load config"),
+    )
+    monkeypatch.setattr(cli_mod, "load_update_prompt_theme", lambda: "dracula")
+
+    def fake_run_check_upgrade(_repository: object, *, theme: str | None) -> None:
+        call["theme"] = theme
+
+    monkeypatch.setattr(cli_mod, "_run_check_upgrade", fake_run_check_upgrade)
+
+    with pytest.raises(SystemExit) as exc_info:
+        cli_mod.run_cli(
+            args,
+            resolve_trusted_folder=lambda: pytest.fail(
+                "check-upgrade should not prompt for trust"
+            ),
+        )
+
+    assert exc_info.value.code == 0
+    assert call["theme"] == "dracula"

tests/cli/test_rename_command.py

@@ -2,7 +2,6 @@ from __future__ import annotations
 
 import json
 from pathlib import Path
-from unittest.mock import MagicMock, patch
 
 import pytest
 
@@ -114,7 +113,7 @@ async def test_resume_picker_shows_renamed_session_title(
 
     assert handled is True
     assert captured_picker is not None
-    assert captured_picker._latest_messages[f"local:{logger.session_id}"] == "New title"
+    assert captured_picker._latest_messages[logger.session_id] == "New title"
 
 
 @pytest.mark.asyncio
@@ -129,34 +128,3 @@ async def test_rename_command_requires_title(tmp_path: Path) -> None:
         assert any(error._error == "Usage: /rename <title>" for error in errors)
 
     assert handled is True
-
-
-@pytest.mark.asyncio
-async def test_rename_command_is_intercepted_for_remote_sessions(
-    tmp_path: Path,
-) -> None:
-    config = build_test_vibe_config(session_logging=_enabled_session_config(tmp_path))
-    app = build_test_vibe_app(config=config)
-
-    async with app.run_test() as pilot:
-        with patch(
-            "vibe.cli.textual_ui.remote.remote_session_manager.RemoteEventsSource"
-        ) as MockSource:
-            remote_source = MagicMock()
-            remote_source.session_id = "remote-session-id"
-            remote_source.is_terminated = False
-            remote_source.is_waiting_for_input = False
-            MockSource.return_value = remote_source
-
-            await app._remote_manager.attach(
-                session_id="remote-session-id", config=config
-            )
-            handled = await app._handle_command("/rename Remote title")
-        await pilot.pause()
-        errors = app.query(ErrorMessage)
-        assert any(
-            error._error == "Renaming is only supported for local sessions."
-            for error in errors
-        )
-
-    assert handled is True

tests/cli/test_session_delete.py

@@ -57,13 +57,11 @@ async def test_session_delete_request_deletes_local_session(
     monkeypatch.setattr(app, "_mount_and_scroll", mount_and_scroll)
 
     await app.on_session_picker_app_session_delete_requested(
-        SessionPickerApp.SessionDeleteRequested(
-            "local:deleted-session", "local", "deleted-session"
-        )
+        SessionPickerApp.SessionDeleteRequested("deleted-session", "deleted-session")
     )
 
     assert deleted_sessions == [("deleted-session", config.session_logging)]
-    assert picker.removed_option_ids == ["local:deleted-session"]
+    assert picker.removed_option_ids == ["deleted-session"]
     assert any(
         isinstance(widget, UserCommandMessage)
         and widget._content
@@ -96,13 +94,11 @@ async def test_session_delete_request_keeps_picker_on_delete_error(
     monkeypatch.setattr(app, "_mount_and_scroll", mount_and_scroll)
 
     await app.on_session_picker_app_session_delete_requested(
-        SessionPickerApp.SessionDeleteRequested(
-            "local:deleted-session", "local", "deleted-session"
-        )
+        SessionPickerApp.SessionDeleteRequested("deleted-session", "deleted-session")
     )
 
     assert picker.removed_option_ids == []
-    assert picker.cleared_pending_option_ids == ["local:deleted-session"]
+    assert picker.cleared_pending_option_ids == ["deleted-session"]
     assert any(
         isinstance(widget, ErrorMessage)
         and widget._error == "Failed to delete session: disk said no"
@@ -140,13 +136,11 @@ async def test_session_delete_request_returns_to_input_when_picker_becomes_empty
     monkeypatch.setattr(app, "_switch_to_input_app", switch_to_input)
 
     await app.on_session_picker_app_session_delete_requested(
-        SessionPickerApp.SessionDeleteRequested(
-            "local:deleted-session", "local", "deleted-session"
-        )
+        SessionPickerApp.SessionDeleteRequested("deleted-session", "deleted-session")
     )
 
     assert switched_to_input is True
-    assert picker.removed_option_ids == ["local:deleted-session"]
+    assert picker.removed_option_ids == ["deleted-session"]
     assert [
         widget._content
         for widget in mounted_widgets
@@ -157,40 +151,6 @@ async def test_session_delete_request_returns_to_input_when_picker_becomes_empty
     ]
 
 
-@pytest.mark.asyncio
-async def test_session_delete_request_rejects_remote_sessions(
-    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
-) -> None:
-    config = build_test_vibe_config(
-        session_logging=_enabled_session_config(tmp_path), enable_connectors=False
-    )
-    app = build_test_vibe_app(config=config)
-    mounted_widgets: list[object] = []
-
-    async def delete_session(
-        session_id: str, session_config: SessionLoggingConfig
-    ) -> None:
-        pytest.fail("remote sessions should not be deleted")
-
-    async def mount_and_scroll(widget: object, after: object | None = None) -> None:
-        mounted_widgets.append(widget)
-
-    monkeypatch.setattr("vibe.cli.textual_ui.app.delete_saved_session", delete_session)
-    monkeypatch.setattr(app, "_mount_and_scroll", mount_and_scroll)
-
-    await app.on_session_picker_app_session_delete_requested(
-        SessionPickerApp.SessionDeleteRequested(
-            "remote:remote-session", "remote", "remote-session"
-        )
-    )
-
-    assert any(
-        isinstance(widget, ErrorMessage)
-        and widget._error == "Deleting remote sessions is not supported."
-        for widget in mounted_widgets
-    )
-
-
 @pytest.mark.asyncio
 async def test_session_delete_request_rejects_current_session(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
@@ -217,12 +177,12 @@ async def test_session_delete_request_rejects_current_session(
 
     await app.on_session_picker_app_session_delete_requested(
         SessionPickerApp.SessionDeleteRequested(
-            f"local:{app.agent_loop.session_id}", "local", app.agent_loop.session_id
+            app.agent_loop.session_id, app.agent_loop.session_id
         )
     )
 
     assert deleted_sessions == []
-    assert picker.cleared_pending_option_ids == [f"local:{app.agent_loop.session_id}"]
+    assert picker.cleared_pending_option_ids == [app.agent_loop.session_id]
     assert any(
         isinstance(widget, ErrorMessage)
         and widget._error == "Deleting the current session is not supported."

tests/cli/test_startup_update_prompt.py

@@ -8,9 +8,20 @@ from unittest.mock import patch
 import pytest
 
 from tests.conftest import build_test_vibe_config
-from vibe.cli.cli import _maybe_run_startup_update_prompt
-from vibe.cli.update_notifier import FileSystemUpdateCacheRepository, UpdateCache
-from vibe.setup.update_prompt.update_prompt_dialog import UpdatePromptResult
+from tests.update_notifier.adapters.fake_update_gateway import FakeUpdateGateway
+from vibe import __version__
+from vibe.cli.cli import _maybe_run_startup_update_prompt, _run_check_upgrade
+from vibe.cli.update_notifier import (
+    FileSystemUpdateCacheRepository,
+    Update,
+    UpdateCache,
+    UpdateGatewayCause,
+    UpdateGatewayError,
+)
+from vibe.setup.update_prompt.update_prompt_dialog import (
+    UpdatePromptMode,
+    UpdatePromptResult,
+)
 
 
 class _BrokenRepository:
@@ -129,6 +140,76 @@ def test_no_op_when_cache_read_raises_oserror() -> None:
     mock_ask.assert_not_called()
 
 
+def test_check_upgrade_fetches_and_prompts_when_update_is_available(
+    repository: FileSystemUpdateCacheRepository,
+) -> None:
+    notifier = FakeUpdateGateway(update=Update(latest_version="999.0.0"))
+
+    with patch(
+        "vibe.cli.cli.ask_update_prompt", return_value=UpdatePromptResult.CONTINUE
+    ) as mock_ask:
+        _run_check_upgrade(repository, update_notifier=notifier, theme="textual-light")
+
+    mock_ask.assert_called_once_with(
+        __version__,
+        "999.0.0",
+        theme="textual-light",
+        prompt_mode=UpdatePromptMode.CHECK_UPGRADE,
+    )
+    assert notifier.fetch_update_calls == 1
+
+
+def test_check_upgrade_cancel_does_not_dismiss_update(
+    repository: FileSystemUpdateCacheRepository,
+) -> None:
+    config = build_test_vibe_config(enable_update_checks=True)
+    notifier = FakeUpdateGateway(update=Update(latest_version="999.0.0"))
+
+    with patch(
+        "vibe.cli.cli.ask_update_prompt", return_value=UpdatePromptResult.CONTINUE
+    ):
+        _run_check_upgrade(repository, update_notifier=notifier)
+
+    cache = asyncio.run(repository.get())
+    assert cache is not None
+    assert cache.dismissed_version is None
+
+    with patch(
+        "vibe.cli.cli.ask_update_prompt", return_value=UpdatePromptResult.CONTINUE
+    ) as mock_ask:
+        _maybe_run_startup_update_prompt(config, repository)
+
+    mock_ask.assert_called_once()
+
+
+def test_check_upgrade_prints_up_to_date_when_no_update_exists(
+    repository: FileSystemUpdateCacheRepository, capsys: pytest.CaptureFixture[str]
+) -> None:
+    notifier = FakeUpdateGateway(update=None)
+
+    with patch("vibe.cli.cli.ask_update_prompt") as mock_ask:
+        _run_check_upgrade(repository, update_notifier=notifier)
+
+    mock_ask.assert_not_called()
+    out = capsys.readouterr().out
+    assert "already up to date" in out
+    assert __version__ in out
+
+
+def test_check_upgrade_exits_one_when_gateway_errors(
+    repository: FileSystemUpdateCacheRepository, capsys: pytest.CaptureFixture[str]
+) -> None:
+    notifier = FakeUpdateGateway(
+        error=UpdateGatewayError(cause=UpdateGatewayCause.REQUEST_FAILED)
+    )
+
+    with pytest.raises(SystemExit) as excinfo:
+        _run_check_upgrade(repository, update_notifier=notifier)
+
+    assert excinfo.value.code == 1
+    assert "Update check failed" in capsys.readouterr().out
+
+
 def test_continue_marks_version_as_dismissed_and_prevents_reprompt(
     repository: FileSystemUpdateCacheRepository,
 ) -> None:

tests/cli/test_ui_teleport_availability.py

@@ -8,10 +8,11 @@ import pytest
 
 from tests.cli.plan_offer.adapters.fake_whoami_gateway import FakeWhoAmIGateway
 from tests.conftest import build_test_vibe_app, build_test_vibe_config
+from tests.constants import OPENAI_BASE_URL
 from vibe.cli.plan_offer.ports.whoami_gateway import WhoAmIPlanType, WhoAmIResponse
 from vibe.cli.textual_ui.widgets.chat_input import ChatInputContainer
 from vibe.core.config import ModelConfig, ProviderConfig, VibeConfig
-from vibe.core.types import Backend, LLMMessage, Role
+from vibe.core.types import Backend
 
 
 def _chat_plan_gateway(*, prompt_switching_to_pro_plan: bool) -> FakeWhoAmIGateway:
@@ -100,46 +101,6 @@ async def test_teleport_command_without_history_sends_early_failure_telemetry(
     ]
 
 
-@pytest.mark.asyncio
-async def test_teleport_command_in_remote_session_sends_early_failure_telemetry(
-    telemetry_events: list[dict[str, Any]],
-) -> None:
-    app = build_test_vibe_app(
-        config=_vibe_code_enabled_config(),
-        plan_offer_gateway=_chat_plan_gateway(prompt_switching_to_pro_plan=False),
-    )
-    app.agent_loop.messages.append(LLMMessage(role=Role.user, content="hello"))
-    app.agent_loop.messages.append(LLMMessage(role=Role.assistant, content="hi"))
-
-    async with app.run_test() as pilot:
-        await _wait_until(
-            pilot.pause,
-            lambda: app.commands.get_command_name("/teleport") == "teleport",
-        )
-
-        await app._remote_manager.attach(session_id="remote-session", config=app.config)
-        await app.on_chat_input_container_submitted(
-            ChatInputContainer.Submitted("/teleport")
-        )
-        await _wait_until(
-            pilot.pause, lambda: len(_teleport_failed_events(telemetry_events)) == 1
-        )
-        await app._remote_manager.detach()
-
-    assert _teleport_failed_events(telemetry_events) == [
-        {
-            "event_name": "vibe.teleport_failed",
-            "properties": {
-                "stage": "remote_session",
-                "error_class": "TeleportRemoteSessionError",
-                "push_required": False,
-                "nb_session_messages": 2,
-                "session_id": app.agent_loop.session_id,
-            },
-        }
-    ]
-
-
 @pytest.mark.asyncio
 async def test_teleport_command_hidden_when_current_key_is_not_eligible() -> None:
     app = build_test_vibe_app(
@@ -215,7 +176,7 @@ async def test_teleport_command_hides_after_switching_to_non_mistral_model(
             ),
             ProviderConfig(
                 name="openai",
-                api_base="https://api.openai.com/v1",
+                api_base=f"{OPENAI_BASE_URL}/v1",
                 api_key_env_var="OPENAI_API_KEY",
                 backend=Backend.GENERIC,
             ),

tests/cli/textual_ui/test_completion_popup.py

@@ -1,12 +0,0 @@
-from __future__ import annotations
-
-from vibe.cli.textual_ui.widgets.chat_input.completion_popup import CompletionPopup
-
-
-def test_rendered_text_length_uses_terminal_cell_width() -> None:
-    # "你" and "🙂" both occupy 2 terminal cells in Rich (+2 for separator).
-    assert CompletionPopup.rendered_text_length("@你", "🙂") == 6
-
-
-def test_rendered_text_length_keeps_description_separator() -> None:
-    assert CompletionPopup.rendered_text_length("@abc", "def") == 8

tests/cli/textual_ui/test_diff_rendering.py

@@ -22,6 +22,11 @@ def _build(
 
 def _render(*args, **kwargs):
     kwargs.setdefault("dark", True)
+    args = list(args)
+    # Tests pass a single start line as an int for readability; the renderer
+    # expects a list of occurrences.
+    if len(args) >= 4 and isinstance(args[3], int):
+        args[3] = [args[3]]
     return render_edit_diff(*args, **kwargs)
 
 
@@ -156,6 +161,32 @@ class TestRenderEditDiff:
         assert any(_plain(w).rstrip().endswith("d") for w in removed)
         assert any(_plain(w).rstrip().endswith("Z") for w in added)
 
+    def test_replace_all_renders_each_occurrence(self) -> None:
+        widgets = render_edit_diff(
+            "foo", "bar", "py", [3, 10, 25], ansi=False, dark=True
+        )
+        removed = [w for w in widgets if "diff-removed" in w.classes]
+        added = [w for w in widgets if "diff-added" in w.classes]
+        assert len(removed) == 3
+        assert len(added) == 3
+
+    def test_replace_all_uses_each_start_line(self) -> None:
+        widgets = render_edit_diff(
+            "foo", "bar", "py", [3, 10, 25], ansi=False, dark=True
+        )
+        joined = "\n".join(_plain(w) for w in widgets)
+        assert "3" in joined
+        assert "10" in joined
+        assert "25" in joined
+
+    def test_replace_all_separates_occurrences_with_gap(self) -> None:
+        widgets = render_edit_diff("foo", "bar", "py", [3, 10], ansi=False, dark=True)
+        assert sum("diff-gap" in w.classes for w in widgets) == 1
+
+    def test_single_occurrence_has_no_gap(self) -> None:
+        widgets = render_edit_diff("foo", "bar", "py", [3], ansi=False, dark=True)
+        assert all("diff-gap" not in w.classes for w in widgets)
+
 
 class TestBorderColors:
     def test_keys_index_into_widgets(self) -> None:

tests/cli/textual_ui/test_message_queue_ui.py

@@ -39,10 +39,12 @@ async def test_no_queue_header_when_empty(vibe_app: VibeApp) -> None:
 async def test_bash_submitted_during_running_bash_is_queued(vibe_app: VibeApp) -> None:
     async with vibe_app.run_test() as pilot:
         chat_input = vibe_app.query_one(ChatInputContainer)
-        chat_input.value = "!sleep 0.3"
+        chat_input.value = "!sleep 1"
         await pilot.press("enter")
 
-        await _wait_until(pilot, lambda: vibe_app._bash_task is not None, timeout=1.0)
+        assert await _wait_until(
+            pilot, lambda: vibe_app._bash_task is not None, timeout=2.0
+        )
 
         chat_input.value = "!echo queued"
         await pilot.press("enter")
@@ -56,6 +58,15 @@ async def test_bash_submitted_during_running_bash_is_queued(vibe_app: VibeApp) -
         queued_bashes = [w for w in vibe_app.query(BashOutputMessage) if w._queued]
         assert len(queued_bashes) == 1
 
+        await pilot.press("ctrl+c")
+        assert await _wait_until(
+            pilot, lambda: len(vibe_app._input_queue) == 0, timeout=2.0
+        )
+        await pilot.press("escape")
+        assert await _wait_until(
+            pilot, lambda: vibe_app._bash_task is None, timeout=5.0
+        )
+
 
 @pytest.mark.asyncio
 async def test_slash_command_rejected_with_warning_when_busy(vibe_app: VibeApp) -> None:

tests/cli/textual_ui/test_quit_confirmation.py

@@ -1,12 +1,13 @@
 from __future__ import annotations
 
+import asyncio
 import time
-from unittest.mock import MagicMock, patch
+from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 
 from tests.conftest import build_test_vibe_app
-from vibe.cli.textual_ui.app import VibeApp
+from vibe.cli.textual_ui.app import VibeApp, _run_app_with_cleanup
 from vibe.cli.textual_ui.quit_manager import QUIT_CONFIRM_DELAY, QuitManager
 
 
@@ -182,3 +183,69 @@ class TestActionDeleteRightOrQuit:
         mock_confirm.assert_called_once_with(
             "Ctrl+D", "1 queued message will be discarded"
         )
+
+
+@pytest.mark.asyncio
+async def test_shutdown_cleanup_cancels_in_flight_tasks(app: VibeApp) -> None:
+    async def _pending() -> None:
+        await asyncio.Event().wait()
+
+    agent_task = asyncio.create_task(_pending())
+    bash_task = asyncio.create_task(_pending())
+    app._agent_task = agent_task
+    app._bash_task = bash_task
+
+    await asyncio.wait_for(app.shutdown_cleanup(), timeout=1.0)
+
+    assert agent_task.cancelled()
+    assert bash_task.cancelled()
+
+
+@pytest.mark.asyncio
+async def test_shutdown_disables_future_queue_drains(app: VibeApp) -> None:
+    app._input_queue.append_prompt("queued")
+
+    await app._begin_shutdown()
+
+    with patch("vibe.cli.textual_ui.message_queue.asyncio.create_task") as create_task:
+        app._queue.start_drain_if_needed()
+
+    create_task.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_begin_shutdown_stops_scheduled_loop_runner(app: VibeApp) -> None:
+    with (
+        patch.object(app._queue, "shutdown", new_callable=AsyncMock) as queue_shutdown,
+        patch.object(app._loop_runner, "stop", new_callable=AsyncMock) as loop_stop,
+    ):
+        await app._begin_shutdown()
+
+    queue_shutdown.assert_awaited_once()
+    loop_stop.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_shutdown_cleanup_flushes_telemetry(app: VibeApp) -> None:
+    with patch.object(
+        app.agent_loop.telemetry_client, "aclose", new_callable=AsyncMock
+    ) as telemetry_aclose:
+        await app.shutdown_cleanup()
+
+    telemetry_aclose.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_run_app_with_cleanup_runs_cleanup_when_run_async_raises(
+    app: VibeApp,
+) -> None:
+    with (
+        patch.object(
+            app, "run_async", new_callable=AsyncMock, side_effect=RuntimeError("boom")
+        ),
+        patch.object(app, "shutdown_cleanup", new_callable=AsyncMock) as cleanup,
+    ):
+        with pytest.raises(RuntimeError, match="boom"):
+            await _run_app_with_cleanup(app)
+
+    cleanup.assert_awaited_once()

