fix(server): keep awaited Claude commands pending until their wake starts#32223

Problem

When a Claude command explicitly marked awaited with t3_background_wait exits after the root turn has ended, its task_notification or replacement roster removes the awaited entry before Claude's wake run exists. Clients can treat the preceding completed run as settled and fire a premature completion alert. This is the native command/monitor case reported on #31087.

Change

Keep completed awaited entries visible in the Claude provider roster until the buffered native wake attaches to its continuation run. Release the retained entry on explicit unawait, stopped notification, query shutdown, or discarded continuation. Task ID reuse starts unawaited. Query replacement also stays blocked while an awaited completion is held, even after the active roster becomes empty. The existing test now covers both notification/roster orderings and these cleanup paths, alongside an unawaited dev server.

Scope and approval

Addresses the native awaited command/monitor subset of #31087. The app-owned delegated-task fixes in #31130 and #31357 change separate projection/delivery paths; their current diffs do not cover this adapter roster.

This qualifies as a focused fix for an obvious ordering bug: preserve an existing explicit wait across the gap before its already-supported continuation starts. It adds no timeout, changes no unawaited-task or stop policy, and changes no model-facing instructions, contracts, or client code. All clients and remote connections consume the same persisted provider roster. This PR does not close the broader delegated-task report.

Verification

Windows, Node v26.12.1, frozen local dependencies. Baseline production module: 32a0531748c44b25072d7495ce0b9bd7e9450a87.

From apps/server, extended the existing test before the production change:

vp test run src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts -t 'holds only the background tasks the agent marks awaited through their wake'

Baseline: 2 failed, 341 skipped. Both orderings expected the completed task to remain awaited: true; both received only the unawaited dev-server entry. Final selection: 349 passed, 1 existing platform skip, 2 files, including all six ordering/cleanup cases. The skipped existing test requires process.getuid, unavailable on Windows.

Attempted vp run check:changed --concurrency 1. Its ignored-input Git scan stalled on this Windows junction layout (#32023); stopped only its retained command session. No check cache was forged. Ran every selected equivalent stage serially, including the same two selected test files from the owning server package (avoiding the unrelated root browser port constraint, #24007):

Every selected equivalent stage exited 0. Typechecks: 3/3 successful (@t3tools/scripts, @t3tools/evals, t3). Format/lint were rerun after the final stopped-notification cleanup.

Integrated runtime verification used fresh isolated state/workspace/Claude-home directories and a deterministic native Claude protocol peer, with no paid model calls. The actual server CLI, Claude SDK subprocess, authenticated HTTP/WebSocket APIs, MCP t3_background_wait, SQLite projection, real background Node process, and continuation worker all ran. The peer withholds the wake's root frames until after the command exits, so the intermediate shell snapshot is checked deterministically.

The baseline runtime loaded the unmodified adapter module from the stated commit; the fixed runtime loaded this branch. The final-source stopped/monitor checks cover explicit-stop cleanup and the other native opaque task kind. All fixture servers were stopped through their captured child handles. No live userdata or baked Vite origins were used.

Limitations: the check:changed wrapper itself did not complete due to #32023. The pre-commit hook repeated checks in parallel and hit oxc-parser: RangeError: Array buffer allocation failed; its format/lint/Knip equivalents had already passed serially, so the hook was disabled for the single commit. This verifies the real server workflow with a deterministic provider peer; it does not claim a Claude cloud-model turn or an audible desktop/mobile notification check.

Query replacement verification

The existing replacement/stop test now also covers a completed awaited task with an empty active roster. On the published head 2ae782e, that added row failed because replacement returned success; the original running-task row passed. Both rows pass with the guard, and all selected equivalent stages were rerun successfully on the final source: 349 tests, the three affected typechecks, GraphQL, format/lint, and both Knip scans.

A separate real-server run changed the thinking option while the completed native command was still awaited. The unchanged published adapter opened a replacement CLI and lost the hold. The fixed adapter blocked the replacement, kept the original process and awaited entry, then consumed the real wake and completed with no pending tasks:

Runtime fixture and reproduction instructions. Both isolated servers were stopped through their captured handles. The follow-up again encountered PR #32023 before wrapper stages; every exact equivalent passed separately. Checks ran serially with one Rust worker and Knip's supported parser transport setting KNIP_DISABLE_RAW_TRANSFER=1. The duplicate parallel commit hook was disabled for this commit only after the unchanged source passed those same checks. Implemented with OpenAI GPT-6 via the Codex harness.