Pharos Release Runner

Trigger And Exclusions

Use only for requested commit, push, publish, release or deployment watching, not read-only reviews or overlapping active writers. Preserve unrelated work; never stash, reset, checkout or delete it.

Classify The Operation

Select commit-only, publish/protected-main PR, deployment watch, or operational acceptance. Inspect status, cached/uncached diffs, history and origin/main; distinguish committed-ahead work, cohesive dirty work, separate themes and unrelated user/agent files. Route intended files before acting.

Mandatory Core

Read Deployment Core Rules. Actual diffs and ownership/theme define commit batches; “commit all” includes cohesive pending work, not clearly unrelated active work.

A push/publish/release request authorizes the necessary release branch and protected-main PR path, never direct main push. Parent alone stages, commits, pushes, merges and judges readiness. Missing readiness blocks publication; successful activation alone is not runtime health.

Branch Reads And Actions

• Commit preparation: read script operational notes and the affected automation registry entries. Keep generated artifacts with their source commit. The pre-commit hook syncs affected autoStage outputs, rejects unsafe unstaged-source overlap, and skips merge/rebase/cherry-pick/revert, an empty index or PHAROS_SKIP_ARTIFACT_HOOK=1; it is neither a test gate nor full convergence. Inspect hook output and re-check the tree after each commit/generator so reviewed and released state match.
• Publish: read Pre-push readiness before every first/replacement push; it is the sole ordered readiness/receipt procedure. Push only with proof, create the PR, wait for required checks and merge with gh pr merge --merge, never squash/rebase. Verify resulting main has two parents and contains the recorded PR head; retain both SHAs. A gate failure routes to pharos-ci-failure-triage: collect every failed leaf, fix all causal defects in one revision, obtain fresh readiness and push once.
• Deployment watch: read CI Deploy Sequence and Monitoring Without Model Polling. Watch Deploy to Cloudflare for the merged SHA; record classifier-selected Pages/Worker surfaces and activation/marker proof. Use one deadline-bounded native GitHub watcher and existing Worker evidence commands, scratch samples and completion evidence, not sleep/status loops or polling-only agents. A missing matching execution at deadline is pending, not an indefinite new watch.
• Operational acceptance: read Operational Acceptance. Cron/scheduler/ingestion/memory/migration work remains pending until its first relevant production observation. Keep deployment and runtime proof separate.
• Authorized delegation: use reviewer prompts for read-only readiness review or dirty-tree classification; parent retains mutation authority and final judgment.

Checks Owned By The Verifier

Assigned verifier checks are authoring feedback, not release proof. For publishing, require the final committed HEAD's fresh passing .tmp/pr-check-receipts/<HEAD>.json under Pre-push readiness; subsequent edits/integration invalidate it. That owner defines runtime/base/head/lane completeness, CI-deferred coverage/Pages, supplemental parity/coverage/Pages checks and generated convergence. GitHub's required PR gate remains authoritative; check:release is only an explicit production rehearsal. There is no pre-push test hook.

Use npm run check:pr -- --explain-receipt for read-only receipt diagnosis: local Git state only, no fetch/check execution/receipt rewrite/staging/push authorization. Diagnostic success is not readiness, even for missing/failed/stale evidence; remote-base freshness remains unknown.

Completion Evidence

Report commits, focused verifier feedback, full readiness/current-HEAD receipt and generated convergence, PR/run/deploy proof, operational acceptance or pending window, excluded dirty files and skipped checks with reasons. Never label pending operational acceptance as success.