tests/cli/textual_ui/test_remote_session_manager.py

@@ -1,286 +0,0 @@
-from __future__ import annotations
-
-from unittest.mock import AsyncMock, MagicMock, patch
-
-import pytest
-
-from vibe.cli.textual_ui.remote.remote_session_manager import RemoteSessionManager
-from vibe.core.tools.builtins.ask_user_question import AskUserQuestionArgs
-from vibe.core.types import WaitingForInputEvent
-
-
-@pytest.fixture
-def manager() -> RemoteSessionManager:
-    return RemoteSessionManager()
-
-
-class TestProperties:
-    def test_is_active_false_by_default(self, manager: RemoteSessionManager) -> None:
-        assert manager.is_active is False
-
-    def test_is_terminated_false_when_inactive(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        assert manager.is_terminated is False
-
-    def test_is_waiting_for_input_false_when_inactive(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        assert manager.is_waiting_for_input is False
-
-    def test_has_pending_input_false_by_default(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        assert manager.has_pending_input is False
-
-    def test_session_id_none_when_inactive(self, manager: RemoteSessionManager) -> None:
-        assert manager.session_id is None
-
-
-class TestAttachDetach:
-    @pytest.mark.asyncio
-    async def test_attach_activates_manager(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        with patch(
-            "vibe.cli.textual_ui.remote.remote_session_manager.RemoteEventsSource"
-        ) as MockSource:
-            mock_source = MagicMock()
-            mock_source.session_id = "test-session-id"
-            MockSource.return_value = mock_source
-
-            config = MagicMock()
-            await manager.attach(session_id="test-session-id", config=config)
-
-            assert manager.is_active is True
-            assert manager.session_id == "test-session-id"
-
-    @pytest.mark.asyncio
-    async def test_detach_cleans_up(self, manager: RemoteSessionManager) -> None:
-        with patch(
-            "vibe.cli.textual_ui.remote.remote_session_manager.RemoteEventsSource"
-        ) as MockSource:
-            mock_source = AsyncMock()
-            mock_source.session_id = "test-id"
-            MockSource.return_value = mock_source
-
-            config = MagicMock()
-            await manager.attach(session_id="test-id", config=config)
-            await manager.detach()
-
-            assert manager.is_active is False
-            assert manager.session_id is None
-            assert manager.has_pending_input is False
-            mock_source.close.assert_called_once()
-
-    @pytest.mark.asyncio
-    async def test_attach_detaches_previous_session(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        with patch(
-            "vibe.cli.textual_ui.remote.remote_session_manager.RemoteEventsSource"
-        ) as MockSource:
-            first_source = AsyncMock()
-            second_source = MagicMock()
-            second_source.session_id = "second-id"
-            MockSource.side_effect = [first_source, second_source]
-
-            config = MagicMock()
-            await manager.attach(session_id="first-id", config=config)
-            await manager.attach(session_id="second-id", config=config)
-
-            first_source.close.assert_called_once()
-            assert manager.session_id == "second-id"
-
-
-class TestValidateInput:
-    @pytest.mark.asyncio
-    async def test_returns_none_when_waiting_for_input(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        with patch(
-            "vibe.cli.textual_ui.remote.remote_session_manager.RemoteEventsSource"
-        ) as MockSource:
-            mock_source = MagicMock()
-            mock_source.is_terminated = False
-            mock_source.is_waiting_for_input = True
-            MockSource.return_value = mock_source
-
-            config = MagicMock()
-            await manager.attach(session_id="id", config=config)
-
-            assert manager.validate_input() is None
-
-    @pytest.mark.asyncio
-    async def test_returns_warning_when_terminated(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        with patch(
-            "vibe.cli.textual_ui.remote.remote_session_manager.RemoteEventsSource"
-        ) as MockSource:
-            mock_source = MagicMock()
-            mock_source.is_terminated = True
-            MockSource.return_value = mock_source
-
-            config = MagicMock()
-            await manager.attach(session_id="id", config=config)
-
-            result = manager.validate_input()
-            assert result is not None
-            assert "ended" in result
-
-    @pytest.mark.asyncio
-    async def test_returns_warning_when_not_waiting_for_input(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        with patch(
-            "vibe.cli.textual_ui.remote.remote_session_manager.RemoteEventsSource"
-        ) as MockSource:
-            mock_source = MagicMock()
-            mock_source.is_terminated = False
-            mock_source.is_waiting_for_input = False
-            MockSource.return_value = mock_source
-
-            config = MagicMock()
-            await manager.attach(session_id="id", config=config)
-
-            result = manager.validate_input()
-            assert result is not None
-            assert "not waiting" in result
-
-
-class TestSendPrompt:
-    @pytest.mark.asyncio
-    async def test_raises_when_inactive_and_required(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        with pytest.raises(RuntimeError, match="No active remote session"):
-            await manager.send_prompt("hello")
-
-    @pytest.mark.asyncio
-    async def test_returns_silently_when_inactive_and_not_required(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        await manager.send_prompt("hello", require_source=False)
-
-    @pytest.mark.asyncio
-    async def test_restores_pending_on_error(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        with patch(
-            "vibe.cli.textual_ui.remote.remote_session_manager.RemoteEventsSource"
-        ) as MockSource:
-            mock_source = AsyncMock()
-            mock_source.send_prompt.side_effect = Exception("connection error")
-            MockSource.return_value = mock_source
-
-            config = MagicMock()
-            await manager.attach(session_id="id", config=config)
-
-            event = WaitingForInputEvent(task_id="t1", label="test")
-            manager.set_pending_input(event)
-
-            with pytest.raises(Exception, match="connection error"):
-                await manager.send_prompt("hello")
-
-            assert manager.has_pending_input is True
-
-
-class TestPendingInput:
-    def test_set_and_cancel_pending_input(self, manager: RemoteSessionManager) -> None:
-        event = WaitingForInputEvent(task_id="t1", label="test")
-        manager.set_pending_input(event)
-        assert manager.has_pending_input is True
-
-        manager.cancel_pending_input()
-        assert manager.has_pending_input is False
-
-
-class TestBuildQuestionArgs:
-    def test_returns_none_with_no_predefined_answers(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        event = WaitingForInputEvent(task_id="t1", label="test")
-        assert manager.build_question_args(event) is None
-
-    def test_returns_none_with_one_predefined_answer(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        event = WaitingForInputEvent(
-            task_id="t1", label="test", predefined_answers=["only one"]
-        )
-        assert manager.build_question_args(event) is None
-
-    def test_returns_args_with_two_predefined_answers(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        event = WaitingForInputEvent(
-            task_id="t1", label="Pick one", predefined_answers=["yes", "no"]
-        )
-        result = manager.build_question_args(event)
-        assert result is not None
-        assert isinstance(result, AskUserQuestionArgs)
-        assert len(result.questions) == 1
-        assert result.questions[0].question == "Pick one"
-        assert len(result.questions[0].options) == 2
-
-    def test_caps_at_four_predefined_answers(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        event = WaitingForInputEvent(
-            task_id="t1",
-            label="Pick",
-            predefined_answers=["a", "b", "c", "d", "e", "f"],
-        )
-        result = manager.build_question_args(event)
-        assert result is not None
-        assert len(result.questions[0].options) == 4
-
-    def test_uses_default_question_when_no_label(
-        self, manager: RemoteSessionManager
-    ) -> None:
-        event = WaitingForInputEvent(task_id="t1", predefined_answers=["a", "b"])
-        result = manager.build_question_args(event)
-        assert result is not None
-        assert result.questions[0].question == "Choose an answer"
-
-
-class TestBuildTerminalMessage:
-    def test_completed_when_no_source(self, manager: RemoteSessionManager) -> None:
-        msg_type, text = manager.build_terminal_message()
-        assert msg_type == "info"
-        assert "completed" in text
-
-    @pytest.mark.asyncio
-    async def test_failed_state(self, manager: RemoteSessionManager) -> None:
-        with patch(
-            "vibe.cli.textual_ui.remote.remote_session_manager.RemoteEventsSource"
-        ) as MockSource:
-            mock_source = MagicMock()
-            mock_source.is_failed = True
-            mock_source.is_canceled = False
-            MockSource.return_value = mock_source
-
-            config = MagicMock()
-            await manager.attach(session_id="id", config=config)
-
-            msg_type, text = manager.build_terminal_message()
-            assert msg_type == "error"
-            assert "failed" in text.lower()
-
-    @pytest.mark.asyncio
-    async def test_canceled_state(self, manager: RemoteSessionManager) -> None:
-        with patch(
-            "vibe.cli.textual_ui.remote.remote_session_manager.RemoteEventsSource"
-        ) as MockSource:
-            mock_source = MagicMock()
-            mock_source.is_failed = False
-            mock_source.is_canceled = True
-            MockSource.return_value = mock_source
-
-            config = MagicMock()
-            await manager.attach(session_id="id", config=config)
-
-            msg_type, text = manager.build_terminal_message()
-            assert msg_type == "warning"
-            assert "canceled" in text.lower()

tests/cli/textual_ui/test_session_picker.py

@@ -19,25 +19,21 @@ def sample_sessions() -> list[ResumeSessionInfo]:
     return [
         ResumeSessionInfo(
             session_id="session-a",
-            source="local",
             cwd="/test",
             title="Session A",
             end_time=(datetime.now(UTC) - timedelta(minutes=5)).isoformat(),
         ),
         ResumeSessionInfo(
             session_id="session-b",
-            source="local",
             cwd="/test",
             title="Session B",
             end_time=(datetime.now(UTC) - timedelta(hours=1)).isoformat(),
         ),
         ResumeSessionInfo(
             session_id="session-c",
-            source="remote",
             cwd="/test",
             title="Session C",
             end_time=(datetime.now(UTC) - timedelta(days=1)).isoformat(),
-            status="RUNNING",
         ),
     ]
 
@@ -45,9 +41,9 @@ def sample_sessions() -> list[ResumeSessionInfo]:
 @pytest.fixture
 def sample_latest_messages() -> dict[str, str]:
     return {
-        "local:session-a": "Help me fix this bug",
-        "local:session-b": "Refactor the authentication module",
-        "remote:session-c": "Add unit tests for the API",
+        "session-a": "Help me fix this bug",
+        "session-b": "Refactor the authentication module",
+        "session-c": "Add unit tests for the API",
     }
 
 
@@ -143,31 +139,19 @@ class TestSessionPickerAppInit:
 
 class TestSessionPickerMessages:
     def test_session_selected_stores_option_id(self) -> None:
-        msg = SessionPickerApp.SessionSelected(
-            "local:test-session-id", "local", "test-session-id"
-        )
-        assert msg.option_id == "local:test-session-id"
-        assert msg.source == "local"
+        msg = SessionPickerApp.SessionSelected("test-session-id", "test-session-id")
+        assert msg.option_id == "test-session-id"
         assert msg.session_id == "test-session-id"
 
-    def test_session_selected_with_full_uuid(self) -> None:
-        session_id = "abc12345-6789-0123-4567-89abcdef0123"
-        option_id = f"remote:{session_id}"
-        msg = SessionPickerApp.SessionSelected(option_id, "remote", session_id)
-        assert msg.option_id == option_id
-        assert msg.source == "remote"
-        assert msg.session_id == session_id
-
     def test_cancelled_can_be_instantiated(self) -> None:
         msg = SessionPickerApp.Cancelled()
         assert isinstance(msg, SessionPickerApp.Cancelled)
 
     def test_session_delete_requested_stores_session_info(self) -> None:
         msg = SessionPickerApp.SessionDeleteRequested(
-            "local:test-session-id", "local", "test-session-id"
+            "test-session-id", "test-session-id"
         )
-        assert msg.option_id == "local:test-session-id"
-        assert msg.source == "local"
+        assert msg.option_id == "test-session-id"
         assert msg.session_id == "test-session-id"
 
 
@@ -199,15 +183,15 @@ class TestSessionPickerSessionRemoval:
         picker = SessionPickerApp(
             sessions=sample_sessions, latest_messages=sample_latest_messages
         )
-        option_list = FakeOptionList(highlighted_option_id="local:session-a")
+        option_list = FakeOptionList(highlighted_option_id="session-a")
         posted_messages: list[object] = []
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
         monkeypatch.setattr(picker, "post_message", posted_messages.append)
 
         picker.action_request_delete()
 
-        assert_delete_state(picker, kind="confirmation", option_id="local:session-a")
-        assert option_list.replaced_prompts[-1].option_id == "local:session-a"
+        assert_delete_state(picker, kind="confirmation", option_id="session-a")
+        assert option_list.replaced_prompts[-1].option_id == "session-a"
         assert (
             "Press D again to delete" in option_list.replaced_prompts[-1].prompt.plain
         )
@@ -222,7 +206,7 @@ class TestSessionPickerSessionRemoval:
         picker = SessionPickerApp(
             sessions=sample_sessions, latest_messages=sample_latest_messages
         )
-        option_list = FakeOptionList(highlighted_option_id="local:session-a")
+        option_list = FakeOptionList(highlighted_option_id="session-a")
         posted_messages: list[object] = []
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
         monkeypatch.setattr(picker, "post_message", posted_messages.append)
@@ -230,14 +214,13 @@ class TestSessionPickerSessionRemoval:
         picker.action_request_delete()
         picker.action_request_delete()
 
-        assert_delete_state(picker, kind="pending", option_id="local:session-a")
-        assert option_list.replaced_prompts[-1].option_id == "local:session-a"
+        assert_delete_state(picker, kind="pending", option_id="session-a")
+        assert option_list.replaced_prompts[-1].option_id == "session-a"
         assert "Deleting..." in option_list.replaced_prompts[-1].prompt.plain
         assert len(posted_messages) == 1
         message = posted_messages[0]
         assert isinstance(message, SessionPickerApp.SessionDeleteRequested)
-        assert message.option_id == "local:session-a"
-        assert message.source == "local"
+        assert message.option_id == "session-a"
         assert message.session_id == "session-a"
 
     def test_delete_confirmation_is_consumed_after_request(
@@ -249,7 +232,7 @@ class TestSessionPickerSessionRemoval:
         picker = SessionPickerApp(
             sessions=sample_sessions, latest_messages=sample_latest_messages
         )
-        option_list = FakeOptionList(highlighted_option_id="local:session-a")
+        option_list = FakeOptionList(highlighted_option_id="session-a")
         posted_messages: list[object] = []
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
         monkeypatch.setattr(picker, "post_message", posted_messages.append)
@@ -259,34 +242,10 @@ class TestSessionPickerSessionRemoval:
         picker.action_request_delete()
 
         assert len(posted_messages) == 1
-        assert_delete_state(picker, kind="pending", option_id="local:session-a")
-        assert option_list.replaced_prompts[-1].option_id == "local:session-a"
+        assert_delete_state(picker, kind="pending", option_id="session-a")
+        assert option_list.replaced_prompts[-1].option_id == "session-a"
         assert "Deleting..." in option_list.replaced_prompts[-1].prompt.plain
 
-    def test_delete_request_shows_feedback_for_remote_sessions(
-        self,
-        sample_sessions: list[ResumeSessionInfo],
-        sample_latest_messages: dict[str, str],
-        monkeypatch: pytest.MonkeyPatch,
-    ) -> None:
-        picker = SessionPickerApp(
-            sessions=sample_sessions, latest_messages=sample_latest_messages
-        )
-        option_list = FakeOptionList(highlighted_option_id="remote:session-c")
-        posted_messages: list[object] = []
-        monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
-        monkeypatch.setattr(picker, "post_message", posted_messages.append)
-
-        picker.action_request_delete()
-
-        assert_delete_state(picker, kind="feedback", option_id="remote:session-c")
-        assert option_list.replaced_prompts[-1].option_id == "remote:session-c"
-        assert (
-            "Can't delete remote session"
-            in option_list.replaced_prompts[-1].prompt.plain
-        )
-        assert posted_messages == []
-
     def test_delete_request_shows_feedback_for_current_session(
         self,
         sample_sessions: list[ResumeSessionInfo],
@@ -298,15 +257,15 @@ class TestSessionPickerSessionRemoval:
             latest_messages=sample_latest_messages,
             current_session_id="session-a",
         )
-        option_list = FakeOptionList(highlighted_option_id="local:session-a")
+        option_list = FakeOptionList(highlighted_option_id="session-a")
         posted_messages: list[object] = []
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
         monkeypatch.setattr(picker, "post_message", posted_messages.append)
 
         picker.action_request_delete()
 
-        assert_delete_state(picker, kind="feedback", option_id="local:session-a")
-        assert option_list.replaced_prompts[-1].option_id == "local:session-a"
+        assert_delete_state(picker, kind="feedback", option_id="session-a")
+        assert option_list.replaced_prompts[-1].option_id == "session-a"
         assert (
             "Can't delete current session"
             in option_list.replaced_prompts[-1].prompt.plain
@@ -315,7 +274,7 @@ class TestSessionPickerSessionRemoval:
 
         picker.action_request_delete()
 
-        assert_delete_state(picker, kind="feedback", option_id="local:session-a")
+        assert_delete_state(picker, kind="feedback", option_id="session-a")
         assert posted_messages == []
 
     def test_pending_delete_blocks_resume_selection(
@@ -327,7 +286,7 @@ class TestSessionPickerSessionRemoval:
         picker = SessionPickerApp(
             sessions=sample_sessions, latest_messages=sample_latest_messages
         )
-        option_list = FakeOptionList(highlighted_option_id="local:session-a")
+        option_list = FakeOptionList(highlighted_option_id="session-a")
         posted_messages: list[object] = []
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
         monkeypatch.setattr(picker, "post_message", posted_messages.append)
@@ -335,15 +294,15 @@ class TestSessionPickerSessionRemoval:
         picker.action_request_delete()
         picker.action_request_delete()
         picker.on_option_list_option_selected(
-            cast(OptionList.OptionSelected, FakeOptionEvent("local:session-a"))
+            cast(OptionList.OptionSelected, FakeOptionEvent("session-a"))
         )
         picker.on_option_list_option_selected(
-            cast(OptionList.OptionSelected, FakeOptionEvent("local:session-b"))
+            cast(OptionList.OptionSelected, FakeOptionEvent("session-b"))
         )
 
         assert len(posted_messages) == 1
         assert isinstance(posted_messages[0], SessionPickerApp.SessionDeleteRequested)
-        assert_delete_state(picker, kind="pending", option_id="local:session-a")
+        assert_delete_state(picker, kind="pending", option_id="session-a")
 
     def test_clear_pending_delete_restores_session_option(
         self,
@@ -354,7 +313,7 @@ class TestSessionPickerSessionRemoval:
         picker = SessionPickerApp(
             sessions=sample_sessions, latest_messages=sample_latest_messages
         )
-        option_list = FakeOptionList(highlighted_option_id="local:session-a")
+        option_list = FakeOptionList(highlighted_option_id="session-a")
         posted_messages: list[object] = []
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
         monkeypatch.setattr(picker, "post_message", posted_messages.append)
@@ -362,9 +321,9 @@ class TestSessionPickerSessionRemoval:
         picker.action_request_delete()
         picker.action_request_delete()
 
-        assert picker.clear_pending_delete("local:session-a") is True
+        assert picker.clear_pending_delete("session-a") is True
         assert picker._delete_state is None
-        assert option_list.replaced_prompts[-1].option_id == "local:session-a"
+        assert option_list.replaced_prompts[-1].option_id == "session-a"
         assert "Help me fix this bug" in option_list.replaced_prompts[-1].prompt.plain
 
     def test_highlighting_another_session_clears_confirmation(
@@ -376,16 +335,16 @@ class TestSessionPickerSessionRemoval:
         picker = SessionPickerApp(
             sessions=sample_sessions, latest_messages=sample_latest_messages
         )
-        option_list = FakeOptionList(highlighted_option_id="local:session-a")
+        option_list = FakeOptionList(highlighted_option_id="session-a")
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
         picker.action_request_delete()
 
         picker.on_option_list_option_highlighted(
-            cast(OptionList.OptionHighlighted, FakeOptionEvent("local:session-b"))
+            cast(OptionList.OptionHighlighted, FakeOptionEvent("session-b"))
         )
 
         assert picker._delete_state is None
-        assert option_list.replaced_prompts[-1].option_id == "local:session-a"
+        assert option_list.replaced_prompts[-1].option_id == "session-a"
         assert "Help me fix this bug" in option_list.replaced_prompts[-1].prompt.plain
 
     def test_highlighting_another_session_clears_delete_feedback(
@@ -395,18 +354,20 @@ class TestSessionPickerSessionRemoval:
         monkeypatch: pytest.MonkeyPatch,
     ) -> None:
         picker = SessionPickerApp(
-            sessions=sample_sessions, latest_messages=sample_latest_messages
+            sessions=sample_sessions,
+            latest_messages=sample_latest_messages,
+            current_session_id="session-c",
         )
-        option_list = FakeOptionList(highlighted_option_id="remote:session-c")
+        option_list = FakeOptionList(highlighted_option_id="session-c")
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
         picker.action_request_delete()
 
         picker.on_option_list_option_highlighted(
-            cast(OptionList.OptionHighlighted, FakeOptionEvent("local:session-b"))
+            cast(OptionList.OptionHighlighted, FakeOptionEvent("session-b"))
         )
 
         assert picker._delete_state is None
-        assert option_list.replaced_prompts[-1].option_id == "remote:session-c"
+        assert option_list.replaced_prompts[-1].option_id == "session-c"
         assert (
             "Add unit tests for the API"
             in option_list.replaced_prompts[-1].prompt.plain
@@ -421,7 +382,7 @@ class TestSessionPickerSessionRemoval:
         picker = SessionPickerApp(
             sessions=sample_sessions, latest_messages=sample_latest_messages
         )
-        option_list = FakeOptionList(highlighted_option_id="local:session-a")
+        option_list = FakeOptionList(highlighted_option_id="session-a")
         posted_messages: list[object] = []
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
         monkeypatch.setattr(picker, "post_message", posted_messages.append)
@@ -444,9 +405,11 @@ class TestSessionPickerSessionRemoval:
         monkeypatch: pytest.MonkeyPatch,
     ) -> None:
         picker = SessionPickerApp(
-            sessions=sample_sessions, latest_messages=sample_latest_messages
+            sessions=sample_sessions,
+            latest_messages=sample_latest_messages,
+            current_session_id="session-c",
         )
-        option_list = FakeOptionList(highlighted_option_id="remote:session-c")
+        option_list = FakeOptionList(highlighted_option_id="session-c")
         posted_messages: list[object] = []
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
         monkeypatch.setattr(picker, "post_message", posted_messages.append)
@@ -471,20 +434,20 @@ class TestSessionPickerSessionRemoval:
         picker = SessionPickerApp(
             sessions=sample_sessions, latest_messages=sample_latest_messages
         )
-        option_list = FakeOptionList(highlighted_option_id="local:session-a")
+        option_list = FakeOptionList(highlighted_option_id="session-a")
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
         picker.action_request_delete()
-        assert_delete_state(picker, kind="confirmation", option_id="local:session-a")
+        assert_delete_state(picker, kind="confirmation", option_id="session-a")
 
-        assert picker.remove_session("local:session-a") is True
+        assert picker.remove_session("session-a") is True
 
         assert [session.option_id for session in picker._sessions] == [
-            "local:session-b",
-            "remote:session-c",
+            "session-b",
+            "session-c",
         ]
-        assert "local:session-a" not in picker._latest_messages
+        assert "session-a" not in picker._latest_messages
         assert picker._delete_state is None
-        assert option_list.removed_option_ids == ["local:session-a"]
+        assert option_list.removed_option_ids == ["session-a"]
 
     def test_remove_missing_session_returns_false(
         self,
@@ -498,7 +461,7 @@ class TestSessionPickerSessionRemoval:
         option_list = FakeOptionList()
         monkeypatch.setattr(picker, "query_one", lambda _selector: option_list)
 
-        assert picker.remove_session("local:missing") is False
+        assert picker.remove_session("missing") is False
 
         assert picker._sessions == sample_sessions
         assert picker._latest_messages == sample_latest_messages

tests/cli/textual_ui/test_user_message_attachments.py

@@ -5,12 +5,14 @@ from weakref import WeakKeyDictionary
 
 from vibe.cli.textual_ui.widgets.messages import UserMessage
 from vibe.cli.textual_ui.windowing.history import build_history_widgets
-from vibe.core.types import ImageAttachment, LLMMessage, Role
+from vibe.core.types import FileImageSource, ImageAttachment, LLMMessage, Role
 
 
 def _att(path: Path, alias: str) -> ImageAttachment:
     path.write_bytes(b"\x89PNG")
-    return ImageAttachment(path=path, alias=alias, mime_type="image/png")
+    return ImageAttachment(
+        source=FileImageSource(path=path), alias=alias, mime_type="image/png"
+    )
 
 
 def test_attachments_footer_singular(tmp_path: Path) -> None:

tests/conftest.py

@@ -5,6 +5,9 @@ from pathlib import Path
 import sys
 from typing import Any
 
+import keyring
+from keyring.backend import KeyringBackend
+import keyring.errors
 import pytest
 import tomli_w
 
@@ -34,6 +37,39 @@ from vibe.core.config.harness_files import (
 from vibe.core.llm.types import BackendLike
 
 
+class _EmptyKeyring(KeyringBackend):
+    """A keyring backend that stores nothing, used to keep tests off the real OS keyring."""
+
+    priority = 1  # type: ignore[assignment]
+
+    def get_password(self, service: str, username: str) -> str | None:
+        return None
+
+    def set_password(self, service: str, username: str, password: str) -> None:
+        return None
+
+    def delete_password(self, service: str, username: str) -> None:
+        raise keyring.errors.PasswordDeleteError(username)
+
+
+@pytest.fixture(autouse=True)
+def _disable_os_keyring() -> Generator[None, None, None]:
+    """Keep the suite off the real OS keyring.
+
+    ``resolve_api_key`` and ``VibeConfig._check_api_key`` now consult the keyring, so
+    without this every config construction would touch the real Keychain. We install an
+    empty backend (rather than patching ``keyring.get_password``) so tests that swap in
+    their own backend via ``keyring.set_keyring`` still work. Tests that exercise keyring
+    behaviour opt in by patching ``keyring.get_password`` / ``set_password`` directly.
+    """
+    original = keyring.get_keyring()
+    keyring.set_keyring(_EmptyKeyring())
+    try:
+        yield
+    finally:
+        keyring.set_keyring(original)
+
+
 def get_base_config() -> dict[str, Any]:
     return {
         "active_model": "devstral-latest",
@@ -143,6 +179,8 @@ def _scratchpad_dir(
 @pytest.fixture(autouse=True)
 def _mock_api_key(monkeypatch: pytest.MonkeyPatch) -> None:
     monkeypatch.setenv("MISTRAL_API_KEY", "mock")
+    monkeypatch.setenv("ANTHROPIC_API_KEY", "mock")
+    monkeypatch.setenv("OPENAI_API_KEY", "mock")
 
