A failed workflow run never reaches the Inbox: it is visible only inside its own graph (and publicRun has no finishedAt or current step) #1706

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: b686bbd3a15447f7afce80aef62d5eaf
Branch: dev/2-workflow-failure-inbox
Base: 742e06c
At: 2026-10-29T14:44:04Z

Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment.

What happens today

A workflow run that fails writes its failure only into the workflow plugin's own state, and nothing surfaces it where the Owner looks:

• plugins/workflow/src/service.js:642-657 (endRun) sets run.status='failed', run.updatedAt and run.failure={code, at, nodeName?, text?}; a step failure also sets next.failure (:500-506).
• One line lands in that graph's messages: kind:'system', tone:'bad' (push() :273-280), serialized by write() :245-251.
• No notification: notify / deliver( hits inside plugins/workflow/src/ = 0.
• core/src/client/inbox.js: grep -c -i workflow = 0 — the Inbox never mentions a run.
• No cross-graph run list anywhere (grep = 0). The sidebar live dot (plugins/workflow/client.js:2034), sidebarPreview (:357-363) and the Settings row preview (service.js:386-389) only describe runs that are running; they never show a failure.
• The only place a failure is readable is Settings → Workflow → that graph → Runs tab (plugins/workflow/client.js:612-618, :1290, failureText :250-257, fed by historyRow :690-700).

So the Owner asking "did last night's card fail, and where" must already know which graph it ran in: 4 steps plus prior knowledge.

Also missing in the payload

publicRun (service.js:89-107) returns status, startedAt, updatedAt, failure? …, but no finishedAt and no explicit current step: steps is an integer count (blankRun :122-123), and the current step has to be derived from running + nodeStatus. A read-only consumer (Inbox) therefore cannot say "failed at 03:12 on node X" without guessing.

Expected

1. A failed run appears in the Inbox as a notification row: graph name, failing node name, reason (failure.nodeName / failure.text / failure.code), when (finishedAt, else updatedAt), and the task it came from when run.task.id exists.
2. publicRun gains two additive fields: finishedAt (null while running) and currentStep ({ id, name } or null) — nothing existing is renamed or removed.
3. No second answer entry: the row reuses STATUSES / InboxCard / inboxAnswer; a failed-run row does not grow its own button set.

Red-first (must be shown on live main, before any fix)

Seed a state with one status:'failed' run carrying a failure, then on live main:

• inboxModel(...) (or the equivalent Inbox card list) contains 0 cards mentioning that run — this is the red;
• after the fix: exactly 1 card, naming graph + node + reason;
• a run that is still running must not appear;
• a failed run whose card is gone (run.task.id absent) must still appear (it is the run that failed, not a task row).

Scope

• core/src/client/inbox.js: add the workflow source from the existing payload (workflow/state → present(): per-graph runs/history/messages, service.js:296-330); reuse the existing statuses and card renderer; the answer path stays inboxAnswer.
• core/src/client/index.js: one action workflowState: () => call('workflow/state') (precedent: core/src/client/workflows.js:47 readWorkflows(call), 8s timeout).
• plugins/workflow/src/service.js: add the two fields to publicRun only (read-only additive).
• Not in scope: any change to the run's write path, any new endpoint, any new store, core/src/shared/workflow-run.js / workflow-store.js, and merging a run into the core task identity (rows carry data-run, never data-task).

Acceptance

1. Red-first exactly as above on live main; the seeded running run stays out.
2. Inbox row names graph + node + reason + time; clicking it opens that graph read-only (no new answer controls).
3. publicRun diff is additive only; every existing consumer of present() still sees the same keys plus the two new ones.
4. TZ=UTC ARCH_DOCS_BASE=<live main> npm run verify exit 0 (report the measured counts); node scripts/ds-bot-patches.mjs check green if core/ mirrors change.

Evidence re-read at live main d5347032 by FlyBot Product (scheduled audit) and re-checked by FlyBot Lead on the same commit.