Fix cursor-cli E2E flakiness and deferred condensation bug

[alishakawaguchi]

Summary

Two targeted fixes discovered during E2E triage:

• Cursor CLI E2E: isolate config dir and fix wait pattern — Parallel cursor-cli E2E tests raced on the shared ~/.config/cursor/cli-config.json file, causing ENOENT errors. Each session now gets an isolated XDG_CONFIG_HOME with a pre-seeded cli-config.json. Also switches WaitFor from the / commands prompt pattern (always visible in the status bar, even during "Thinking") to the Add a follow-up text that only appears after the agent finishes.

• Fix deferred condensation when transcript is empty at commit time — When Cursor commits mid-turn before flushing its transcript, condensation silently skipped creating the checkpoint. The checkpoint ID was not recorded, so finalizeAllTurnCheckpoints at stop time had nothing to retry. Now: (1) condenseAttempted flag ensures the checkpoint ID is always appended to TurnCheckpointIDs even on failure, and (2) finalizeAllTurnCheckpoints catches ErrCheckpointNotFound and falls back to CondenseSession with the now-available transcript.

Test plan

• New tests: TestPostCommit_ActiveSession_RecordsTurnCheckpointIDs_OnCondensationFailure, TestFinalizeAllTurnCheckpoints_DeferredCondensation, TestFinalizeAllTurnCheckpoints_DeferredCondensation_TranscriptStillEmpty
• mise run test:ci passes
• Cursor CLI E2E tests pass without ENOENT races

🤖 Generated with Claude Code

[cursor]

PR Summary

Medium Risk
Touches git hook/session checkpointing logic and introduces a new retry path on turn finalization, which could affect when/if checkpoints are created; changes are localized and covered by new tests.

Overview
Improves Cursor CLI E2E stability by running each tmux session with its own XDG_CONFIG_HOME (pre-seeded cli-config.json) to avoid parallel config races, and by waiting for a post-completion UI marker (Add a follow-up) instead of the always-visible prompt pattern.

Fixes a deferred-condensation edge case in manual commit hooks: PostCommit now records TurnCheckpointIDs when condensation is attempted (even if it fails), and finalizeAllTurnCheckpoints now retries missing checkpoints by calling CondenseSession on ErrCheckpointNotFound.

Adds focused tests covering checkpoint ID recording on condensation failure and deferred checkpoint creation (including the “still empty transcript” no-crash path).

^{Written by Cursor Bugbot for commit 6bbb43e. Configure here.}

[Copilot]

Pull request overview

Adds tooling and documentation to improve debugging/triage of flaky E2E runs (especially Cursor), including a helper script for downloading CI artifacts and new .claude skill/plugin docs for a triage→fix workflow.

Changes:

• Add scripts/download-e2e-artifacts.sh to fetch and normalize GitHub Actions E2E artifacts locally.
• Update Cursor E2E agent bootstrap to pre-create the Cursor config directory to avoid runtime ENOENT failures.
• Add/update E2E triage/debug/implement skill + plugin documentation and refresh E2E README guidance.

Reviewed changes

Copilot reviewed 15 out of 16 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
scripts/download-e2e-artifacts.sh	New helper script to locate a run (latest/ID/URL), download artifacts, flatten wrapper dirs, and write .run-info.json.
e2e/agents/cursor_cli.go	Create Cursor config directory during bootstrap to reduce flaky failures.
e2e/README.md	Document local artifact download + new triage workflow references.
cmd/entire/cli/strategy/common_test.go	Minor formatting/alignment in tests.
.claude/skills/e2e/triage-ci.md	New CI triage procedure doc (download artifacts or rerun locally, classify flaky vs real-bug).
.claude/skills/e2e/implement.md	New procedure doc for applying fixes and verifying via scoped E2E runs.
.claude/skills/e2e/debug.md	Update/normalize debug procedure doc formatting.
.claude/skills/e2e/SKILL.md	Add orchestrator skill definition for the E2E triage→implement pipeline.
.claude/skills/agent-integration/*.md	Update references to use /e2e:debug instead of the old command name.
.claude/plugins/e2e/**	Add local plugin command wrappers + plugin metadata and README.

[Copilot]

The redirection 2>&1 >&2 doesn’t reliably enforce the contract that only the final absolute path is written to stdout; depending on evaluation order it can still leak output to stdout. Redirect the command’s stdout to stderr explicitly (and keep stderr on stderr) so callers can safely capture stdout.