 
 @pytest.fixture(autouse=True)
@@ -262,6 +300,9 @@ def build_test_vibe_config(**kwargs) -> VibeConfig:
     )
     if kwargs.get("models"):
         kwargs.setdefault("active_model", kwargs["models"][0].alias)
+    # Connectors trigger a real HTTP discovery on agent construction; off by
+    # default so tests don't pay for it. Connector tests pass enable_connectors=True.
+    kwargs.setdefault("enable_connectors", False)
     return VibeConfig(
         session_logging=resolved_session_logging,
         enable_update_checks=resolved_enable_update_checks,

tests/constants.py

@@ -0,0 +1,14 @@
+from __future__ import annotations
+
+MISTRAL_BASE_URL = "https://api.mistral.ai"
+CHAT_COMPLETIONS_PATH = "/v1/chat/completions"
+CONNECTORS_BOOTSTRAP_PATH = "/v1/connectors/bootstrap"
+
+ANTHROPIC_BASE_URL = "https://api.anthropic.com"
+ANTHROPIC_MESSAGES_PATH = "/v1/messages"
+
+OPENAI_BASE_URL = "https://api.openai.com"
+OPENAI_RESPONSES_PATH = "/v1/responses"
+
+TELEPORT_SESSIONS_PATH = "/api/v1/code/sessions"
+TELEPORT_COMPLETE_URL = "https://chat.example.com/code/project-id/web-session-id"

tests/core/nuage/test_remote_events_source.py

@@ -1,301 +0,0 @@
-from __future__ import annotations
-
-from unittest.mock import AsyncMock, MagicMock, patch
-
-from pydantic import BaseModel, ValidationError
-import pytest
-
-from tests.conftest import build_test_vibe_config
-from vibe.core.agent_loop import AgentLoopStateError
-from vibe.core.nuage.exceptions import ErrorCode, WorkflowsException
-from vibe.core.nuage.remote_events_source import RemoteEventsSource
-from vibe.core.nuage.streaming import StreamEvent
-
-_SESSION_ID = "test-session"
-
-
-def _make_source(**kwargs) -> RemoteEventsSource:
-    config = build_test_vibe_config(enabled_tools=kwargs.pop("enabled_tools", []))
-    return RemoteEventsSource(session_id=_SESSION_ID, config=config, **kwargs)
-
-
-def _make_retryable_exc(msg: str) -> WorkflowsException:
-    return WorkflowsException(message=msg, code=ErrorCode.GET_EVENTS_STREAM_ERROR)
-
-
-def _make_stream_event(
-    broker_sequence: int | None = None, data: dict | None = None
-) -> StreamEvent:
-    return StreamEvent(data=data or {}, broker_sequence=broker_sequence)
-
-
-class TestIsRetryableStreamDisconnect:
-    def test_peer_closed_connection(self) -> None:
-        source = _make_source()
-        exc = _make_retryable_exc("Peer closed connection without response")
-        assert source._is_retryable_stream_disconnect(exc) is True
-
-    def test_incomplete_chunked_read(self) -> None:
-        source = _make_source()
-        exc = _make_retryable_exc("Incomplete chunked read during streaming")
-        assert source._is_retryable_stream_disconnect(exc) is True
-
-    def test_non_retryable_message(self) -> None:
-        source = _make_source()
-        exc = WorkflowsException(
-            message="some other error", code=ErrorCode.GET_EVENTS_STREAM_ERROR
-        )
-        assert source._is_retryable_stream_disconnect(exc) is False
-
-    def test_wrong_error_code(self) -> None:
-        source = _make_source()
-        exc = WorkflowsException(
-            message="peer closed connection",
-            code=ErrorCode.POST_EXECUTIONS_SIGNALS_ERROR,
-        )
-        assert source._is_retryable_stream_disconnect(exc) is False
-
-
-async def _async_gen_from_list(items):
-    for item in items:
-        yield item
-
-
-async def _async_gen_raise(items, exc):
-    for item in items:
-        yield item
-    raise exc
-
-
-class _FakeStream:
-    def __init__(self, payloads=None, exc=None):
-        self._payloads = payloads or []
-        self._exc = exc
-        self._closed = False
-
-    def __aiter__(self):
-        return self._iterate().__aiter__()
-
-    async def _iterate(self):
-        for p in self._payloads:
-            yield p
-        if self._exc is not None:
-            raise self._exc
-
-    async def aclose(self):
-        self._closed = True
-
-
-class TestStreamRemoteEventsRetry:
-    @pytest.mark.asyncio
-    async def test_retries_on_retryable_disconnect(self) -> None:
-        source = _make_source()
-        exc = _make_retryable_exc("peer closed connection")
-
-        call_count = 0
-
-        def make_stream(_params):
-            nonlocal call_count
-            call_count += 1
-            return _FakeStream(exc=exc)
-
-        mock_client = MagicMock()
-        mock_client.stream_events = make_stream
-        source._client = mock_client
-
-        with patch("asyncio.sleep", new_callable=AsyncMock):
-            events = [e async for e in source._stream_remote_events()]
-
-        assert events == []
-        assert call_count == 4  # 1 initial + 3 retries
-
-    @pytest.mark.asyncio
-    async def test_stops_after_max_retry_count(self) -> None:
-        source = _make_source()
-        exc = _make_retryable_exc("incomplete chunked read")
-
-        call_count = 0
-
-        def make_stream(_params):
-            nonlocal call_count
-            call_count += 1
-            return _FakeStream(exc=exc)
-
-        mock_client = MagicMock()
-        mock_client.stream_events = make_stream
-        source._client = mock_client
-
-        with patch("asyncio.sleep", new_callable=AsyncMock):
-            events = [e async for e in source._stream_remote_events()]
-
-        assert events == []
-        assert call_count == 4
-
-    @pytest.mark.asyncio
-    async def test_resets_retry_count_on_successful_event(self) -> None:
-        source = _make_source()
-        exc = _make_retryable_exc("peer closed connection")
-
-        successful_event = _make_stream_event(broker_sequence=0, data={})
-        call_count = 0
-
-        def make_stream(_params):
-            nonlocal call_count
-            call_count += 1
-            if call_count <= 2:
-                return _FakeStream(payloads=[successful_event], exc=exc)
-            return _FakeStream(exc=exc)
-
-        mock_client = MagicMock()
-        mock_client.stream_events = make_stream
-        source._client = mock_client
-
-        with (
-            patch("asyncio.sleep", new_callable=AsyncMock),
-            patch.object(source, "_normalize_stream_event", return_value=None),
-        ):
-            events = [e async for e in source._stream_remote_events()]
-
-        assert events == []
-        # call 1: success + exc -> retry_count = 1
-        # call 2: success (reset) + exc -> retry_count = 1
-        # call 3: exc -> retry_count = 2
-        # call 4: exc -> retry_count = 3
-        # call 5: exc -> retry_count = 4 > 3 -> break
-        assert call_count == 5
-
-    @pytest.mark.asyncio
-    async def test_non_retryable_raises_agent_loop_state_error(self) -> None:
-        source = _make_source()
-        exc = WorkflowsException(
-            message="something bad", code=ErrorCode.TEMPORAL_CONNECTION_ERROR
-        )
-
-        mock_client = MagicMock()
-        mock_client.stream_events = lambda _: _FakeStream(exc=exc)
-        source._client = mock_client
-
-        with pytest.raises(AgentLoopStateError):
-            async for _ in source._stream_remote_events():
-                pass
-
-
-class TestStreamRemoteEventsIdleBoundary:
-    @pytest.mark.asyncio
-    async def test_stops_on_idle_boundary(self) -> None:
-        source = _make_source()
-        event_data = _make_stream_event(broker_sequence=0, data={})
-        sentinel_event = MagicMock()
-
-        mock_client = MagicMock()
-        mock_client.stream_events = lambda _: _FakeStream(payloads=[event_data])
-        source._client = mock_client
-
-        workflow_event = MagicMock()
-
-        with (
-            patch.object(
-                source, "_normalize_stream_event", return_value=workflow_event
-            ),
-            patch.object(
-                source, "_consume_workflow_event", return_value=[sentinel_event]
-            ),
-            patch.object(source, "_is_idle_boundary", return_value=True) as mock_idle,
-        ):
-            events = [
-                e
-                async for e in source._stream_remote_events(stop_on_idle_boundary=True)
-            ]
-
-        assert events == [sentinel_event]
-        mock_idle.assert_called_once_with(workflow_event)
-
-    @pytest.mark.asyncio
-    async def test_continues_past_idle_boundary_when_disabled(self) -> None:
-        source = _make_source()
-        event1 = _make_stream_event(broker_sequence=0, data={})
-        event2 = _make_stream_event(broker_sequence=1, data={})
-        sentinel1 = MagicMock()
-        sentinel2 = MagicMock()
-
-        mock_client = MagicMock()
-        mock_client.stream_events = lambda _: _FakeStream(payloads=[event1, event2])
-        source._client = mock_client
-
-        workflow_event = MagicMock()
-        call_count = 0
-
-        def consume_side_effect(_evt):
-            nonlocal call_count
-            call_count += 1
-            return [sentinel1] if call_count == 1 else [sentinel2]
-
-        with (
-            patch.object(
-                source, "_normalize_stream_event", return_value=workflow_event
-            ),
-            patch.object(
-                source, "_consume_workflow_event", side_effect=consume_side_effect
-            ),
-            patch.object(source, "_is_idle_boundary", return_value=True),
-        ):
-            events = [
-                e
-                async for e in source._stream_remote_events(stop_on_idle_boundary=False)
-            ]
-
-        assert events == [sentinel1, sentinel2]
-
-
-class TestBrokerSequenceTracking:
-    @pytest.mark.asyncio
-    async def test_next_start_seq_updated(self) -> None:
-        source = _make_source()
-        assert source._next_start_seq == 0
-
-        event1 = _make_stream_event(broker_sequence=5, data={})
-        event2 = _make_stream_event(broker_sequence=10, data={})
-
-        mock_client = MagicMock()
-        mock_client.stream_events = lambda _: _FakeStream(payloads=[event1, event2])
-        source._client = mock_client
-
-        with patch.object(source, "_normalize_stream_event", return_value=None):
-            events = [e async for e in source._stream_remote_events()]
-
-        assert events == []
-        assert source._next_start_seq == 11
-
-    @pytest.mark.asyncio
-    async def test_none_broker_sequence_not_updated(self) -> None:
-        source = _make_source()
-        source._next_start_seq = 5
-
-        event = _make_stream_event(broker_sequence=None, data={})
-
-        mock_client = MagicMock()
-        mock_client.stream_events = lambda _: _FakeStream(payloads=[event])
-        source._client = mock_client
-
-        with patch.object(source, "_normalize_stream_event", return_value=None):
-            events = [e async for e in source._stream_remote_events()]
-
-        assert events == []
-        assert source._next_start_seq == 5
-
-
-def test_consume_workflow_event_validation_error_is_logged_and_ignored() -> None:
-    class _InvalidPayload(BaseModel):
-        required: str
-
-    source = _make_source()
-    with pytest.raises(ValidationError) as exc_info:
-        _InvalidPayload.model_validate({})
-
-    source._translator.consume_workflow_event = MagicMock(side_effect=exc_info.value)
-
-    with patch("vibe.core.nuage.remote_events_source.logger.warning") as mock_warning:
-        events = source._consume_workflow_event(MagicMock())
-
-    assert events == []
-    mock_warning.assert_called_once()

tests/core/nuage/test_workflows_client.py

@@ -1,248 +0,0 @@
-from __future__ import annotations
-
-import json
-from unittest.mock import AsyncMock, MagicMock, patch
-
-import pytest
-
-from vibe.core.nuage.client import WorkflowsClient
-from vibe.core.nuage.exceptions import ErrorCode, WorkflowsException
-from vibe.core.nuage.streaming import StreamEvent, StreamEventsQueryParams
-from vibe.core.nuage.workflow import WorkflowExecutionStatus
-
-
-def _make_client() -> WorkflowsClient:
-    return WorkflowsClient(base_url="http://localhost:8080", api_key="test-key")
-
-
-def _valid_event_payload() -> dict:
-    return {
-        "stream": "test-stream",
-        "timestamp_unix_nano": 1000000,
-        "data": {"key": "value"},
-    }
-
-
-class TestParseSSEData:
-    def test_valid_json_returns_stream_event(self) -> None:
-        client = _make_client()
-        payload = _valid_event_payload()
-        result = client._parse_sse_data(json.dumps(payload), event_type=None)
-        assert isinstance(result, StreamEvent)
-        assert result.stream == "test-stream"
-        assert result.data == {"key": "value"}
-
-    def test_error_event_type_raises(self) -> None:
-        client = _make_client()
-        payload = {"some": "data"}
-        with pytest.raises(WorkflowsException) as exc_info:
-            client._parse_sse_data(json.dumps(payload), event_type="error")
-        assert exc_info.value.code == ErrorCode.GET_EVENTS_STREAM_ERROR
-        assert "Stream error from server" in exc_info.value.message
-
-    def test_error_key_in_json_raises(self) -> None:
-        client = _make_client()
-        payload = {"error": "something went wrong"}
-        with pytest.raises(WorkflowsException) as exc_info:
-            client._parse_sse_data(json.dumps(payload), event_type=None)
-        assert exc_info.value.code == ErrorCode.GET_EVENTS_STREAM_ERROR
-        assert "something went wrong" in exc_info.value.message
-
-    def test_error_event_type_with_non_dict_parsed(self) -> None:
-        client = _make_client()
-        with pytest.raises(WorkflowsException) as exc_info:
-            client._parse_sse_data(json.dumps("a plain string"), event_type="error")
-        assert "a plain string" in exc_info.value.message
-
-    def test_malformed_json_raises(self) -> None:
-        client = _make_client()
-        with pytest.raises(json.JSONDecodeError):
-            client._parse_sse_data("{not valid json", event_type=None)
-
-
-class TestIterSSEEvents:
-    @pytest.mark.asyncio
-    async def test_parses_data_lines(self) -> None:
-        client = _make_client()
-        payload = _valid_event_payload()
-        lines = [f"data: {json.dumps(payload)}"]
-        response = AsyncMock()
-        response.aiter_bytes = _async_byte_iter(lines)
-
-        events = [e async for e in client._iter_sse_events(response)]
-        assert len(events) == 1
-        assert events[0].stream == "test-stream"
-
-    @pytest.mark.asyncio
-    async def test_skips_empty_lines_and_comments(self) -> None:
-        client = _make_client()
-        payload = _valid_event_payload()
-        lines = ["", ": this is a comment", f"data: {json.dumps(payload)}", ""]
-        response = AsyncMock()
-        response.aiter_bytes = _async_byte_iter(lines)
-
-        events = [e async for e in client._iter_sse_events(response)]
-        assert len(events) == 1
-
-    @pytest.mark.asyncio
-    async def test_parses_event_type_and_passes_to_parse(self) -> None:
-        client = _make_client()
-        payload = {"error": "server broke"}
-        lines = ["event: error", f"data: {json.dumps(payload)}"]
-        response = AsyncMock()
-        response.aiter_bytes = _async_byte_iter(lines)
-
-        with pytest.raises(WorkflowsException) as exc_info:
-            _ = [e async for e in client._iter_sse_events(response)]
-        assert exc_info.value.code == ErrorCode.GET_EVENTS_STREAM_ERROR
-
-    @pytest.mark.asyncio
-    async def test_resets_event_type_after_data_line(self) -> None:
-        client = _make_client()
-        payload = _valid_event_payload()
-        lines = [
-            "event: custom_type",
-            f"data: {json.dumps(payload)}",
-            f"data: {json.dumps(payload)}",
-        ]
-        response = AsyncMock()
-        response.aiter_bytes = _async_byte_iter(lines)
-
-        with patch.object(
-            client, "_parse_sse_data", wraps=client._parse_sse_data
-        ) as mock_parse:
-            events = [e async for e in client._iter_sse_events(response)]
-            assert len(events) == 2
-            assert mock_parse.call_args_list[0].args[1] == "custom_type"
-            assert mock_parse.call_args_list[1].args[1] is None
-
-    @pytest.mark.asyncio
-    async def test_skips_non_data_non_event_lines(self) -> None:
-        client = _make_client()
-        payload = _valid_event_payload()
-        lines = ["id: 123", "retry: 5000", f"data: {json.dumps(payload)}"]
-        response = AsyncMock()
-        response.aiter_bytes = _async_byte_iter(lines)
-
-        events = [e async for e in client._iter_sse_events(response)]
-        assert len(events) == 1
-
-    @pytest.mark.asyncio
-    async def test_parse_failure_logs_warning_and_continues(self) -> None:
-        client = _make_client()
-        payload = _valid_event_payload()
-        lines = ["data: {not valid json}", f"data: {json.dumps(payload)}"]
-        response = AsyncMock()
-        response.aiter_bytes = _async_byte_iter(lines)
-
-        with patch("vibe.core.nuage.client.logger") as mock_logger:
-            events = [e async for e in client._iter_sse_events(response)]
-            assert len(events) == 1
-            mock_logger.warning.assert_called_once()
-
-
-def _setup_mock_client(client: WorkflowsClient, mock_response: AsyncMock) -> None:
-    mock_stream = AsyncMock()
-    mock_stream.__aenter__ = AsyncMock(return_value=mock_response)
-    mock_stream.__aexit__ = AsyncMock(return_value=False)
-    mock_http = AsyncMock()
-    mock_http.stream = lambda *args, **kwargs: mock_stream
-    client._client = mock_http
-
-
-class TestStreamEvents:
-    @pytest.mark.asyncio
-    async def test_yields_stream_events(self) -> None:
-        client = _make_client()
-        payload = _valid_event_payload()
-        lines = [f"data: {json.dumps(payload)}"]
-
-        mock_response = AsyncMock()
-        mock_response.raise_for_status = lambda: None
-        mock_response.aiter_bytes = _async_byte_iter(lines)
-        _setup_mock_client(client, mock_response)
-
-        params = StreamEventsQueryParams(workflow_exec_id="wf-1", start_seq=0)
-        events = [e async for e in client.stream_events(params)]
-        assert len(events) == 1
-        assert isinstance(events[0], StreamEvent)
-
-    @pytest.mark.asyncio
-    async def test_reraises_workflows_exception(self) -> None:
-        client = _make_client()
-
-        mock_response = AsyncMock()
-        mock_response.raise_for_status = lambda: None
-        mock_response.aiter_bytes = _async_byte_iter([
-            "event: error",
-            'data: {"error": "stream error"}',
-        ])
-        _setup_mock_client(client, mock_response)
-
-        params = StreamEventsQueryParams(workflow_exec_id="wf-1")
-        with pytest.raises(WorkflowsException) as exc_info:
-            _ = [e async for e in client.stream_events(params)]
-        assert exc_info.value.code == ErrorCode.GET_EVENTS_STREAM_ERROR
-
-    @pytest.mark.asyncio
-    async def test_wraps_other_exceptions_in_workflows_exception(self) -> None:
-        client = _make_client()
-
-        mock_response = AsyncMock()
-        mock_response.raise_for_status.side_effect = RuntimeError("connection lost")
-        _setup_mock_client(client, mock_response)
-
-        params = StreamEventsQueryParams(workflow_exec_id="wf-1")
-        with pytest.raises(WorkflowsException) as exc_info:
-            _ = [e async for e in client.stream_events(params)]
-        assert exc_info.value.code == ErrorCode.GET_EVENTS_STREAM_ERROR
-        assert "Failed to stream events" in exc_info.value.message
-
-
-class TestGetWorkflowRuns:
-    @pytest.mark.asyncio
-    async def test_sends_current_user_filter_by_default(self) -> None:
-        client = _make_client()
-        mock_response = MagicMock()
-        mock_response.raise_for_status.return_value = None
-        mock_response.json.return_value = {"executions": [], "next_page_token": None}
-
-        mock_http = AsyncMock()
-        mock_http.get.return_value = mock_response
-        client._client = mock_http
-
-        await client.get_workflow_runs(
-            workflow_identifier="workflow-1",
-            page_size=10,
-            status=[WorkflowExecutionStatus.RUNNING],
-        )
-
-        mock_http.get.assert_awaited_once()
-        call_params = mock_http.get.call_args.kwargs["params"]
-        assert call_params["user_id"] == "current"
-        assert call_params["workflow_identifier"] == "workflow-1"
-        assert call_params["page_size"] == 10
-
-    @pytest.mark.asyncio
-    async def test_allows_overriding_user_id(self) -> None:
-        client = _make_client()
-        mock_response = MagicMock()
-        mock_response.raise_for_status.return_value = None
-        mock_response.json.return_value = {"executions": [], "next_page_token": None}
-
-        mock_http = AsyncMock()
-        mock_http.get.return_value = mock_response
-        client._client = mock_http
-
-        await client.get_workflow_runs(user_id="user-123")
-
-        call_params = mock_http.get.call_args.kwargs["params"]
-        assert call_params["user_id"] == "user-123"
-
-
-def _async_byte_iter(lines: list[str]):
-    async def _iter():
-        for line in lines:
-            yield f"{line}\n".encode()
-
-    return _iter

tests/core/session/test_image_snapshot.py

@@ -1,11 +1,17 @@
 from __future__ import annotations
 
+import base64
 import hashlib
 from pathlib import Path
 
 import pytest
 
-from vibe.core.session.image_snapshot import ImageSnapshotError, snapshot_image
+from vibe.core.session.image_snapshot import (
+    ImageSnapshotError,
+    snapshot_image,
+    snapshot_image_bytes,
+)
+from vibe.core.types import MAX_IMAGE_BYTES, FileImageSource, InlineImageSource
 
 PNG_BYTES = b"\x89PNG\r\n\x1a\n" + b"\x00" * 16
 
@@ -19,10 +25,11 @@ def test_snapshot_image_copies_to_attachments_dir(tmp_path: Path) -> None:
     att = snapshot_image(src, alias="screenshot.png", session_dir=session_dir)
 
     digest = hashlib.sha1(PNG_BYTES, usedforsecurity=False).hexdigest()
-    assert att.path == (session_dir / "attachments" / f"{digest}.png").resolve()
+    assert isinstance(att.source, FileImageSource)
+    assert att.source.path == (session_dir / "attachments" / f"{digest}.png").resolve()
     assert att.alias == "screenshot.png"
     assert att.mime_type == "image/png"
-    assert att.path.read_bytes() == PNG_BYTES
+    assert att.source.path.read_bytes() == PNG_BYTES
 
 
 def test_snapshot_image_is_idempotent_on_same_bytes(tmp_path: Path) -> None:
@@ -35,7 +42,9 @@ def test_snapshot_image_is_idempotent_on_same_bytes(tmp_path: Path) -> None:
     att_a = snapshot_image(src_a, alias="a.png", session_dir=session_dir)
     att_b = snapshot_image(src_b, alias="b.png", session_dir=session_dir)
 
-    assert att_a.path == att_b.path
+    assert isinstance(att_a.source, FileImageSource)
+    assert isinstance(att_b.source, FileImageSource)
+    assert att_a.source.path == att_b.source.path
     assert sum(1 for _ in (session_dir / "attachments").iterdir()) == 1
 
 
@@ -45,7 +54,8 @@ def test_snapshot_image_returns_source_when_session_dir_is_none(tmp_path: Path)
 
     att = snapshot_image(src, alias="screenshot.png", session_dir=None)
 
-    assert att.path == src.resolve()
+    assert isinstance(att.source, FileImageSource)
+    assert att.source.path == src.resolve()
     assert att.alias == "screenshot.png"
 
