feat(j5): peer messages reach every running Codex GPT 6 turn mid-turn

[Jacksondr5]

Problem

A peer's message is delivered into a running turn only when the receiver runs Codex Astra; every other model queues it until the turn ends. Luna and Sol take messages during work the same way Astra does, so an agent on either of them works from stale guidance for the rest of a long turn.

What changed

The exception in divergence D4 now covers every Codex GPT 6 model. gpt6PeerSteeringRun (was astraPeerSteeringRun) accepts a run whose resolved model name starts gpt-6- or gpt-6., which is Astra, Luna, Sol and their point releases such as gpt-6.1-sol, with or without Codex's openai. qualifier. Everything else about the rule is unchanged: the receiver must be a Codex thread whose session reports active steering, only peer messages steer, and platform notices still queue.

The match is on the name's prefix, so a future GPT 6 model is covered without another change, and so is a custom Codex model named gpt-6-…. The old test listed gpt-6-astra-custom as not eligible; under "all GPT 6 models" it is.

The send result's backlog notice follows the same function, so a send to a running Luna or Sol turn no longer reports a queue it will not wait in.

The definitions (A2A, substrate, Crews, the register of divergences, the convergence watchlist), the user docs for peer messages and personas, and FORK.md case 26 are rewritten to say GPT 6.

Live testing

Pass on 3888c6a5c4, default setup, live turns on Codex gpt-6-luna and Claude Haiku only. Report: test-reports/j5-gpt6-peer-steering.md.

• A held Luna turn took a peer message into the running turn: the steer effect succeeded, no second run was queued, the send result carried no deliveryNotice, and the file the message asked for was written by the same run.
• A held Haiku turn queued the same kind of message: the send result carried a deliveryNotice, and the message ran as its own turn after release.
• An idle Luna receiver started a normal new turn.
• After a restart at rest, the three deliveries still read as delivered and a further message to Luna was delivered.

Not exercised live:

• Astra, Sol and gpt-6.1-sol. The tester was limited to the models authorized for live turns. Astra's steering was live-tested when the exception was first made; Sol is covered by the integration test only.
• The second commit, 9baac934da, which makes openai.-qualified names eligible and updates docs. It is covered by the integration test.
• Desktop, mobile, peering and relay: the change is in the server's delivery path and none of them alter it.

Upstream impact

No upstream-owned file is edited. This widens divergence D4 in docs/j5/product/upstream.md, on Jackson's decision of 2026-10-12, and FORK.md case 26 is updated to match.

Checklist

• One concern: the description has no "also"
• Tests cover the changed behavior (backend changes ship with focused tests)
• UI changes: no UI in this PR
• Live testing: the section above says what ran live and what did not
• Upstream-owned files: none edited; FORK.md case 26 text updated
• Upstream product: the decision is Jackson's of 2026-10-12, recorded in D4
• Surfaces: server delivery path only. Providers: Codex only by design; Claude and the rest still queue. No contract, client or entry-point change
• Docs: definitions under docs/j5/product/ and user docs rewritten

Claude Opus 5.5, Claude Code harness.

🤖 Generated with Claude Code

[coderabbitai]

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

---

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

---

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

_{Comment @coderabbitai help to get the list of available commands.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
j5-code	Ready	Preview	Oct 12, 2026 7:39pm UTC