Suggested change

	gh run download "$run_id" --dir "$dest" 2>&1 >&2 || die "Failed to download artifacts. They may have expired (retention: 7 days)."
	gh run download "$run_id" --dir "$dest" 1>&2 || die "Failed to download artifacts. They may have expired (retention: 7 days)."

[Copilot]

This hard-codes the Linux-style config path ~/.config/cursor. To keep local E2E runs working across OSes (and respect XDG_CONFIG_HOME), prefer deriving the base config dir with os.UserConfigDir() and then appending cursor.

Suggested change

	home, err := os.UserHomeDir()
	if err != nil {
	return fmt.Errorf("get home dir: %w", err)
	}
	dir := filepath.Join(home, ".config", "cursor")
	cfgDir, err := os.UserConfigDir()
	if err != nil {
	return fmt.Errorf("get user config dir: %w", err)
	}
	dir := filepath.Join(cfgDir, "cursor")

[Copilot]

The README lists .github/workflows/e2e-triage.yml, but there’s no such workflow file in .github/workflows/ in this branch. Either add the workflow in this PR or remove/adjust this bullet so the documentation doesn’t point at a non-existent file.

Suggested change

- **`.github/workflows/e2e-triage.yml`** — Auto-triggers on E2E failure via `workflow_run`. Runs Claude Code (Opus) to download artifacts, classify failures, and create PRs (flaky) or issues (real bugs).

[alishakawaguchi]

@cursor review

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

[claude]

🟡 When deferred condensation succeeds via CondenseSession at line 2247, the result is discarded (_) and no session state is updated afterward — unlike the normal path through condenseAndUpdateState which updates BaseCommit, StepCount, CheckpointTranscriptStart, LastCheckpointID, clears FilesTouched, and schedules shadow branch deletion. Additionally, nil is passed for committedFiles, causing the 1:1 file filtering (lines 158-177 of manual_commit_condensation.go) to be skipped entirely, so the deferred checkpoint will have ALL session files in FilesTouched instead of just the files from the specific commit. These are minor metadata quality issues in a rare edge-case recovery path, but could theoretically cause duplicate condensation if a manual commit occurs before a new turn starts.

Extended reasoning...

What the bug is

The deferred condensation path in finalizeAllTurnCheckpoints (lines 2244-2261) has two related deficiencies:

1. Nil committedFiles: CondenseSession is called with nil for committedFiles (line 2247). The filtering logic at lines 158-177 of manual_commit_condensation.go checks len(committedFiles) > 0, so with nil the entire filter block is skipped. The resulting checkpoint will have ALL files from the session transcript in FilesTouched, not just the files from the specific commit linked via the Entire-Checkpoint trailer.

2. No state updates after success: When CondenseSession succeeds, only a log message is emitted. The CondenseResult is discarded (_). Compare with condenseAndUpdateState (lines 1117-1134) which updates: state.BaseCommit, state.StepCount, state.CheckpointTranscriptStart, state.LastCheckpointID, clears state.FilesTouched, and schedules shadow branch deletion.

The specific code path

This fires when: (1) Cursor commits mid-turn before flushing the transcript, (2) PostCommit condensation fails (empty transcript), (3) the checkpoint ID is recorded in TurnCheckpointIDs via the new condenseAttempted flag, (4) at HandleTurnEnd, finalizeAllTurnCheckpoints tries UpdateCommitted which returns ErrCheckpointNotFound, and (5) falls back to CondenseSession.

Step-by-step proof for the multi-commit scenario (bug_008)

1. PostCommit C1: Condensation fails (empty transcript). condenseAttempted=true, condensed=false. state.BaseCommit stays at B0. Shadow branch at B0 is preserved (uncondensedActiveOnBranch prevents deletion). TurnCheckpointIDs=[cp1].
2. PostCommit C2 (same turn): state.BaseCommit still B0, shadow at B0 resolves correctly. Condensation succeeds. condenseAndUpdateState updates state.BaseCommit=C2, state.CheckpointTranscriptStart=totalLines, state.FilesTouched=nil. Shadow at B0 added to shadowBranchesToDelete and deleted. TurnCheckpointIDs=[cp1, cp2].
3. HandleTurnEnd: finalizeAllTurnCheckpoints runs. For cp2, UpdateCommitted succeeds. For cp1, gets ErrCheckpointNotFound. Falls back to CondenseSession(ctx, repo, cp1, state, nil) — but state has been mutated: BaseCommit=C2, CheckpointTranscriptStart points past C2. getShadowBranchNameForCommit(C2, ...) looks for the wrong shadow branch. The checkpoint cp1 is created with data from the current (mutated) state, not the state at commit C1 time.

