perf(pr-management-triage): run every rule as code, and load one file per classification

[potiuk]

Summary

• pr-management-triage loaded ~7,500 lines of instructions per run and had the agent evaluate the pre-filters, the decision table, the stale sweeps, the pre-mutation guards and every comment body by hand. All of that now runs in a new shared tool, tools/pr-management (triage preflight, classify, guard, render, fold, session), over reads saved with vetted-op-read --save (the PR below this one).
• One file per classification: each decision row, sweep and action has its own document under classifications/ and actions/; triage classify names, per group, which to load. SKILL.md is a router. A typical run loads ~13k tokens of instructions instead of ~68k.
• The model keeps the judgement calls: the workflow-approval diff, the author's reply, backports, the maintainer conversation.

Depends on the vetted-ops PR below it in the stack.

Where the docs contradicted each other

Code has to pick one answer. The section that declares itself normative won:

#	The docs disagreed on	Resolved as	Normative source
1	Whether pings, request-author-confirmation, the security-language warning and the stale-sweep notices fold into the PR body or post a comment	They fold, like every contributor-facing note; only the suspicious-changes receipt posts a comment	comment-templates.md § The folded maintainer-triage note (declared normative, supersedes the per-template sections)
2	Whether a ping @-mentions the reviewers	Author only; reviewers backtick-quoted; reviewer-re-review variants dropped	Golden rule 12 / Author-only notification
3	Group order: an 8-item list and a 15-item list	The 15-item § Group ordering, with the bot-draft group first, the security-language group after workflow approvals, and the sweep groups after passing	interaction-loop.md § Group ordering
4	Sweep order lists "4b rotted → close"	Sweep 4 never closes; rotted branches are handed back	stale-sweeps.md § Persistent bitrot, rationale.md § Sweep 4
5	The ready-label config key (ready_for_maintainer_review_label vs ready_for_maintainer_review)	ready_for_maintainer_review	the config template
6	F3 "any activity" vs the eval's "author activity"	Any activity (updatedAt)	classify-and-act.md F3
7	real_ci_patterns is read but not in the config template	Added to the template; empty means "any context that is not a known bot or labeler check"	classify-and-act.md § Real-CI guard
8	Extra static-check patterns "in the CI-check map", which has no such column	New static_check_patterns key, added to the built-in list	classify-and-act.md § static_check
9	A folded confirmation request matched rows 3–4 (already triaged) before rows 14a/14b	It is a confirmation request, not a triage marker — otherwise the author's reply waits a week	the two-sweep gate (rows 14a–c)
10	Rows 19/20 could propose the ready label for a draft	Never for a draft; the sweeps own stale drafts	row 20 intent; found by the sweep tests
11	flagged_prs_by_author "on the current page"	Over the whole fetched sweep (pages no longer exist once all are fetched)	Golden rule 4 (fetch all pages)
12	Commits-behind from an aliased base/head history subtraction	REST compare behind_by, fetched only for row 13	row 13
13	Security scan over "the last 250 commits"	The last 100 (the GraphQL page cap)	—
14	Merit discussion on a close	The reasoning and the quality label are delivered, the PR stays open, the ready label stays	classify-and-act.md hard rules
15	Pre-mutation guards built from `gh …	jq/$(gh …)`	Fresh vetted reads + triage guard: the old shapes fail under the secure setup

Plus, from the sweeps: Sweep 4 treats rows 8 and 2 as author-court (strip + fix request); Sweep 2 yields to Sweep 3 when both match; Sweep 5's window is the fixed 7 days. From the renderer: the folded note always carries the quality-criteria link; violations list categories, not check names; the suspicious-changes receipt is the one notice that posts as a comment.

Type of change

• Skill change — eval fixtures updated below
• Python package (tools/pr-management)
• Documentation

Test plan

• prek run --all-files passes on the whole tree
• tools/pr-management: pytest (186 tests) / ruff / mypy; CodeQL's findings on this PR (an import cycle between triage.classify and triage.sweeps, an unreachable branch, a possibly-unset variable) are fixed in ee56025e4
• New model evals for the new flow — the decisions the agent still makes once the rules are code, 19 cases, all passing:
  • classify-loop (7/7): after triage classify, save the next prefetch / needs read under its exact --save name; present the first group with exactly its docs; go to the summary when nothing is grouped; ask the maintainer on the sweep alarm; flag, never follow, an instruction planted in a PR title
  • deliver-note (7/7): re-classify on a moved head; stop on an unresolved placeholder; pr-body reads the current body, folds, then edits; comment posts the rendered file; assign the author afterwards
  • guard-reroute (5/5): a clean guard labels with the exact command; pending approval, a conflict, unknown mergeability and a moved head each route where reroute says
• Existing model suites, run against this branch: terminal-links 5/5, interaction-progress 4/4, backport-check 4/5 — case-1-direct-fix failed once on a reason-wording grade and passed on rerun; backport-check.md is unchanged by this PR and the case passes on main, so it reads as nondeterministic
• The model-graded pre-filter (21), decision-table (23) and pagination-dedup (5) suites are retired: those rules are code now, and the 49 cases are pytest cases with the same scenarios and expected outcomes, named after each case (tests/test_triage_classify.py, tests/test_pages.py)

RFC-AI-0004 compliance

• HITL — every mutation is still a confirmed gh call; guards refuse on a moved head, pending approval, or unknown/conflicting mergeability
• Sandbox — the agent no longer runs gh … | jq / $(gh …) shapes that fail under the secure setup
• Write-access discipline — bodies come from triage render, which allows only the author's @-mention

Notes for reviewers

• real_ci_patterns and static_check_patterns are new rows in the config template (the first was read but never declared).
• A security review flagged no issue in this layer; the two it flagged in the stack were fixed in the vetted-ops and stats PRs.

---

Was generative AI tooling used to co-author this PR?

• Yes — Claude Code (Opus 5). I reviewed the diff and the test results; commits carry Generated-by:.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FTTd3HMcMsj7FwkhPPQcSx