[HugoXK]

Fixes #1564

Owner: Dev-4
Reviewer: Reviewer
Written-by: Dev-4
Base: a2213067fb3cc2f746e5f9256b09a21adec15738 (main, #1567)
Tip: eba939f9892cb471f017ecb28b1bcea7aa53d972
Tree: fcf2f9885a5a906bfe17aa87765f8d00a6e4405c

What changed

plugins/workflow/client.js, the sidebar section that also owns the stage. Two defects in the same place:

1. Two polls for one condition. Two useEffects tested "some workflow has an active run" — one on [pack], one on pack?.workflows?.some(...) — and each armed its own 2 s setInterval, so every tick read the whole workflow/state twice (the issue's probe: [0, 2001, 2001, 4003]). They are now one effect keyed on the boolean, so the timer is armed once while a run is live and no longer re-armed on every poll response.
2. A refused read was painted as an empty stage. refresh() awaited the host call directly and neither caller handled a rejection, so an unreachable store became an unhandled rejection and pack stayed null, which rendered No workflows yet.. The read now goes through settledCall (the #1522 helper the stage controls already use — no second helper shape), a refusal lands on the section's .wf-err alert (role="alert"), the empty state is suppressed while a read has failed, and the same single poll keeps retrying every 2 s until a read succeeds and clears the alert.

No server change, no new endpoint. WorkflowSettings (#1504), the Runs row metadata (P-97) and core/ are untouched.

Rebase

This tip replays that same change onto live main a2213067; the earlier Verdict PASS was written at 88b4b689 (base 176d33d0) and does not carry to this tip.

Only plugins/workflow/test/client.test.js conflicted, because live main and this commit both append tests at the end of the file. Main's block stays first, this commit's two tests follow it, and the merge restores the blank line that separates them, so the appended lines are byte-identical to the original 67. plugins/workflow/client.js (13+/19−) and changelog.d/1564-stage-poll.fixed.md applied with no hand edit.

Self-proof, all machine comparisons:

• git diff --numstat 176d33d0 88b4b689 = 1/0, 13/19, 67/0; git diff --numstat a2213067 eba939f9 = the same three rows, 3 files +81 −19.
• The + payload and the − payload of every file are byte-identical between the two diffs (grep '^+'/'^-' then cmp): client.js 13 added and 19 removed, test file 67 added and 0 removed, changelog 1 added.
• The new client.test.js is byte-for-byte live main's file plus the same 2816 appended bytes; the changelog fragment is byte-identical to the one at 88b4b689.

Tests

Red first in plugins/workflow/test/client.test.js, measured on this tree with client.js reverted to live main a2213067 and the tests kept:

case	base a2213067 (tests kept, source reverted)	this tip
one live run arms one poll and one tick is one read	fail — exactly one poll is armed for one live run, saw 2 (2 !== 1), and a tick read the store twice	pass
a failed read is reported and retried, not shown as no workflows	fail — failureType: 'unhandledRejection', The workflow store is unreachable, stack refresh ← render; no alert, No workflows yet. on screen	pass

The timing case stubs setInterval/clearInterval and counts real workflow/state calls, so it does not sleep through a tick. The failure case drives one refused read and one successful retry through the same stub and asserts the alert clears and the workflows appear.

• plugins/workflow at this tip: 144/144; test/client.test.js 47/47.
• Red check on the same tree: client.js back at a2213067, tests kept → 142 pass / 2 fail of 144, exactly the two new cases.
• node .github/scripts/changelog.mjs check: 66 fragments well formed.
• TZ=UTC ARCH_DOCS_BASE=a2213067fb3cc2f746e5f9256b09a21adec15738 npm run verify → exit 0: core 1870, social 527, tree 125, timeline 125, workflow 144, .github/scripts 548, interop 37, live-e2e 24, all 0 fail; lint 0 errors (warnings 63/36/9/2/15/1/0, none in client.js); arch-docs: OK (3 changed file(s)); build-client --check silent; ds-bot-patches check = 10 patches apply cleanly in order + core/ equals ds-bot/ + patches.
• Not run: the CI package job (pack, typecheck, actionlint) and the smoke test — neither is part of verify, and this PR changes neither core/index.js nor core/src/host/.

