[Szer]

Problem

The project agent workflow (.github/workflows/project.yml, .github/prompts/project.md) was extremely spammy:

• The prompt defaulted to bumping an existing open issue with a zero-information **Project assessment bump (date)** Still relevant. comment on every run. Deprecate dead pre-event-sourcing tables (rename first, drop later) + delete dead code referencing them #330 got 19 such bumps in 6 weeks, 9 of them after the maintainer had already commented with a linked PR. Historically ReminderService uses invalid Task casts that fail at runtime #28/NotificationService uses invalid Task cast that fails at runtime #29 got 28 bumps each.
• Every run created a throwaway "Daily project assessment YYYY-MM-DD" orchestration issue (33/month), each getting 2 boilerplate comments and an auto-close.
• Daily cadence was overkill for a repo-checkout-only tech-debt sweep.
• The model had repeatedly ignored explicit prompt rules (e.g. --body instead of --body-file), so prompt-only enforcement wasn't reliable — this needed a mechanical backstop.

Changes

1. Prompt rewrite (.github/prompts/project.md): deleted the "bump if exists" rule and its template. Rule 2 (search before creating) now defaults to doing nothing when a matching open issue exists — open already means "still relevant". New rule 3: a comment on an existing finding is allowed only to close it as resolved, or to add evidence that actually changed — never to restate what's already there — with hard exclusions (never comment if a human commented, a PR references it, or the last agent comment is <14 days old). All references to the orchestration issue are gone; the agent's final action is now writing its summary to $AGENT_SUMMARY_FILE, and it's told $AGENT_LOG_ISSUE exists and must never be touched.

2. Mechanical enforcement (scripts/agent-shim/gh): a bash shim placed first on PATH for the agent step. It resolves the real gh, and refuses (REFUSED by project-agent gh shim: ..., exit 1) gh issue comment when the issue is the log issue, has a human comment, is referenced by a PR (via the timeline API), or was bot-commented on within AGENT_COMMENT_THROTTLE_DAYS (14d); refuses gh issue close on the log issue. Everything else passes through. Every call appends one line to $AGENT_ACTIONS_LOG (ALLOWED/REFUSED/PASSTHROUGH) — this log is the evidence the shim was actually live inside the codex sandbox. scripts/agent-shim/test.sh exercises all 7 rules against a fake gh (19/19 assertions pass, no network). Both scripts are shellcheck-clean.

3. No more per-run orchestration issue (.github/workflows/project.yml): removed the cleanup job and per-run issue creation. Added "Ensure project agent log issue" (finds-or-creates one long-lived, pinned, open Project agent log issue) and a PATH/self-test step that fails the run if gh doesn't resolve to the shim. Post-agent steps (if: always()) assert the agent wrote a summary file and that the actions log is non-empty (empty = shim never reached inside the sandbox = hard failure, never a silent "clean day"), publish both to the step summary, and post to the log issue only when the actions log contains at least one ALLOWED line. Clean runs post nothing anywhere.

4. Weekly cadence: cron 37 4 * * 1, workflow renamed Weekly Project Assessment. Updated workflow-canary.yml's stall threshold (7d2h) and agent-retry.yml's workflow-name match. Grepped the tracked tree for remaining "Daily Project Assessment" references — none left outside these files.

Verification

• actionlint clean on all three touched workflows; shellcheck clean on both shim scripts; scripts/agent-shim/test.sh 19/19 pass; scripts/check-comment-ratio.sh --cached passes (comment blocks capped at 2 lines everywhere, including the shim scripts).
• Dispatched one real run from this branch: https://github.com/Szer/bots/actions/runs/35024351898 (success, ~1m30s). Observed:
  • Self-test: gh resolves to the shim: /home/runner/work/bots/bots/scripts/agent-shim/gh ✅
  • agent_posted=true, shim_reached=true (actions log non-empty) ✅
  • The agent's own reasoning trace: "There's also the open issue Deprecate dead pre-event-sourcing tables (rename first, drop later) + delete dead code referencing them #330 regarding dead tables, which I shouldn't duplicate" — it found Deprecate dead pre-event-sourcing tables (rename first, drop later) + delete dead code referencing them #330 (which has a human comment + a linked PR) and, under the new default-silence rule, made zero gh issue comment/create/close calls on it. The only gh call made all run was gh issue list --state all --label project ... (a search), which the shim logged as PASSTHROUGH issue list and let through — proving the shim was live inside the sandbox.
  • Because the actions log had no ALLOWED line, "Post to log issue when actions were taken" correctly posted nothing: No ALLOWED actions this run — nothing to post.
  • Net result: the prompt rewrite alone was sufficient this run to avoid a Deprecate dead pre-event-sourcing tables (rename first, drop later) + delete dead code referencing them #330 bump — the shim's REFUSED path wasn't mechanically exercised live (nothing was attempted to refuse), but it is covered by the 19/19 offline test.sh assertions (human-comment, PR-reference, and throttle cases all assert REFUSED).

Follow-ups (not in this PR)

• product.yml/product.md have the identical per-run orchestration-issue pattern; deliberately out of scope here.
• The gh shim is reusable — product.yml could adopt it the same way if its own bump/comment behavior ever needs the same mechanical backstop.

🤖 Generated with Claude Code