feat(voice): send berd-call transcripts to Codex tasks

[johnmatthewtennant]

Summary

Adds berd-call start --codex, which delivers a running call's transcript straight into the Codex task that started it. The agent no longer needs a wrapper process to read --stream output.

• The call finds its task from CODEX_THREAD_ID. It checks that Codex Desktop owns the task, then connects to the local Codex IPC socket (CODEX_MONITOR_SOCKET, then $CODEX_HOME/ipc/ipc.sock, then ~/.codex/ipc/ipc.sock). The socket must belong to the current user, and its directory must be private to that user.
• Each record is a bracketed header followed by the text, for example [3 USER], [4 HANDOFF handoff-a-4], [LIFECYCLE]. Records are delivered in order, steering the active turn or starting a new one when the task is idle. Failed deliveries retry, and records queued close together are sent as one batch. The first delivery includes short guidance: how to reply with berd-call speak --re CURSOR, and how to resolve handoffs with --resolves HANDOFF_ID.
• The records are the same ones --stream emits: user and spokesperson speech, handoffs, restart lifecycle, and speech_result rows from non-blocking speech. --non-blocking now accepts --codex as its delivery path. On exit the task receives a final lifecycle record, either a clean stop that should not be restarted or the failure message.

Reviewer-reproducible example

From a Codex Desktop task terminal, start an Expert-Spokesperson call:

berd-call start --codex --port 5340 --voice Aaron --language en-US --mode expert-spokesperson &

Say "Please ask the expert to look up the current git branch name of this worktree." The task receives:

[1 USER]
Please ask the expert to look up the current git branch name of this worktree.

[2 HANDOFF handoff-external-2]
Look up the current git branch name of this worktree and return the branch name.

Answer it with berd-call speak --port 5340 --re 2 --resolves handoff-external-2 "The worktree is on jtennant/berd-call-codex.". The Spokesperson speaks the answer and the task receives [5 SPOKESPERSON] with the spoken text. Then berd-call stop --port 5340 delivers:

[LIFECYCLE]
Berd Call stopped after a stop request. This is a clean shutdown, not a crash. Do not restart it unless asked.

For non-blocking speech, start a conventional call with berd-call start --codex --non-blocking --port 5342 --voice Aaron --language en-US & and say "Please reply." The task receives [1 USER] with that text. berd-call speak --port 5342 --re 1 "Delivery check received." returns {"requestId": 2, "status": "accepted"} immediately, and a completed reply delivers nothing further.

Invalid invocations fail before a call starts:

$ berd-call start --stream --codex --voice Aaron --language en-US
choose one of --stream or --codex for transcript delivery
$ env -u CODEX_THREAD_ID berd-call start --codex --voice Aaron --language en-US
berd-call start failed: --codex requires CODEX_THREAD_ID; start the call from a Codex task

[johnmatthewtennant]

🤖 Review dispositions at 4f86740 (fixed in 573b329):

• Accepted — Ctrl-C drops the terminal lifecycle record. SIGINT/SIGTERM now route through the same stop request as berd-call stop; the session child runs in its own process group. Verified live: sending SIGINT to the process group printed the clean-stop lifecycle record and exited 0.
• Accepted — terminal record is Codex-only. --stream now prints the terminal lifecycle row too, and the stop text is one constant.
• Declined — remove multi-record batching. Batching keeps a burst of utterances as one Codex turn rather than several steered turns; the stable batch identity is what makes retries idempotent.
• Declined — use ACP/berdctl instead of Codex IPC. The target is native delivery into a Codex task, which ACP does not reach.
• Declined — relay shutdown budget modeled twice. The caller-side wait is a hard cap for a worker that may be blocked mid-delivery; low value to restructure.
• Declined — record shape exists twice. Readability nit; not worth churn in this PR.

[morgmart]

🤖 Automated code review

Engineering found one blocking reliability issue and one non-blocking signal-handling issue. Supplied GitHub evidence is structurally valid and all captured checks for the exact head SHA passed, but required checks still govern merge readiness. Recommend REQUEST_CHANGES because a blocking finding remains.

Deterministic publication result: 1 blocking and 1 non-blocking inline finding(s) publishable; 0 duplicate(s) suppressed; 0 blocking screenshot-evidence requirement(s) in this review body.

[morgmart]

🤖 Automated code review

Engineering completed a fresh static review of the exact PR comparison and found no new publishable issues. Two previously reported issues remain suppressed because each resolved thread has a substantive human reply. Supplied GitHub evidence is structurally valid; some checks are still in progress, so required checks continue to govern merge readiness. Recommend APPROVE under the publication policy.

Deterministic publication result: 0 blocking and 0 non-blocking inline finding(s) publishable; 2 duplicate(s) suppressed; 0 blocking screenshot-evidence requirement(s) in this review body.

Pending checks: 4 check(s) are not complete.

This approval reflects the completed code review only; merge readiness remains governed by the repository's required checks.

The merge-base changed after approval.

[morgmart]

🤖 Automated code review

Engineering completed a fresh static review of the exact PR comparison and found no new publishable issues. Two previously reported issues remain suppressed because each resolved thread has a substantive human reply. Supplied GitHub evidence is structurally valid; Transcript virtualization is failing while the other captured checks pass, so required checks still govern merge readiness. Recommend APPROVE under the publication policy.

Deterministic publication result: 0 blocking and 0 non-blocking inline finding(s) publishable; 2 duplicate(s) suppressed; 0 blocking screenshot-evidence requirement(s) in this review body.

Completed checks currently reporting failure:

• Transcript virtualization: failure

Pending checks: 1 check(s) are not complete.

This approval reflects the completed code review only; merge readiness remains governed by the repository's required checks.