Keep manager responses visible across child lifecycle steers

[ryanbbrown]

What was wrong

A child can send a detailed report while its parent manager turn is active. The parent can then show a complete assistant response before BB accepts the child's normal lifecycle notification as a system auto-steer in the same turn. When the turn completes, that visible response can disappear.

The failure has two connected causes:

1. Completed thread-view projection treated accepted child lifecycle system steers as foldable housekeeping. It moved the directly preceding assistant response into the collapsed Worked for summary.
2. Pi did not close the active assistant stream at each assistant message_start. If Pi emitted two assistant messages in one steered turn with no tool boundary, both could use one canonical item ID. The later completion could replace the earlier text, and resolved-delta pruning could then remove the only canonical copy of the first response.

The first cause affects completed timelines from all providers. The second affects Pi and makes this a high-severity canonical-data-loss bug, not only a display bug. Histories already pruned before this fix are not repaired; a native provider transcript can still contain text that BB's canonical event history no longer has.

Minimal reproduced order:

1. A child agent tell starts or joins the parent turn with a detailed report.
2. The parent emits a detailed assistant response.
3. BB accepts child-completed as a system auto-steer in that turn.
4. Pi emits a second assistant response with no tool boundary after the first response.
5. The turn completes, the timeline refetches, and resolved deltas are pruned.

Before this change, the first response became collapsed or was lost from canonical Pi history. It must remain a top-level row with its own item ID.

What changed

• Completed-turn projection now treats accepted child-needs-attention, child-completed, child-failed, child-interrupted, and child-outcome-batch system steers as visible exchange boundaries.
• Projection preserves only a directly adjacent assistant or error row, so it does not move a response across commands or other activity.
• Unlabeled, reconnect, pending, rejected, and other housekeeping steers remain folded.
• The Pi translator now closes an accumulated assistant stream at the next assistant message_start. The shared assembler then gives later assistant output a new canonical item ID. Retry and empty-abort boundaries retain their existing safe behavior.
• Added Pi translator, completed thread-view, cproxy-shaped, Terminal-Bench-shaped, public timeline, source-order, pruning, and housekeeping control regressions.
• Bumped HOST_DAEMON_PROTOCOL_VERSION from 143 to 144 because daemon-emitted canonical events can now use separate Pi assistant item IDs.

This does not weaken pruning and does not add a compatibility fallback. It is not retroactive: it prevents new loss but cannot reconstruct canonical text that an older daemon already merged and pruned.

How you verified

Tests added for this change fail on the old behavior and pass with the fix.

• pnpm exec turbo run build typecheck lint --output-logs=new-only: 88/88 tasks passed after pnpm install --frozen-lockfile restored missing managed-worktree dependencies; the lockfile did not change.
• pnpm exec turbo run test --output-logs=new-only: 68/68 tasks passed. The app result was 402 files, 3,096 passed tests, and 3 skipped tests.
• One earlier full-test run hit an unrelated PromptBoxInternal.test.tsx timeout. The full file then passed three forced runs at 101/101 tests each, and the full repository rerun passed.
• Plugin SDK npm-version guard, app bundle budgets, and bb-app tarball smoke passed. The tarball smoke ran 12/12 tasks.
• git diff --check passed. Every changed file passes Prettier.
• Root pnpm run format:check still stops on the pre-existing unquoted-colon parse errors in .github/ISSUE_TEMPLATE/bug.yml:71 and .github/ISSUE_TEMPLATE/feature.yml:42; neither file is changed by this PR.
• One independent Codex, Claude, and GLM review-panel cycle found source-order and coverage gaps. Those findings were fixed before the full validation. No second review cycle was run.

Disposable live QA used a separate SQLite data directory and ports 63894-63896. It used production Pi translation and assembly with synthetic Pi 0.84 events, real persistence and pruning, the public timeline route, and the React app. No provider API was called. The exact sequence remained top-level before completion, after completion, after full browser/timeline refetch, after pruning two resolved deltas, and after a full disposable server/app restart. The assistant row IDs remained distinct:

• ...|turn:qa-pi-t1|parent:root|item:qa-pi-i1
• ...|turn:qa-pi-t1|parent:root|item:qa-pi-i2

The accepted unlabeled reconnect control remained folded. Browser page errors were empty. QA did not exercise the process-local two-second notification batch or live WebSocket ingestion because fixture insertion does not publish realtime hub events; automated tests cover the accepted lifecycle event producer shape.

Related work

Related but not duplicate: #1355, #1656, #1657, #1981, #897, #1224, and #1834.

AGENT GENERATED: by GPT-5.6 Sol