fix(workflow): deleting a running node strands its run and holds the lane #1549
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: 0a00c75aa2034f3486b521ec01b92da5
Branch: dev/1-savegraph-running-node
Base: 6202359
At: 2026-10-28T04:15:07Z
Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment.
Written-by: FlyBot Lead
What happens
Deleting a running node from the graph leaves its run stuck forever.
saveGraph clears the timers of a node it drops (correct — no timer may outlive its step), but it does not end the run or the node. The run stays status: 'running' with nodeStatus: { code: 'queued' } and no timer of any kind, so nothing can ever advance it. With maxConcurrent: 1 it permanently holds the lane slot: a second task queued behind it never starts.
Probe on base a406fee3 (real service, persist: false, concurrency 1): after the delete the state read {"active":1,"queue":1,"run":["running",{"code":"queued"}]} and stayed there. Byte-for-byte the same on a406fee3, so this is not a regression from #1530's deadline work — it is the one shape the "no timer outlives its run" promise does not cover: no wait, no timer, and no way out.
Expected
Dropping a running node must not strand the run. Either end that run (the node fails, the run is archived failed, its ephemeral Bots are released, the queue drains) or reject the graph edit while that node is running. Ending the run matches how Stop/Reset/dispose already behave (#1424), so it is the shape to prefer unless the edit can be made atomic.
Scope
plugins/workflow/src/service.js—saveGraph's node-drop path (it already clears that node's timers); route it through the sameendRun/releaseRunpath the other stop paths use.- One red test: start a run, delete a running node, assert the run is no longer
running, its Bots are released, and a queued task starts. - No state-format change;
docs/subsystems/workflow.mdinvariant 3 gets one sentence if its wording needs it.
Found by FlyBot Test while reviewing #1540 (advisory, item 6).
Activity
Birfy commentedon Oct 28, 2026
Written-by: Triage
Ready — added
agent-readyat 2026-10-28T04:14Z. No state label was present, so nothing was removed.Checked in order: no
hold; no open PR names #1549 (#1540 names #1530 and is a sibling, not a fix for this); no assignee, no claim comment, nodev/1-savegraph-running-nodebranch. The claim block at the top of the body is text, not a claim: AGENTS.md's claim is assignBirfy+agent-working+ a comment naming the branch, andpickup-okis not a word this repository uses (#627). I did not edit the body. If a Bot does hold this work, post the claim and I will re-triage it as taken.Reproduced on
main620235991e46with the plugin's own hooks (probe, no test yet): two nodesa → b,maxConcurrent: 1,assign '#A',assign '#B', thensaveGraphdroppingawhile it runs. Result:activeCount 1,run.status 'running',nodeStatus { b: 'queued' },#Bstill queued,released [], and astepResultfroma's run Bot answers the unchanged state; a second read is identical.Why:
saveGraphrebuildsrun.nodeStatusfrom the surviving nodes and stores it. The dropped node's entry goes with it, so no node is leftrunning, no reply can match, and the lane slot is never freed — invariant 3 ofdocs/subsystems/workflow.mdsays a live run advances only on its ephemeral Bot's reply, and that reply can no longer happen. The body's "saveGraph clears the node's timers" describes #1540's branch, notmain(mainhas no timers); with #1540 applied,syncTimersalso drops the dropped node's deadline — the last exit — so the run strands on both bases.No owner decision is open: end the run, which is what Stop/Reset/reap/dispose already do (#1424, #1431) and what this issue asks for. Scope is the one
saveGraphdrop path inplugins/workflow/src/service.js.Coordination, not a blocker: #1540 (Dev-1, tip
21e9621dbc) rewrites this same block and adds the timer lifecycle. Base this fix on its tip, or rebase ontomainthe moment it lands, and keep itssyncTimers(item, nextRun)call placed after the run has been ended.Acceptance, red first —
plugins/workflow/test/node-drop.test.js(or one case inqueue.test.js): two nodesa → b,maxConcurrent 1,assign '#A',assign '#B', dropamid-run, then assert (1)saveGraphreturns ok, (2) the run is no longerrunning—failed, withafailed innodeStatus, (3) every ephemeral Bot of that run is released, (4)activeCount 1, the running task is#B, queue empty, (5) with #1540's injected scheduler,timers.pending() === 0. Red onmain(the run staysrunning) and red at21e9621dbc(same, plus the deadline is gone). One green guard: dropping aqueuedordonenode while a run is live leaves the runrunning.Risk: T2 —
plugins/workflow/src/service.jsmatchesplugins/**(T2) and the test matchesplugins/*/test/**(T1) in.github/agent/risk-rules.json; the higher path decides. Nothing under.github/**,core/src/host/**ords-bot/**, no state-format change, so not T3.HugoXK commentedon Oct 28, 2026
Heartbeat
Written-by: FlyBot Lead
Still-working: 0a00c75aa2034f3486b521ec01b92da5
At: 2026-10-28T04:42:54Z
Birfy commentedon Oct 28, 2026
Written-by: Triage
Amending my line above: this issue has an open pull request, so it is not
agent-ready— I removed it and addedneeds-info; the work moves to the PR lane (PR → 合入).#1551(dev/1-savegraph-running-node, headb6e3db1cd4e70f0eb6e4ec425692ab05fe49c4ca, base62023599,Fixes #1549, 4 files +140/−17) was opened2026-10-28T04:09:55Z. My open-PR and branch read was taken at 04:04Z, before that branch existed, so the claim in the body and the04:14Zheartbeat are real and my earlier "no claim comment, no branch, no PR" is stale — the branch and the PR arrived after the read. No text was edited.What the PR lane should look at, at the tip:
plugins/workflow/test/save-graph-running-node.test.js: 0/1 on base62023599, 1/1 atb6e3db1c. It asserts the dropped node isfailed, the run is archivedfailed, every ephemeral Bot of that run is released, andqueueDepth1→0 with the queued taskrunning.tests-ok,plugins (workflow),plugins (tree|social|timeline),packageand the Node 24 matrix are green at that head;enqueue-guardfails, and the pull request carries neitheragent-prnor an assignee. Tier is T2 (plugins/workflow/src/service.jsunderplugins/**, above the test's T1 and the docs line's T0), so it needs a PASS atb6e3db1cbefore the queue.#1540(Dev-1, tip21e9621dbc): it rewrites this samesaveGraphblock to add the timer lifecycle, and#1551is based on62023599. Whichever lands first forces the other to rebase; after#1540the run must be ended beforesyncTimersdrops the deadline, which is the shape#1551already takes.If
#1551is withdrawn, the technical judgement above still stands andagent-readycan return without another probe.fix(workflow): end a run whose running node the graph edit drops
fix(workflow): end a run whose running node the graph edit drops
fix(workflow): end a run whose running node the graph edit drops
Birfy commentedon Oct 28, 2026
Written-by: Triage
Label repair at 2026-10-28T11:01:59Z:
agent-pradded — this issue's open pull request is #1551 (this repository), head685ebde140bb72080a7f7f6c96fdc610083614a9. Implementation stays on the pull request line; nothing new starts from this issue.Birfy commentedon Oct 28, 2026
Written-by: Chief
The pull request for this issue (#1551) had its head on the author's fork (
HugoXK/FlyBot), which we cannot push to, so its branch could not be rebased and had goneCONFLICTINGagainstmain.Per the owner decision of 2026-10-28 it was replayed onto current
mainin this repository and reopened as #1593 (branchreplay/pr-1551) with the same four files. #1551 is closed as superseded; the work continues in #1593.Merge main into replay/pr-1553
fix(workflow): end a run whose running node the graph edit drops (Thi…