agentHost: stop a failed git probe from stalling Agent Merge

[Benjamin Christopher Simmonds (benibenj)]

Fixes a class of bug where a session with Agent Merge enabled sits idle forever without ever binding to its pull request, so review comments and CI failures never reach it.

Found while investigating a real session (copilotcli:/7535d422-…, PR #331760). Agent Merge was enabled and running — it had injected its autonomous configuration and the host was keeping the session resident for it:

[AgentService] Skipping idle eviction for a session held by Agent Merge: copilotcli:/7535d422-…

…but every evaluation died on the first check, exactly backstopInterval apart:

10:57:06 [AgentMergeController] Evaluation started: session=copilotcli:/7535d422-…
10:57:06 [AgentMergeController] Waiting for a current branch: session=copilotcli:/7535d422-…
11:07:06 [AgentMergeController] Evaluation started: session=copilotcli:/7535d422-…
11:07:06 [AgentMergeController] Waiting for a current branch: session=copilotcli:/7535d422-…

Its agentMerge.controller state had no target, confirming it never got far enough to bind the PR — so it never subscribed to PR events, and a review comment and a failing check both went unnoticed.

Root cause

The session's persisted git state had lost its branch, while the worktree itself was perfectly healthy:

// session_metadata['agentHost.git']
{ "baseBranchName": "main" }   // no branchName, no hasGitHubRemote, no uncommittedChanges

_computeSessionGitState runs its probes independently and fills fields best-effort. parseGitStatusV2 always yields uncommittedChanges: 0 for any non-empty output, so its absence means git status itself returned undefined. The repository root is cached, so the function still returned an object — just a branch-less one. _setSessionGitState replaces persisted git state wholesale, so that object overwrote the good branch.

The failure left no trace: _runGit only logged when stderr was non-empty, and a timeout kill produces none.

Why it never recovered

Three separate paths that should have repaired it did not:

1. AgentMergeController._evaluate calls attachSessionGitHubPullRequest (which begins with refreshSessionGitState), but only after the if (!branchName) return bail — so its own repair path was unreachable.
2. The lazy refresh on subscribe only fired when git state was entirely undefined; a partial state counted as "already computed" and masked it.
3. The file-monitor refresh needs watch interest plus an actual file change. A session held resident by Agent Merge alone — no client watching, no edits landing — gets neither.

Changes

• agentHostGitService — return undefined when the status probe fails, so callers keep the state they already had instead of persisting a misleading one.
• agentMergeController — refresh git state before giving up on the branch. This also recovers sessions already stuck in this state, which matters because the bad state is persisted.
• agentService — treat a branch-less state as missing in the subscribe-time refresh.
• agentHostGitService — log git failures that produce no stderr (warn for timeouts/signals, trace for the --quiet probes' expected non-zero exits), so this can't fail invisibly again.

Validation

agentMergeController.test.ts gains a regression test that reproduces the stall. Reverting just the controller change makes it fail the same way the real session did:

1) AgentMergeController
     recovers a session whose persisted git state lost its branch:
   Error: Timeout of 5000ms exceeded.

The git-probe change is covered by a new integration test, and was verified end to end against the compiled service: with the fix the degraded probe returns undefined and logs Not reporting session git state because git status failed; without it, it returns {} — the same shape as the corruption observed on disk.

Full agentHost unit suite: 5840 passing. The 11 failures are pre-existing and unrelated (sessionPermissions, agentHostDebugLogs, agentHostRequestService, copilotAgent, localAgentHostMetadata).

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Prevents Agent Merge from stalling indefinitely when persisted session git state loses its branch (e.g., after a failed git status probe), by avoiding persistence of misleading partial git state and adding a repair path that refreshes git state when a branch is missing.

Changes:

• Add a recovery path in AgentMergeController that refreshes session git state when no branch is available.
• Make AgentHostGitService return undefined (instead of partial state) when git status fails, and improve logging for silent failures/timeouts.
• Add/extend tests to cover the “probe fails → no git state persisted” behavior and the Agent Merge recovery behavior.

Show a summary per file

File	Description
src/vs/platform/agentHost/test/node/agentMergeController.test.ts	Adds a regression test ensuring Agent Merge recovers when persisted git state is missing a branch.
src/vs/platform/agentHost/test/node/agentHostGitService.integrationTest.ts	Adds an integration test verifying that a failed status probe yields no git state (prevents persisting partial state).
src/vs/platform/agentHost/node/agentService.ts	Treats “git state with no branchName” as missing to trigger lazy git-state repair for watchers.
src/vs/platform/agentHost/node/agentMergeController.ts	Adds branch resolution + repair (_resolveCurrentBranch) to unblock Agent Merge when branch is missing.
src/vs/platform/agentHost/node/agentHostGitService.ts	Avoids returning/persisting partial git state when git status fails; improves failure logging in _runGit.

Review details

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

• Files reviewed: 5/5 changed files
• Comments generated: 2
• Review effort level: Lite