auto-dev run — ai-migration-kit, 2026-09-05/06

Supervisor: one session, N=3, tiered (mid by default, top reserved). The run also released v2.1.0 and updated the plugin cache from 2.0.0, so its own workers changed mid-run.

Outcome

   
PRs merged 35 (27 supervised; the rest Renovate)
Issues closed 29
Issues open at stop 13
Follow-ups filed during the run #455, #468, #469, #470, #471

Stopped on the user’s instruction, not on a drained queue. Left in QUEUE: #378, #386, #441, #447. Held: #144, #156, #432 (large), #442 (tracking parent — a person closes or rescopes it), #2 (Renovate, bot-regenerated).

#444 is open on purpose. Its Task 4 retitles six published GitHub releases. The worker was told not to run gh release edit and instead recorded the six commands in PR #466’s body; the PR says Part of #444 — not closing it. An outward-facing act stayed with the owner.

Needs manual sweep

Two worktrees hold unpushed commits and were left in place rather than pruned:

  • .claude/worktrees/agent-ac71fbd0bb0c969a2 — fix/455-base-verdict-report-line, 9 commits ahead of main, never pushed. #455 landed as a squash, so this is redundant history.
  • .claude/worktrees/agent-ad5b173de2a5d8d22 — work-447-journal-index, 11 commits ahead of main, never pushed, and #447 is still open. Deleting this worktree destroys work no remote holds.

Cost

160 sessions (1 orchestrator, 159 sub-agents) · 2.07B tokens · $1,328 list-equivalent (subscription: rate-limit budget, not cash).

   
tokens / supervised merge ~76.7M
$equiv / supervised merge ~$49
orchestrator share 35% ($459 from one session)
cache read 97.3% of all tokens

By model: opus 11 sessions / $632 · sonnet 124 / $591 · haiku 16 / $8.

The orchestrator share came in above the 33% this skill records as its measured benchmark, from a single session, which is more than every top-tier worker combined. Per Step 6 that is not a curiosity: it says the counted compaction cadence did not fire often enough over a run this long.

Boundary findings

None. No worker reported a passage failing the untrusted-input boundary.

Lessons

decision verdict events distinct inputs programs flag
ci.verdict clear 35 31 1  
merge.step4 merge 7 5 1  
merge.step4 sync 18 15 1 systematic
merge.step4 fix-check 1 1 1  
merge.step4 wait 4 4 1  
merge.step4 ready 2 2 1  
merge.step4 review 4 4 1  
ci.verdict pending 194 72 1  

merge.step4 sync carries the systematic flag, but reads as a property of the fleet rather than a defect: with three workers merging in parallel a branch is nearly always behind main when it is first looked at. ci.verdict is the row that should be uncomfortable — 194 pending for 35 clear, roughly 5.5 polls per verdict, each one a turn that re-reads the whole context.

lessons:

  • category: tool-economy evidence: tally row ci.verdict pending 194 events / 72 distinct inputs against ci.verdict clear 35 candidate: Replace merge-pr’s poll-until-final loop with a single blocking wait on the run, so a CI verdict costs one turn instead of the ~5.5 it costs today.
  • category: steering evidence: usage_report — orchestrator 35% of $equiv from one session, above the 33% benchmark this skill records; the session compacted once across 27 supervised merges candidate: Make Step 4’s counted compaction cadence fire off the merge counter without the supervisor’s judgement, since a supervisor that never feels bloated never compacts.
  • category: automated-checks evidence: PR #463 — four independent review angles (A, B, C, G) each reported the owner_repo_dash="$owner-$repo" collision; the worker flipped the PR ready with it unfixed and its report line said READY candidate: A worker must not flip a PR ready while a CONFIRMED review finding is unaddressed — either fix it, or name it in the report as declined with a reason.
  • category: information-access evidence: #391 (commit 493104c) and #444 (2 commits), both rate-limit-killed workers whose work was invisible to gh pr view --json commits because it was never pushed candidate: On resume, the supervisor must inspect the killed worker’s worktree for unpushed commits before re-dispatching, because GitHub structurally cannot show them.
  • category: automated-checks evidence: PR #437 merged green — gh issue list --paginate is not a real flag, and main’s survey returned 0 rows for ~90 minutes because tests/survey/test.sh’s gh stub accepted any flag candidate: Every command stub in the suite should reject flags the real binary does not accept, so a fixture cannot pass by being more permissive than production.
  • category: coding-standards evidence: PR #449 (a non-quote-aware ${cmd%%<<*} reproducing the class #440 exists to close) and PR #465 (a false claim that gh pr create needs a touched-paths list), both caught by the worker’s own review axes candidate: Record in the review doctrine that the two axes’ main yield is defects the fix introduces, not defects in the code under repair — twice in one run that was the only catch.

What the run got wrong

  • I merged the --paginate regression myself. CI was green; the stub was the reason.
  • I told #391’s worker “no implementation was ever committed” when an unpushed commit sat in an orphaned worktree.
  • I argued plan-freshness.sh needed a rewritten parser, wrote it, and measured 65/70 against 70/70. Then a fourth false-STALE shape appeared and I withdrew the closure recommendation I had posted on #441. Both reversals came from running something, not from reasoning about it.