 
@@ -69,3 +79,54 @@ def test_snapshot_image_normalizes_jpg_to_jpeg_mime(tmp_path: Path) -> None:
     att = snapshot_image(src, alias="photo.jpg", session_dir=None)
 
     assert att.mime_type == "image/jpeg"
+
+
+def test_snapshot_image_bytes_inlines_when_session_dir_is_none() -> None:
+    att = snapshot_image_bytes(
+        PNG_BYTES, alias="pasted.png", mime_type="image/png", session_dir=None
+    )
+
+    assert isinstance(att.source, InlineImageSource)
+    assert att.source.data == base64.b64encode(PNG_BYTES).decode("ascii")
+    assert att.alias == "pasted.png"
+    assert att.mime_type == "image/png"
+
+
+def test_snapshot_image_bytes_writes_file_when_session_dir_is_set(
+    tmp_path: Path,
+) -> None:
+    session_dir = tmp_path / "session"
+
+    att = snapshot_image_bytes(
+        PNG_BYTES, alias="pasted.png", mime_type="image/png", session_dir=session_dir
+    )
+
+    digest = hashlib.sha1(PNG_BYTES, usedforsecurity=False).hexdigest()
+    assert isinstance(att.source, FileImageSource)
+    assert att.source.path == (session_dir / "attachments" / f"{digest}.png").resolve()
+    assert att.source.path.read_bytes() == PNG_BYTES
+
+
+def test_snapshot_image_bytes_rejects_unsupported_mime() -> None:
+    with pytest.raises(ImageSnapshotError):
+        snapshot_image_bytes(
+            PNG_BYTES, alias="x", mime_type="image/tiff", session_dir=None
+        )
+
+
+def test_snapshot_image_rejects_oversized_file(tmp_path: Path) -> None:
+    src = tmp_path / "big.png"
+    src.write_bytes(b"\x89PNG\r\n\x1a\n" + b"\x00" * MAX_IMAGE_BYTES)
+
+    with pytest.raises(ImageSnapshotError):
+        snapshot_image(src, alias="big.png", session_dir=None)
+
+
+def test_snapshot_image_bytes_rejects_oversized_data() -> None:
+    with pytest.raises(ImageSnapshotError):
+        snapshot_image_bytes(
+            b"x" * (MAX_IMAGE_BYTES + 1),
+            alias="big.png",
+            mime_type="image/png",
+            session_dir=None,
+        )

tests/core/test_agent_loop_refresh_config.py

@@ -0,0 +1,47 @@
+from __future__ import annotations
+
+import pytest
+
+from tests.conftest import build_test_agent_loop, build_test_vibe_config
+from tests.stubs.fake_mcp_registry import FakeMCPRegistry
+from vibe.core.config import MCPHttp, MCPOAuth, MCPStreamableHttp, VibeConfig
+from vibe.core.tools.mcp import AuthStatus
+
+
+def test_refresh_config_reconciles_mcp_registry_status(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    kept = MCPHttp(name="kept", transport="http", url="http://kept:1")
+    removed = MCPHttp(name="removed", transport="http", url="http://removed:1")
+    registry = FakeMCPRegistry()
+    agent_loop = build_test_agent_loop(
+        config=build_test_vibe_config(mcp_servers=[kept, removed]),
+        mcp_registry=registry,
+    )
+    refreshed_config = build_test_vibe_config(mcp_servers=[kept])
+
+    monkeypatch.setattr(VibeConfig, "load", staticmethod(lambda: refreshed_config))
+    agent_loop.refresh_config()
+
+    assert registry.status() == {"kept": AuthStatus.STATIC}
+
+
+def test_refresh_config_does_not_mark_undiscovered_oauth_server_ok(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    oauth = MCPStreamableHttp(
+        name="linear",
+        transport="streamable-http",
+        url="https://mcp.example.com/mcp",
+        auth=MCPOAuth(type="oauth", scopes=["read"]),
+    )
+    registry = FakeMCPRegistry()
+    agent_loop = build_test_agent_loop(
+        config=build_test_vibe_config(mcp_servers=[]), mcp_registry=registry
+    )
+    refreshed_config = build_test_vibe_config(mcp_servers=[oauth])
+
+    monkeypatch.setattr(VibeConfig, "load", staticmethod(lambda: refreshed_config))
+    agent_loop.refresh_config()
+
+    assert registry.status() == {"linear": AuthStatus.NEEDS_AUTH}

tests/core/test_backend_error.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import pytest
 
+from tests.constants import CHAT_COMPLETIONS_PATH
 from vibe.core.llm.exceptions import BackendError, PayloadSummary
 
 
@@ -24,7 +25,7 @@ def _make_error(
 ) -> BackendError:
     return BackendError(
         provider="test-provider",
-        endpoint="/v1/chat/completions",
+        endpoint=CHAT_COMPLETIONS_PATH,
         status=status,
         reason="some reason",
         headers=headers or {},

tests/core/test_build_attachment_counts.py

@@ -4,7 +4,7 @@ from pathlib import Path
 
 from vibe.core.telemetry.build_metadata import build_attachment_counts
 from vibe.core.telemetry.types import AttachmentKind
-from vibe.core.types import ImageAttachment, LLMMessage, Role
+from vibe.core.types import FileImageSource, ImageAttachment, LLMMessage, Role
 
 
 def _msg(images: list[ImageAttachment] | None) -> LLMMessage:
@@ -13,7 +13,9 @@ def _msg(images: list[ImageAttachment] | None) -> LLMMessage:
 
 def _image() -> ImageAttachment:
     return ImageAttachment(
-        path=Path("/tmp/x.png"), alias="x.png", mime_type="image/png"
+        source=FileImageSource(path=Path("/tmp/x.png")),
+        alias="x.png",
+        mime_type="image/png",
     )
 
 

tests/core/test_cache_store.py

@@ -0,0 +1,109 @@
+from __future__ import annotations
+
+from pathlib import Path
+import tomllib
+
+from vibe.core.cache_store import FileSystemVibeCodeCacheStore
+
+
+class TestFileSystemVibeCodeCacheStore:
+    def test_reads_valid_toml_section(self, tmp_path: Path) -> None:
+        cache_path = tmp_path / "cache.toml"
+        cache_path.write_text('[update_cache]\nlatest_version = "1.0.0"\n')
+        store = FileSystemVibeCodeCacheStore(cache_path)
+
+        result = store.read_section("update_cache")
+
+        assert result["latest_version"] == "1.0.0"
+
+    def test_returns_empty_dict_when_file_is_missing(self, tmp_path: Path) -> None:
+        store = FileSystemVibeCodeCacheStore(tmp_path / "missing.toml")
+
+        assert store.read_section("update_cache") == {}
+
+    def test_returns_empty_dict_when_file_is_corrupted(self, tmp_path: Path) -> None:
+        cache_path = tmp_path / "cache.toml"
+        cache_path.write_text("{bad toml")
+        store = FileSystemVibeCodeCacheStore(cache_path)
+
+        assert store.read_section("update_cache") == {}
+
+    def test_writes_new_file(self, tmp_path: Path) -> None:
+        cache_path = tmp_path / "cache.toml"
+        store = FileSystemVibeCodeCacheStore(cache_path)
+
+        store.write_section("feedback", {"last_shown_at": 100.0})
+
+        with cache_path.open("rb") as f:
+            data = tomllib.load(f)
+        assert data["feedback"]["last_shown_at"] == 100.0
+
+    def test_merges_with_existing(self, tmp_path: Path) -> None:
+        cache_path = tmp_path / "cache.toml"
+        cache_path.write_text('[update_cache]\nlatest_version = "1.0.0"\n')
+        store = FileSystemVibeCodeCacheStore(cache_path)
+
+        store.write_section("feedback", {"last_shown_at": 200.0})
+
+        with cache_path.open("rb") as f:
+            data = tomllib.load(f)
+        assert data["update_cache"]["latest_version"] == "1.0.0"
+        assert data["feedback"]["last_shown_at"] == 200.0
+
+    def test_merges_within_section_and_leaves_other_sections_alone(
+        self, tmp_path: Path
+    ) -> None:
+        cache_path = tmp_path / "cache.toml"
+        cache_path.write_text(
+            "[update_cache]\n"
+            'latest_version = "1.0.0"\n'
+            "stored_at_timestamp = 1\n"
+            'seen_whats_new_version = "1.0.0"\n\n'
+            "[feedback]\n"
+            "last_shown_at = 100.0\n"
+        )
+        store = FileSystemVibeCodeCacheStore(cache_path)
+
+        store.write_section(
+            "update_cache", {"latest_version": "2.0.0", "stored_at_timestamp": 2}
+        )
+
+        with cache_path.open("rb") as f:
+            data = tomllib.load(f)
+        assert data["update_cache"]["latest_version"] == "2.0.0"
+        assert data["update_cache"]["stored_at_timestamp"] == 2
+        assert data["update_cache"]["seen_whats_new_version"] == "1.0.0"
+        assert data["feedback"]["last_shown_at"] == 100.0
+
+    def test_reads_section(self, tmp_path: Path) -> None:
+        cache_path = tmp_path / "cache.toml"
+        cache_path.write_text("[feedback]\nlast_shown_at = 100\n")
+        store = FileSystemVibeCodeCacheStore(cache_path)
+
+        assert store.read_section("feedback") == {"last_shown_at": 100}
+
+    def test_returns_empty_dict_for_missing_section(self, tmp_path: Path) -> None:
+        store = FileSystemVibeCodeCacheStore(tmp_path / "cache.toml")
+
+        assert store.read_section("feedback") == {}
+
+    def test_writes_section(self, tmp_path: Path) -> None:
+        cache_path = tmp_path / "cache.toml"
+        store = FileSystemVibeCodeCacheStore(cache_path)
+
+        store.write_section("feedback", {"last_shown_at": 100})
+
+        with cache_path.open("rb") as f:
+            data = tomllib.load(f)
+        assert data["feedback"]["last_shown_at"] == 100
+
+    def test_replaces_non_table_section_when_writing(self, tmp_path: Path) -> None:
+        cache_path = tmp_path / "cache.toml"
+        cache_path.write_text('feedback = "not-a-table"\n')
+        store = FileSystemVibeCodeCacheStore(cache_path)
+
+        store.write_section("feedback", {"last_shown_at": 100})
+
+        with cache_path.open("rb") as f:
+            data = tomllib.load(f)
+        assert data["feedback"]["last_shown_at"] == 100

tests/core/test_config_builder.py

@@ -31,6 +31,9 @@ class FakeLayer(ConfigLayer[RawConfig]):
     async def _build_config_snapshot(self) -> LayerConfigSnapshot:
         return LayerConfigSnapshot(data=dict(self._data), fingerprint="fp")
 
+    async def _save_to_store(self, _next_config: RawConfig) -> str:
+        raise NotImplementedError
+
 
 class UntrustedFakeLayer(FakeLayer):
     async def _check_trust(self) -> bool:

tests/core/test_config_layer.py

@@ -3,18 +3,24 @@ from __future__ import annotations
 import asyncio
 from typing import Any
 
+from jsonpointer import JsonPointerException
 from pydantic import BaseModel, ValidationError
 import pytest
 
 from vibe.core.config.layer import (
     ConfigLayer,
+    ConfigPatchApplicationError,
     LayerImplementationError,
     RawConfig,
     TrustNotResolvedError,
     UntrustedLayerError,
 )
-from vibe.core.config.patch import ConfigPatch
-from vibe.core.config.types import ConcurrencyConflictError, LayerConfigSnapshot
+from vibe.core.config.patch import AddOperationPatch, ConfigPatch, ReplaceOperationPatch
+from vibe.core.config.types import (
+    ConcurrencyConflictError,
+    ConflictStrategy,
+    LayerConfigSnapshot,
+)
 
 
 class StubLayer(ConfigLayer[BaseModel]):
@@ -45,6 +51,9 @@ class StubLayer(ConfigLayer[BaseModel]):
             data=dict(self._data), fingerprint=f"fp-{self.read_count}"
         )
 
+    async def _save_to_store(self, _next_config: BaseModel) -> str:
+        raise NotImplementedError("StubLayer.apply() is not implemented")
+
 
 class ObservableStubLayer(StubLayer):
     """Stub that records _on_trust_changed calls."""
@@ -57,6 +66,18 @@ class ObservableStubLayer(StubLayer):
         self.trust_changes.append((old, new))
 
 
+class WritableStubLayer(StubLayer):
+    def __init__(self, **kwargs: Any) -> None:
+        super().__init__(**kwargs)
+        self.writes: list[dict[str, Any]] = []
+
+    async def _save_to_store(self, _next_config: BaseModel) -> str:
+        next_data = _next_config.model_dump()
+        self.writes.append(next_data)
+        self._data = next_data
+        return f"write-fp-{len(self.writes)}"
+
+
 class SampleSchema(BaseModel):
     name: str
     count: int = 0
@@ -91,6 +112,9 @@ async def test_default_check_trust_returns_false() -> None:
         async def _build_config_snapshot(self) -> LayerConfigSnapshot:
             return LayerConfigSnapshot(data={}, fingerprint="fp")
 
+        async def _save_to_store(self, _next_config: BaseModel) -> str:
+            raise NotImplementedError
+
     layer = DefaultTrustLayer(name="default")
     result = await layer.resolve_trust()
     assert result is False
@@ -236,7 +260,7 @@ async def test_load_returns_data() -> None:
     layer = StubLayer(data={"key": "value"})
     result = await layer.load()
     assert isinstance(result, RawConfig)
-    assert result.model_extra == {"key": "value"}
+    assert result.model_dump() == {"key": "value"}
     assert layer.fingerprint == "fp-1"
 
 
@@ -246,7 +270,7 @@ async def test_load_auto_resolves_trust() -> None:
     assert layer.is_trusted is None
     result = await layer.load()
     assert layer.is_trusted is True
-    assert result.model_extra == {"a": 1}
+    assert result.model_dump() == {"a": 1}
     assert layer.fingerprint == "fp-1"
 
     await layer.resolve_trust()
@@ -311,21 +335,21 @@ async def test_invalidate_cache_causes_reload() -> None:
 async def test_revoke_grant_cycle_refreshes_data() -> None:
     layer = StubLayer(data={"v": 1})
     result1 = await layer.load()
-    assert result1.model_extra == {"v": 1}
+    assert result1.model_dump() == {"v": 1}
 
     await layer.revoke_trust()
     layer._data = {"v": 2}
     await layer.grant_trust()
 
     result2 = await layer.load()
-    assert result2.model_extra == {"v": 2}
+    assert result2.model_dump() == {"v": 2}
 
 
 @pytest.mark.asyncio
 async def test_resolve_trust_clears_data_on_revocation() -> None:
     layer = StubLayer(data={"v": 1})
     result1 = await layer.load()
-    assert result1.model_extra == {"v": 1}
+    assert result1.model_dump() == {"v": 1}
 
     # External revocation via resolve_trust (not revoke_trust)
     layer._stub_trusted = False
@@ -339,7 +363,7 @@ async def test_resolve_trust_clears_data_on_revocation() -> None:
     await layer.resolve_trust()
 
     result2 = await layer.load()
-    assert result2.model_extra == {"v": 2}
+    assert result2.model_dump() == {"v": 2}
 
 
 @pytest.mark.asyncio
@@ -350,7 +374,7 @@ async def test_load_returns_deep_copy() -> None:
     result1.model_extra["items"].append("mutated")
 
     result2 = await layer.load()
-    assert result2.model_extra == {"items": ["a", "b"]}
+    assert result2.model_dump() == {"items": ["a", "b"]}
     assert layer.read_count == 1
 
 
@@ -384,7 +408,7 @@ async def test_default_schema_preserves_extras() -> None:
     layer = StubLayer(data={"anything": "goes"})
     result = await layer.load()
     assert isinstance(result, RawConfig)
-    assert result.model_extra == {"anything": "goes"}
+    assert result.model_dump() == {"anything": "goes"}
 
 
 @pytest.mark.asyncio
@@ -423,6 +447,9 @@ async def test_concurrent_loads_serialize() -> None:
                 data={"v": self.read_count}, fingerprint=f"fp-{self.read_count}"
             )
 
+        async def _save_to_store(self, _next_config: BaseModel) -> str:
+            raise NotImplementedError
+
     layer = SlowLayer()
     results = await asyncio.gather(layer.load(), layer.load(), layer.load())
     assert layer.read_count == 1
@@ -438,8 +465,155 @@ async def test_fingerprint_returns_none_before_load() -> None:
 @pytest.mark.asyncio
 async def test_apply_not_implemented() -> None:
     layer = StubLayer()
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+
     with pytest.raises(NotImplementedError):
-        await layer.apply(ConfigPatch(fingerprint="fp-1"))
+        await layer.apply(ConfigPatch(fingerprint=fingerprint))
+
+
+@pytest.mark.asyncio
+async def test_apply_operation_failure_wrapped() -> None:
+    original_data = {"tools": {"disabled_tools": ["bash"]}}
+    layer = WritableStubLayer(data=original_data)
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+
+    with pytest.raises(
+        ConfigPatchApplicationError, match="Layer 'stub': failed to apply patch"
+    ) as exc_info:
+        await layer.apply(
+            ConfigPatch(
+                AddOperationPatch(path="/tools/enabled_tools/-", value="read"),
+                fingerprint=fingerprint,
+            )
+        )
+
+    assert exc_info.value.layer_name == "stub"
+    assert isinstance(exc_info.value.__cause__, JsonPointerException)
+    assert "enabled_tools" in str(exc_info.value.__cause__)
+    assert layer.fingerprint == fingerprint
+    assert (await layer.load()).model_dump() == original_data
+    assert layer.writes == []
+
+
+@pytest.mark.asyncio
+async def test_apply_operation_failure_after_successful_operation_is_atomic() -> None:
+    original_data = {"active_model": "old", "tools": {"disabled_tools": ["bash"]}}
+    layer = WritableStubLayer(data=original_data)
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+
+    with pytest.raises(ConfigPatchApplicationError):
+        await layer.apply(
+            ConfigPatch(
+                ReplaceOperationPatch(path="/active_model", value="new"),
+                AddOperationPatch(path="/tools/enabled_tools/-", value="read"),
+                fingerprint=fingerprint,
+            )
+        )
+
+    assert layer.fingerprint == fingerprint
+    assert (await layer.load()).model_dump() == original_data
+    assert layer.writes == []
+
+
+@pytest.mark.asyncio
+async def test_apply_schema_validation_failure_wrapped() -> None:
+    original_data = {"name": "test", "count": 1}
+    layer = WritableStubLayer(output_schema=SampleSchema, data=original_data)
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+
+    with pytest.raises(
+        ConfigPatchApplicationError, match="Layer 'stub': failed to apply patch"
+    ) as exc_info:
+        await layer.apply(
+            ConfigPatch(
+                ReplaceOperationPatch(path="/count", value={"bad": "value"}),
+                fingerprint=fingerprint,
+            )
+        )
+
+    assert exc_info.value.layer_name == "stub"
+    assert isinstance(exc_info.value.__cause__, ValidationError)
+    assert layer.fingerprint == fingerprint
+    assert (await layer.load()).model_dump() == original_data
+    assert layer.writes == []
+
+
+@pytest.mark.asyncio
+async def test_apply_cancel_rejects_stale_fingerprint() -> None:
+    layer = WritableStubLayer(data={"key": "old"})
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+
+    await layer.apply(
+        ConfigPatch(
+            ReplaceOperationPatch(path="/key", value="first"), fingerprint=fingerprint
+        )
+    )
+
+    with pytest.raises(ConcurrencyConflictError) as exc_info:
+        await layer.apply(
+            ConfigPatch(
+                ReplaceOperationPatch(path="/key", value="second"),
+                fingerprint=fingerprint,
+            )
+        )
+
+    assert exc_info.value.expected_fp == fingerprint
+    assert exc_info.value.actual_fp == layer.fingerprint
+    assert (await layer.load()).model_dump() == {"key": "first"}
+
+
+@pytest.mark.asyncio
+async def test_apply_replace_accepts_stale_fingerprint() -> None:
+    layer = WritableStubLayer(data={"key": "old"})
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+
+    await layer.apply(
+        ConfigPatch(
+            ReplaceOperationPatch(path="/key", value="first"), fingerprint=fingerprint
+        )
+    )
+    await layer.apply(
+        ConfigPatch(
+            ReplaceOperationPatch(path="/key", value="second"), fingerprint=fingerprint
+        ),
+        on_conflict=ConflictStrategy.REPLACE,
+    )
+
+    assert (await layer.load()).model_dump() == {"key": "second"}
+    assert layer.writes == [{"key": "first"}, {"key": "second"}]
+
+
+@pytest.mark.asyncio
+async def test_save_to_store_failure_wrapped() -> None:
+    class BrokenApplyLayer(StubLayer):
+        async def _save_to_store(self, _next_config: BaseModel) -> str:
+            raise OSError("disk full")
+
+    layer = BrokenApplyLayer(data={"key": "old"})
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+
+    with pytest.raises(LayerImplementationError, match="_save_to_store") as exc_info:
+        await layer.apply(
+            ConfigPatch(
+                ReplaceOperationPatch(path="/key", value="new"), fingerprint=fingerprint
+            )
+        )
+
+    assert isinstance(exc_info.value.__cause__, OSError)
 
 
 @pytest.mark.asyncio
@@ -499,6 +673,9 @@ class FakeLocalUserLayer(ConfigLayer[UserConfigSchema]):
     async def _build_config_snapshot(self) -> LayerConfigSnapshot:
         return LayerConfigSnapshot(data=dict(self._data), fingerprint="user-fp")
 
+    async def _save_to_store(self, _next_config: UserConfigSchema) -> str:
+        raise NotImplementedError
+
 
 @pytest.mark.asyncio
 async def test_scenario_local_user_layer_always_trusted() -> None:
@@ -544,6 +721,9 @@ class FakeLocalProjectLayer(ConfigLayer[BaseModel]):
     async def _build_config_snapshot(self) -> LayerConfigSnapshot:
         return LayerConfigSnapshot(data=dict(self._data), fingerprint="project-fp")
 
+    async def _save_to_store(self, _next_config: BaseModel) -> str:
+        raise NotImplementedError
+
 
 @pytest.mark.asyncio
 async def test_scenario_local_project_layer_trust_lifecycle() -> None:
@@ -561,7 +741,7 @@ async def test_scenario_local_project_layer_trust_lifecycle() -> None:
     await layer.grant_trust()
     assert trust_store == {"/tmp/my-project": True}
     result = await layer.load()
-    assert result.model_extra == project_data
+    assert result.model_dump() == project_data
 
     # 3. New instance with same trust store — loads directly
     layer2 = FakeLocalProjectLayer(
@@ -570,7 +750,7 @@ async def test_scenario_local_project_layer_trust_lifecycle() -> None:
     assert layer2.is_trusted is None
     result2 = await layer2.load()
     assert layer2.is_trusted is True
-    assert result2.model_extra == project_data
+    assert result2.model_dump() == project_data
 
     # 4. Revoke trust — removed from store, load raises
     await layer2.revoke_trust()

tests/core/test_config_load_dotenv.py

@@ -20,7 +20,7 @@ def test_skips_missing_file(tmp_path: Path) -> None:
     assert environ == {"EXISTING": "1"}
 
 
-def test_sets_and_overrides_values(tmp_path: Path) -> None:
+def test_adds_missing_values_without_overriding_existing(tmp_path: Path) -> None:
     env_path = tmp_path / ".env"
     _write_env_file(
         env_path,
@@ -42,11 +42,23 @@ def test_sets_and_overrides_values(tmp_path: Path) -> None:
 
     load_dotenv_values(env_path=env_path, environ=environ)
 
-    assert environ["MISTRAL_API_KEY"] == "new-key"
-    assert environ["HTTPS_PROXY"] == "https://local-proxy:8080"
-    assert environ["OTHER"] == "from-env"
+    # An explicit process/shell value wins over the .env file.
+    assert environ["MISTRAL_API_KEY"] == "old-key"
+    assert environ["HTTPS_PROXY"] == "old-https"
+    assert environ["OTHER"] == "keep"
+    assert environ["FOO"] == "keep"
+    # Keys absent from the process env are still loaded from the .env file.
     assert environ["NEW_KEY"] == "added"
-    assert environ["FOO"] == "replace"
+
+
+def test_adds_dotenv_value_when_process_env_is_empty(tmp_path: Path) -> None:
+    env_path = tmp_path / ".env"
+    _write_env_file(env_path, "MISTRAL_API_KEY=file-key\n")
+    environ = {"MISTRAL_API_KEY": ""}
+
+    load_dotenv_values(env_path=env_path, environ=environ)
+
+    assert environ["MISTRAL_API_KEY"] == "file-key"
 
 
 def test_ignores_empty_values(tmp_path: Path) -> None:

tests/core/test_config_mcp_auth.py

@@ -1,11 +1,12 @@
 from __future__ import annotations
 
-import logging
+from unittest.mock import AsyncMock, MagicMock, patch
 
 from pydantic import ValidationError
 import pytest
 
 from vibe.core.config import MCPHttp, MCPOAuth, MCPStaticAuth, MCPStreamableHttp
+from vibe.core.tools.mcp import AuthStatus
 from vibe.core.tools.mcp.registry import MCPRegistry
 
