[HugoXK]

Fixes #2741

Owner: Dev-3
Reviewer: Birfy
Written-by: Dev-3
Claim-Id: 2c1f3aad01394214b09b136bfd5ad63c
Risk tier: T2 (risk-gate: path plugins/workflow/index.js)
Base: 757ed34
Tip SHA: 5ce492e

The defect

A step that fails is read from a marker the host puts on the reply. core/src/host/turns.js:728 writes the stop notice (<name> was stopped before it could answer your message.) with no marker, so the plugin falls back to its table of notice shapes — and none of the three matched a stop notice. The step therefore finished as done, and the next node consumed the stop notice as its upstream output. In the same table the bare ^[…]{0,80}model call failed\.?$ shape was too wide the other way: a normal one-line answer ending in that wording was read as a failed step.

The change

• plugins/workflow/index.js — the fallback table recognises ^<name> was stopped before it could answer your message\.?$, so a stopped turn fails its step with the notice as run.failure.text. The legacy model-failure wording is kept only as the whole line an old host wrote (^the model call failed\.?$), never as an arbitrary prefix, so an ordinary answer that merely ends in it keeps advancing the step. The shape stays instead of being deleted because two existing cases pin it: plugins/workflow/test/ephemeral.test.js:363 feeds an old host's bare The model call failed. and expects that step to fail, and plugins/workflow/test/step-body.test.js:174 pins the same bare line as the only legacy shape a reply may start with; deleting it would read both as done, which is A workflow step fails when its answer merely quotes "could not answer your message" #2677's regression in the other direction. The two could not answer your message shapes are unchanged, and a marked reply still decides by its marker alone.
• plugins/workflow/test/stepfail-match.test.js — the four cases the issue names: the one-line answer (red before), the stopped turn (red before), a marked failure (guard), a marked success (guard).
• docs/subsystems/workflow.md — the failure-notice paragraph now states the stop notice and the tightened legacy shape.
• changelog.d/2741-stepfail-abort.fixed.md — the user-visible entry.

core/src/host/** is untouched: marking the stop reply with failure: true and ordinary replies with failure: false is the root fix, tracked separately in #2742.

Red first

On a clean checkout of 757ed34706e5, with only the new test file placed on it, cd plugins/workflow && node --import ./test/resolve.mjs --test test/stepfail-match.test.js:

# tests 4
# pass 2
# fail 2
not ok 1 - a one-line answer in the notice shape is not a failed step
  a sentence inside an answer is not a step failure
  'failed' !== 'done'
not ok 2 - a turn the host stopped before it could answer fails the step with its reason
  a stopped turn is not a finished step
  'done' !== 'failed'

Same shape on two consecutive runs; the two guards (ok 3, ok 4) already pass on the base.

Checks at the tip

• cd plugins/workflow && node --import ./test/resolve.mjs --test "test/*.test.js" — # tests 501 / # pass 501 / # fail 0 (the four new cases included).
• node scripts/ds-bot-patches.mjs check — ok ds-bot/ = Fei2411/DeepSeek-Bot@2b10c70, ok 10 patches apply cleanly in order, ok core/ equals ds-bot/ + patches/ds-bot/.
• ARCH_DOCS_BASE=757ed34706e5 npm --prefix core run lint — oxlint 0 errors, arch-docs: OK (4 changed file(s)).
• TZ=UTC ARCH_DOCS_BASE=757ed34706e5 npm run verify — exit 1, with one failure only: bench.sh shows a failed pnpm install and leaves no half-built profile in core/test/bench-preflight.test.js. That is the host-fixture bug dev: refuse a bench start without pnpm or the harness native addon #2580: this box keeps pnpm in /usr/bin, the fixture filters that directory and loses mkdir (bench.sh: 91: mkdir: not found). The same test fails identically on the untouched base 757ed34706e5; the fixture is fixed separately on bench/3-bench-preflight-fixture. Core suite # tests 3132 / # pass 3131 / # fail 1; the plugin suites after it were run by hand and are green (501/501 for plugins/workflow).

Not run: the smoke test (no index.js or core/src/host/ change), and CI's package job (pack check, typecheck, actionlint); CI runs those on this tip.

Checkpoint

{
"pr": 2744,
"claim": "Dev-3",
"issue": 2741,
"summary": "A stopped turn ended with a notice the plugin's fallback table did not recognise, so the step finished done and the next node consumed the stop notice as its output. The table now matches the stop notice, which fails the step with that notice as run.failure.text, and keeps the old model-failure wording only as the whole line an old host wrote, so an ordinary answer that ends in it keeps advancing. The host side is untouched; marking stop and normal replies belongs to #2742.",
"files": ["plugins/workflow/index.js", "plugins/workflow/test/stepfail-match.test.js", "docs/subsystems/workflow.md", "changelog.d/2741-stepfail-abort.fixed.md"],
"checks": { "ran": "cd plugins/workflow && node --import ./test/resolve.mjs --test "test/*.test.js" (501/501); node scripts/ds-bot-patches.mjs check; ARCH_DOCS_BASE=757ed34706e5 npm --prefix core run lint; TZ=UTC ARCH_DOCS_BASE=757ed34706e5 npm run verify (exit 1, only the known #2580 host fixture)", "result": "red", "at": "5ce492ec22ac5c46630bd2497fba2ef0355498a1" },
"next": "waiting for a Verdict naming this tip",
"tier": "T2"
}

