fix(chat): show why a settled subagent failed in its detail panel

[Yi-111-a]

Summary

When a delegate ends failed, timed_out, or aborted, its detail panel shows a step stream that just stops. The status capsule says Failed, the elapsed time is there, and nothing says why. A delegate that dies before emitting any message row has no other carrier for its reason, so the panel is a list of steps and no explanation.

Reported as #161.

This renders the reason as an error card at the foot of the dock.

Why the reason was missing

The reason already exists at runtime — SubagentRunResult.error: { code, message } — and delegationSummary in packages/agent-runtime/src/runtime.ts already spreads it onto the delegation roster entry. It simply had no reader on this surface:

• Task returns the moment a delegate starts (ADR 0089), so its own tool result is pinned at running forever, no matter how the delegate ended. SubagentDetail renders that row.
• SubagentRun (the delegate's own row collection) carries only agentName + items — no error field.
• The roster entries are attached to the lifecycle rows (TaskWait / TaskList / TaskStop), which are deliberately not topology nodes.

So this is a presentation gap, not a runtime gap: the error was already being delivered to the renderer and then discarded. No runtime, IPC, storage, or tool-result change is needed.

What changed

• collectDelegationFailures(items) in src/lib/subagent-topology.ts walks the lifecycle rows' details.delegations[] and details.stopped[] and returns delegationId → { code, message }. It is deliberately the same shape as the existing collectDelegationStatuses (same iteration, same last-write-wins rule, same "the Task row is not a carrier" exclusion), so the status and the reason can never disagree about which row wins.
• SubagentDetail gains an optional delegationFailures prop and renders a SubagentFailureCard at the foot of the panel. SubagentPanel computes the map from selected.turnActivityItems — the same array it already feeds to collectDelegationStatuses and collectDelegationTimings.
• The card reuses the transcript's existing error-card visual language and vocabulary rather than inventing a subagent-specific one: error.code uses the same errors.<code> lookup (AssistantErrorMessage), the summary falls back to the localized chat.subagentStatus.* outcome when a code is not registered, and the raw provider message sits behind a disclosure with a copy control.

No new i18n keys. All eight locales already carry chat.subagentStatus.* (all eight outcomes), chat.showErrorDetails / chat.hideErrorDetails, and chat.copyErrorDetails — the card only reuses them.

Two details worth calling out, since both would be easy to get wrong:

• The card follows a non-success terminal outcome, not the error field. A delegation can carry an error and still be completed, and a still-running delegate has no terminal outcome to explain. The gate is failure && outcome !== "completed" && outcome !== "running".
• The disclosure control is in the heading, not inside the collapsed region. The parent error card puts its toggle in message-error-actions for exactly this reason; hiding the details must not take away the control that brings them back.

Documentation

• spec/04-ux/08-component-spec.md §5.7 gains the failure card, the summary fallback, and the terminal-outcome gate.
• E2E-198 (the subagent dock scenario) gains step 6 and the matching expectation, and its status records the new unit coverage.
• spec/08-meta/decisions-log.md records D382, mirrored in docs/zh-CN/.

Note on the zh-CN mirrors: docs/zh-CN/spec/04-ux/08-component-spec.md is missing the entire §5.7, and docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md does not have E2E-198 at all — both predate this change. I did not translate a whole pre-existing missing section here, since that would be an unrelated change (AGENTS.md R2); happy to do it as a separate PR if useful. The decisions log mirror is current, so its entry is synchronized.

Validation

• node --test test/subagent-topology.test.mjs test/subagent-panel.test.mjs — 23 passed, 0 failed.
  • 5 new behavioural tests in subagent-topology.test.mjs: a failure read from a TaskWait roster entry, one read from TaskStop's stopped[], last-write-wins across a repeated row, entries that report no error at all (status only, error: {}, whitespace-only message, missing delegationId), and the Task row that must never carry one.
  • 1 new source-contract test in subagent-panel.test.mjs, following that file's existing pattern for this panel: the prop is wired, the gate is the terminal outcome, and the toggle precedes the region it collapses.
• node scripts/check-style-tokens.mjs — style tokens OK.
• The rest of apps/desktop/test/*.test.mjs reports 24 failures on this machine, all of them ERR_MODULE_NOT_FOUND: Cannot find package '@pi-desktop/shared' — this checkout has no node_modules because pnpm install cannot complete on this network (metadata timeouts; it has been retrying for several minutes). Those files are unrelated to this change and fail identically on unmodified main. CI is the authoritative check for pnpm typecheck / lint / the full workspace suite.

If the toggle placement or the summary fallback should follow a different convention in this panel, say so and I will adjust.

---

Numbering note: the decision is D382. D381 is taken by #191 (the selectable Windows
shell change), which itself moved up from D380 after upstream 8ce8ed36 claimed it.
Both decisions logs follow. The sibling PRs are #191 at D381 and #193 at D383.

[vercel]

@Yi-111-a is attempting to deploy a commit to the vastsa's projects Team on Vercel.

A member of the Team first needs to authorize it.

[vastsa]

Principle

Sound. Merging this PR.

Issue #161 is real on current main: a delegate that ends failed / timed_out / aborted shows a Failed capsule and a step stream, and nothing else. SubagentRunResult.error already travels on the lifecycle roster (delegationSummary spreads record.result?.error), but the dock renders the Task row, which ADR 0089 pins at running. This is a presentation gap, not a runtime/IPC/storage change.

The approach is the right one:

• Collect error: { code, message } from the same lifecycle rows, with the same last-write-wins rule, as collectDelegationStatuses.
• Gate the card on a non-success terminal outcome, not on the error field alone.
• Reuse the transcript error card (summary, code, disclosure + copy) and existing i18n keys.
• Topology node and sidebar share SubagentDetail via the dock, so both entry points get the card.

Landing

main had moved (#191 / D381), so the decisions logs conflicted. I merged main into this branch (keep D381 then D382) and copied the tool-row message before the isDelegationActivityItem guard — the same never narrowing that collectDelegationStatuses already documents. Typecheck now passes.

Validation

• node --test test/subagent-topology.test.mjs test/subagent-panel.test.mjs — 23 passed
• node scripts/check-style-tokens.mjs — OK
• CI after the landing commits: Docs locale pair, JS typecheck/lint/test, Rust host-core — pass. Vercel is unrelated deploy-auth.

Follow-up (not merge blockers)

• Optional issue ask (failed step/tool name) is still out of scope.
• zh-CN is missing the pre-existing §5.7 / E2E-198 sections; correctly not invented here.
• A reconstructed leftover aborted with no roster error still has no card.

Thank you — clean diagnosis and a small, matching collector.

[vastsa]

Merged as 4d967576 on main, preserving the author commits plus two landing commits on top (93dfb5d7 merge of main for the D381/D382 decisions-log conflict, 79fa90f4 typecheck fix for the Task-row guard).