integrate-worker

name	integrate-worker
description	Integrate a worker agent's diff onto fresh trunk — apply, verify repros by hand, cross-check interacting mechanisms, full suite, ship + queue. Use after any worktree worker (fix cycle, hunt follow-up) returns a diff file.

Integrating a worker diff

Battle-tested protocol from the 2026-06 autonomous runs (~30 PRs integrated). Each step exists because skipping it once caused a real failure.

Steps

1. Fresh trunk, always. jj git fetch && jj new 'trunk()' then git reset --quiet — the colocated git index goes stale after jj working-copy switches and makes git apply --3way fail with "does not match index".

2. Apply with 3-way fallback. git apply --3way /tmp/<worker>.diff. Conflicts mean the worker's base predates trunk changes — resolve by intent, not by side: a worker editing code that trunk deleted usually means the worker's edit is subsumed (keep the deletion); append-append conflicts in test TOML files keep both case sets.

3. Re-verify the repros yourself. Build (bash scripts/rue build) and run every repro the worker claims to fix, on this checkout, by execution. A worker's "full suite green" was true in its worktree against its base — not here. Workers have shipped fixes verified against stale drafts of trunk; only integration-time re-verification catches that.

4. Write the cross-mechanism test. If this worker and another (landed or in-flight) built interacting mechanisms — e.g. one added per-field drop flags while another added drop-on-overwrite — write a test that exercises the combination NOW. The workers' own suites cannot see the seam; twice in one night, individually-green workers were jointly wrong (double-drops).

5. A tightening change compiles the example corpus BEFORE the suite. ./buck2 build //:cli-staged-programs builds all ten multi-file example programs. Any change that makes the compiler reject, trap, or error where it previously accepted needs this, and canaries are not enough: a RUE-1786 over-rejection passed examples/harbor and its own mutation check, and was caught only by examples/gazette — an outer borrow self receiver holding a loan while an inner call takes the same root inout, which is sound because a receiver is address-passed. Minutes against ~20 for the full suite, and it is exactly where over-rejection surfaces. A mutation check proves the fix DOES something; it says nothing about what the fix breaks.

6. Commit, then full suite + format. ./test.sh must exit 0; bash scripts/rue fmt. Commit before starting the suite and do not touch the checkout while it runs — switching branches mid-run once produced a fabricated E0583: file not found for module comptime citing a line that existed only on trunk, and cost an hour of chasing a bug that was not there. Read the banner: === TEST SUITE: PASSED === is now asserted rather than inferred, and a killed run says INTERRUPTED (SIG..) instead (RUE-1782). Do NOT judge by the tally alone — Buck can return cached results, so a run stopped partway can print a complete-looking Tests finished: ... Fail 0 and still not have finished. The banner is what distinguishes them.

7. Ship + queue. jj describe (subject + why + what was verified), jj git push -c @, gh pr create --repo rue-language/rue --base trunk --head steveklabnik:<bookmark> with one Fixes RUE-NN per line (a comma list only closes the first), gh pr merge <n> --auto. Use Part of RUE-NN when items remain — and remember "Part of" strands the issue In Progress (sweep afterwards per CLAUDE.md).

8. Queue bounces (DIRTY after a sibling merges): jj rebase -s <change> -d 'trunk()', resolve keeping BOTH PRs' semantics, re-run step 3's repros plus the sibling's, full suite, push the bookmark, re-arm --auto. Never force-push a branch that is actively queued without re-arming.

   On MERGE_CONFLICT, check trunk for the issue ID first: git log origin/trunk --format='%h %s' | grep -E '^[0-9a-f]+ [^:]*RUE-NN'. Anchor on the SUBJECT — plain --grep=RUE-NN searches the whole message and happily matches commits that merely cite the issue in their body, which is a false "it already landed" on exactly the question you are asking.

   A dequeue is not always a real conflict — RUE-1766 landed on trunk independently while a PR carrying it sat in the queue, and every conflicting file was that now-redundant half. There the resolution is to DROP the redundant half and keep only what is still unique to your branch, not to reconcile two spellings of one fix; then confirm the remaining half is still unlanded before assuming the PR has anything left to say. The same collision shows up in other disguises — mergeable_state: dirty on a PR you have not queued yet, or 3-way conflicts at apply time — so make the trunk check reflexive whenever a diff fights you in a file it should own.

Lane discipline (when dispatching parallel workers)

Check open PRs for each target issue ID before dispatching, not just Linear status. Linear tells you what is Done; it does not tell you what is in flight. In one night four issues were fixed in parallel by other agents while workers were already on them — RUE-1711 (across four separate PRs), RUE-1745, RUE-1646 and RUE-1766 — costing two branch rebuilds, one merge-queue dequeue, one duplicate issue filed, and most of a worker's output twice over. A search_pull_requests for the issue ID, or a scan of list_pull_requests --state open, costs one call per cluster and catches all of it. Re-check at integration time too: a fix can land while your worker is running.

Parallelize across disjoint crates, never across the same hot file. The conflict magnets — rue-air/src/sema/analysis.rs, the rue-error E-code/preview registry, and both codegen/*/cfg_lower.rs — should be serialized (land one, then dispatch the next in that lane). Most integration pain this project has hit (stale-base 3-way conflicts, the stabilization --preview miss) was two workers in the same hot file. A worker's cluster should name its crate-set; two clusters in the same hot lane go in consecutive cycles, not the same one.

Keeping the tree tidy

• Prefer Fixes RUE-NN over Part of whenever the PR actually closes the issue — Part of strands it In Progress and you must sweep it later.
• After a batch of merges, run scripts/jj-tidy — it deletes orphaned worktree-wf_*/cycle* branches, merged push-* bookmarks, and abandons dangling changes (safe: git protects checked-out branches; only unbookmarked non-@ heads are abandoned). Without it, jj log fills with dozens of dead heads within a session.
• For disk pressure, run scripts/rue storage clean (guard-gated, reclaims stale Buck outputs host-wide; it never deletes worktrees or source files) — never blanket rm -rf .claude/worktrees, which races running workers.

Worker-prompt invariants this protocol assumes

Workers were told: reproduce first, refutations are as valuable as fixes, file ownership is by region, output is git diff <base> HEAD to an OUTFILE plus a .meta summary. If integrating a diff produced under weaker rules, treat every claim as unverified.