 HTTP_TRANSPORTS = [
@@ -183,10 +184,8 @@ def test_oauth_scopes_empty_list_allowed() -> None:
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize(("cls", "transport"), HTTP_TRANSPORTS)
-async def test_registry_skips_oauth_servers_with_gated_warning(
-    cls: type[MCPHttp | MCPStreamableHttp],
-    transport: str,
-    caplog: pytest.LogCaptureFixture,
+async def test_registry_marks_oauth_servers_without_tokens_as_needing_auth(
+    cls: type[MCPHttp | MCPStreamableHttp], transport: str
 ) -> None:
     srv = cls.model_validate({
         "name": "linear",
@@ -195,19 +194,23 @@ async def test_registry_skips_oauth_servers_with_gated_warning(
         "auth": {"type": "oauth", "scopes": ["read"]},
     })
     registry = MCPRegistry()
+    storage = MagicMock()
+    storage.get_tokens = AsyncMock(return_value=None)
+    storage.delete_tokens = AsyncMock()
+    storage.delete_client_info = AsyncMock()
 
-    with caplog.at_level(logging.WARNING, logger="vibe"):
+    with (
+        patch("vibe.core.tools.mcp.registry.KeyringTokenStorage", return_value=storage),
+        patch(
+            "vibe.core.tools.mcp.registry.Fingerprint.load",
+            new=AsyncMock(return_value=None),
+        ),
+        patch("vibe.core.tools.mcp.registry.Fingerprint.delete", new=AsyncMock()),
+    ):
         first = await registry.get_tools_async([srv])
-
-    assert first == {}
-    assert (
-        "OAuth support for MCP servers is not yet enabled; coming in a future release"
-        in caplog.text
-    )
-
-    caplog.clear()
-    with caplog.at_level(logging.WARNING, logger="vibe"):
         second = await registry.get_tools_async([srv])
 
+    assert first == {}
     assert second == {}
-    assert "OAuth support" not in caplog.text
+    assert registry.needs_auth == {"linear"}
+    assert registry.status()["linear"] == AuthStatus.NEEDS_AUTH

tests/core/test_config_orchestrator.py

@@ -23,6 +23,9 @@ class FakeLayer(ConfigLayer[RawConfig]):
     async def _build_config_snapshot(self) -> LayerConfigSnapshot:
         return LayerConfigSnapshot(data=dict(self._data), fingerprint="fp")
 
+    async def _save_to_store(self, _next_config: RawConfig) -> str:
+        raise NotImplementedError
+
 
 class SimpleSchema(ConfigSchema):
     value: Annotated[str, WithReplaceMerge()] = "default"

tests/core/test_config_otel.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import pytest
 
+from tests.constants import ANTHROPIC_BASE_URL
 from vibe.core.config import OtelSpanExporterConfig, ProviderConfig, VibeConfig
 from vibe.core.types import Backend
 
@@ -62,7 +63,7 @@ class TestOtelSpanExporterConfig:
             update={
                 "providers": [
                     ProviderConfig(
-                        name="anthropic", api_base="https://api.anthropic.com/v1"
+                        name="anthropic", api_base=f"{ANTHROPIC_BASE_URL}/v1"
                     )
                 ]
             }
@@ -80,6 +81,18 @@ class TestOtelSpanExporterConfig:
         assert result is not None
         assert result.endpoint == "https://api.mistral.ai/telemetry/v1/traces"
 
+    def test_resolves_api_key_from_keyring(
+        self, vibe_config: VibeConfig, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        # Key stored only in the OS keyring (no env var) must still authenticate OTEL.
+        monkeypatch.delenv("MISTRAL_API_KEY", raising=False)
+        monkeypatch.setattr(
+            "keyring.get_password", lambda service, username: "sk-keyring"
+        )
+        result = vibe_config.otel_span_exporter_config
+        assert result is not None
+        assert result.headers == {"Authorization": "Bearer sk-keyring"}
+
     def test_returns_none_and_warns_when_api_key_missing(
         self,
         vibe_config: VibeConfig,

tests/core/test_config_patch_ops.py

@@ -1,135 +1,84 @@
 from __future__ import annotations
 
-from collections.abc import Callable
-from dataclasses import FrozenInstanceError
 from typing import Any, get_args
 
+from pydantic import ValidationError
 import pytest
 
 from vibe.core.config import (
-    AppendToList,
-    DeleteField,
-    PatchOp,
-    RemoveFromList,
-    SetField,
+    AddOperationPatch,
+    RemoveOperationPatch,
+    ReplaceOperationPatch,
 )
-from vibe.core.config.patch import ConfigPatch
+from vibe.core.config.patch import ConfigPatch, PatchOp
+
+
+@pytest.mark.parametrize(
+    ("operation", "expected"),
+    [
+        (
+            AddOperationPatch(path="/tools/disabled_tools/0", value="bash"),
+            {"op": "add", "path": "/tools/disabled_tools/0", "value": "bash"},
+        ),
+        (
+            ReplaceOperationPatch(path="/active_model", value="devstral-small"),
+            {"op": "replace", "path": "/active_model", "value": "devstral-small"},
+        ),
+        (
+            RemoveOperationPatch(path="/tools/deprecated_setting"),
+            {"op": "remove", "path": "/tools/deprecated_setting"},
+        ),
+    ],
+)
+def test_json_patch_operations_convert_to_json_patch_payload(
+    operation: AddOperationPatch | ReplaceOperationPatch | RemoveOperationPatch,
+    expected: dict[str, Any],
+) -> None:
+    assert operation.to_json_patch() == expected
 
 
 def test_patch_op_union_contains_all_operations() -> None:
-    assert get_args(PatchOp) == (SetField, AppendToList, RemoveFromList, DeleteField)
-
-
-def test_set_field_accepts_top_level_key() -> None:
-    op = SetField("active_model", "devstral-small")
-
-    assert op.key == "active_model"
-    assert op.value == "devstral-small"
-
-
-def test_set_field_accepts_nested_key() -> None:
-    op = SetField("models.providers", {"mistral": {"region": "eu"}})
-
-    assert op.key == "models.providers"
-
-
-def test_append_to_list_accepts_nested_key() -> None:
-    op = AppendToList("tools.disabled_tools", ("bash", "python"))
-
-    assert op.key == "tools.disabled_tools"
-    assert op.items == ("bash", "python")
-
-
-def test_remove_from_list_accepts_nested_key() -> None:
-    op = RemoveFromList("models.available_models", ("codestral-latest",))
-
-    assert op.key == "models.available_models"
-    assert op.values == ("codestral-latest",)
-
-
-def test_delete_field_accepts_nested_key() -> None:
-    op = DeleteField("tools.deprecated_setting")
-
-    assert op.key == "tools.deprecated_setting"
+    assert set(get_args(PatchOp.__value__)) == {
+        AddOperationPatch,
+        ReplaceOperationPatch,
+        RemoveOperationPatch,
+    }
 
 
 @pytest.mark.parametrize(
     "factory",
     [
-        lambda key: SetField(key, "value"),
-        lambda key: AppendToList(key, ("value",)),
-        lambda key: RemoveFromList(key, ("value",)),
-        lambda key: DeleteField(key),
+        lambda path: AddOperationPatch(path=path, value="value"),
+        lambda path: ReplaceOperationPatch(path=path, value="value"),
+        lambda path: RemoveOperationPatch(path=path),
     ],
 )
-@pytest.mark.parametrize(
-    "invalid_key", ["", ".active_model", "active_model.", "tools..bash"]
-)
-def test_patch_operations_reject_invalid_key_paths(
-    factory: Callable[[str], object], invalid_key: str
-) -> None:
-    with pytest.raises(ValueError, match="dot-separated path|must not be empty"):
-        factory(invalid_key)
+def test_json_patch_operations_reject_non_pointer_paths(factory: Any) -> None:
+    with pytest.raises(ValidationError, match="valid JSON Pointer"):
+        factory("tools.disabled_tools")
 
 
 @pytest.mark.parametrize(
     "factory",
     [
-        lambda key: SetField(key, "value"),
-        lambda key: AppendToList(key, ("value",)),
-        lambda key: RemoveFromList(key, ("value",)),
-        lambda key: DeleteField(key),
+        lambda path: AddOperationPatch(path=path, value="value"),
+        lambda path: ReplaceOperationPatch(path=path, value="value"),
+        lambda path: RemoveOperationPatch(path=path),
     ],
 )
-def test_patch_operations_reject_non_string_keys(
-    factory: Callable[[Any], object],
-) -> None:
-    with pytest.raises(TypeError, match="Patch operation key must be a string"):
-        factory(1)
+def test_json_patch_operations_reject_invalid_escapes(factory: Any) -> None:
+    with pytest.raises(ValidationError, match="valid JSON Pointer"):
+        factory("/tools/~2")
 
 
-def test_append_to_list_rejects_non_tuple_items() -> None:
-    bad_items: Any = ["bash"]
+def test_json_patch_operations_accept_slash_prefixed_paths() -> None:
+    op = ReplaceOperationPatch(path="/", value={"active_model": "devstral-small"})
 
-    with pytest.raises(TypeError, match="AppendToList.items must be a tuple"):
-        AppendToList("tools.disabled_tools", bad_items)
-
-
-def test_remove_from_list_rejects_non_tuple_values() -> None:
-    bad_values: Any = ["bash"]
-
-    with pytest.raises(TypeError, match="RemoveFromList.values must be a tuple"):
-        RemoveFromList("tools.disabled_tools", bad_values)
-
-
-def test_patch_operations_are_frozen() -> None:
-    op = SetField("active_model", "devstral-small")
-
-    with pytest.raises(FrozenInstanceError):
-        op.__setattr__("key", "models.active_model")
-
-
-def test_scenario_mini_vibe_patch_operations() -> None:
-    operations: list[PatchOp] = [
-        SetField("active_model", "devstral-small"),
-        AppendToList("tools.disabled_tools", ("bash",)),
-        RemoveFromList("models.available_models", ("codestral-latest",)),
-        DeleteField("tools.deprecated_setting"),
-    ]
-
-    assert operations == [
-        SetField("active_model", "devstral-small"),
-        AppendToList("tools.disabled_tools", ("bash",)),
-        RemoveFromList("models.available_models", ("codestral-latest",)),
-        DeleteField("tools.deprecated_setting"),
-    ]
-
-
-# --- ConfigPatch ---
+    assert op.path == "/"
 
 
 def test_config_patch_stores_operations_and_metadata() -> None:
-    op = SetField("active_model", "devstral-small")
+    op = ReplaceOperationPatch(path="/active_model", value="devstral-small")
     patch = ConfigPatch(op, fingerprint="fp-1", reason="test")
 
     assert patch.operations == [op]
@@ -145,9 +94,9 @@ def test_config_patch_defaults() -> None:
 
 
 def test_config_patch_accepts_multiple_operations() -> None:
-    ops: list[PatchOp] = [
-        SetField("active_model", "devstral-small"),
-        AppendToList("tools.disabled_tools", ("bash",)),
+    ops = [
+        ReplaceOperationPatch(path="/active_model", value="devstral-small"),
+        AddOperationPatch(path="/tools/disabled_tools/-", value="bash"),
     ]
     patch = ConfigPatch(*ops, fingerprint="fp-1")
 
@@ -155,18 +104,23 @@ def test_config_patch_accepts_multiple_operations() -> None:
 
 
 def test_config_patch_add_appends_operations() -> None:
-    patch = ConfigPatch(SetField("active_model", "devstral-small"), fingerprint="fp-1")
-    patch.add(DeleteField("tools.deprecated_setting"))
+    patch = ConfigPatch(
+        ReplaceOperationPatch(path="/active_model", value="devstral-small"),
+        fingerprint="fp-1",
+    )
+    patch.add(RemoveOperationPatch(path="/tools/deprecated_setting"))
 
     assert patch.operations == [
-        SetField("active_model", "devstral-small"),
-        DeleteField("tools.deprecated_setting"),
+        ReplaceOperationPatch(path="/active_model", value="devstral-small"),
+        RemoveOperationPatch(path="/tools/deprecated_setting"),
     ]
 
 
 def test_config_patch_add_returns_self() -> None:
     patch = ConfigPatch(fingerprint="fp-1")
-    result = patch.add(SetField("active_model", "devstral-small"))
+    result = patch.add(
+        ReplaceOperationPatch(path="/active_model", value="devstral-small")
+    )
 
     assert result is patch
 
@@ -174,8 +128,8 @@ def test_config_patch_add_returns_self() -> None:
 def test_config_patch_add_accepts_multiple_operations() -> None:
     patch = ConfigPatch(fingerprint="fp-1")
     patch.add(
-        SetField("active_model", "devstral-small"),
-        AppendToList("tools.disabled_tools", ("bash",)),
+        ReplaceOperationPatch(path="/active_model", value="devstral-small"),
+        AddOperationPatch(path="/tools/disabled_tools/-", value="bash"),
     )
 
     assert len(patch.operations) == 2
@@ -184,42 +138,52 @@ def test_config_patch_add_accepts_multiple_operations() -> None:
 def test_config_patch_add_is_chainable() -> None:
     patch = (
         ConfigPatch(fingerprint="fp-1")
-        .add(SetField("active_model", "devstral-small"))
-        .add(DeleteField("tools.deprecated_setting"))
+        .add(ReplaceOperationPatch(path="/active_model", value="devstral-small"))
+        .add(RemoveOperationPatch(path="/tools/deprecated_setting"))
     )
 
     assert len(patch.operations) == 2
 
 
-def test_config_patch_describe_set_field() -> None:
-    patch = ConfigPatch(SetField("active_model", "devstral-small"), fingerprint="fp-1")
-
-    assert patch.describe() == ["set 'active_model' = 'devstral-small'"]
-
-
-def test_config_patch_describe_append_to_list() -> None:
+def test_config_patch_to_json_patch_from_wrappers() -> None:
     patch = ConfigPatch(
-        AppendToList("tools.disabled_tools", ("bash", "python")), fingerprint="fp-1"
-    )
-
-    assert patch.describe() == ["append to 'tools.disabled_tools': ['bash', 'python']"]
-
-
-def test_config_patch_describe_remove_from_list() -> None:
-    patch = ConfigPatch(
-        RemoveFromList("models.available_models", ("codestral-latest",)),
+        ReplaceOperationPatch(path="/active_model", value="devstral-small"),
+        AddOperationPatch(path="/tools/disabled_tools/-", value="bash"),
+        RemoveOperationPatch(path="/tools/deprecated_setting"),
         fingerprint="fp-1",
     )
 
-    assert patch.describe() == [
-        "remove from 'models.available_models': ['codestral-latest']"
+    assert patch.to_json_patch() == [
+        {"op": "replace", "path": "/active_model", "value": "devstral-small"},
+        {"op": "add", "path": "/tools/disabled_tools/-", "value": "bash"},
+        {"op": "remove", "path": "/tools/deprecated_setting"},
     ]
 
 
-def test_config_patch_describe_delete_field() -> None:
-    patch = ConfigPatch(DeleteField("tools.deprecated_setting"), fingerprint="fp-1")
+def test_config_patch_describe_add_operation() -> None:
+    patch = ConfigPatch(
+        AddOperationPatch(path="/tools/disabled_tools/-", value="bash"),
+        fingerprint="fp-1",
+    )
 
-    assert patch.describe() == ["delete 'tools.deprecated_setting'"]
+    assert patch.describe() == ["add '/tools/disabled_tools/-' = 'bash'"]
+
+
+def test_config_patch_describe_replace_operation() -> None:
+    patch = ConfigPatch(
+        ReplaceOperationPatch(path="/active_model", value="devstral-small"),
+        fingerprint="fp-1",
+    )
+
+    assert patch.describe() == ["replace '/active_model' = 'devstral-small'"]
+
+
+def test_config_patch_describe_remove_operation() -> None:
+    patch = ConfigPatch(
+        RemoveOperationPatch(path="/tools/deprecated_setting"), fingerprint="fp-1"
+    )
+
+    assert patch.describe() == ["remove '/tools/deprecated_setting'"]
 
 
 def test_config_patch_describe_empty_returns_empty_list() -> None:
@@ -230,28 +194,28 @@ def test_config_patch_describe_empty_returns_empty_list() -> None:
 
 def test_config_patch_describe_multiple_operations() -> None:
     patch = ConfigPatch(
-        SetField("active_model", "devstral-small"),
-        AppendToList("tools.disabled_tools", ("bash",)),
-        DeleteField("tools.deprecated_setting"),
+        ReplaceOperationPatch(path="/active_model", value="devstral-small"),
+        AddOperationPatch(path="/tools/disabled_tools/-", value="bash"),
+        RemoveOperationPatch(path="/tools/deprecated_setting"),
         fingerprint="fp-1",
     )
 
     assert patch.describe() == [
-        "set 'active_model' = 'devstral-small'",
-        "append to 'tools.disabled_tools': ['bash']",
-        "delete 'tools.deprecated_setting'",
+        "replace '/active_model' = 'devstral-small'",
+        "add '/tools/disabled_tools/-' = 'bash'",
+        "remove '/tools/deprecated_setting'",
     ]
 
 
 def test_scenario_build_patch_incrementally() -> None:
     patch = ConfigPatch(fingerprint="fp-abc", reason="/model command")
-    patch.add(SetField("active_model", "devstral-small"))
-    patch.add(AppendToList("tools.disabled_tools", ("bash",)))
+    patch.add(ReplaceOperationPatch(path="/active_model", value="devstral-small"))
+    patch.add(AddOperationPatch(path="/tools/disabled_tools/-", value="bash"))
 
     assert patch.fingerprint == "fp-abc"
     assert patch.reason == "/model command"
     assert len(patch.operations) == 2
     assert patch.describe() == [
-        "set 'active_model' = 'devstral-small'",
-        "append to 'tools.disabled_tools': ['bash']",
+        "replace '/active_model' = 'devstral-small'",
+        "add '/tools/disabled_tools/-' = 'bash'",
     ]

tests/core/test_fingerprint.py

@@ -4,7 +4,11 @@ from pathlib import Path
 
 import pytest
 
-from vibe.core.config.fingerprint import capture_stable_file, create_dict_fingerprint
+from vibe.core.config.fingerprint import (
+    capture_stable_file,
+    create_dict_fingerprint,
+    create_file_fingerprint,
+)
 from vibe.core.config.types import ConcurrencyConflictError
 
 
@@ -67,6 +71,32 @@ class TestCaptureStableFile:
                 pass
 
 
+class TestCreateFileFingerprint:
+    def test_captures_file_state(self, tmp_working_directory: Path) -> None:
+        path = tmp_working_directory / "config.toml"
+        path.write_text("key = 1")
+
+        with path.open("rb") as file:
+            first_fingerprint = create_file_fingerprint(file)
+        with path.open("rb") as file:
+            second_fingerprint = create_file_fingerprint(file)
+
+        assert isinstance(first_fingerprint, str)
+        assert first_fingerprint
+        assert first_fingerprint == second_fingerprint
+
+    def test_changes_when_file_changes(self, tmp_working_directory: Path) -> None:
+        path = tmp_working_directory / "config.toml"
+        path.write_text("key = 1")
+        with path.open("rb") as file:
+            first_fingerprint = create_file_fingerprint(file)
+
+        path.write_text("key = 2")
+
+        with path.open("rb") as file:
+            assert create_file_fingerprint(file) != first_fingerprint
+
+
 class TestCreateDictFingerprint:
     def test_empty_dict_returns_stable_non_empty_token(self) -> None:
         first_fingerprint = create_dict_fingerprint({})

tests/core/test_image_attachment.py

@@ -0,0 +1,58 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+from pydantic import ValidationError
+import pytest
+
+from vibe.core.types import FileImageSource, ImageAttachment, InlineImageSource
+
+
+def test_migrates_legacy_flat_path_shape() -> None:
+    att = ImageAttachment.model_validate({
+        "path": "/tmp/a.png",
+        "alias": "a.png",
+        "mime_type": "image/png",
+    })
+
+    assert isinstance(att.source, FileImageSource)
+    assert att.source.path == Path("/tmp/a.png")
+
+
+def test_migrates_legacy_flat_data_shape() -> None:
+    att = ImageAttachment.model_validate({
+        "data": "Zm9v",
+        "alias": "pasted.png",
+        "mime_type": "image/png",
+    })
+
+    assert isinstance(att.source, InlineImageSource)
+    assert att.source.data == "Zm9v"
+
+
+def test_source_path_construction() -> None:
+    att = ImageAttachment(
+        source=FileImageSource(path=Path("/tmp/a.png")),
+        alias="a.png",
+        mime_type="image/png",
+    )
+
+    assert isinstance(att.source, FileImageSource)
+    assert att.source.path == Path("/tmp/a.png")
+
+
+def test_file_source_round_trips_through_json() -> None:
+    att = ImageAttachment(
+        source=FileImageSource(path=Path("/tmp/a.png")),
+        alias="a.png",
+        mime_type="image/png",
+    )
+
+    dumped = att.model_dump(exclude_none=True, mode="json")
+    assert dumped["source"] == {"kind": "file", "path": "/tmp/a.png"}
+    assert ImageAttachment.model_validate(dumped) == att
+
+
+def test_rejects_attachment_without_source() -> None:
+    with pytest.raises(ValidationError):
+        ImageAttachment.model_validate({"alias": "a.png", "mime_type": "image/png"})

tests/core/test_llm_message_merge.py

@@ -4,27 +4,43 @@ from pathlib import Path
 
 import pytest
 
-from vibe.core.types import ImageAttachment, LLMMessage, Role
+from vibe.core.types import (
+    FileImageSource,
+    ImageAttachment,
+    LLMMessage,
+    Role,
+    UserDisplayContentMetadata,
+)
 
 
 @pytest.fixture()
 def image_a(tmp_path: Path) -> ImageAttachment:
     p = tmp_path / "a.png"
     p.write_bytes(b"\x89PNG")
-    return ImageAttachment(path=p, alias="a.png", mime_type="image/png")
+    return ImageAttachment(
+        source=FileImageSource(path=p), alias="a.png", mime_type="image/png"
+    )
 
 
 @pytest.fixture()
 def image_b(tmp_path: Path) -> ImageAttachment:
     p = tmp_path / "b.png"
     p.write_bytes(b"\x89PNG")
-    return ImageAttachment(path=p, alias="b.png", mime_type="image/png")
+    return ImageAttachment(
+        source=FileImageSource(path=p), alias="b.png", mime_type="image/png"
+    )
 
 
 def _msg(content: str, images: list[ImageAttachment] | None = None) -> LLMMessage:
     return LLMMessage(role=Role.assistant, content=content, images=images)
 
 
+def _display_content(host: str) -> UserDisplayContentMetadata:
+    return UserDisplayContentMetadata(
+        version="1.0.0", host=host, content=[{"type": "text", "text": host}]
+    )
+
+
 def test_merge_prefers_self_images_when_present(
     image_a: ImageAttachment, image_b: ImageAttachment
 ) -> None:
@@ -49,3 +65,26 @@ def test_merge_preserves_explicit_empty_self_images_over_other(
 def test_merge_yields_none_when_both_sides_are_none() -> None:
     merged = _msg("hi") + _msg(" there")
     assert merged.images is None
+
+
+def test_merge_prefers_self_user_display_content_when_present() -> None:
+    self_display_content = _display_content("mistral-vscode")
+    other_display_content = _display_content("mistral-jetbrains")
+
+    merged = LLMMessage(
+        role=Role.user, content="hi", user_display_content=self_display_content
+    ) + LLMMessage(
+        role=Role.user, content=" there", user_display_content=other_display_content
+    )
+
+    assert merged.user_display_content == self_display_content
+
+
+def test_merge_falls_back_to_other_user_display_content() -> None:
+    other_display_content = _display_content("mistral-vscode")
+
+    merged = LLMMessage(role=Role.user, content="hi") + LLMMessage(
+        role=Role.user, content=" there", user_display_content=other_display_content
+    )
+
+    assert merged.user_display_content == other_display_content

tests/core/test_mcp_oauth.py

@@ -4,6 +4,8 @@ import asyncio
 from collections.abc import Callable, Iterator
 from contextlib import suppress
 import socket
+from types import TracebackType
+from unittest.mock import patch
 import urllib.parse
 
 import httpx
@@ -11,6 +13,7 @@ import keyring
 from keyring.backend import KeyringBackend
 import keyring.backends.fail
 import keyring.errors
+from mcp.client.auth import OAuthFlowError
 from mcp.shared.auth import OAuthClientInformationFull, OAuthToken
 import pytest
 import respx
@@ -21,6 +24,7 @@ from vibe.core.auth.mcp_oauth import (
     LoopbackCallbackHandler,
     MCPOAuthError,
     MCPOAuthHeadlessError,
+    MCPOAuthLoginFailed,
     MCPOAuthPortInUse,
     build_oauth_provider,
     perform_oauth_login,
@@ -366,6 +370,26 @@ class TestBuildOAuthProvider:
         )
         assert provider.context.client_metadata_url == "https://vibe.example/cm.json"
 
