reflect: sibling-transcript check, template-identical caution, commit-as-you-go

[EdbertChan]

Summary

Three findings from running /reflect against Invoker's fleet / a71f331 (12 jobs) CI-repair incident (2026-08-17), where at least 4 uncoordinated Invoker workflows independently worked the same failing job over ~7 hours (none landed), and two prior reflect passes on that same incident crashed non-zero before their own drafted skill edits were ever committed.

• Step 3: check for a sibling reflect's surviving transcript before fanning out lenses. A crashed sibling's real synthesis work is only visible in its raw JSONL (never committed), so without this check a later reflect re-derives the same facts from zero instead of reading it as prior art.
• Step 3 (lens table): caution for History (and any lens comparing multiple fix attempts) that agreement between commits isn't corroboration when both were rendered from the same static formula/template against the same anchor commit — directly observed as two commits diffing to nothing but SHA/timestamp/task-ID, converging on the same wrong answer (a regression that reversed an already-merged prior fix, uncited by either commit).
• Step 6: commit each applied skill edit immediately, not batched to the end. Directly observed as the cause of losing a sibling reflect's two drafted CLAUDE.md findings — its closing-summary turn was blocked by an unrelated Stop hook with no further turn, and the edits were never committed.

The step-6 addition also folds in the still-unmerged fix from branch reflect-check-for-unlanded-prior-drafts (check git history for an unlanded prior reflect commit before drafting a new one) rather than leaving it to bit-rot separately, and explicitly notes the gap it doesn't cover (an uncommitted sibling draft with no git trace at all) — which the new step-3 transcript check closes.

Review Claim

Each addition traces to a directly-observed, cited fact from this reflect pass's own investigation (git commits, PR numbers, and a sibling transcript's own tail), not a speculative rewrite.

Test plan

• Read the full resulting skills/reflect/SKILL.md diff for tone/structure consistency with the rest of the doc.
• Confirmed via gh pr list that no existing open PR already covers this before opening.
• No script/test changes in this PR (prose-only); skills/reflect/scripts/tests/ untouched.

🤖 Generated with Claude Code

---

Note

Low Risk
Prose-only updates to the reflect skill workflow; no runtime, auth, or script changes.

Overview
Process hardening for reflect so lessons aren’t re-derived or lost when passes crash or duplicate each other.

Step 3 adds a pre–lens-fan-out check for sibling reflect transcripts on the same incident (ls + keyword grep under ~/.claude/projects/), with guidance to read a crashed sibling’s tail as prior art—covering cases where nothing was ever committed (complementing step 6’s git-based unlanded-draft search).

A new History lens caution warns that two commits agreeing on the same fix is not independent corroboration when both were rendered from the same static template against the same anchor; reviewers should diff commit bodies (ignoring SHA/timestamp/task IDs) before treating agreement as evidence.

Step 6 instructs checking git for unlanded prior reflect commits on the same skill section before drafting again, and committing each skill edit immediately instead of batching at the end, so Stop-hook blocks or late crashes don’t leave drafted edits only in the transcript.

^{Reviewed by Cursor Bugbot for commit f6df334. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit f6df334. Configure here.}

[cursor]

Sibling check searches wrong paths

High Severity

The prescribed sibling check ls ~/.claude/projects/ | grep -F <incident-or-branch-keyword> only matches encoded project directory names. Per step 1 and the skill’s own scanners, transcripts are *.jsonl files inside those directories, and incident keywords usually appear in file content—not directory names. Sibling reflects on the same cwd share one project dir, so this command systematically misses the prior-art transcripts the step exists to find.

 

^{Reviewed by Cursor Bugbot for commit f6df334. Configure here.}