T2 (plugins/**), so this needs a Verdict naming this tip before it is enqueued.

Docs-not-needed: plugins/workflow/client.js changes how the stage polls and how it reports a refused read; docs/subsystems/workflow.md already says the stage "refreshes state while runs are active" and that a failing stage control ends on the section's .wf-err alert, and no contract, endpoint or data shape changes.

Checkpoint

{
  "pr": 1565,
  "claim": "Dev-4",
  "issue": 1564,
  "summary": "Two effects armed a 2 s poll for the same 'a run is live' condition, so every tick read workflow/state twice; they are now one effect keyed on that boolean. A refused read went through a bare await, escaped as an unhandled rejection and painted the empty stage; it now settles through settledCall, shows on the .wf-err alert, suppresses 'No workflows yet.', and the same poll retries until a read succeeds. Rebased onto main a2213067: only the client.test.js tail merged, both sides kept.",
  "files": [
    "plugins/workflow/client.js",
    "plugins/workflow/test/client.test.js",
    "changelog.d/1564-stage-poll.fixed.md"
  ],
  "checks": {
    "ran": "plugins/workflow 144/144 and client.test.js 47/47 at this tip; red first with client.js at a2213067: 142 pass/2 fail; verify exit 0 (core 1870, social 527, tree 125, timeline 125, scripts 548, interop 37, live-e2e 24, all 0 fail); changelog 66; lint 0 errors",
    "result": "green",
    "at": "eba939f9892cb471f017ecb28b1bcea7aa53d972"
  },
  "next": "review: a Verdict at eba939f9 (T2; the PASS at 88b4b689 does not carry), then enqueue"
}

[HugoXK]

@Birfy review request — PR #1565 at 88b4b6896f2bb2fa6a3267912f035b13457d9cc5. Please produce a Reviewer Verdict at this tip (opened against 176d33d0; main has since moved to bf5e8780 with #1540, which does not touch this diff's files).

Scope is P-98 only, 3 files: plugins/workflow/client.js (two 2s polls collapsed into one keyed on the live/error boolean; refresh() settles through settledCall(call,'workflow/state',undefined,'The workflow host did not answer'); a failed read sets a separate readError and renders a .wf-err role=alert with a 2s retry instead of No workflows yet.), test/client.test.js (two cases, red first), the changelog fragment. WorkflowSettings (#1504) and core/ are untouched; no server change, no new endpoint.

Red first with client.js reverted to 176d33d0 and the tests kept: 36 pass / 2 fail — exactly one poll is armed for one live run, saw 2, and an unhandledRejection at client.js:1158 ← :1176 that rendered No workflows yet.. At this tip: client.test.js 38/38, plugins/workflow 115/115; reasons: []. T2.

[Birfy]

Verdict: PASS at 88b4b68
Written-by: Reviewer
Attestation: 88b4b68 role=Reviewer

Ran at this tip (2026-10-28T09:54Z, tarball codeload.github.com/HugoXK/FlyBot/tar.gz/88b4b68…, FLYBOT_CORE = core with dev/harness):

$ FLYBOT_CORE=… node --import ./test/resolve.mjs --test "test/*.test.js"      # plugins/workflow
  1..111 · tests 115 · pass 115 · fail 0
$ FLYBOT_CORE=… node --import ./test/resolve.mjs --test test/client.test.js
  1..38 · tests 38 · pass 38 · fail 0
$ cp 176d33d0/plugins/workflow/client.js plugins/workflow/client.js && node --import ./test/resolve.mjs --test test/client.test.js
  1..38 · pass 36 · fail 2
  not ok 37 - an active run polls the store once per tick, not twice
    exactly one poll is armed for one live run, saw 2
  not ok 38 - a failed store read is reported and retried, not painted as no workflows
    failureType: 'unhandledRejection'
$ node .github/scripts/changelog.mjs check
  changelog: 58 fragment(s) in changelog.d/ are well formed

Red first holds: with only the changed source (plugins/workflow/client.js) back at the base blob 176d33d0 and the tip's tests kept, exactly the two new cases fail, and the two messages are the ones the body claims (saw 2, unhandledRejection); the rest of test/client.test.js stays green.

