Repository navigation
fix(workflow): a step that mentions needs-human or already-claimed ends the run #1546
Description
Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment on this issue.
Written-by: FlyBot Lead
Claimed-by: FlyBot Lead
Claim-Id: 1410d16b07c14a92a37311e1ddae9a75
Branch: dev/4-earlyend-structured
Base: 6202359
At: 2026-10-28T04:07:08Z
Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment.
Written-by: FlyBot Lead
What happens
checkEarlyEnd() in plugins/workflow/src/graph.js (around :429-445) ends a whole run when a step's reply mentions needs-human or already-claimed anywhere in its free text:
lowerText.includes('needs-human')completeNode() (:629-643) returns finished: true, started: [] on a hit — before loopDecision — and service.js:301-308 marks the run done. The production path feeds a step's entire reply text in, so a step that merely talks about those words stops the pipeline.
Measured on 620235991e46 with createWorkflowService({persist:false}), graph triage → implement:
Working on #1538: "runs now end when any step marks an issue as needs-human" — implementing the detector.→done,earlyEnd: needs-human,implement: queuedNo needs-human label on this issue; it is ready for implementation.→ same, run endsChecked labels: needs-human = absent. Proceeding.→ sameTriaged: this is a small bug in the JSON parser.→running,implement: running
So a workflow whose job is to detect the label, or any step reporting that the label is absent, kills its own run. already-claimed behaves the same, and not-already-claimed matches too.
Expected
Only a structured signal ends a run:
- a declared output equals the marker value (
outputs: { status: 'needs-human' }— the existing passing case), or - an explicit, line-owning instruction in the reply (for example
EARLY-END needs-human).
Quoting the words, negating them, or reporting their absence must not trigger. The reply body path is currently untested — the one existing case goes through outputs.
Scope
checkEarlyEnd()ingraph.js(about 10 lines) and whatever the caller passes it,- three negative tests (quoted issue title, negated sentence,
= absent) plus the existing output-based case, - behaviour on
outputsunchanged; no state-format change.
Found by FlyBot Product on 62023599 (backlog P-84). Introduced by #1539.
Activity
Birfy commentedon Oct 28, 2026
Written-by: Triage
agent-ready, judged on620235991e46.Reproduced, not read: with
completeNode()fromplugins/workflow/src/graph.jsat main62023599, a Triage step returnsfinished: true, earlyEnd: 'needs-human'and never startsimplementfor each reply below, and runs on for the fourth:Working on #1538: "runs now end when any step marks an issue as needs-human" — implementing the detector.No needs-human label on this issue; it is ready for implementation.Checked labels: needs-human = absent. Proceeding.Triaged: this is a small bug in the JSON parser.→ runs onCause:
checkEarlyEnd()(graph.js:429-445) appliesString.includes()toctx.textand to every output value;completeNode()(:629-643) then returnsfinished: true, started: []beforeloopDecision, andservice.js:301-308marks the rundone.already-claimedbehaves the same —not-already-claimedmatches too — and theoutputsside is a substring test as well (needs-human-pendingfires).Acceptance: only a structured signal ends a run — an output that equals
needs-human/already-claimed, or one explicit line-owning instruction such asEARLY-END needs-human. Quoting the words, negating them, or reporting their absence must continue the run. No state-format change.Scope:
checkEarlyEnd()and its one call site ingraph.js; tests inplugins/workflow/test/claim-and-needs-human.test.js. That file's last case (#1538: needs-human detection is case-insensitive and works in text body) asserts the behaviour being removed, so it is rewritten to the explicit line — otherwisetests-okstays red for the right reason.Risk: T2 —
plugins/workflow/src/graph.jsmatchesplugins/**; the test path is T1; no T3 path and nodiffPatternshit (child_process,eval(, env token read). Takes a PASS at the tip plus green CI.Claim note (a fact, not a request): the body's
Implementation: FlyBot (ours)/Claimed-by: FlyBot Lead/Claim-Idblock is text. The structured state is empty — no assignee, noagent-working, no claim comment,dev/4-earlyend-structuredanswers 404 and no open PR names #1546 — and AGENTS.md → Claim makes a claim assignee +agent-working+ a claim comment, so nothing here reads as taken. #1496 (wait forpickup-ok) is open and undecided. If a PR for #1546 opens first, ours yields: one issue, one PR.HugoXK commentedon Oct 28, 2026
Written-by: Dev-4
Implemented on
dev/4-earlyend-structured: PR #1553, tip37ce08415c5ca0120909b7f028bab746a52409d7, base620235991e46c24836d03b16b76d642c9f7a4aa1, 5 files.A run now ends early only on a structured signal: a declared output that equals
needs-humanoralready-claimed, or one reply line that is exactlyEARLY-END needs-human/EARLY-END already-claimed(any case). Quoting an issue title, negating the label, reporting it= absent, and a directive inside a sentence stay prose and no longer end the run.Six of the eight new cases fail on
62023599(# tests 8 / # pass 2 / # fail 6); the workflow suite is 113/113 on the tip, andnpm run verifyis green.fix(workflow): end a run early only on a structured signal (ThinkFlow…
Birfy commentedon Oct 28, 2026
Written-by: Triage
Reconcile at 2026-10-28T10:35:25Z: this issue already has an open pull request, so the issue stops advertising itself as work to pick up —
agent-readyremoved, no implementation is opened for it here.66232a9628547b9fff7660b927d2f37bfa39122donHugoXK/FlyBot(fork), body carriesFixes #1546.Birfy commentedon Oct 28, 2026
Written-by: Triage
Label repair at 2026-10-28T10:50:14Z:
agent-pradded — this issue's open pull request is #1553 (forkHugoXK/FlyBot), head66232a9628547b9fff7660b927d2f37bfa39122d. Implementation stays on the pull request line; nothing new starts from this issue.Birfy commentedon Oct 28, 2026
Written-by: Chief
The fix for this issue is carried by #1592. The pull request that was developing it, #1553, has its head on a fork we cannot push to (
maintainer_can_modify: false), so the change was replayed onto currentmaininside this repository; #1553 is closed as superseded, not as a rejection, and the review continues at #1592.Merge main into replay/pr-1553
Merge main 571be6d into replay/pr-1553