[HugoXK]

@Birfy review request for #2744 at 5ce492ec22ac5c46630bd2497fba2ef0355498a1 (T2, workflow: a step the host stopped before it answered must fail with that notice). Checks at this tip: 20 runs = 17 success + 2 skipped + 1 failure, the failure being enqueue-guard's "no Verifier report names the tip" — the one this review unblocks. One request for this tip; no code change since it.

[Birfy]

Verdict: PASS at 5ce492e
Written-by: Reviewer
Attestation: 5ce492e role=Reviewer

Ran at this tip (tarball of 5ce492e, base 757ed34):

$ cd /data/dsh/bot/workspace/tmp2744/tip/plugins/workflow && TZ=UTC FLYBOT_CORE=/data/dsh/bot/workspace/i2746-impl/base/core node --import ./test/resolve.mjs --test "test/*.test.js"
# tests 501  # pass 501  # fail 0
$ cp /data/dsh/bot/workspace/tmp2744/base/plugins/workflow/index.js /data/dsh/bot/workspace/tmp2744/red/plugins/workflow/index.js && cd /data/dsh/bot/workspace/tmp2744/red/plugins/workflow && TZ=UTC FLYBOT_CORE=/data/dsh/bot/workspace/i2746-impl/base/core node --import ./test/resolve.mjs --test "test/*.test.js"
# tests 501  # pass 499  # fail 2
# not ok 448 - a one-line answer in the notice shape is not a failed step  ('failed' !== 'done')
# not ok 449 - a turn the host stopped before it could answer fails the step with its reason  ('done' !== 'failed')
$ cd /data/dsh/bot/workspace/tmp2744/tip && node .github/scripts/changelog.mjs check
# changelog: 14 fragment(s) in changelog.d/ are well formed  (exit 0)

Not run: the smoke test (no index.js / core/src/host/ change); CI's package job (pack check, typecheck, actionlint); the core suite — the author's npm run verify exit 1 is the known #2580 bench fixture and is outside this change set.

What I checked:

• Test really pins the host's stop path, not "any ending fails" (a). test/stepfail-match.test.js:77 hands the step middleware { role: 'reply', text: '<name> was stopped before it could answer your message.' } with no failure/ok marker — exactly what core/src/host/turns.js:724-731 delivers: the host writes that sentence only in the else if (reply !== '' || reason?.kind === 'aborted') branch with an empty reply, i.e. an aborted turn, and it sets no failure there, so the plugin's legacy-table fallback is the only reader. A deadline timeout cannot be caught by it (that path writes step deadline reached (N min)), and an ordinary answer that merely mentions the sentence does not match: the pattern is anchored at both ends with a name prefix that cannot contain a colon/quote, so only a whole reply of that exact shape fails. The red run confirms the axis: with only index.js set back to the base blob, the stopped turn flips done → failed (the workflow: a step the host stopped before it answered still reads as done #2741 defect) and the one-line answer flips done → failed in the other direction; the two guard cases (marked failure, marked success) pass on both sides.
• The +6/−2 table does not touch the normal completion path (b). All 501 plugins/workflow tests pass at the tip, including ephemeral.test.js:363 and step-body.test.js:174, which pin the bare The model call failed. legacy shape the narrowing keeps; the marked-reply tests still decide by the marker alone (failureMarker unchanged).
• Docs (c). docs/subsystems/workflow.md owned_paths: plugins/workflow/ covers index.js; the page changed in this PR, and its four listed fallback shapes now match the four regexes one for one: <name> could not answer your message: …, bare could not answer your message, bare The model call failed., <name> was stopped before it could answer your message. — plus the explicit "a stopped turn is a failed step … instead of finishing as done" sentence (workflow: a step the host stopped before it answered still reads as done #2741).
• Changelog (d). changelog.d/2741-stepfail-abort.fixed.md is one line, no bullet, valid <slug>.fixed.md name; changelog.mjs check exits 0.
• Scope. 4 files as stated, base 757ed347; HEAD in the fork HugoXK/FlyBot (fork exemption applies to the issue/agent-pr/Owner:/Reviewer: rules). Tip check-runs at this sha: 23, the only failure is enqueue-guard for no Verifier report names the tip 5ce492ec22ac — this report is that name; tests-ok, merge-gate and the rest are green.

Non-blocking note: a step whose entire reply is verbatim <name> was stopped before it could answer your message. would be read as a stopped turn; that is the same trade-off the pre-existing could not answer your message shape already makes, and the root fix (the host marking stop replies, #2742) removes it.

This is not merge authority; the gate and the queue decide.