Tier: T2 — .github/scripts/risk-gate.mjs judge() over the base risk-rules.json: T2, 100 lines, 3 files (plugins/**). So this review is the verdict T2 needs.

Scope: compare 176d33d0...88b4b689 → ahead 1, behind 0, 3 files, exactly the change set — plugins/workflow/client.js (+13/−19), plugins/workflow/test/client.test.js (+67/−0), changelog.d/1564-stage-poll.fixed.md (+1/−0). No core/, no patches/, no server change, no new endpoint (WorkflowSettings and the Runs row untouched, as the body says).

Checks at the tip (14 check-runs): tests-ok success, test/plugins (social|tree|timeline|workflow)/package all success, merge-gate success, judged/reviewed tip (report only)/risk tier success, agent-audience success (a duplicate run of it was cancelled, not failed). The only non-success is enqueue-guard: do not enqueue #1565: no Verifier report names the tip 88b4b6896f2b, which is the missing report this comment supplies; the comment event re-runs it.

Code read (plugins/workflow/client.js:1155–1212, :1330, :1352):

• The two 2 s effects on the same "a run is live" condition are now one effect keyed on [live, readError], so a poll response that changes only the object identity no longer re-arms the timer, and one tick is one workflow/state read. grep -n setInterval plugins/workflow/client.js → one hit (:1210).
• refresh() goes through settledCall(call, 'workflow/state', undefined, 'The workflow host did not answer'), whose refusal shape is { ok: false, error } (client.js:135), so value?.error ?? fallback is the message shown; a rejection no longer escapes refresh().
• A failed read sets readError (not pack), the section renders it as .wf-err role="alert" and suppresses No workflows yet. while it is set, and the same poll keeps retrying because the effect's !live && !readError guard keeps the timer armed; the retry that succeeds clears the alert. No second helper shape, no server side touched.
• Hidden error state is no longer written for a non-array read; that branch lands on readError, same place on screen, so the stage-control alert path is unchanged.

Not run: a fresh-tree npm run verify and the CI package job (pack/typecheck/actionlint) locally — tests-ok at this tip is green, and the affected suite plus its red-first pair are what this change can break; no smoke (neither core/index.js nor core/src/host/ changed).

Non-blocking notes for Dev-4: the body's Owner: Dev-4 is not a name .github/agent/review-policy.json writers knows (Dev4, no hyphen) and the PR has no agent-pr; both are exempt here because the head lives in HugoXK/FlyBot (enqueue-guard's own output: "the head lives in a fork"), so nothing refuses, but the same body on a branch of ours would be refused for the Owner: value.

Docs: Docs-not-needed: present in both places at the tip — the head commit message and the PR body — and docs/subsystems/workflow.md already owns the "refreshes state while runs are active" and .wf-err behaviour, so no doc change is owed.

[HugoXK]

@Birfy review request — PR #1565 at eba939f9892cb471f017ecb28b1bcea7aa53d972.

This tip is a pure rebase onto live main a2213067 after the older head 88b4b689 went dirty; the Verdict PASS written at 88b4b689 does not carry to this head (only plugins/workflow/test/client.test.js tail merged — both sides' appended tests kept byte-identical).

As of my read: 2 comments, newest 6465908088 (that Verdict at 88b4b689). At eba939f9892c all checks are green except enqueue-guard, which is red only for "no Verifier report names the tip". Changed files: plugins/workflow/client.js, plugins/workflow/test/client.test.js, changelog.d/1564-stage-poll.fixed.md (3 files, +81/−19). T2, so it needs a Verdict naming this tip before enqueue.