Addressing the refutation

One verifier refuted the state update issue, arguing: (a) the comment at lines 2277-2281 says "Do NOT update CheckpointTranscriptStart here", (b) shadow branches are not orphaned because PostCommit didnt update BaseCommit, (c) its best-effort by design. However, these arguments assume the normal deferred path where PostCommit condensation failed for ALL checkpoints in the turn. In the multi-commit scenario where a later condensation succeeds, condenseAndUpdateState DOES update BaseCommit and DOES delete the shadow branch. The comment is correct that updating CheckpointTranscriptStart would break carry-forward in the normal case, but the deferred path should at minimum handle the case where the CondenseResult is used (not discarded) so TotalTranscriptLines is available if needed. The "best-effort" framing is fair — creating any checkpoint is better than none — but the metadata imprecision (wrong FilesTouched, potentially wrong file state in multi-commit turns) is worth noting.

Impact

The practical impact is limited because: (1) this is a recovery path for an already-rare edge case (Cursor committing before transcript flush), (2) the core transcript data in the checkpoint is correct, (3) the FilesTouched imprecision is metadata quality rather than functional breakage, and (4) the multi-commit scenario (first condensation fails, later succeeds) is extremely narrow. The worst case is a checkpoint with a superset of files in FilesTouched and potentially stale CheckpointTranscriptStart that could cause sessionHasNewContent to trigger redundant condensation on a subsequent commit.

How to fix

For the nil committedFiles: resolve the committed files from the checkpoint IDs commit (via the Entire-Checkpoint trailer) using filesChangedInCommit. For the missing state updates: after successful deferred condensation, update at minimum CheckpointTranscriptStart from the CondenseResult.TotalTranscriptLines, and schedule the shadow branch for cleanup. For the multi-commit scenario, the proper fix would require storing per-checkpoint metadata (BaseCommit, committedFiles) alongside TurnCheckpointIDs at record time, which is a more significant design change.

[claude]

🟡 Nit: cursorConfigDir() leaks the temp directory created by os.MkdirTemp if the subsequent os.MkdirAll (line 85) or os.WriteFile (line 88) fails. Adding os.RemoveAll(tmp) before the error returns on lines 86 and 89 would fix this.

Extended reasoning...

What the bug is

The cursorConfigDir() function at e2e/agents/cursor_cli.go:79-92 creates a temporary directory via os.MkdirTemp on line 80, then performs two subsequent operations: os.MkdirAll to create a cursor subdirectory (line 85) and os.WriteFile to seed a config file (line 88). If either of these operations fails, the function returns an error without cleaning up the tmp directory.

Why existing code doesn't prevent it

The caller in startInteractiveSession only receives the configDir path on success and cleans it up via os.RemoveAll(configDir) in two places: the error path after NewTmuxSession fails, and via s.OnClose() on success. But when cursorConfigDir() itself returns an error, the caller never receives the path, so no cleanup occurs.

Step-by-step proof

1. os.MkdirTemp("", "cursor-e2e-config-*") succeeds, creating e.g. /tmp/cursor-e2e-config-123456
2. os.MkdirAll(filepath.Join(tmp, "cursor"), 0o755) fails (e.g., filesystem full, permissions issue)
3. Function returns ("", error) — the caller gets no path to clean up
4. /tmp/cursor-e2e-config-123456 remains on disk with no reference to it

Impact

This is very minor in practice. This is E2E test infrastructure code, the failure scenario requires the filesystem to fail between two closely-spaced operations, the leaked directory is tiny, and the OS cleans /tmp periodically. However, it is new code introduced in this PR and the fix is trivial.

How to fix

Add os.RemoveAll(tmp) before each error return after the MkdirTemp call:

if err := os.MkdirAll(cursorDir, 0o755); err != nil {
    os.RemoveAll(tmp)
    return "", fmt.Errorf("mkdir cursor config: %w", err)
}
if err := os.WriteFile(filepath.Join(cursorDir, "cli-config.json"), []byte("{}"), 0o644); err != nil {
    os.RemoveAll(tmp)
    return "", fmt.Errorf("seed cli-config.json: %w", err)
}