[rawphp]

Closes #48

Contract

• done: unassignedClaims skips issues the pass knows runs for and issues it already judged not the factory's at the same updatedAt; a history read that fails for that issue alone is reported for that issue and the scan goes on; the pass prints one error line per such issue; an issue that is the factory's is still moved to factory:proposed as before.
• out: no memo of "ours" judgements; no change to CLAIM_GRACE_MS or releaseClaim; no new files.
• risk: none
• test: vp test run apps/server/src/factory/tracker/github.test.ts apps/server/src/factory/pass.test.ts apps/server/src/factory/index.test.ts → pass (325 tests)

Design decisions settled in this run:

• The memo key is repo, issue, state label and updatedAt (the state label is added to the issue's key so an issue carrying both labels is judged per label).
• One ClaimMemo is built in liveWorld, which is built once per process, so a judgement lasts for the life of the factory process.
• Review finding: a GitHubRateLimited, GitHubTimedOut or GitHubSpawnError on a history read would fail every following call too, so it ends the repo's scan as a failing list call does (one "cannot list the unassigned claims" line), instead of spending one refused call per remaining issue every pass. Only GitHubFailed, GitHubParse and GitHubOutputTooLarge are isolated per issue. This narrows "a history read that fails" in criterion 3 to the failures that are the issue's own; the budget protection the issue asks for is why.

Acceptance criteria

• No history read for a known issue — github.test.ts "unassignedClaims reads no history for an issue the pass knows a run for"; pass.test.ts "the issues with runs are handed to the scan so their history is not read"
• Judgement remembered while updatedAt is unchanged, one history call per issue over two scans — github.test.ts "unassignedClaims remembers an issue judged not the factory's while updatedAt is unchanged"; index.test.ts "a judgement that an issue is not the factory's lasts across passes of one world"
• A failing history read is that issue's error only; the pass prints one error line naming it — github.test.ts "unassignedClaims: a history read that fails is an error for that issue only" and "… an unparseable history is that issue's error and the next is still read"; index.test.ts "a history read that fails is reported to the pass with the issue and the message"; pass.test.ts "an issue whose label history cannot be read is one error line and the others are still moved"
• Behaviour for the factory's own cut-short claim unchanged — existing pass.test.ts "a triage or building issue with no run and no assignee is moved to proposed" and index.test.ts "a cut-short claim is commented on, then moved to proposed" still pass; all 952 tests under apps/server/src/factory pass

Gate

• vp test run apps/server/src/factory → exit 0 (31 files, 952 tests)
• vp run --filter t3 typecheck → exit 0

Models

• maker: claude-sonnet-5-5 (no models block in .dw/config.yml; roster default)
• checker: claude-opus-5-5

Review

• cycles: 2
• fixed: cycle 1 (correctness + reliability, one shared action): a rate-limited or timed-out history read no longer continues to the next issue; it ends the scan. Cycle 2 (testing, 12 mutants, 2 survivors): added index.test.ts cases proving the memo lives for the world and that unread reaches the pass; also the two nits (an unparseable history case, a run in another repo left out of known).
• leftovers: none

Lessons

• Read the branch name from git, not from the launch snapshot. Symptom: the first push failed with "src refspec does not match any". Cause: the harness's git status snapshot named a different branch from the one the worktree was on. Default: run git branch --show-current at Isolate and use that name on every push.

• Never cd into the skill directory to read a reference. Symptom: the session's working directory flipped to the skill checkout and back, and a later compound command ran from the wrong place. Default: read reference files by absolute path with cat; every product command starts with cd <worktree>.

• Pass --repo <owner/repo> to gh pr create in a fork. Symptom: the draft PR opened on the upstream parent (pingdotgg/t3code#20137, closed within a minute) instead of the watched repo. Cause: a checkout with an upstream remote makes gh default to the parent repository. Default: always pass --repo with the issue's repo and --head with the worktree's own branch name, then check the URL's owner before linking.

Commits

• aa444d356 test(factory): cut-short scan skips known issues, remembers judgements, isolates failures — Make slice (red)
• a8fc4b268 fix(factory): cut-short scan reads no history for known or remembered issues — Make slice (green)
• bc58f39a6 test(factory): a rate-limited or timed-out history read ends the cut-short scan — Check cycle 1 (red)
• ba10b6f22 fix(factory): only a call's own failure is isolated in the cut-short scan — Check cycle 1 (green)
• ac95256fd test(factory): cover the cut-short scan's wiring and the call-specific failures — Check cycle 2 (test-only, no red phase: behaviour already worked)

Files

• apps/server/src/factory/tracker/github.ts — ClaimMemo, ClaimScan, isCallSpecific; unassignedClaims takes known and memo, isolates an issue's own failure
• apps/server/src/factory/pass.ts — ClaimScanResult, Known.issues; the pass hands known issues to the scan and prints one error line per unread issue
• apps/server/src/factory/index.ts — one ClaimMemo per liveWorld; maps tracker errors to messages for the pass
• apps/server/src/factory/tracker/github.test.ts, pass.test.ts, index.test.ts — the cases above

Labels: all seven factory:* labels already existed; none created. The worktree was fast-forwarded from e4bcce9 to origin/main (9e80956) at Isolate.

[rawphp]

Factory gate: passed (attempt 1, run by the factory in a fresh checkout of the pull request's head ac95256 merged with the base branch tip 9e80956)

Command	Exit	Time
factory/scripts/check.sh	0	37.9 s

[rawphp]

Factory gate: passed again (attempt 2, run by the factory in a fresh checkout of the pull request's head 8e81ea2 merged with the base branch tip 11492a7)

Command	Exit	Time
factory/scripts/check.sh	0	44.7 s