Best possible solution:

Undraft after maintainer review, rerun current-head changed/targeted validation, then merge the queue-first completion announce path if checks remain green.

Do we have a high-confidence way to reproduce the issue?

Yes. Source inspection gives a high-confidence reproduction path: current main sends completion announces direct-first and retries gateway timeout as transient before queue fallback; I did not run a live busy-session reproduction in this read-only pass.

Is this the best way to solve the issue?

Yes. The PR uses existing queue/steer and fallback machinery rather than adding a new routing system, but it still needs maintainer review and current-head validation before merge.

What I checked:

• Protected PR state: GitHub API shows the PR is open, draft, on head f9eb7d9, and labeled maintainer. (f9eb7d993c26)
• Current main direct-first completion dispatch: Current main calls direct() first when expectsCompletionMessage is true, then attempts queue fallback only after direct delivery fails. (src/agents/subagent-announce-dispatch.ts:94, 990f931a2efc)
• Current main retries gateway timeouts: Current main treats gateway timeout as transient and retries through runAnnounceDeliveryWithRetry. (src/agents/subagent-announce-delivery.ts:202, 990f931a2efc)
• PR queue-first dispatch: The PR head changes dispatch to always try queue-primary before direct-fallback. (src/agents/subagent-announce-dispatch.ts:78, f9eb7d993c26)
• PR forced completion queue path: The PR adds forceCompletionQueue, steering with steeringMode: "all" and queue settings independent of channel queue mode for completion announces. (src/agents/subagent-announce-delivery.ts:477, f9eb7d993c26)
• Docs and changelog updated: The PR documents queue-first completion resilience and adds a user-facing changelog entry for the agents/subagents fix. Public docs: docs/tools/subagents.md. (docs/tools/subagents.md:83, f9eb7d993c26)

Likely related people:

• steipete: Recent main-branch history on subagent announce delivery, dispatch, queue behavior, route helpers, and docs; also authored multiple current-main commits in this exact surface before this PR. (role: recent maintainer and adjacent owner; confidence: high; commits: 9e9df8f2c578, 33e527d1fcc2, 30a2b3049ae0; files: src/agents/subagent-announce-delivery.ts, src/agents/subagent-announce-dispatch.ts, src/agents/subagent-announce-queue.ts)
• Udjin79: Introduced the busy-session announce queue deferral behavior that this PR builds on when forcing completion delivery through the queue path. (role: related queue behavior contributor; confidence: medium; commits: 1841dd9977cb; files: src/agents/subagent-announce-queue.ts, src/agents/subagent-announce-delivery.ts)
• DolencLuka: Contributed thread-bound subagent completion fallback behavior in the same dispatch/delivery surface. (role: related completion fallback contributor; confidence: medium; commits: 325e5e921fd6; files: src/agents/subagent-announce-delivery.ts, src/agents/subagent-announce-dispatch.ts)
• vincentkoc: Recent history includes fallback subagent completion delivery and subagent docs work in the reviewed area. (role: adjacent completion delivery and docs contributor; confidence: medium; commits: a1b656705992, bf2c992a8627, 1427c3a78d80; files: src/agents/subagent-announce-delivery.ts, docs/tools/subagents.md)

Remaining risk / open question:

• Latest-head validation is not proven by the PR body: the linked Testbox run is cancelled and belongs to an older SHA, while current head check-runs are mostly skipped.
• This read-only review did not execute the targeted tests or a live busy-requester reproduction.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 990f931a2efc.