fix(subagents): harden lifecycle, completion delivery, and transcript UX across the stack

[Rinat S (rsolmano)]

Summary

Post-merge hardening for the subagents stack (#302/#303/#304), from a full-feature review with adversarial verification of every finding. Fixes the still-open #303 review thread (child settings parity) plus confirmed defects across all three layers: lost aborts, silently-dropped background completions, false-permanent transcript errors, an unrestricted-tools parsing hole, and missing regression pins.

Changes

packages/pi-delegation

• An abort landing in pi prompt() preflight (before streaming starts) is now re-delivered from the run's event subscription — previously the child streamed to full completion, billed, returned as aborted with the completed finalText, and delete/archive flows blocked on the whole natural run.
• finalText derives from the run's own turn_end-captured assistant message (index-scan fallback retained) — a mid-run compaction no longer misreads a completed run as "(no output)".
• steer() rejects with typed not-running unless a run is queued/running; aborted-before-start runs clear the pi queue, so a stale steer can no longer contaminate the child's next run.
• Lifecycle listeners and onUpdate callbacks are exception-isolated — a throwing embedder callback can no longer reject a completing run or abandon a disposal cascade.
• onUpdate now always delivers a queued details object before running (makes the UI's queued state live).
• New DelegationBindings.settingsManager seam; typed unknown-model error; a rejected fallback ModelRuntime.create() is evicted instead of cached forever.

packages/pi-subagents

• Definition frontmatter supports YAML block lists for tools:/skills: and fails closed: a list-valued key with zero parseable items skips the definition instead of granting pi's default read/bash/edit/write toolset.
• serviceFor refuses to recreate the fallback service after session_shutdown; unknown-id error wording covers disposed children, not just host restarts.
• The round-2 session_shutdown → disposeChildrenOf fix is now regression-pinned; the detached-survives-abort test is gate-driven instead of sleep-timed.

packages/server + packages/contracts

• Children get the host's buildSessionSettings via the new binding, and the curated child extension set includes the oversized-image guard — closes the open feat(host): embed pi-delegation/pi-subagents in the engine host #303 review thread (packaged-app child image reads). The pin exposed and fixes a latent parent-side bug too: pi's SettingsManager.reload() dropped the autoResize:false override during every session assembly.
• readChildTranscript consults the run registry before throwing: a queued/unflushed child returns { messages: [], status } instead of a false-permanent NOT_FOUND (pi creates the session file only at the first assistant message).
• The host emits session_shutdown (idempotent) before disposing/trashing in delete flows, so completions arriving during the unbounded trashFile window are suppressed instead of triggering a billed turn on a chat being deleted.
• An undelivered-completion sweep (on agent_settled and post-reload, guarded by hasQueuedMessages and the transcript-as-delivery-record) redelivers background completions lost to queue drains or mid-run session.reload().
• New subagent.abort wire method (PROTOCOL_VERSION 54), lineage-checked like subagent.getTranscript.

apps/web

• The transcript dialog gains a Stop control for live background runs (wired to subagent.abort); background-ack cards note that runs survive a chat Stop.
• Hydrated non-terminal background acks probe run status once and swap to a "no longer running" state after a host restart, instead of promising a completion forever.
• Aborted foreground runs show a warning state (data-outcome) instead of the green success check; the queued card state now renders on the live path.

e2e

• New no-agent spec subagents-ui.spec.ts (hydrated three-state Agent cards, completion turn, disk-transcript dialog, restart-reconciliation lost state) — the subagent UI previously had zero commit-gate coverage. New fixture helper seeds child transcripts under the delegation root.
• The pure-pi smoke cleans up its copied credentials on every exit path.

Deliberately not changed (recorded design decisions): per-child session GC ("no per-parent GC in V1"), within-lineage provider-mirror semantics, WorkspaceProvider.dispose sync typing, vanilla-pi /reload child disposal.

Testing

• Every behavior fix carries a regression test verified to fail on the unfixed code (fix reverted, test observed red, fix restored).
• bun run test — 14/14 tasks green (server 870, web 909, pi-delegation 44, pi-subagents 46+).
• bun run typecheck, bun run lint, bun run check:deps, bun run check:boundaries, bun run check:seams — green.
• Full no-agent e2e suite: all 8 shards passed (132s), including the new subagents-ui.spec.ts.
• Diff audits: zero suppressions, zero new code comments.

[jetbrains-air]

Approved — ready to merge.

[jetbrains-air]

Approved — ready to merge.

[jetbrains-air]

Approved — ready to merge.

[jetbrains-air]

Approved — ready to merge.