flowb

name	flowb
description	Hand a finished plan to an autonomous Codex (GPT-5.6 Sol) executor and stay reachable as advisor over agmsg. Invoke ONLY when the user explicitly types /flowb.
disable-model-invocation	true

Flow B: autonomous Codex implementation with an agmsg consult line

In this flow you (the Claude session) never spawn implementation subagents. The executor is an external Codex CLI process running GPT-5.6 Sol. You are the architect: you write the plan, answer questions, and verify the result.

Steps

1. Write the plan to a file: put the requirements in PLAN.md as verifiable bullet points (see the template in this repo). Always include a "done means" condition, and a mandatory consult step so the executor asks before guessing.
2. Join the agmsg team (once per project): ~/.agents/skills/agmsg/scripts/join.sh <team> architect claude-code "$(pwd)" and ~/.agents/skills/agmsg/scripts/join.sh <team> builder codex "<workdir>"
3. Launch Codex in the background (stdin MUST be closed):

   codex exec -m gpt-5.6-sol --sandbox workspace-write --skip-git-repo-check \
     "Read PLAN.md and implement every item, making the tests green. When the spec is ambiguous, ask via \
      ~/.agents/skills/agmsg/scripts/send.sh <team> builder architect '<question>' and poll \
      ~/.agents/skills/agmsg/scripts/inbox.sh <team> builder for the answer. \
      Run tests with python3 -m pytest. When finished, send DONE with the test result." < /dev/null
   

4. Watch the consult line: poll inbox.sh <team> architect every few seconds while Codex runs (a Monitor loop works well). Answer questions immediately with send.sh.
5. Verify independently: after DONE arrives, diff the changes and re-run the tests yourself before reporting completion.

Hard rules

• codex exec without < /dev/null hangs forever waiting for stdin. Never omit it.
• The Codex sandbox blocks outbound network (e.g. PyPI). Instruct locally-runnable test commands.
• Treat every "sent" or "done" from the executor as a claim, not a fact. Verify against the agmsg DB and by running the tests yourself.
• If the executor is unreachable or fails repeatedly, stop and report honestly instead of pretending progress.