[Dong1017]

Governing Issue and delivery slice

Refs #284

• Accepted scope / delegation: the one implementation PR named in the D1–D4 decision record and the Issue body: the label list, the rules, template labels and the nightly drift report. Dong1017 delegated decisions in this Issue line to the acting agent on 2026-11-01, except label deletions, which Dong1017 sees first.
• What this PR delivers: .github/labels.json with the 21 target labels; a Labels section in docs/agent-collaboration.md with the group table, the status and priority rules, the skill-eval:execute revocation rule and the status-block convention; kind:task and kind:acceptance on the two Issue templates, plus a status block in the scoped-task template; and a read-only drift step in the nightly safety-net job, backed by tools/check_labels.py and its tests. The guide's opening no longer says "no new automation, permissions"; it now says the conventions add no merge gates or CI lanes, and the checks they mention are reports in the existing lanes.
• What it does not deliver: any change to repository labels, edits to [Roadmap] InferRSI: reusable knowledge, verifiable inference optimization #61, a workflow that writes labels, or a status block in the campaign form (form markdown does not reach the Issue body; the acting agent adds the block when it edits the Issue). The label changes run after merge, as listed under Remaining work.

Knowledge authorship declaration

Not applicable: this PR changes no path protected by docs/knowledge-model-policy.md.

Changed artifacts and evidence

Requirement (#284 completion criteria)	Actual change	Check / evidence and result	Remaining gap
1. Every label listed with name, group and description; names have the form group:value; each group's meaning and who applies it is documented	.github/labels.json; docs/agent-collaboration.md Labels section	test_repository_manifest_is_valid; review	#61 "Issue organization" link to the Labels section, after merge
2. Repository labels match labels.json	none in this PR	the nightly drift step	label changes after merge
3. Templates apply labels	scoped-task.md applies kind:task; campaign-acceptance.yml applies kind:acceptance instead of the nonexistent acceptance	test_templates_and_configs_reference_only_listed_labels	none
4. Nightly read-only drift report, including labels referenced by templates and tools/skill_eval_*_config.json	ci-nightly.yml safety-net: gh api lists labels with an issues: read token, then tools/check_labels.py compares names and descriptions and checks every template and config reference; printed lines start with label drift so a label name cannot form a workflow command	tests/test_check_labels.py	the first nightly run after merge
5. Three-line status block on open kind:task/acceptance/tracking/rfc Issues	convention in the Labels section; status lines at the top of the scoped-task template	review	applying blocks to existing open Issues
6. Pending-decision view	the Labels section links is:open label:"status:proposed"	review	linking the view from #61, after status:proposed exists

Design notes:

• labels.json holds the target state only; the one-time rename and deletion list is in this PR body and task: agent-maintained issue system: one label scheme, auto-labeling and a pending-decision view #284, not in the manifest.
• The drift step fails the safety-net job, like validate_evidence.py, so drift is visible. No step writes labels.
• Template front matter is parsed with PyYAML, which the nightly and ready lanes already install through .[dev].
• The one-kind:* rule exempts Issues created before the label change, as D4 records: bug: Drama3 does not fuse NPU normalization operators #166, fix: source and device identity accept evidence that cannot support it #178, fix: run state and comparison verdicts can report support the artifacts do not establish #179 and fix: knowledge bundle reads can escape the skills root and store protection depends on a .git entry #180 keep two kind:* labels after bug becomes kind:bug.
• Descriptions of kept labels are unchanged except where the decision changed their meaning: priority:* now states how soon the next action is needed, not a delivery week.

Checks and dependencies

• Checks actually run (result and tested scope): ruff check --no-cache tools/check_labels.py tests/test_check_labels.py (ruff 0.14.10) passed. The gh api ... --jq '.[] | {name, description}' output format was confirmed against the live repository: one compact JSON object per line, null for an empty description.
• Failed, skipped, timed-out or not-run checks (if applicable): pytest was not run locally; the ready lane runs the full suite. The drift step has not run; until the post-merge label changes are applied, it reports the renames, creations, description updates and deletions listed below and fails the nightly job.
• Dependencies / shared interfaces / integration assumptions (if applicable): safety-net gains job-level permissions: contents: read, issues: read; its existing steps need only contents: read. tools/skill_eval_auto_config.json and tools/skill_eval_historical_config.json are read, not changed; skill-eval:execute keeps its name and description.

Remaining work and Issue closure

• Remaining Issue requirements / follow-up: after merge, the acting agent:
  1. renames accepted → status:accepted, bug → kind:bug and help wanted → help:wanted, with the descriptions in labels.json;
  2. creates area:usability, status:proposed and status:deferred, and updates the priority:P0/P1/P2 descriptions;
  3. adds kind:docs to task: distill local-dream Android diffusion contracts with AREX repo-to-skill #234;
  4. shows Dong1017 the final deletion list (documentation, accessibility, duplicate, enhancement, good first issue, invalid, question, wontfix), then deletes those labels;
  5. links the Labels section and the pending-decision view from [Roadmap] InferRSI: reusable knowledge, verifiable inference optimization #61, and records the before/after label list on task: agent-maintained issue system: one label scheme, auto-labeling and a pending-decision view #284.
• Goal-acceptance owner or accepted closure delegation: Dong1017
• Goal ready for acceptance? Not claimed

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-01T13:10:44.101882Z	1f35624	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d5a31947bf

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Qualify the revocation guarantee for already-checked requests

If the label is removed after a request's last live-label check but before its provider HTTP send, that request can still start. controlled_execute.check() performs further ledger reads after live_label(), and anthropic_agent.send_message() obtains its grant before separately calling opener.open(); src/inferrsi/model/request_guard.py:5-6 explicitly documents this race. Consequently, limiting the exception to requests already sent gives operators an incorrect stop-spending guarantee. State that subsequent live checks enforce revocation and include requests that passed their last check but have not yet been sent.

Useful? React with 👍 / 👎.

[Dong1017]

Agreed; request_guard.py lines 5-6 document that race. Fixed in 1f35624: every check after the removal blocks its request, and a request that passed its last check before the removal can still be sent.

[Dong1017]

LGTM for this implementation slice at 1f35624. The prior revocation-wording finding is fixed: the documentation now explicitly retains the race for requests that passed the last check before label removal but have not yet sent, consistent with request_guard.py. The author reply alone was not used as clearance; the one-file correction was checked against the guard's stated contract.

Current-head Ready CI passed all six jobs on the merge with main 7b6d00a, and the existing independent review completed without remaining findings.

This delivers the manifest, conventions, templates and read-only drift checker. Live label migration, the separately confirmed deletion list, first nightly drift evidence and Issue #284 acceptance remain with the existing owner.