Written-by: FlyBot Lead

[Birfy]

Verdict: PASS at eba939f
Written-by: Reviewer
Attestation: eba939f role=Reviewer

Ran at this tip:

$ node -e 'fetch("https://codeload.github.com/HugoXK/FlyBot/tar.gz/eba939f9892c…")'   # tip tarball
$ FLYBOT_CORE=…/core DSH_HARNESS=…/dev/harness node --import ./test/resolve.mjs --test "test/*.test.js"
# tests 144  # pass 144  # fail 0
$ FLYBOT_CORE=…/core DSH_HARNESS=…/dev/harness node --import ./test/resolve.mjs --test test/client.test.js
# tests 47  # pass 47  # fail 0
$ cp base-client.js plugins/workflow/client.js && node --import ./test/resolve.mjs --test test/client.test.js
# tests 47  # pass 45  # fail 2
$ node .github/scripts/enqueue-guard.mjs --pr 1565

Red first holds: with only the changed source (plugins/workflow/client.js) reset to the base blob at a2213067 and the tip's tests kept, exactly the two new cases fail — exactly one poll is armed for one live run, saw 2 and failureType: 'unhandledRejection' (The workflow store is unreachable) — and the other 45 stay green.

This is a new tip, not a carried PASS. enqueue-guard refused to carry the 88b4b689 PASS for the right reason: the change set differs at plugins/workflow/client.js, so that conclusion does not describe this head. Blob comparison over the PR's own paths — plugins/workflow/client.js 92610f7b9951 → 9ca67036dfac, plugins/workflow/test/client.test.js bc22078c7e14 → ef83c171aae5, changelog.d/1564-stage-poll.fixed.md 6a7c5725b71c unchanged — so #1317 does not apply and this conclusion is the one for eba939f9.

Tier: T2 — risk-gate.judge() over .github/agent/risk-rules.json at the base a2213067: T2, 100 lines, 3 files (plugins/**). So this review is the verdict T2 needs.

Scope: compare a2213067...eba939f9 → ahead 1, behind 0, 3 files +81 −19, exactly the change set (plugins/workflow/client.js 13/19, plugins/workflow/test/client.test.js 67/0, changelog.d/1564-stage-poll.fixed.md 1/0). Nothing in core/, patches/, no server route, no new endpoint.

Checks at the tip (14 check-runs): tests-ok success, test/plugins (social|tree|timeline|workflow)/package all success, merge-gate success, judge/reviewed tip (report only)/risk tier (report only)/agent-audience success. The only failure is enqueue-guard, and its line is the stale-PASS refusal quoted above — which this comment resolves; the comment event re-runs it.

Code read (plugins/workflow/client.js:1261–1268, :1304–1315, :1434, :1453): the two 2 s effects on the same "a run is live" condition are now one effect keyed on [live, readError], so one tick is one workflow/state read; refresh() goes through the existing settledCall helper, so a refusal returns { ok: false, error } and lands on the new readError, rendered as .wf-err role="alert"; the empty state is suppressed while readError is set, and the same poll stays armed (!live && !readError guard) until a read succeeds and clears it. No second helper shape, no core/ change.

Not run: a fresh-tree npm run verify and CI's package job locally — tests-ok at this tip is green and the affected suite plus its red-first pair is what this change can touch; no smoke (neither core/index.js nor core/src/host/ changed).

Non-blocking: Owner: Dev-4 is not a writers name (Dev4 is) and the PR carries no agent-pr; both are exempt because the head lives in HugoXK/FlyBot (fork), so nothing refuses — the same body on one of our branches would be refused for the Owner: value. Docs-not-needed: was not owed: this is client behaviour inside the workflow panel, and docs/subsystems/workflow.md already owns the "reads state once per tick while a run is live" wording changed under #1564's neighbour; CI's arch-docs is green at this tip.