ci(risk tier): grade on the glaeda-route App's budget; create risk labels only when missing

Merged
TE
25d ago
+194-10

risk evidence now does its reads on the manaflow-glaeda-route App's budget, the way merge freeze already does, and creates the risk labels only when one is missing.

Why. Every workflow shares one GITHUB_TOKEN installation budget. On 2026-10-26 the janitor read 4514 of 5000 left at 07:18Z and 0 at 07:52Z, and the budget ran out again before 08:48Z. That second drain failed the nightly-next promotion request (tick run 40294661953). Before this change, each risk evidence run (24 to 35 an hour) spent about eight GITHUB_TOKEN calls:

  • the files, the pull request and the compare;
  • four gh label create calls, each answering 422 because the label already exists;
  • a gh pr view for the labels;
  • for high-risk PRs, also the comments, reviews and label events.

What changes (main's copy, which is the one GitHub runs for pull_request_target):

  • Token mint: a Mint the glaeda-route token step asks for contents, issues and pull-requests, all read-only. It is continue-on-error, like merge-freeze.yml.
  • Grade: runs with GH_TOKEN=$READ_TOKEN. If the mint failed or the App's read is refused, it grades again on GITHUB_TOKEN. Output and summary go to temp files first, so a failed first attempt leaves no half verdict.
  • Label:
    • The labels read goes on the App, with GITHUB_TOKEN as the fallback.
    • Removes and adds stay on GITHUB_TOKEN. Its label events start no workflow, so the step doesn't grade itself again.
    • The four labels are created only when --add-label fails, then the add is tried again.
  • Result: in the usual run (tier unchanged), GITHUB_TOKEN calls go from about 8 to 0.

Not changed:

  • No permission, environment, trigger or concurrency change.
  • The gate logic and ENFORCE_FROM are untouched.
  • No supply-chain or permissions test is loosened.

Tests: tests/test_ci_workflow_run_api_budget.py (wired in ci-guards) has a new RiskTierCalls class. It runs both steps as written, with a recording gh and a stand-in risk_tier.py, and checks:

  • reads use the App token;
  • a refused App read falls back with exactly one verdict;
  • an unchanged label costs only one App read;
  • a tier change makes two writes on GITHUB_TOKEN and creates no label;
  • a missing label is created and then added.

Also ran locally: test_ci_fork_runner_routing, test_ci_required_checks_have_no_path_filters, test_ci_pr_media_privacy, test_ci_production_secrets_protected_env, test_ci_workflow_permissions_are_job_scoped, test_ci_signing_secrets_need_release_environment and test_ci_self_hosted_guard.sh. All pass.

Overlap:

  • cmux#25203 (gt-quota, same lane) touches this step too. Its risk-tier part will be dropped in favour of this PR.
  • cmux#25301 changes feat-cmux-next's copy of risk-tier.yml, which pull_request_target never runs. See the comment at the top of main's copy.

Risk: low. If the App token is missing or refused, every path falls back to today's GITHUB_TOKEN behaviour. Only a missing label costs more than today: four creates plus a second add, once.

Agent session: agent session

🤖 Generated with Claude Code


View with [code]smith Autofix with [code]smith
Need help on this PR? Tag @codesmith-bot with what you need.

Summary by CodeRabbit

  • New Features
    • Pull requests can now be labeled with low, normal, or high risk tiers. A separate label marks high-risk pull requests reviewed by a maintainer.
  • Improvements
    • Risk grading and label checks can fall back to the workflow’s standard token when an optional app token is unavailable or a read fails. Labels are created only when needed.
  • Tests
    • Added coverage for token fallback, risk-label updates, and missing-label handling.