+    @pytest.mark.asyncio
+    async def test_client_id_is_exposed_as_fallback_client_info(
+        self, memory_keyring: _MemoryKeyring
+    ) -> None:
+        srv = _oauth_server(client_id="pre-registered-client")
+
+        async def on_url(_url: str) -> None:
+            return None
+
+        async def cb() -> tuple[str, str | None]:
+            return "code", None
+
+        provider = build_oauth_provider(
+            srv, redirect_handler=on_url, callback_handler=cb
+        )
+        client_info = await provider.context.storage.get_client_info()
+
+        assert client_info is not None
+        assert client_info.client_id == "pre-registered-client"
+
     @pytest.mark.asyncio
     async def test_rejects_static_auth(self, memory_keyring: _MemoryKeyring) -> None:
         from vibe.core.config import MCPStaticAuth
@@ -388,6 +412,39 @@ class TestBuildOAuthProvider:
 
 
 class TestPerformOAuthLogin:
+    @pytest.mark.asyncio
+    async def test_oauth_flow_error_becomes_login_failed(
+        self, memory_keyring: _MemoryKeyring
+    ) -> None:
+        srv = _oauth_server(name="demo")
+
+        class OAuthFlowFailingClient:
+            def __init__(self, *args: object, **kwargs: object) -> None:
+                pass
+
+            async def __aenter__(self) -> OAuthFlowFailingClient:
+                return self
+
+            async def __aexit__(
+                self,
+                exc_type: type[BaseException] | None,
+                exc: BaseException | None,
+                traceback: TracebackType | None,
+            ) -> None:
+                pass
+
+            async def get(self, _url: str) -> None:
+                raise OAuthFlowError("cancelled")
+
+        async def on_url(_url: str) -> None:
+            pass
+
+        with patch(
+            "vibe.core.auth.mcp_oauth.httpx.AsyncClient", new=OAuthFlowFailingClient
+        ):
+            with pytest.raises(MCPOAuthLoginFailed, match="cancelled"):
+                await perform_oauth_login(srv, on_url=on_url)
+
     @pytest.mark.asyncio
     async def test_full_flow_persists_tokens_and_fingerprint(
         self, memory_keyring: _MemoryKeyring

tests/core/test_message_list.py

@@ -0,0 +1,56 @@
+from __future__ import annotations
+
+from vibe.core.types import LLMMessage, MessageList, Role
+
+
+def test_update_system_prompt_replaces_existing_system_slot() -> None:
+    messages = MessageList(
+        initial=[
+            LLMMessage(role=Role.system, content="old"),
+            LLMMessage(role=Role.user, content="hi"),
+        ]
+    )
+
+    messages.update_system_prompt("new")
+
+    assert len(messages) == 2
+    assert messages[0].role == Role.system
+    assert messages[0].content == "new"
+    assert messages[1].content == "hi"
+
+
+def test_update_system_prompt_inserts_without_clobbering_when_no_system() -> None:
+    messages = MessageList(
+        initial=[
+            LLMMessage(role=Role.user, content="Hello"),
+            LLMMessage(role=Role.assistant, content="Hi there!"),
+        ]
+    )
+
+    messages.update_system_prompt("system prompt")
+
+    assert len(messages) == 3
+    assert messages[0].role == Role.system
+    assert messages[1].content == "Hello"
+    assert messages[2].content == "Hi there!"
+
+
+def test_update_system_prompt_inserts_into_empty_list() -> None:
+    messages = MessageList()
+
+    messages.update_system_prompt("system prompt")
+
+    assert len(messages) == 1
+    assert messages[0].role == Role.system
+
+
+def test_update_system_prompt_notifies_only_when_requested() -> None:
+    observed: list[LLMMessage] = []
+    messages = MessageList(observer=observed.append)
+
+    messages.update_system_prompt("silent")
+    assert observed == []
+
+    messages.update_system_prompt("loud", notify=True)
+    assert len(observed) == 1
+    assert observed[0].content == "loud"

tests/core/test_overrides_layer.py

@@ -3,7 +3,6 @@ from __future__ import annotations
 import pytest
 
 from vibe.core.config.layers.overrides import OverridesLayer
-from vibe.core.config.patch import ConfigPatch
 
 
 @pytest.mark.asyncio
@@ -20,13 +19,6 @@ async def test_always_trusted() -> None:
     assert await layer.resolve_trust() is True
 
 
-@pytest.mark.asyncio
-async def test_apply_raises_not_implemented() -> None:
-    layer = OverridesLayer(data={})
-    with pytest.raises(NotImplementedError, match="M2"):
-        await layer.apply(ConfigPatch(fingerprint="fp-1"))
-
-
 @pytest.mark.asyncio
 async def test_default_name() -> None:
     layer = OverridesLayer(data={})

tests/core/test_project_config_layer.py

@@ -6,7 +6,6 @@ import pytest
 
 from vibe.core.config.layer import UntrustedLayerError
 from vibe.core.config.layers.project import ProjectConfigLayer
-from vibe.core.config.patch import ConfigPatch
 from vibe.core.config.types import MISSING_CONFIG_FILE_FINGERPRINT
 from vibe.core.paths._vibe_home import GlobalPath
 from vibe.core.trusted_folders import trusted_folders_manager
@@ -73,13 +72,6 @@ async def test_default_name(tmp_working_directory: Path) -> None:
     assert layer.name == "project-toml"
 
 
-@pytest.mark.asyncio
-async def test_apply_raises_not_implemented(tmp_working_directory: Path) -> None:
-    layer = ProjectConfigLayer(path=tmp_working_directory)
-    with pytest.raises(NotImplementedError, match="M2"):
-        await layer.apply(ConfigPatch(fingerprint="fp-1"))
-
-
 @pytest.mark.asyncio
 async def test_trust_uses_path_parent_for_resolution(
     tmp_working_directory: Path,

tests/core/test_remote_agent_loop.py

@@ -1,937 +0,0 @@
-from __future__ import annotations
-
-from typing import Any
-from unittest.mock import patch
-
-from tests.conftest import build_test_vibe_config
-from vibe.core.nuage.events import (
-    CustomTaskCanceled,
-    CustomTaskCanceledAttributes,
-    CustomTaskCompleted,
-    CustomTaskCompletedAttributes,
-    CustomTaskInProgress,
-    CustomTaskInProgressAttributes,
-    CustomTaskStarted,
-    CustomTaskStartedAttributes,
-    JSONPatchAdd,
-    JSONPatchAppend,
-    JSONPatchPayload,
-    JSONPatchReplace,
-    JSONPayload,
-)
-from vibe.core.nuage.remote_events_source import RemoteEventsSource
-from vibe.core.types import (
-    AssistantEvent,
-    ReasoningEvent,
-    Role,
-    ToolCallEvent,
-    ToolResultEvent,
-    ToolStreamEvent,
-    UserMessageEvent,
-    WaitingForInputEvent,
-)
-
-_EXEC_ID = "session-123"
-
-
-def _make_loop(enabled_tools: list[str] | None = None) -> RemoteEventsSource:
-    config = build_test_vibe_config(enabled_tools=enabled_tools or [])
-    return RemoteEventsSource(session_id=_EXEC_ID, config=config)
-
-
-def _started(
-    task_id: str, task_type: str, payload: dict[str, Any]
-) -> CustomTaskStarted:
-    return CustomTaskStarted(
-        event_id=f"evt-{task_id}-start",
-        workflow_exec_id=_EXEC_ID,
-        attributes=CustomTaskStartedAttributes(
-            custom_task_id=task_id,
-            custom_task_type=task_type,
-            payload=JSONPayload(value=payload),
-        ),
-    )
-
-
-def _completed(
-    task_id: str, task_type: str, payload: dict[str, Any]
-) -> CustomTaskCompleted:
-    return CustomTaskCompleted(
-        event_id=f"evt-{task_id}-done",
-        workflow_exec_id=_EXEC_ID,
-        attributes=CustomTaskCompletedAttributes(
-            custom_task_id=task_id,
-            custom_task_type=task_type,
-            payload=JSONPayload(value=payload),
-        ),
-    )
-
-
-def _in_progress(
-    task_id: str, task_type: str, patches: list[Any]
-) -> CustomTaskInProgress:
-    return CustomTaskInProgress(
-        event_id=f"evt-{task_id}-progress",
-        workflow_exec_id=_EXEC_ID,
-        attributes=CustomTaskInProgressAttributes(
-            custom_task_id=task_id,
-            custom_task_type=task_type,
-            payload=JSONPatchPayload(value=patches),
-        ),
-    )
-
-
-def _canceled(task_id: str, task_type: str, reason: str = "") -> CustomTaskCanceled:
-    return CustomTaskCanceled(
-        event_id=f"evt-{task_id}-cancel",
-        workflow_exec_id=_EXEC_ID,
-        attributes=CustomTaskCanceledAttributes(
-            custom_task_id=task_id, custom_task_type=task_type, reason=reason
-        ),
-    )
-
-
-def test_consume_wait_for_input_event_emits_waiting_event() -> None:
-    loop = _make_loop()
-    event = _started(
-        "wait-task-1",
-        "wait_for_input",
-        {
-            "task_id": "wait-task-1",
-            "input_schema": {"title": "ChatInput"},
-            "label": "What next?",
-        },
-    )
-
-    emitted_events = loop._consume_workflow_event(event)
-
-    assert len(emitted_events) == 2
-    assistant_event = emitted_events[0]
-    waiting_event = emitted_events[1]
-    assert isinstance(assistant_event, AssistantEvent)
-    assert assistant_event.content == "What next?"
-    assert isinstance(waiting_event, WaitingForInputEvent)
-    assert waiting_event.task_id == "wait-task-1"
-    assert waiting_event.label == "What next?"
-    assert waiting_event.predefined_answers is None
-
-
-def test_consume_agent_input_keeps_repeated_text_across_distinct_turns() -> None:
-    loop = _make_loop()
-    first_event = _completed(
-        "input-1", "AgentInputState", {"input": {"message": [{"text": "continue"}]}}
-    )
-    second_event = _completed(
-        "input-2", "AgentInputState", {"input": {"message": [{"text": "continue"}]}}
-    )
-
-    assert loop._consume_workflow_event(first_event) == []
-    assert loop._consume_workflow_event(second_event) == []
-
-    assert [msg.content for msg in loop.messages if msg.role == Role.user] == [
-        "continue",
-        "continue",
-    ]
-
-
-def test_wait_for_input_emits_predefined_answers_and_user_message() -> None:
-    loop = _make_loop()
-    started = _started(
-        "wait-task-1",
-        "wait_for_input",
-        {
-            "input_schema": {
-                "title": "ChatInput",
-                "properties": {
-                    "message": {
-                        "examples": [
-                            [{"type": "text", "text": "Python"}],
-                            [{"type": "text", "text": "JavaScript"}],
-                            [{"type": "text", "text": "Other"}],
-                        ]
-                    }
-                },
-            },
-            "label": "Which language?",
-        },
-    )
-    completed = _completed(
-        "wait-task-1",
-        "wait_for_input",
-        {"input": {"message": [{"type": "text", "text": "Python"}]}},
-    )
-
-    started_events = loop._consume_workflow_event(started)
-    completed_events = loop._consume_workflow_event(completed)
-
-    assistant_event = next(
-        event for event in started_events if isinstance(event, AssistantEvent)
-    )
-    waiting_event = next(
-        event for event in started_events if isinstance(event, WaitingForInputEvent)
-    )
-    assert assistant_event.content == "Which language?"
-    assert waiting_event.predefined_answers == ["Python", "JavaScript"]
-    user_event = next(
-        event for event in completed_events if isinstance(event, UserMessageEvent)
-    )
-    assert user_event.content == "Python"
-
-
-def test_tool_events_update_stats_and_messages() -> None:
-    loop = _make_loop(enabled_tools=["todo"])
-    started = _started(
-        "tool-task-1",
-        "AgentToolCallState",
-        {"name": "todo", "tool_call_id": "call-1", "kwargs": {"action": "read"}},
-    )
-    completed = _completed(
-        "tool-task-1",
-        "AgentToolCallState",
-        {
-            "name": "todo",
-            "tool_call_id": "call-1",
-            "kwargs": {"action": "read"},
-            "output": {"total_count": 0},
-        },
-    )
-
-    started_events = loop._consume_workflow_event(started)
-    completed_events = loop._consume_workflow_event(completed)
-
-    assert any(isinstance(event, ToolCallEvent) for event in started_events)
-    result_event = next(
-        event for event in completed_events if isinstance(event, ToolResultEvent)
-    )
-    assert result_event.error is None
-    assert result_event.cancelled is False
-    assert result_event.tool_call_id == "call-1"
-    assert loop.stats.tool_calls_agreed == 1
-    assert loop.stats.tool_calls_succeeded == 1
-    assert loop.stats.tool_calls_failed == 0
-    tool_messages = [msg for msg in loop.messages if msg.role == Role.tool]
-    assert len(tool_messages) == 1
-    assert tool_messages[0].tool_call_id == "call-1"
-
-
-def test_ask_user_question_tool_emits_assistant_question() -> None:
-    loop = _make_loop(enabled_tools=["ask_user_question"])
-    started = _started(
-        "tool-task-question",
-        "AgentToolCallState",
-        {
-            "name": "ask_user_question",
-            "tool_call_id": "call-question",
-            "kwargs": {
-                "questions": [
-                    {
-                        "question": "Which file type should I create?",
-                        "options": [{"label": "Python"}, {"label": "JavaScript"}],
-                    }
-                ]
-            },
-        },
-    )
-
-    events = loop._consume_workflow_event(started)
-
-    assistant_event = next(
-        event for event in events if isinstance(event, AssistantEvent)
-    )
-    tool_call_event = next(
-        event for event in events if isinstance(event, ToolCallEvent)
-    )
-    assert assistant_event.content == "Which file type should I create?"
-    assert tool_call_event.tool_call_id == "call-question"
-
-
-def test_ask_user_question_invalid_args_are_logged_and_ignored() -> None:
-    loop = _make_loop(enabled_tools=["ask_user_question"])
-    started = _started(
-        "tool-task-question",
-        "AgentToolCallState",
-        {
-            "name": "ask_user_question",
-            "tool_call_id": "call-question",
-            "kwargs": {"questions": [{}]},
-        },
-    )
-
-    with patch(
-        "vibe.core.nuage.remote_workflow_event_translator.logger.warning"
-    ) as mock_warning:
-        events = loop._consume_workflow_event(started)
-
-    assert any(isinstance(event, ToolCallEvent) for event in events)
-    assert not any(isinstance(event, AssistantEvent) for event in events)
-    mock_warning.assert_called_once()
-
-
-def test_ask_user_question_wait_for_input_completion_emits_tool_result() -> None:
-    loop = _make_loop(enabled_tools=["ask_user_question"])
-    ask_started = _started(
-        "tool-task-question",
-        "AgentToolCallState",
-        {
-            "name": "ask_user_question",
-            "tool_call_id": "call-question",
-            "kwargs": {
-                "questions": [{"question": "Which type of file?", "options": []}]
-            },
-        },
-    )
-    wait_started = _started(
-        "wait-task-1",
-        "wait_for_input",
-        {"input_schema": {"title": "ChatInput"}, "label": "Which type of file?"},
-    )
-    wait_completed = _completed(
-        "wait-task-1",
-        "wait_for_input",
-        {
-            "input_schema": {"title": "ChatInput"},
-            "label": "Which type of file?",
-            "input": {"message": [{"type": "text", "text": "Python"}]},
-        },
-    )
-
-    loop._consume_workflow_event(ask_started)
-    loop._consume_workflow_event(wait_started)
-    completed_events = loop._consume_workflow_event(wait_completed)
-
-    tool_result = next(
-        (e for e in completed_events if isinstance(e, ToolResultEvent)), None
-    )
-    assert tool_result is not None
-    assert tool_result.tool_call_id == "call-question"
-    user_message = next(e for e in completed_events if isinstance(e, UserMessageEvent))
-    assert user_message.content == "Python"
-
-
-def test_working_events_without_tool_call_id_render_remote_progress_row() -> None:
-    loop = _make_loop()
-    started = _started(
-        "working-1",
-        "working",
-        {"title": "Creating sandbox", "content": "initializing", "toolUIState": None},
-    )
-    completed = _completed(
-        "working-1",
-        "working",
-        {
-            "title": "Creating sandbox",
-            "content": "sandbox created",
-            "toolUIState": None,
-        },
-    )
-
-    started_events = loop._consume_workflow_event(started)
-    completed_events = loop._consume_workflow_event(completed)
-
-    assert started_events == []
-    assert any(isinstance(event, ToolCallEvent) for event in completed_events)
-    assert any(isinstance(event, ToolStreamEvent) for event in completed_events)
-    result_event = next(
-        event for event in completed_events if isinstance(event, ToolResultEvent)
-    )
-    assert result_event.tool_name == "Creating sandbox"
-    assert result_event.tool_call_id == "working-1"
-
-
-def test_working_events_with_tool_call_id_wait_for_real_tool_call() -> None:
-    loop = _make_loop()
-    working_started = _started(
-        "working-tool-1",
-        "working",
-        {
-            "title": "Executing write_file",
-            "content": "writing file",
-            "toolUIState": {"toolCallId": "call-write"},
-        },
-    )
-    tool_started = _started(
-        "tool-task-1",
-        "AgentToolCallState",
-        {
-            "name": "write_file",
-            "tool_call_id": "call-write",
-            "kwargs": {
-                "path": "hello_world.js",
-                "content": "console.log('Hello, World!');",
-            },
-        },
-    )
-
-    working_events = loop._consume_workflow_event(working_started)
-    tool_events = loop._consume_workflow_event(tool_started)
-
-    assert any(isinstance(event, ToolCallEvent) for event in working_events)
-    assert any(isinstance(event, ToolStreamEvent) for event in working_events)
-    assert not any(isinstance(event, ToolCallEvent) for event in tool_events)
-    assert not any(isinstance(event, ToolResultEvent) for event in tool_events)
-
-
-def test_working_task_promoted_to_real_tool_call_does_not_create_duplicate_row() -> (
-    None
-):
-    loop = _make_loop(enabled_tools=["write_file"])
-    working_started = _started(
-        "working-tool-1",
-        "working",
-        {
-            "title": "Writing file",
-            "content": '# hello.py\n\nprint("Hello, World!")',
-            "toolUIState": None,
-        },
-    )
-    working_promoted = _completed(
-        "working-tool-1",
-        "working",
-        {
-            "title": "Executing write_file",
-            "content": "",
-            "toolUIState": {
-                "type": "file",
-                "toolCallId": "call-write",
-                "operations": [
-                    {
-                        "type": "create",
-                        "uri": "/workspace/hello.py",
-                        "content": 'print("Hello, World!")',
-                    }
-                ],
-            },
-        },
-    )
-    agent_tool_completed = _completed(
-        "tool-task-1",
-        "AgentToolCallState",
-        {
-            "name": "write_file",
-            "tool_call_id": "call-write",
-            "kwargs": {"path": "hello.py", "content": 'print("Hello, World!")'},
-            "output": {
-                "path": "/workspace/hello.py",
-                "bytes_written": 22,
-                "content": 'print("Hello, World!")',
-            },
-        },
-    )
-
-    assert loop._consume_workflow_event(working_started) == []
-
-    promoted_events = loop._consume_workflow_event(working_promoted)
-    assert len([e for e in promoted_events if isinstance(e, ToolCallEvent)]) == 1
-    assert not any(isinstance(e, ToolStreamEvent) for e in promoted_events)
-    assert any(isinstance(e, ToolResultEvent) for e in promoted_events)
-
-    completed_events = loop._consume_workflow_event(agent_tool_completed)
-    assert not any(isinstance(e, ToolCallEvent) for e in completed_events)
-    assert not any(isinstance(e, ToolResultEvent) for e in completed_events)
-
-
-def test_idle_boundary_waits_for_open_tool_results() -> None:
-    loop = _make_loop(enabled_tools=["write_file"])
-    working_started = _started(
-        "working-tool-1",
-        "working",
-        {
-            "title": "Executing write_file",
-            "content": "writing file",
-            "toolUIState": {"toolCallId": "call-write"},
-        },
-    )
-    idle_candidate = _completed("input-task-1", "AgentInputState", {"input": None})
-    tool_completed = _completed(
-        "tool-task-1",
-        "AgentToolCallState",
-        {
-            "name": "write_file",
-            "tool_call_id": "call-write",
-            "kwargs": {
-                "path": "hello_world.js",
-                "content": "console.log('Hello, World!');",
-            },
-            "output": {
-                "path": "/workspace/hello_world.js",
-                "bytes_written": 29,
-                "content": "console.log('Hello, World!');",
-            },
-        },
-    )
-    idle_after_tool = _completed("input-task-2", "AgentInputState", {"input": None})
-
-    working_events = loop._consume_workflow_event(working_started)
-    assert any(isinstance(event, ToolCallEvent) for event in working_events)
-
-    loop._consume_workflow_event(idle_candidate)
-    assert loop._is_idle_boundary(idle_candidate) is False
-
-    tool_events = loop._consume_workflow_event(tool_completed)
-    assert not any(isinstance(event, ToolCallEvent) for event in tool_events)
-    assert any(isinstance(event, ToolResultEvent) for event in tool_events)
-
-    loop._consume_workflow_event(idle_after_tool)
-    assert loop._is_idle_boundary(idle_after_tool) is True
-
-
-def test_send_user_message_tool_is_not_rendered() -> None:
-    loop = _make_loop()
-    started = _started(
-        "tool-task-send-user-message",
-        "AgentToolCallState",
-        {
-            "name": "send_user_message",
-            "tool_call_id": "call-send",
-            "kwargs": {"message": "hello"},
-        },
-    )
-    completed = _completed(
-        "tool-task-send-user-message",
-        "AgentToolCallState",
-        {
-            "name": "send_user_message",
-            "tool_call_id": "call-send",
-            "kwargs": {"message": "hello"},
-            "output": {"success": True, "error": None},
-        },
-    )
-
-    assert loop._consume_workflow_event(started) == []
-    assert loop._consume_workflow_event(completed) == []
-
-
-def test_send_user_message_working_events_are_not_rendered() -> None:
-    loop = _make_loop()
-    started = _started(
-        "working-send-user-message",
-        "working",
-        {
-            "title": "Executing send_user_message",
-            "content": "Hello!",
-            "toolUIState": {"toolCallId": "call-send-working"},
-        },
-    )
-    completed = _completed(
-        "working-send-user-message",
-        "working",
-        {
-            "title": "Executing send_user_message",
-            "content": "Hello!",
-            "toolUIState": {"toolCallId": "call-send-working"},
-        },
-    )
-
-    assert loop._consume_workflow_event(started) == []
-    assert loop._consume_workflow_event(completed) == []
-
-
-def test_remote_bash_uses_known_tool_display_even_when_disabled_locally() -> None:
-    loop = _make_loop(enabled_tools=["write_file"])
-    started = _started(
-        "tool-task-bash",
-        "AgentToolCallState",
-        {
-            "name": "bash",
-            "tool_call_id": "call-bash",
-            "kwargs": {"command": "cat hello.py | wc -c"},
-        },
-    )
-    completed = _completed(
-        "tool-task-bash",
-        "AgentToolCallState",
-        {
-            "name": "bash",
-            "tool_call_id": "call-bash",
-            "kwargs": {"command": "cat hello.py | wc -c"},
-            "output": {
-                "command": "cat hello.py | wc -c",
-                "stdout": "22\n",
-                "stderr": "",
-                "returncode": 0,
-            },
-        },
-    )
-
-    started_events = loop._consume_workflow_event(started)
-    completed_events = loop._consume_workflow_event(completed)
-
-    tool_call_event = next(
-        event for event in started_events if isinstance(event, ToolCallEvent)
-    )
-    result_event = next(
-        event for event in completed_events if isinstance(event, ToolResultEvent)
-    )
-
-    assert tool_call_event.tool_name == "bash"
-    assert tool_call_event.tool_class.get_name() == "bash"
-    assert tool_call_event.args is not None
-    assert tool_call_event.args.command == "cat hello.py | wc -c"  # type: ignore[attr-defined]
-    assert result_event.result is not None
-    assert result_event.result.command == "cat hello.py | wc -c"  # type: ignore[attr-defined]
-    assert result_event.result.stdout == "22\n"  # type: ignore[attr-defined]
-
-
-def test_canceled_tool_marks_cancelled_and_failed_stats() -> None:
-    loop = _make_loop(enabled_tools=["todo"])
-    loop._task_state["tool-task-2"] = {
-        "name": "todo",
-        "tool_call_id": "call-2",
-        "kwargs": {"action": "read"},
-    }
-    canceled = _canceled(
-        "tool-task-2", "AgentToolCallState", reason="user interrupted tool"
-    )
-
-    events = loop._consume_workflow_event(canceled)
-
-    result_event = next(event for event in events if isinstance(event, ToolResultEvent))
-    assert result_event.cancelled is True
-    assert result_event.error == "Canceled: user interrupted tool"
-    assert loop.stats.tool_calls_failed == 1
-    assert loop.stats.tool_calls_succeeded == 0
-
-
-def test_working_thinking_type_emits_assistant_events() -> None:
-    loop = _make_loop()
-    started = _started(
-        "thinking-1",
-        "working",
-        {"type": "thinking", "title": "Thinking", "content": "", "toolUIState": None},
-    )
-    in_progress = _in_progress(
-        "thinking-1", "working", [JSONPatchAppend(path="/content", value="Hello!")]
-    )
-    completed = _completed(
-        "thinking-1",
-        "working",
-        {
-            "type": "thinking",
-            "title": "Thinking",
-            "content": "Hello!",
-            "toolUIState": None,
-        },
-    )
-
-    started_events = loop._consume_workflow_event(started)
-    progress_events = loop._consume_workflow_event(in_progress)
-    completed_events = loop._consume_workflow_event(completed)
-
-    assert started_events == []
-    assert len(progress_events) == 1
-    assert isinstance(progress_events[0], ReasoningEvent)
-    assert progress_events[0].content == "Hello!"
-    assert completed_events == []
-
-
-def test_working_bash_progress_without_tool_call_id_streams_command_output() -> None:
-    loop = _make_loop(enabled_tools=["write_file"])
-    started = _started(
-        "working-bash-1",
-        "working",
-        {"type": "tool", "title": "Planning", "content": "", "toolUIState": None},
-    )
-    in_progress = _in_progress(
-        "working-bash-1",
-        "working",
-        [
-            JSONPatchAdd(
-                path="/toolUIState",
-                value={
-                    "type": "command",
-                    "command": "ls -la /workspace",
-                    "result": {
-                        "status": "success",
-                        "output": "total 4\ndrwxrwxrwx 2 root root 4096 Mar 20 10:18 .\ndrwxr-xr-x 1 root root   80 Mar 20 10:18 ..\n",
-                    },
-                },
-            ),
-            JSONPatchReplace(path="/title", value="Executing bash"),
-            JSONPatchReplace(path="/content", value=""),
-        ],
-    )
-
-    started_events = loop._consume_workflow_event(started)
-    progress_events = loop._consume_workflow_event(in_progress)
-
-    assert started_events == []
-    tool_call_event = next(
-        event for event in progress_events if isinstance(event, ToolCallEvent)
-    )
-    tool_stream_event = next(
-        event for event in progress_events if isinstance(event, ToolStreamEvent)
-    )
-
-    assert tool_call_event.tool_name == "bash"
-    assert tool_call_event.tool_class.get_name() == "bash"
-    assert tool_call_event.tool_call_id == "working-bash-1"
-    assert tool_stream_event.tool_name == "bash"
-    assert tool_stream_event.tool_call_id == "working-bash-1"
-    assert "command: ls -la /workspace" in tool_stream_event.message
-    assert "total 4" in tool_stream_event.message
-    assert "drwxrwxrwx 2 root root 4096" in tool_stream_event.message
-
-
-def test_working_completed_with_tool_call_id_emits_tool_result() -> None:
-    loop = _make_loop(enabled_tools=["write_file"])
-    working_started = _started(
-        "working-tool-1",
-        "working",
-        {
-            "title": "Executing write_file",
-            "content": "",
-            "toolUIState": {"toolCallId": "call-write-solo"},
-        },
-    )
-    working_completed = _completed(
-        "working-tool-1",
-        "working",
-        {
-            "title": "Executing write_file",
-            "content": "",
-            "toolUIState": {
-                "type": "file",
-                "toolCallId": "call-write-solo",
-                "operations": [
-                    {
-                        "type": "create",
-                        "uri": "/workspace/hello.py",
-                        "content": 'print("Hello, World!")',
-                    }
-                ],
-            },
-        },
-    )
-
-    started_events = loop._consume_workflow_event(working_started)
-    assert any(isinstance(e, ToolCallEvent) for e in started_events)
-
-    completed_events = loop._consume_workflow_event(working_completed)
-    result_events = [e for e in completed_events if isinstance(e, ToolResultEvent)]
-    assert len(result_events) == 1
-    assert result_events[0].error is None
-    assert result_events[0].tool_call_id == "call-write-solo"
-
-
-def test_working_completed_with_tool_call_id_emits_error_result() -> None:
-    loop = _make_loop(enabled_tools=["write_file"])
-    working_started = _started(
-        "working-tool-2",
-        "working",
-        {
-            "title": "Executing write_file",
-            "content": "",
-            "toolUIState": {"toolCallId": "call-write-err"},
-        },
-    )
-    working_completed = _completed(
-        "working-tool-2",
-        "working",
-        {
-            "title": "Executing write_file",
-            "content": "Error: Permission denied.",
-            "toolUIState": {
-                "type": "file",
-                "toolCallId": "call-write-err",
-                "operations": [],
-            },
-        },
-    )
-
-    loop._consume_workflow_event(working_started)
-    completed_events = loop._consume_workflow_event(working_completed)
-
-    result_events = [e for e in completed_events if isinstance(e, ToolResultEvent)]
-    assert len(result_events) == 1
-    assert result_events[0].error is not None
-    assert result_events[0].tool_call_id == "call-write-err"
-
-
-def test_json_patch_with_array_index_preserves_list_structure() -> None:
-    loop = _make_loop()
-    started = _started(
-        "msg-1",
-        "assistant_message",
-        {"contentChunks": [{"type": "text", "text": "Hello"}]},
-    )
-    in_progress = _in_progress(
-        "msg-1",
-        "assistant_message",
-        [JSONPatchReplace(path="/contentChunks/0/text", value="Hello world")],
-    )
-
-    loop._consume_workflow_event(started)
-    progress_events = loop._consume_workflow_event(in_progress)
-
-    assert len(progress_events) == 1
-    assert isinstance(progress_events[0], AssistantEvent)
-    assert progress_events[0].content == " world"
-
-
-def test_steer_input_events_are_suppressed() -> None:
-    loop = _make_loop()
-    steer_started = _started(
-        "steer-1",
-        "wait_for_input",
-        {"input_schema": {"title": "ChatInput"}, "label": "Send a message to steer..."},
-    )
-    steer_completed = _completed(
-        "steer-1",
-        "wait_for_input",
-        {
-            "input_schema": {"title": "ChatInput"},
-            "label": "Send a message to steer...",
-            "input": None,
-        },
-    )
-
-    assert loop._consume_workflow_event(steer_started) == []
-    assert loop._translator.pending_input_request is not None
-    assert loop._translator.pending_input_request.task_id == "steer-1"
-    assert loop._consume_workflow_event(steer_completed) == []
-    assert loop._translator.pending_input_request is None
-
-
-def test_steer_input_allows_user_submission() -> None:
-    loop = _make_loop()
-    steer_started = _started(
-        "steer-1",
-        "wait_for_input",
-        {"input_schema": {"title": "ChatInput"}, "label": "Send a message to steer..."},
-    )
-
-    assert loop._consume_workflow_event(steer_started) == []
-    assert loop.is_waiting_for_input
-
-    loop._translator.pending_input_request = None
-
-    steer_completed = _completed(
-        "steer-1",
-        "wait_for_input",
-        {
-            "input_schema": {"title": "ChatInput"},
-            "label": "Send a message to steer...",
-            "input": {"message": [{"type": "text", "text": "do X instead"}]},
-        },
-    )
-    events = loop._consume_workflow_event(steer_completed)
-    assert any(isinstance(e, UserMessageEvent) for e in events)
-    user_event = next(e for e in events if isinstance(e, UserMessageEvent))
-    assert user_event.content == "do X instead"
-
-
-def test_steer_does_not_overwrite_regular_pending_input() -> None:
-    loop = _make_loop()
-    regular_started = _started(
-        "regular-1",
-        "wait_for_input",
-        {"input_schema": {"title": "ChatInput"}, "label": "Enter your message"},
-    )
-    loop._consume_workflow_event(regular_started)
-    assert loop._translator.pending_input_request is not None
-    assert loop._translator.pending_input_request.task_id == "regular-1"
-
-    steer_started = _started(
-        "steer-1",
-        "wait_for_input",
-        {"input_schema": {"title": "ChatInput"}, "label": "Send a message to steer..."},
-    )
-    loop._consume_workflow_event(steer_started)
-    assert loop._translator.pending_input_request.task_id == "regular-1"
-
-
-def test_invalid_steer_start_registers_task_for_terminal_handling() -> None:
-    loop = _make_loop()
-    steer_started = _started(
-        "steer-1", "wait_for_input", {"label": "Send a message to steer..."}
-    )
-
-    with patch("vibe.core.nuage.remote_events_source.logger.warning") as mock_warning:
-        assert loop._consume_workflow_event(steer_started) == []
-
-    mock_warning.assert_called_once()
-    assert "steer-1" in loop._translator._steer_task_ids
-    assert "steer-1" in loop._translator._invalid_steer_task_ids
-    assert loop._translator.pending_input_request is None
-
-
-def test_invalid_steer_completion_does_not_clear_regular_prompt() -> None:
-    loop = _make_loop()
-    regular_started = _started(
-        "regular-1",
-        "wait_for_input",
-        {"input_schema": {"title": "ChatInput"}, "label": "Pick an option"},
-    )
-    loop._consume_workflow_event(regular_started)
-
-    steer_started = _started(
-        "steer-1", "wait_for_input", {"label": "Send a message to steer..."}
-    )
-    assert loop._consume_workflow_event(steer_started) == []
-    assert loop._translator.pending_input_request is not None
-    assert loop._translator.pending_input_request.task_id == "regular-1"
-
-    steer_completed = _completed(
-        "steer-1",
-        "wait_for_input",
-        {"label": "Send a message to steer...", "input": None},
-    )
-    events = loop._consume_workflow_event(steer_completed)
-    assert events == []
-    assert loop._translator.pending_input_request is not None
-    assert loop._translator.pending_input_request.task_id == "regular-1"
-
-
-def test_invalid_steer_cancellation_does_not_clear_regular_prompt() -> None:
-    loop = _make_loop()
-    regular_started = _started(
-        "regular-1",
-        "wait_for_input",
-        {"input_schema": {"title": "ChatInput"}, "label": "Pick an option"},
-    )
-    loop._consume_workflow_event(regular_started)
-
-    steer_started = _started(
-        "steer-1", "wait_for_input", {"label": "Send a message to steer..."}
-    )
-    assert loop._consume_workflow_event(steer_started) == []
-    assert loop._translator.pending_input_request is not None
-    assert loop._translator.pending_input_request.task_id == "regular-1"
-
-    events = loop._consume_workflow_event(_canceled("steer-1", "wait_for_input"))
-    assert events == []
-    assert loop._translator.pending_input_request is not None
-    assert loop._translator.pending_input_request.task_id == "regular-1"
-
-
-def test_steer_completion_is_preserved_while_regular_prompt_pending() -> None:
-    loop = _make_loop()
-    regular_started = _started(
-        "regular-1",
-        "wait_for_input",
-        {"input_schema": {"title": "ChatInput"}, "label": "Pick an option"},
-    )
-    loop._consume_workflow_event(regular_started)
-
-    steer_started = _started(
-        "steer-1",
-        "wait_for_input",
-        {"input_schema": {"title": "ChatInput"}, "label": "Send a message to steer..."},
-    )
-    loop._consume_workflow_event(steer_started)
-
-    steer_completed = _completed(
-        "steer-1",
-        "wait_for_input",
-        {
-            "input_schema": {"title": "ChatInput"},
-            "label": "Send a message to steer...",
-            "input": {"message": [{"type": "text", "text": "user steer msg"}]},
-        },
-    )
-    events = loop._consume_workflow_event(steer_completed)
-
-    user_event = next(e for e in events if isinstance(e, UserMessageEvent))
-    assert user_event.content == "user steer msg"
-    assert loop._translator.pending_input_request is not None
-    assert loop._translator.pending_input_request.task_id == "regular-1"

