Skip to content

2026-07-13-clean-sweep-self-run.md

Latest commit

 

History

History
82 lines (68 loc) · 5.31 KB

2026-07-13-clean-sweep-self-run.md

File metadata and controls

82 lines (68 loc) · 5.31 KB
Raw

Run report — clean-sweep self-run, 2026-07-13

First recorded run of a mission from this catalog. Mission: clean-sweep, source=doc-claims ("the README lies — verify and fix it"), target: this repo itself. Proof tier earned: self-run (external-run requires a repo that is not this catalog).

field value
coordinator Claude (interactive session), tubeworm worktree
BASE ravidsrk/sweep-base @ 41fec8f (preflight.py OK — first live exercise)
baseline validate.py green; 25/25 tests green
runtime Orca orchestration (task-create → spawn_worker.sh → dispatch --inject → check --wait)
workers 4 dispatched: triage (claude, ro) · builder (claude, rw) · reviewer (codex, rw — cross-vendor) · integrator/conductor (claude, rw)
reviewer_mode cross-vendor (codex reviewed claude's fix; blind-fix expectation written pre-diff)
outcome DRY — re-enumeration found zero open claims

Findings ledger (final)

id title VERIFIED CLASS PR reviewed_sha MERGED CLOSED
F-1 README badge: 19 tests (actual 25) CONFIRMED-FALSE real-bug #9 ed540a0 c85bbd3 yes
F-2 README badge: v0.1.1 (actual 0.2.0) CONFIRMED-FALSE real-bug #9 ed540a0 c85bbd3 yes
F-3 README layout: "25 architecture tests" mislabel CONFIRMED-FALSE real-bug #9 ed540a0 c85bbd3 yes
F-4 getting-started: nonexistent ../../playbooks/ mechanism CONFIRMED-FALSE real-bug #9 ed540a0 c85bbd3 yes
F-5 getting-started troubleshooting: same stale mechanism CONFIRMED-FALSE real-bug #9 ed540a0 c85bbd3 yes
F-6 CHANGELOG: duplicated body-less 0.1.1 heading CONFIRMED (defect) real-bug #9 ed540a0 c85bbd3 yes

All six folded into one PR per the build-change sizing seam (sub-10-line fixes fold into a neighbor; one file never splits). Closes cite the shared merge SHA c85bbd3, ancestry-verified on origin/ravidsrk/sweep-base by the conductor and independently by the coordinator.

Pipeline evidence

  1. Enumerate (coordinator): countable claims vs reality; 3 candidates.
  2. Skeptic-triage (worker, ro): candidates verified with pasted commands; sweep of README/ARCHITECTURE/AGENTS/docs found 2 more (S-1/S-2 — install docs described a ../../playbooks/ reference mechanism that no longer exists); CHANGELOG defect surfaced as an aside and folded in at FREEZE by audited coordinator decision. 0 refuted, 0 needs-human.
  3. Freeze: F-1..F-6.
  4. Build (worker, rw, own worktree off BASE): one commit ed540a0, author Ravindra Kumar, no trailers; negative control = revert reproduced all six false claims, then reset.
  5. Build-blind review (worker, codex, cross-vendor): blind-fix expectations written BEFORE the diff was opened; PASS on spec-fidelity, standards, and truth axes; reviewed_sha recorded; reviewer artifacts verified non-identical to worker output.
  6. Integrate + conduct (worker): PR #9 → BASE, baseRefName asserted, headRefOid == reviewed_sha freshness check, merge commit c85bbd3, ancestry + grep verification, branch deleted.
  7. Re-enumerate (coordinator): badges = 25/0.2.0, layout label fixed, zero ../../playbooks/ claims, exactly one 0.1.1 heading → DRY.

Deviations and lessons (recorded, not hidden)

  • spawn_worker.sh exit-3'd twice on workers that were actually fine: the first heartbeat can land seconds after the third 40s poll. Both were left running per the liveness policy (activity = alive); consider a longer default HB_POLL_SECS for claude workers.
  • One real dispatch failure: the first codex reviewer terminal wedged with no heartbeat (dispatch marked failed). Doctor path worked as written: evidence line → task-update → ready → fresh terminal → success on attempt 1/3.
  • One unrelated global task existed in runtime state; scoped out via the ledger per liveness-resume (counted-but-untouched).
  • The run-closure commit itself tripped a validator false-positive: the proof_evidence: path matched the path-prefixed-protocol rule, which had been written broader than its purpose. The rule now flags path-prefixed PROTOCOL names only and requires other paths to exist — with two new fixture tests. A run that exercises its own guards finds their edges; that is the point.

Run-close integrity inventory (sha256)

5166ab89c1e60c6f94946d25e498db0037d4337f2e1abf4f6b9b341f284abd24  triage.md
f4d58053900679ace43d6bbd0401c297cfb1e46ee00877cf8ab8f2c65e54a892  build-manifest.json
673d9e6458c4df207e555f8f0e70903fbe6e7751b4bddb263bb25c23b2958476  review-blind-fix.md
121e2aa7cf94e3497977d0131a0f4717c72ccc529aaa01584fd848ea0017cb1d  review.md
3316a4cee0d339869e3c603f872604be5cc883bcb93c40887c8e21d628f6b8a3  merge.md

Artifacts live in the coordinator's run directory; Orca provenance holds the task/dispatch lifecycle (task_97bda1191e39, task_3cba63915e28, task_1541d83e1a13, task_87cda3091e8f).

Gates

  • Refuted/duplicate batch gate: not needed (0 refuted, 0 duplicates).
  • BASE → main promotion: open, human-owned — this run stops at the promotion PR.