docs(skills): task-email-audit — round-2 learnings (conditional no-sends, 401-rejected rows, retry resets)

[sweetmantech]

Prep so tomorrow's post-deploy audit runs without operator feedback — net-negative diff (+25/−51): the new learnings are funded by removing what the shipped fixes made obsolete.

Added (as tightened caveats)

1. Classify runs sent / correct no-send / failed — conditional prompts (e.g. the rewritten roster-wide new-video sweep) make no-send a success; read the run's final narration before judging.
2. rejected ≠ always guard-blocked — expired ephemeral keys (runs >~20 min) log 401 retries as rejected in bursts; triage via the trace before reporting guard stats.
3. Part counts can shrink — workflow-step retries reset the persisted assistant message.
4. Vercel-login-expired fallback on the prod TRIGGER_SECRET_KEY step.

Removed (obsolete after chat#1829 + chat#1833 shipped)

• §4 "Blocked empties" listing section — the §1 headline already counts rejected rows; interpretation now lives in the caveat.
• §5 "Helper adoption" — the helper shipped and is a registered default skill; the adoption metric served its rollout purpose.
• The dated 2026-07-01 "reference point" anecdote, the pre-fix Resend-scraping fallback method (no one audits pre-06-30 windows), and baseline percentages (the linked issues carry the history).
• Sections renumbered (§6/§7 → §4/§5), all cross-refs updated.

CI green: check_resolvable, run_resolver_eval (structural), validate_manifests.

🤖 Generated with Claude Code

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR updates the "Caveats" section of the recoup-internal-task-email-audit SKILL.md documentation with expanded guidance on interpreting audit data: conditional no-send outcomes, rejected email log causes, chat_messages part-count fluctuations, and vercel login token error handling.

Changes

SKILL.md Caveats Update

Layer / File(s)	Summary
Expanded audit interpretation caveats skills/recoup-internal-task-email-audit/SKILL.md	Adds guidance for classifying conditional "no send" outcomes, distinguishing sandbox/API 401 rejections from guard-block rejections, noting chat_messages part-count shrinkage from durable workflow retries, and handling expired vercel login for TRIGGER_SECRET_KEY.

Estimated code review effort: 1 (Trivial) | ~3 minutes

Possibly related PRs

• recoupable/skills#69: Introduced the original Caveats section in the same SKILL.md file that this PR expands.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main docs update and the key learnings added to the task-email-audit skill.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/audit-skill-round2-learnings

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}