feat(control): in-place workflow reruns (GitHub-faithful) + attempt history

[Bnjoroge1]

Re-run workflow attempts in place (GitHub-faithful)

Base: Bnjoroge/lock-refactor · Head: gha-reruns-port

Gate green on this head: fmt-check clean, clippy -D warnings clean, full lib 848 passed/0 failed/3 ignored at --test-threads=1, 11 integration binaries all 0 failed, preloop-cli 206/0, focused migration+rerun lite 15/0 + PG 14/0.

Draft because of merge order, not code state: the schema migration below has been ported to the refinery migration branch (#372, V2026100503). Before merge this branch must rebase onto #372 and drop its in-branch migration plumbing (migrate_3_to_4/migrate_4_to_5, SCHEMA_VERSION bumps). The feature code and tests are final.

Problem

preloop's "re-run" resubmitted the recorded submission as a brand-new run: new run_id, run_attempt always 1, the old run left terminal. GitHub re-runs in place — same run id, run_attempt incremented, the previous attempt archived to history — and "Re-run failed jobs" / per-job re-runs only re-execute the failed cone. Tools (gh run rerun, Checks-page rerequests, the REST shims) expect that.

Design

• One write transaction resets the mode-selected jobs on the same run_id: attempt snapshot into run_history/job_history keyed by run_attempt, runs.run_attempt += 1, github.run_attempt bumped in the stored context, fresh jobId/timeline.id/requestId per selected job.
• Modes: All, Failed (failed/cancelled/timed-out seeds + dependent closure), Job(id) (+ dependents); reusable-caller subtrees reset wholesale; expanded matrix parents are never selected (their legs are).
• Re-admission mirrors submit_run step for step: workflow concurrency, if: (needs-less inline, needs-gated in promotion), platform/pool labels, environment gate (re-armed per attempt), job concurrency, max-parallel.
• Jobs outside the selection keep status + outputs, so dependents see the carried-forward needs context.
• Surfaces: native POST /api/v1/runs/:id/rerun {mode, job_id?} (absent body = all), CLI preloop rerun [RUN_ID] [--failed|--job ID], and GitHub-compat POST /repos/{o}/{r}/actions/runs/{id}/rerun, /rerun-failed-jobs, /actions/jobs/{job_id}/rerun, /actions/runs/{id}/cancel (dispatch auth, actions: write; 201/202/403/404/409/422; /repos/... errors carry GitHub's message + documentation_url beside preloop's error).
• check_run.rerequested (job mode) and check_suite.rerequested (all mode) now re-run in place; check-run reporting is scoped to the reset set so carried-forward jobs keep their existing checks.
• Archiver hold: PRELOOP_RERUN_WINDOW_DAYS (default 30, 0 disables) keeps a completed run with ≥1 failed/cancelled/timed-out job in the live tables (60 s grace, 5 s ticks; retention still wins). A fully archived run falls back to the old behavior for all (new run from the recorded submission) and 409 for failed/job.
• Schema migration: run_history/job_history gain run_attempt in their primary keys, backfilling run_attempt = 1 for pre-change rows and preserving identities. The tested SQL lives in the migrations branch as V2026100503 (refinery, forward-only); this branch's own version-stamp plumbing will be dropped at rebase.

Conflict points

• Route all GitHub check-run updates through the outbox and one background sender #365 (check-run outbox): github.rs report_check_run_* / check_run_outbox.rs vs the new rerun_checks.rs; github.rs::process_check_run_rerequest region.
• feat(server): GitHub-sourced environment protection (rebased + environment_url, unknown envs, review audit) #367 (environment gates): runs.rs::approve_job, control/types.rs (RerunRun.environment_rules), lite|pg/rerun.rs environment evaluation, control/backend.rs trait.
• trigger-compat: github.rs delivery handler (check_suite hook), dispatch.rs resolve helpers, runs.rs submit path. A rerun must NOT emit a second requested (workflow_run emitter stays submit-only) and bumps run_attempt before claim/settle hooks read it.
• Migrations branch (feat(control): versioned refinery migrations, boot ledger guard, store CLI #372): rebase drops in-branch migration plumbing; control/tests.rs suite additions keep.

Known gaps

• In-place rehydration of fully archived runs is a follow-up (all starts a new run; partial modes 409).
• Per-attempt log separation: logs stay per run (all attempts accumulate), not per attempt like github.com's UI.

---

Summary by cubic

Rerunning a completed workflow now replays it in place on the same run id, like GitHub's "Re-run jobs": run_attempt increments, the previous attempt is snapshotted into run_history/job_history, and selected jobs are reset and re-admitted through the same concurrency, if:, platform, and environment-gate chain a fresh submit uses. Previously a rerun started a brand-new run with a new run_id.

• Three modes mirror GitHub's buttons: all jobs, failed/cancelled jobs plus dependents, and one job plus dependents.
• Re-run surfaces: POST /api/v1/runs/:id/rerun (absent body = all), the new preloop rerun CLI, GitHub-compatible /repos/.../actions/runs/{id}/rerun, /rerun-failed-jobs, /actions/jobs/{job_id}/rerun, and /cancel shims, plus in-place handling of check_run/check_suite rerequested. The shims return GitHub-shaped errors (message + documentation_url) alongside preloop's error key and reject rerun bodies with enable_debug_logging/enable_debugger as 422.
• Jobs outside the selection keep their results and outputs, so dependents see the carried-forward needs context.
• A fully archived run can only be re-run in full, which starts a new run (old behavior); partial modes return 409. PRELOOP_RERUN_WINDOW_DAYS (default 30, 0 disables) holds completed runs with failed jobs in the live tables so the in-place path stays reachable; retention still wins.
• Schema: run_history/job_history add run_attempt to their primary keys. SQLite 4→5 and PostgreSQL 5→6 migrate on first open — one rollback-safe transaction — backfilling pre-change rows to attempt 1 while keeping their identities.

^{Written for commit e65da3f. Summary will update on new commits.}

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}