[purp]

Summary

Connect the shared labeling helper to the managed classifier through one issue/PR runner. A number or URL produces a validated preview; explicitly authorized application reuses the saved proposal and exact preview, rechecks current evidence and eligibility, and preserves the helper's checkpoint and partial-failure recovery.

Related Issue

Closes #5411.

Based on #5407 (f2f0fae3a876d27e9848568baef89cff7630533e), with #5355's shared-helper branch (af08bc89249083f16dcb56b43ed56d1a6446a7e1) merged as a dependency. The parent branches live in the fork, so this draft targets main. Review the runner changes alone.

Changes

• Add a preview-first one-item entry point with saved-preview --apply, human --force, and credential-free --request-preview; consolidate the temporary adapter CLI.
• Stop deterministic preparation outcomes before model invocation. Support bounded evidence-expansion rounds using trusted checkout instructions and the current shared result contract, with an invocation-wide budget and deadline.
• Delegate eligibility, classification validation, exact mutations, checkpoint verification and local ambiguity formatting to the shared helper. Persist session IDs before polling and retain execution references and uncertain outcomes without automatically replaying writes.
• Document setup, concise reports, recovery and live prerequisites; register offline tests and code ownership. Automatic dispatch and coordination across entry points remain follow-up work.

Release Notes

Testing

• 245 labeling tests pass: uv run --no-project --python /usr/bin/python3 --offline python -m unittest discover -s tasks/scripts -p '*label*test.py'. These include actual-helper orchestration with fake GitHub/CMA, deterministic skips, expansion, exact preview application, stale eligibility/content, verified checkpoints, uncertain creation/writes and recovery.
• 28 accepted-work and 43 review-readiness Python tests pass; 19 JavaScript issue-acceptance tests pass.
• Read-only preparation of live feat(review): connect the end-to-end labeling runner #5411 returns ready, classifier_required: true, with accepted/eligible work. No live CMA sessions or label writes were performed.
• Task TOML, Python syntax, skill metadata/inventories, local links/anchors, crate inventory, new-file SPDX headers and staged diff checks pass.
• Ruff, Markdownlint and the mise wrapper remain unverified: required tool downloads were blocked by network restrictions. Direct stdlib test commands above passed. Actionlint and external skill discovery were not run.
• Observed live CMA configuration, previews and authorized one-item applications remain pending. This draft does not claim completion of that live-validation criterion.

Checklist

• Follows Conventional Commits
• Commits are signed off (DCO)
• Internal integration documentation and labeling skill updated

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.