tests/core/test_resolve_api_key.py

@@ -0,0 +1,138 @@
+from __future__ import annotations
+
+import keyring
+from keyring.errors import KeyringError
+import pytest
+
+from tests.conftest import build_test_vibe_config
+from vibe.core.config import MissingAPIKeyError, ProviderConfig, resolve_api_key
+from vibe.core.llm.backend.mistral import MistralBackend
+from vibe.core.types import Backend
+
+
+def test_resolve_returns_env_value_without_consulting_keyring(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("CUSTOM_API_KEY", "env-key")
+
+    def _fail(service: str, username: str) -> str | None:
+        raise AssertionError("keyring must not be consulted when env is set")
+
+    monkeypatch.setattr(keyring, "get_password", _fail)
+
+    assert resolve_api_key("CUSTOM_API_KEY") == "env-key"
+
+
+def test_resolve_falls_back_to_keyring_when_env_unset(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("CUSTOM_API_KEY", raising=False)
+    monkeypatch.setattr(
+        keyring, "get_password", lambda service, username: "keyring-key"
+    )
+
+    assert resolve_api_key("CUSTOM_API_KEY") == "keyring-key"
+
+
+def test_resolve_returns_none_when_env_unset_and_keyring_empty(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("CUSTOM_API_KEY", raising=False)
+    monkeypatch.setattr(keyring, "get_password", lambda service, username: None)
+
+    assert resolve_api_key("CUSTOM_API_KEY") is None
+
+
+def test_resolve_returns_none_when_keyring_raises(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("CUSTOM_API_KEY", raising=False)
+
+    def _unavailable(service: str, username: str) -> str | None:
+        raise KeyringError("no keyring")
+
+    monkeypatch.setattr(keyring, "get_password", _unavailable)
+
+    assert resolve_api_key("CUSTOM_API_KEY") is None
+
+
+def test_resolve_returns_none_for_empty_env_key(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    def _fail(service: str, username: str) -> str | None:
+        raise AssertionError("keyring must not be consulted for an empty env key")
+
+    monkeypatch.setattr(keyring, "get_password", _fail)
+
+    assert resolve_api_key("") is None
+
+
+def test_check_api_key_accepts_keyring_only_key(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("MISTRAL_API_KEY", raising=False)
+    monkeypatch.setattr(
+        keyring, "get_password", lambda service, username: "keyring-key"
+    )
+
+    # Should not raise MissingAPIKeyError despite the env var being unset.
+    config = build_test_vibe_config()
+
+    assert config.get_active_provider().api_key_env_var == "MISTRAL_API_KEY"
+
+
+def test_check_api_key_raises_when_neither_env_nor_keyring(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("MISTRAL_API_KEY", raising=False)
+    monkeypatch.setattr(keyring, "get_password", lambda service, username: None)
+
+    with pytest.raises(MissingAPIKeyError):
+        build_test_vibe_config()
+
+
+def test_mistral_backend_reads_keyring_only_key(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("MISTRAL_API_KEY", raising=False)
+    monkeypatch.setattr(
+        keyring, "get_password", lambda service, username: "keyring-key"
+    )
+    provider = ProviderConfig(
+        name="mistral",
+        api_base="https://api.mistral.ai/v1",
+        api_key_env_var="MISTRAL_API_KEY",
+        backend=Backend.MISTRAL,
+    )
+
+    backend = MistralBackend(provider)
+
+    assert backend._api_key == "keyring-key"
+
+
+def test_vibe_code_api_key_resolves_from_keyring(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("MISTRAL_API_KEY", raising=False)
+    monkeypatch.setattr(
+        keyring, "get_password", lambda service, username: "keyring-key"
+    )
+
+    config = build_test_vibe_config()
+
+    assert config.vibe_code_api_key == "keyring-key"
+
+
+def test_vibe_code_api_key_empty_when_unresolved(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("MISTRAL_API_KEY", raising=False)
+    monkeypatch.setattr(
+        keyring, "get_password", lambda service, username: "keyring-key"
+    )
+    config = build_test_vibe_config()
+
+    # Nothing resolves the key anymore; the property must return "" (not None).
+    monkeypatch.setattr(keyring, "get_password", lambda service, username: None)
+
+    assert config.vibe_code_api_key == ""

tests/core/test_telemetry_send.py

@@ -510,16 +510,6 @@ class TestTelemetryClient:
             "nb_session_messages": 4,
         }
 
-    def test_send_remote_resume_requested_payload(
-        self, telemetry_events: list[dict[str, Any]]
-    ) -> None:
-        config = build_test_vibe_config(enable_telemetry=True)
-        client = TelemetryClient(config_getter=lambda: config)
-        client.send_remote_resume_requested(session_id="remote-123")
-        assert len(telemetry_events) == 1
-        assert telemetry_events[0]["event_name"] == "vibe.remote_resume_requested"
-        assert telemetry_events[0]["properties"] == {"session_id": "remote-123"}
-
     def test_send_teleport_failed_payload_includes_error_details(
         self, telemetry_events: list[dict[str, Any]]
     ) -> None:

tests/core/test_teleport_nuage.py

@@ -5,6 +5,7 @@ import json
 import httpx
 import pytest
 
+from tests.constants import TELEPORT_SESSIONS_PATH
 from vibe.core.teleport.errors import ServiceTeleportError
 from vibe.core.teleport.nuage import (
     NuageClient,
@@ -52,7 +53,7 @@ async def test_start_posts_nuage_request() -> None:
         response = await nuage.start(_request())
 
     assert seen_request is not None
-    assert str(seen_request.url) == "https://chat.example.com/api/v1/code/sessions"
+    assert str(seen_request.url) == f"https://chat.example.com{TELEPORT_SESSIONS_PATH}"
     assert seen_request.headers["authorization"] == "Bearer api-key"
     assert seen_request.headers["content-type"] == "application/json"
     assert json.loads(seen_request.content) == {

tests/core/test_teleport_service.py

@@ -14,6 +14,7 @@ import pytest
 import zstandard
 
 from tests.conftest import build_test_vibe_config
+from tests.constants import TELEPORT_COMPLETE_URL, TELEPORT_SESSIONS_PATH
 from vibe.core.teleport.errors import (
     ServiceTeleportError,
     ServiceTeleportNotSupportedError,
@@ -59,7 +60,7 @@ def _mock_handler() -> Any:
                 "webSessionId": "web-session-id",
                 "projectId": "project-id",
                 "status": "running",
-                "url": "https://chat.example.com/code/project-id/web-session-id",
+                "url": TELEPORT_COMPLETE_URL,
             },
         )
 
@@ -218,7 +219,7 @@ class TestTeleportServiceExecute:
                     "webSessionId": "web-session-id",
                     "projectId": "project-id",
                     "status": "running",
-                    "url": "https://chat.example.com/code/project-id/web-session-id",
+                    "url": TELEPORT_COMPLETE_URL,
                 },
             )
 
@@ -247,10 +248,8 @@ class TestTeleportServiceExecute:
         assert isinstance(events[0], TeleportCheckingGitEvent)
         assert isinstance(events[1], TeleportStartingWorkflowEvent)
         assert isinstance(events[2], TeleportCompleteEvent)
-        assert (
-            events[2].url == "https://chat.example.com/code/project-id/web-session-id"
-        )
-        assert seen_url == "https://chat.example.com/api/v1/code/sessions"
+        assert events[2].url == TELEPORT_COMPLETE_URL
+        assert seen_url == f"https://chat.example.com{TELEPORT_SESSIONS_PATH}"
         assert seen_body is not None
         assert seen_body["message"] == {
             "role": "user",

tests/core/test_text.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from vibe.core.utils.text import snippet_start_line
+from vibe.core.utils.text import snippet_start_line, snippet_start_lines
 
 
 class TestSnippetStartLine:
@@ -27,3 +27,26 @@ class TestSnippetStartLine:
 
     def test_blank_snippet(self) -> None:
         assert snippet_start_line("hello", "\n") is None
+
+
+class TestSnippetStartLines:
+    def test_single_occurrence(self) -> None:
+        assert snippet_start_lines("a\nb\nc", "b") == [2]
+
+    def test_all_occurrences(self) -> None:
+        assert snippet_start_lines("x\ny\nx\nz\nx", "x") == [1, 3, 5]
+
+    def test_repeated_on_same_line(self) -> None:
+        assert snippet_start_lines("x x x", "x") == [1, 1, 1]
+
+    def test_non_overlapping(self) -> None:
+        assert snippet_start_lines("aaaa", "aa") == [1, 1]
+
+    def test_multiline_snippet_occurrences(self) -> None:
+        assert snippet_start_lines("a\nb\nc\na\nb", "a\nb") == [1, 4]
+
+    def test_not_found(self) -> None:
+        assert snippet_start_lines("hello\nworld", "missing") == []
+
+    def test_blank_snippet(self) -> None:
+        assert snippet_start_lines("hello", "\n") == []

tests/core/test_user_config_layer.py

@@ -1,18 +1,31 @@
 from __future__ import annotations
 
+import os
 from pathlib import Path
+import tomllib
+from uuid import uuid4
 
 import pytest
 
-from vibe.core.config.layer import LayerImplementationError
+from vibe.core.config.fingerprint import create_file_fingerprint
+from vibe.core.config.layer import LayerImplementationError, LayerNotLoadedError
 from vibe.core.config.layers.user import UserConfigLayer
-from vibe.core.config.patch import ConfigPatch
+from vibe.core.config.patch import (
+    AddOperationPatch,
+    ConfigPatch,
+    RemoveOperationPatch,
+    ReplaceOperationPatch,
+)
 from vibe.core.config.types import MISSING_CONFIG_FILE_FINGERPRINT
 
 
+def random_config_file_name() -> str:
+    return f"config-{uuid4().hex}.toml"
+
+
 @pytest.mark.asyncio
 async def test_reads_toml_file(tmp_working_directory: Path) -> None:
-    path = tmp_working_directory / "config.toml"
+    path = tmp_working_directory / random_config_file_name()
     path.write_text('active_model = "mistral-large"\ncount = 42\n')
 
     layer = UserConfigLayer(path=path, name="user-toml")
@@ -25,7 +38,7 @@ async def test_reads_toml_file(tmp_working_directory: Path) -> None:
 
