fix(retro): warn on stale date windows

[jbetala7]

Fixes #1624

Summary

• Move the /retro date/window guard into a dedicated pre-flight before normal data gathering.
• Use a window-aware stale-base threshold: warn when the latest default-branch commit gap is greater than the current window length (gap > window_days), not a fixed 3 days.
• Print the system date, current window length, latest origin default-branch commit date/gap, and commit count before narrative generation.
• Apply the same pre-flight to /retro compare and to each discovered repo in /retro global; global mode surfaces affected repos before aggregate totals.
• Warn for mismatched today anchors and empty windows so /retro does not produce a clean-looking empty report from bad date/window evidence.

Tests

• bun run gen:skill-docs --host all
• bun test test/gen-skill-docs.test.ts -t 'retro SKILL.md warns on stale base or empty windows'
• bun run scripts/gen-skill-docs.ts --dry-run --host claude
• git diff --check
• git diff --check origin/main...HEAD
• git merge-tree --write-tree HEAD origin/main
• Temporary 9-day-gap/7d proof: latest origin/main commit dated 2026-05-12, system date 2026-05-21, RETRO_WINDOW_DAYS: 7, RETRO_WINDOW_COMMIT_COUNT: 0, emitted both STALE-BASE WARNING and EMPTY-WINDOW WARNING.

Collision checks

• Open PR search for 1624 retro stale window returned only this PR.
• Open PR search for STALE-BASE WARNING returned only this PR.
• Open PR searches for RETRO_WINDOW_DAYS and retro global empty window found no competing PR.
• Fresh origin/main merge-tree completed cleanly.

[poppinpixels]

Issue author. Four questions from #1624 I don’t see resolved in this PR:

• Threshold: fixed 3d, or window-aware (gap > window_days)?
• Scope: does this extend to /retro compare and /retro global? Global mode’s multi-repo discovery has the same failure mode.
• Placement: Step 1 inline, or dedicated pre-flight?
• Reproducer: tested against the 9-day-gap case in /retro silently produces empty/misleading output when 'today' anchor is wrong (no stale-base warning) #1624, or a synthetic empty-window?

[xg-gh-25]

Smart catch on the stale-window issue — this is exactly the kind of silent failure that undermines trust in retro tools.

Why this matters:
When /retro runs with an outdated anchor date, it produces plausible-looking but meaningless output. Users assume the tool is working, waste time analyzing fake trends, and only realize the issue days later when retrospectives don't align with reality.

Implementation suggestion:
Your warning approach is solid. I'd recommend:

1. Threshold: Flag if today - anchor_date > 48 hours (aggressive but prevents most drift)
2. Visibility: Surface the warning above the output, not just in logs — make it impossible to miss
3. Auto-correction option: Offer a quick action: "Anchor is stale. Re-run with today's date?" (reduces friction)

The tricky part is handling legitimate use cases where someone intentionally runs historical retros (e.g., auditing last month). Maybe:

• Default: auto-warn
• Escape hatch: --allow-stale-anchor flag for intentional historical runs

Have you considered caching the last anchor date per workspace to detect drift automatically?

---

Posted by SwarmAI — an agentic community engagement system

[jbetala7]

Addressed in 79e4724d.

• Threshold: changed from fixed > 3d to window-aware gap > window_days. The pre-flight now resolves default 7d, Nd, Nw * 7, and ceil(Nh / 24) with a minimum of 1 day, then emits RETRO_WINDOW_DAYS.
• Scope: yes, this now explicitly covers normal /retro, /retro compare current-window metrics, and /retro global per discovered repo. Global mode runs the same pre-flight after repo/default-branch detection and lists affected repos before aggregate totals.
• Placement: moved to a dedicated ### Date/window pre-flight section before Step 1, rather than leaving it inline inside raw data gathering.
• Reproducer: tested the 9-day-gap/7d case from /retro silently produces empty/misleading output when 'today' anchor is wrong (no stale-base warning) #1624 with a temporary git repo, not only a synthetic empty-window. The proof had system date 2026-05-21, latest origin/main commit 2026-05-12, RETRO_WINDOW_DAYS: 7, RETRO_WINDOW_COMMIT_COUNT: 0, and emitted both STALE-BASE WARNING and EMPTY-WINDOW WARNING.

Verification also passed: bun run gen:skill-docs --host all, focused generated-doc regression, Claude dry-run freshness, git diff --check, and merge-tree against fresh origin/main. I updated the PR body with the exact proof and collision searches.

[poppinpixels]

Good turnaround. Brief read on each:

1. Window-aware gap > window_days matches what my local patch does. Better than fixed 3d.
2. Scope extension to compare and global per discovered repo closes the gap I flagged.
3. Pre-flight section reads cleaner than the inline option.
4. The 9-day case with system date 2026-05-21, commit 2026-05-12, both warnings emitted is exactly the reproducer.

Will verify against my local repo and reproducer.

The three failing E2E Eval checks (e2e-review, llm-judge, report) — related to this change, or pre-existing flakes?

[jbetala7]

Not related to this change from what I can see.

• llm-judge: the job starts with Secret source: None, then the Anthropic SDK fails before the judged eval can run because no API auth is available. That is an external PR/secrets issue, not branch logic.
• e2e-review: the failing case is /retro default branch detection, but the eval comparison reports FAIL -> FAIL, Status: 3 unchanged, and No regressions. So it is the existing baseline failure, not a new regression from this branch.
• report: it downloads the eval artifacts and builds the PR comment body, then fails only while posting the report comment with GraphQL: Resource not accessible by integration (addComment). That is the external PR token permission issue.

The branch-specific checks remain green: actionlint, build-image, check-freshness, windows-free-tests, and the rest of the E2E suites all passed. The local verification for this change also passed as noted above.