Target: dodopayments/chimely (upstream, READ access) via the fork ravidsrk/chimely.
Source: tracker — 10 open issues (#34-#39, #55, #56, #58, #61). T0 2026-07-15T17:40Z.
Coordinator: manual Orca loop (no orchestration run). No merge conductor (no merge rights).
This is the run the mission was extracted from. It began as a clean-sweep invocation and surfaced,
in flight, why upstream contribution is a distinct mission: clean-sweep's definition of done is a
merged SHA, which is unreachable on a repo you do not control.
| Issue | Class | Terminal state |
|---|---|---|
| #34 | buildable | PR #81 (docs) open; parked on a 1-word imprecision, self-noted |
| #55 | already-has-PR | PR #84 (alt to #79) + assist comment on #79 |
| #56 | already-has-PR | PR #85 (alt to #76) + assist comment on #76 |
| #58 | already-has-PR | PR #83 (alt to #77) + assist comment on #77 |
| #61 | already-has-PR | PR #82 (alt to #72) + assist comment on #72 |
| #35, #36 | externally-covered | maintainer PRs #74/#73; no hollow comment |
| #37, #38, #39 | needs-human | product/schema forks named |
Deliverable: 5 upstream PRs (#81-#85) + 4 review-assist comments carrying 10 confirmed findings, each quoting the target PR's own diff. Merge left to the maintainers (the intended terminal).
- Two-denominator enumeration is load-bearing. The initial triage searched code but not the
upstream OPEN PRs; a parallel contributor already had PRs for 6 of 10 issues, discovered ~2h into
building. This is why
upstream-contributionmakes the per-issue PR search a FREEZE-gating step. - The contribution decision is real.
complement, not compete(assist),alternative(cross- linked PR), andstand-down(externally-covered) were all exercised, each logged as a taste gate. - Terminal is PR-open, not merged. Every unit closed on an open, internally-reviewed, etiquette-conformant PR or a quoted assist comment — never a merge the fleet cannot perform.
- The two-reviewer pipeline (codex + greptile) earned its cost. Each caught real defects the other passed (a focus-trap tabindex escape; a missing log-scrub boundary test). Both the fixed alternative PR and the finding on the parallel PR reached the maintainers.
Every PR drew a Greptile review after it opened. Treating "PR open" as terminal would have left that
feedback unanswered, so the run added a follow-up loop, now part of upstream-contribution.md: watch
each PR, triage every thread against the current head (one was already resolved by an earlier fix
round — answered, not re-fixed), fix the valid ones as fix rounds on the same branch, and answer every
thread. The FOLLOWED_UP ledger flag closes a unit only when no thread is left hanging.
A dodopayments maintainer pushed back on #84 the next day: a parallel reimplementation of an open PR is not a good use of anyone's time — "I'd rather not maintain two competing PRs for the same issue." All four alternative PRs (#82/#83/#84/#85) were closed in deference the same day, with the findings already preserved as assist comments and the branches kept for cherry-picking. The protocol change this feeds: opening an alternative PR is MAINTAINER-gated — offered inside the assist comment, opened only on invitation — not a self-serve taste gate as this run originally treated it.
- Codex workers emit no Orca heartbeats;
spawn_workerexit 3 is a false negative for them — verify by pane read. Cost two false "no-boot" diagnoses. - A builder pane survived its own
worker_doneand pushed an out-of-band commit mid-review. Close unit panes atworker_done, do not merely consume the message.
Evidence: PRs and comments live on dodopayments/chimely (#81-#85, comments on #72/#76/#77/#79). The
run's ledger, per-unit SHA-bound manifests, review artifacts, and integrity inventory were retained in
the fork worktree at close.