 @pytest.mark.asyncio
 async def test_always_trusted(tmp_working_directory: Path) -> None:
-    path = tmp_working_directory / "config.toml"
+    path = tmp_working_directory / random_config_file_name()
     path.write_text('key = "value"\n')
 
     layer = UserConfigLayer(path=path, name="user-toml")
@@ -37,7 +50,7 @@ async def test_always_trusted(tmp_working_directory: Path) -> None:
 
 @pytest.mark.asyncio
 async def test_missing_file_returns_empty(tmp_working_directory: Path) -> None:
-    path = tmp_working_directory / "nonexistent.toml"
+    path = tmp_working_directory / random_config_file_name()
     layer = UserConfigLayer(path=path, name="user-toml")
     data = await layer.load()
     assert data.model_extra == {}
@@ -45,16 +58,241 @@ async def test_missing_file_returns_empty(tmp_working_directory: Path) -> None:
 
 
 @pytest.mark.asyncio
-async def test_apply_raises_not_implemented(tmp_working_directory: Path) -> None:
-    path = tmp_working_directory / "config.toml"
+async def test_apply_raises_when_file_does_not_exist(
+    tmp_working_directory: Path,
+) -> None:
+    path = tmp_working_directory / random_config_file_name()
     layer = UserConfigLayer(path=path, name="user-toml")
-    with pytest.raises(NotImplementedError, match="M2"):
-        await layer.apply(ConfigPatch(fingerprint="fp-1"))
+
+    await layer.load()
+    assert layer.fingerprint == MISSING_CONFIG_FILE_FINGERPRINT
+
+    with pytest.raises(LayerNotLoadedError, match="loaded before applying patches"):
+        await layer.apply(
+            ConfigPatch(
+                AddOperationPatch(path="/active_model", value="mistral-large"),
+                fingerprint=MISSING_CONFIG_FILE_FINGERPRINT,
+            )
+        )
+
+    assert not path.exists()
+
+
+@pytest.mark.asyncio
+async def test_apply_sets_field_and_refreshes_cache(
+    tmp_working_directory: Path,
+) -> None:
+    path = tmp_working_directory / random_config_file_name()
+    path.write_text("""\
+active_model = "old"
+
+[tools]
+disabled_tools = ["bash", "python"]
+deprecated_setting = true
+""")
+    layer = UserConfigLayer(path=path, name="user-toml")
+
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+
+    await layer.apply(
+        ConfigPatch(
+            ReplaceOperationPatch(path="/active_model", value="new"),
+            AddOperationPatch(path="/tools/enabled_tools", value=["read"]),
+            AddOperationPatch(path="/tools/disabled_tools/-", value="node"),
+            RemoveOperationPatch(path="/tools/disabled_tools/0"),
+            RemoveOperationPatch(path="/tools/deprecated_setting"),
+            fingerprint=fingerprint,
+        )
+    )
+
+    expected_data = {
+        "active_model": "new",
+        "tools": {"disabled_tools": ["python", "node"], "enabled_tools": ["read"]},
+    }
+    with path.open("rb") as file:
+        assert tomllib.load(file) == expected_data
+
+    cached_data = layer._state.data
+    assert cached_data is not None
+    assert cached_data.model_extra == expected_data
+    assert layer.fingerprint != fingerprint
+
+
+@pytest.mark.asyncio
+async def test_apply_cache_fingerprint_matches_written_file(
+    tmp_working_directory: Path,
+) -> None:
+    path = tmp_working_directory / random_config_file_name()
+    path.write_text("")
+    layer = UserConfigLayer(path=path, name="user-toml")
+
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+
+    await layer.apply(
+        ConfigPatch(
+            AddOperationPatch(path="/active_model", value="mistral-large"),
+            fingerprint=fingerprint,
+        )
+    )
+
+    with path.open("rb") as file:
+        assert layer.fingerprint == create_file_fingerprint(file)
+
+
+@pytest.mark.asyncio
+async def test_apply_uses_unique_temp_file(tmp_working_directory: Path) -> None:
+    path = tmp_working_directory / random_config_file_name()
+    fixed_tmp_path = tmp_working_directory / f".{path.name}.tmp"
+    path.write_text("")
+    fixed_tmp_path.write_text("stale")
+    layer = UserConfigLayer(path=path, name="user-toml")
+
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+
+    await layer.apply(
+        ConfigPatch(
+            AddOperationPatch(path="/active_model", value="mistral-large"),
+            fingerprint=fingerprint,
+        )
+    )
+
+    assert fixed_tmp_path.read_text() == "stale"
+    assert list(tmp_working_directory.glob(f".{path.name}.*.tmp")) == []
+    with path.open("rb") as file:
+        assert tomllib.load(file) == {"active_model": "mistral-large"}
+
+
+def test_atomic_replace_preserves_replacement_fingerprint(
+    tmp_working_directory: Path,
+) -> None:
+    path = tmp_working_directory / random_config_file_name()
+    replacement = tmp_working_directory / f".{path.name}.tmp"
+    path.write_text("key = 1")
+    replacement.write_text("key = 2")
+
+    with replacement.open("rb") as file:
+        replacement_fingerprint = create_file_fingerprint(file)
+
+    os.replace(replacement, path)
+
+    with path.open("rb") as file:
+        assert create_file_fingerprint(file) == replacement_fingerprint
+
+
+@pytest.mark.asyncio
+async def test_apply_raises_when_layer_is_not_loaded(
+    tmp_working_directory: Path,
+) -> None:
+    path = tmp_working_directory / random_config_file_name()
+    layer = UserConfigLayer(path=path, name="user-toml")
+
+    with pytest.raises(LayerNotLoadedError, match="loaded before applying patches"):
+        await layer.apply(
+            ConfigPatch(
+                AddOperationPatch(path="/active_model", value="mistral-large"),
+                fingerprint=MISSING_CONFIG_FILE_FINGERPRINT,
+            )
+        )
+
+
+@pytest.mark.asyncio
+async def test_apply_raises_when_cache_is_invalidated(
+    tmp_working_directory: Path,
+) -> None:
+    path = tmp_working_directory / random_config_file_name()
+    path.write_text('active_model = "old"\n')
+    layer = UserConfigLayer(path=path, name="user-toml")
+
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+    await layer.invalidate_cache()
+
+    with pytest.raises(LayerNotLoadedError, match="loaded before applying patches"):
+        await layer.apply(
+            ConfigPatch(
+                ReplaceOperationPatch(path="/active_model", value="new"),
+                fingerprint=fingerprint,
+            )
+        )
+
+
+@pytest.mark.asyncio
+async def test_apply_raises_when_parent_directory_does_not_exist(
+    tmp_working_directory: Path,
+) -> None:
+    path = tmp_working_directory / "nested" / random_config_file_name()
+    layer = UserConfigLayer(path=path, name="user-toml")
+
+    await layer.load()
+
+    with pytest.raises(LayerNotLoadedError, match="loaded before applying patches"):
+        await layer.apply(
+            ConfigPatch(
+                AddOperationPatch(path="/active_model", value="mistral-large"),
+                fingerprint=MISSING_CONFIG_FILE_FINGERPRINT,
+            )
+        )
+
+    assert not path.exists()
+
+
+@pytest.mark.asyncio
+async def test_commit_sets_missing_nested_field(tmp_working_directory: Path) -> None:
+    path = tmp_working_directory / random_config_file_name()
+    path.write_text("[models]\n")
+    layer = UserConfigLayer(path=path, name="user-toml")
+
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+
+    await layer.apply(
+        ConfigPatch(
+            AddOperationPatch(path="/models/active_model", value="mistral-large"),
+            fingerprint=fingerprint,
+        )
+    )
+
+    with path.open("rb") as file:
+        assert tomllib.load(file) == {"models": {"active_model": "mistral-large"}}
+
+
+@pytest.mark.asyncio
+async def test_apply_overwrites_external_file_changes(
+    tmp_working_directory: Path,
+) -> None:
+    path = tmp_working_directory / random_config_file_name()
+    path.write_text('active_model = "old"\n')
+    layer = UserConfigLayer(path=path, name="user-toml")
+
+    await layer.load()
+    fingerprint = layer.fingerprint
+    assert isinstance(fingerprint, str)
+    path.write_text('active_model = "external"\n')
+
+    await layer.apply(
+        ConfigPatch(
+            ReplaceOperationPatch(path="/active_model", value="new"),
+            fingerprint=fingerprint,
+        )
+    )
+
+    with path.open("rb") as file:
+        assert tomllib.load(file) == {"active_model": "new"}
+    data = await layer.load()
+    assert data.model_extra == {"active_model": "new"}
 
 
 @pytest.mark.asyncio
 async def test_nested_toml_structure(tmp_working_directory: Path) -> None:
-    path = tmp_working_directory / "config.toml"
+    path = tmp_working_directory / random_config_file_name()
     path.write_text("""\
 [models]
 active_model = "test"
@@ -72,7 +310,7 @@ provider = "p"
 
 @pytest.mark.asyncio
 async def test_invalid_toml_raises(tmp_working_directory: Path) -> None:
-    path = tmp_working_directory / "bad.toml"
+    path = tmp_working_directory / random_config_file_name()
     path.write_text("this is not valid = = = toml [[[")
     layer = UserConfigLayer(path=path, name="user-toml")
     with pytest.raises(LayerImplementationError, match="_build_config_snapshot"):
@@ -81,7 +319,7 @@ async def test_invalid_toml_raises(tmp_working_directory: Path) -> None:
 
 @pytest.mark.asyncio
 async def test_force_reload_reads_fresh_data(tmp_working_directory: Path) -> None:
-    path = tmp_working_directory / "config.toml"
+    path = tmp_working_directory / random_config_file_name()
     path.write_text('value = "first"\n')
     layer = UserConfigLayer(path=path, name="user-toml")
 
@@ -107,7 +345,7 @@ async def test_force_reload_reads_fresh_data(tmp_working_directory: Path) -> Non
 
 @pytest.mark.asyncio
 async def test_empty_toml_file(tmp_working_directory: Path) -> None:
-    path = tmp_working_directory / "empty.toml"
+    path = tmp_working_directory / random_config_file_name()
     path.write_text("")
     layer = UserConfigLayer(path=path, name="user-toml")
     data = await layer.load()

tests/core/test_user_display_content.py

@@ -0,0 +1,131 @@
+from __future__ import annotations
+
+import math
+from types import SimpleNamespace
+
+from pydantic import ValidationError
+import pytest
+
+from vibe.core.types import LLMMessage, Role, UserDisplayContentMetadata
+
+
+def _metadata() -> UserDisplayContentMetadata:
+    return UserDisplayContentMetadata(
+        version="1.0.0",
+        host="mistral-vscode",
+        content=[
+            {"type": "text", "text": "Look at "},
+            {
+                "type": "workspace_mention",
+                "kind": "file",
+                "uri": "file:///repo/src/app.ts",
+                "name": "app.ts",
+            },
+        ],
+    )
+
+
+def test_accepts_valid_metadata_and_preserves_host_owned_content() -> None:
+    metadata = UserDisplayContentMetadata.model_validate({
+        "version": "1.0.0",
+        "host": "mistral-vscode",
+        "content": [
+            {"type": "text", "text": "Look at "},
+            {
+                "type": "workspace_mention",
+                "kind": "file",
+                "uri": "file:///repo/src/app.ts",
+                "name": "app.ts",
+                "automatic": False,
+                "nested": {"line": 12, "tags": ["source", None]},
+            },
+        ],
+    })
+
+    assert metadata.version == "1.0.0"
+    assert metadata.host == "mistral-vscode"
+    assert metadata.content == [
+        {"type": "text", "text": "Look at "},
+        {
+            "type": "workspace_mention",
+            "kind": "file",
+            "uri": "file:///repo/src/app.ts",
+            "name": "app.ts",
+            "automatic": False,
+            "nested": {"line": 12, "tags": ["source", None]},
+        },
+    ]
+
+
+def test_strips_metadata_strings_without_stripping_host_owned_content() -> None:
+    metadata = UserDisplayContentMetadata.model_validate({
+        "version": " 1.0.0 ",
+        "host": " mistral-vscode ",
+        "content": [{"type": "text", "text": " keep spaces "}],
+    })
+
+    assert metadata.version == "1.0.0"
+    assert metadata.host == "mistral-vscode"
+    assert metadata.content == [{"type": "text", "text": " keep spaces "}]
+
+
+@pytest.mark.parametrize(
+    "payload",
+    [
+        {"version": "   ", "host": "mistral-vscode", "content": []},
+        {"version": 2, "host": "mistral-vscode", "content": []},
+        {"version": "1.0.0", "host": "   ", "content": []},
+        {"version": "1.0.0", "host": "mistral-vscode", "content": ["plain text"]},
+        {"version": "1.0.0", "host": "mistral-vscode", "content": {"type": "text"}},
+        {
+            "version": "1.0.0",
+            "host": "mistral-vscode",
+            "content": [{"type": "text"}],
+            "unexpected": True,
+        },
+        {
+            "version": "1.0.0",
+            "host": "mistral-vscode",
+            "content": [{"type": "text", "value": object()}],
+        },
+        {
+            "version": "1.0.0",
+            "host": "mistral-vscode",
+            "content": [{"type": "text", "value": math.nan}],
+        },
+    ],
+)
+def test_rejects_invalid_metadata(payload: object) -> None:
+    with pytest.raises(ValidationError):
+        UserDisplayContentMetadata.model_validate(payload)
+
+
+def test_llm_message_round_trips_user_display_content() -> None:
+    metadata = _metadata()
+    message = LLMMessage(
+        role=Role.user, content="Look at app.ts", user_display_content=metadata
+    )
+
+    dumped = message.model_dump(exclude_none=True, mode="json")
+    loaded = LLMMessage.model_validate(dumped)
+
+    assert dumped["user_display_content"] == metadata.model_dump(mode="json")
+    assert loaded.user_display_content == metadata
+
+
+def test_llm_message_keeps_old_sessions_without_user_display_content_valid() -> None:
+    message = LLMMessage.model_validate({"role": "user", "content": "hello"})
+
+    assert message.user_display_content is None
+
+
+def test_llm_message_object_adapter_preserves_user_display_content() -> None:
+    metadata = _metadata()
+
+    message = LLMMessage.model_validate(
+        SimpleNamespace(
+            role=Role.user, content="Look at app.ts", user_display_content=metadata
+        )
+    )
+
+    assert message.user_display_content == metadata

tests/core/test_vibe_config_schema.py

@@ -27,6 +27,7 @@ async def test_full_toml_to_vibe_config_schema(tmp_path: Path) -> None:
         """\
 vim_keybindings = true
 api_timeout = 300.0
+api_retry_max_elapsed_time = 120.0
 active_model = "codestral"
 disabled_tools = ["bash"]
 default_agent = "plan"
@@ -54,6 +55,7 @@ provider = "mistral"
 
     assert config.vim_keybindings is True
     assert config.api_timeout == 300.0
+    assert config.api_retry_max_elapsed_time == 120.0
     assert config.active_model == "codestral"
     assert config.models[0].alias == "codestral"
     assert "bash" in config.disabled_tools

tests/core/tools/builtins/test_edit.py

@@ -230,21 +230,21 @@ def test_get_result_display() -> None:
     assert "foo.py" in display.message
 
 
-def test_ui_start_line_not_part_of_model_contract() -> None:
+def test_ui_start_lines_not_part_of_model_contract() -> None:
     result = EditResult(file="/x", message="m", old_string="a", new_string="b")
-    result._ui_start_line = 42
+    result._ui_start_lines = [42]
 
-    assert result.ui_start_line == 42
-    assert "ui_start_line" not in result.model_dump()
-    assert "_ui_start_line" not in result.model_dump()
-    assert "ui_start_line" not in result.model_dump_json()
-    assert "ui_start_line" not in EditResult.model_fields
-    assert "ui_start_line" not in EditResult.model_json_schema().get("properties", {})
-    assert "ui_start_line" not in dict(result)
+    assert result.ui_start_lines == [42]
+    assert "ui_start_lines" not in result.model_dump()
+    assert "_ui_start_lines" not in result.model_dump()
+    assert "ui_start_lines" not in result.model_dump_json()
+    assert "ui_start_lines" not in EditResult.model_fields
+    assert "ui_start_lines" not in EditResult.model_json_schema().get("properties", {})
+    assert "ui_start_lines" not in dict(result)
 
 
 @pytest.mark.asyncio
-async def test_ui_start_line_computed_at_edit_site(
+async def test_ui_start_lines_computed_at_edit_site(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ) -> None:
     monkeypatch.chdir(tmp_path)
@@ -255,11 +255,11 @@ async def test_ui_start_line_computed_at_edit_site(
         edit.run(EditArgs(file_path="f.txt", old_string="gamma", new_string="GAMMA"))
     )
 
-    assert result.ui_start_line == 3
+    assert result.ui_start_lines == [3]
 
 
 @pytest.mark.asyncio
-async def test_ui_start_line_set_for_pure_deletion(
+async def test_ui_start_lines_set_for_pure_deletion(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ) -> None:
     monkeypatch.chdir(tmp_path)
@@ -270,4 +270,38 @@ async def test_ui_start_line_set_for_pure_deletion(
         edit.run(EditArgs(file_path="f.txt", old_string="remove\n", new_string=""))
     )
 
-    assert result.ui_start_line == 3
+    assert result.ui_start_lines == [3]
+
+
+@pytest.mark.asyncio
+async def test_ui_start_lines_lists_all_occurrences_for_replace_all(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.chdir(tmp_path)
+    _write(tmp_path, "f.txt", "x\ntgt\ny\ntgt\nz\ntgt\n")
+    edit = _make_edit()
+
+    result = await collect_result(
+        edit.run(
+            EditArgs(
+                file_path="f.txt", old_string="tgt", new_string="TGT", replace_all=True
+            )
+        )
+    )
+
+    assert result.ui_start_lines == [2, 4, 6]
+
+
+@pytest.mark.asyncio
+async def test_ui_start_lines_single_entry_without_replace_all(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.chdir(tmp_path)
+    _write(tmp_path, "f.txt", "a\nuniq\nb\n")
+    edit = _make_edit()
+
+    result = await collect_result(
+        edit.run(EditArgs(file_path="f.txt", old_string="uniq", new_string="UNIQ"))
+    )
+
+    assert result.ui_start_lines == [2]

tests/e2e/agent_loop_characterization/support.py

@@ -0,0 +1,178 @@
+from __future__ import annotations
+
+from collections.abc import Callable
+import io
+import json
+from pathlib import Path
+import time
+import tomllib
+from typing import Any
+
+import pexpect
+import tomli_w
+
+from tests.e2e.common import strip_ansi, wait_for_rendered_text
+from tests.e2e.mock_server import ChatCompletionsRequestPayload, StreamingMockServer
+
+APPROVAL_INPUT_GRACE_PERIOD_S = 0.65
+
+
+def assistant_text_chunks(text: str, *, created: int = 100) -> list[dict[str, object]]:
+    return [
+        StreamingMockServer.build_chunk(
+            created=created,
+            delta={"role": "assistant", "content": text},
+            finish_reason=None,
+        ),
+        StreamingMockServer.build_chunk(
+            created=created + 1,
+            delta={},
+            finish_reason="stop",
+            usage={"prompt_tokens": 3, "completion_tokens": 4},
+        ),
+    ]
+
+
+def single_tool_call_chunks(
+    *, call_id: str, tool_name: str, arguments: dict[str, Any], created: int = 10
+) -> list[dict[str, object]]:
+    return [
+        StreamingMockServer.build_chunk(
+            created=created,
+            delta=StreamingMockServer.build_tool_call_delta(
+                call_id=call_id,
+                tool_name=tool_name,
+                arguments=json.dumps(arguments, separators=(",", ":")),
+            ),
+            finish_reason=None,
+        ),
+        StreamingMockServer.build_chunk(
+            created=created + 1,
+            delta={},
+            finish_reason="tool_calls",
+            usage={"prompt_tokens": 3, "completion_tokens": 4},
+        ),
+    ]
+
+
+def multi_tool_call_chunks(
+    tool_calls: list[tuple[str, str, dict[str, Any]]], *, created: int = 10
+) -> list[dict[str, object]]:
+    return [
+        StreamingMockServer.build_chunk(
+            created=created,
+            delta={
+                "role": "assistant",
+                "tool_calls": [
+                    {
+                        "index": index,
+                        "id": call_id,
+                        "type": "function",
+                        "function": {
+                            "name": tool_name,
+                            "arguments": json.dumps(arguments, separators=(",", ":")),
+                        },
+                    }
+                    for index, (call_id, tool_name, arguments) in enumerate(tool_calls)
+                ],
+            },
+            finish_reason=None,
+        ),
+        StreamingMockServer.build_chunk(
+            created=created + 1,
+            delta={},
+            finish_reason="tool_calls",
+            usage={"prompt_tokens": 3, "completion_tokens": 4},
+        ),
+    ]
+
+
+def _messages(payload: ChatCompletionsRequestPayload) -> list[dict[str, Any]]:
+    raw_messages = payload.get("messages")
+    assert raw_messages is not None
+    return [dict(message) for message in raw_messages]
+
+
+def assert_tool_result_contains(
+    payload: ChatCompletionsRequestPayload, *, call_id: str, expected: str
+) -> None:
+    matching_messages = [
+        message
+        for message in _messages(payload)
+        if message.get("role") == "tool" and message.get("tool_call_id") == call_id
+    ]
+    assert len(matching_messages) == 1
+    content = matching_messages[0].get("content")
+    assert isinstance(content, str)
+    assert expected in content, content
+
+
+def assert_assistant_tool_call_present(
+    payload: ChatCompletionsRequestPayload, *, call_id: str, tool_name: str
+) -> None:
+    for message in _messages(payload):
+        if message.get("role") != "assistant":
+            continue
+        tool_calls = message.get("tool_calls")
+        if not isinstance(tool_calls, list):
+            continue
+        for tool_call in tool_calls:
+            if not isinstance(tool_call, dict):
+                continue
+            function = tool_call.get("function")
+            if not isinstance(function, dict):
+                continue
+            if tool_call.get("id") == call_id and function.get("name") == tool_name:
+                return
+    raise AssertionError(f"Tool call {call_id!r} for {tool_name!r} was not present.")
+
+
+def assert_message_content_present(
+    payload: ChatCompletionsRequestPayload, *, role: str, expected: str
+) -> None:
+    assert any(
+        message.get("role") == role and expected in str(message.get("content", ""))
+        for message in _messages(payload)
+    )
+
+
+def wait_for_request_count_while_draining_child_output(
+    child: pexpect.spawn,
+    captured: io.StringIO,
+    request_count_getter: Callable[[], int],
+    *,
+    expected_count: int,
+    timeout: float,
+) -> None:
+    start = time.monotonic()
+    while time.monotonic() - start < timeout:
+        if request_count_getter() >= expected_count:
+            return
+        try:
+            child.expect(r"\S", timeout=0.05)
+        except pexpect.TIMEOUT:
+            pass
+    rendered_tail = strip_ansi(captured.getvalue())[-1200:]
+    raise AssertionError(
+        f"Timed out waiting for {expected_count} backend request(s).\n\n"
+        f"Rendered tail:\n{rendered_tail}"
+    )
+
+
+def answer_approval(
+    child: pexpect.spawn, captured: io.StringIO, *, tool_name: str, key: str
+) -> None:
+    wait_for_rendered_text(
+        child, captured, needle=f"Permission for the {tool_name} tool", timeout=10
+    )
+    wait_for_rendered_text(child, captured, needle="Deny", timeout=10)
+    time.sleep(APPROVAL_INPUT_GRACE_PERIOD_S)
+    child.send(key)
+    child.send("\r")
+
+
+def set_tool_denylist(vibe_home: Path, tool_name: str, patterns: list[str]) -> None:
+    config_path = vibe_home / "config.toml"
+    config = tomllib.loads(config_path.read_text(encoding="utf-8"))
+    config.setdefault("tools", {}).setdefault(tool_name, {})["denylist"] = patterns
+    config_path.write_bytes(tomli_w.dumps(config).encode())