agent-all-codex

/agent-all (Codex port)

Runs the cost-unrestricted multi-agent pipeline using Codex CLI primitives. Two dispatch paths supported, chosen at preflight:

• Preferred (when available): Codex's agent hook type — fires on parent-→-subagent invocation, enables true parallel fan-out.
• Fallback: sequential invocation of .codex/skills/<role>/SKILL.md per wave task. Slower but works on any Codex CLI version.

Usage

/agent-all-codex "add user signup form"
/agent-all-codex docs/tasks/12-fix-login.md
/agent-all-codex "fix flaky test" --loop --max-iter=5
/agent-all-codex docs/tasks/x.md --no-pr --wave-size=large

Flags

Same as Claude Code: --loop, --max-iter=<N>, --max-cost=<USD>, --wave-size=small|medium|large, --no-pr, --no-brainstorm, --resume, --force, --yes, --break-condition=<spec>, --reconfigure.

When --loop is set, Phase 0 prompts the user interactively (via Codex's ask_user) for the break-condition preset (test-auto / visual-qa / Custom shell / Composite) and offers to save the choice to .agent-all.json. Use --break-condition=<spec> to skip the prompt for one invocation, or --reconfigure to re-prompt even when a non-default value already lives in config.

Additional Codex-specific:

• --dispatch=agent-hook|sequential — force the dispatch strategy (default: auto-detect at preflight).

Pipeline

Phase	File	Purpose
0	phases/0-preflight.md	git + .codex/skills + config + dispatch-strategy detect
1	phases/1-intent.md	brainstorm (chat) OR load task file
2	phases/2-plan.md	draft plan via apply_patch
3	phases/3-dispatch.md	agent hook fan-out OR sequential skill invocations
4	phases/4-gate.md	dispatch reviewer (same strategy as Phase 3)
5	phases/5-pr.md	shell_command: git branch push + gh pr create
6	phases/6-loop.md	breakCondition shell + state.iter re-entry

Rules

1. You orchestrate; phases are source of truth. Read each phase file before running it.
2. State lives in .agent-all-state.json. Atomic write via apply_patch (temp + rename).
3. Delegate, don't reimplement. Phase 3 uses agent hook (or sequential skill calls); Phase 4 same; Phase 5 uses shell_command.
4. Loop is opt-in. Without --loop, Phase 6 is a no-op.
5. Hard caps: --max-iter clamped to 50; --max-cost enforced after each wave via Codex's session-cost API (if exposed; else best-effort).

Codex primitive map

Action	Codex primitive
Read file	implicit (model reads file directly)
Write file	apply_patch
Shell (one-shot)	shell_command
Shell (long-running)	exec_command (keeps PTY)
Dispatch subagent	agent hook (preferred) OR invoke .codex/skills/<role>/SKILL.md
Prompt user	ask_user
Persist plan/state	apply_patch to .agent-all-state.json

On error

Same as Claude port. Additional Codex-specific notes:

• If agent hook not registered AND --dispatch=agent-hook was forced: abort with Codex agent hook unavailable; rerun without --dispatch=agent-hook.
• If .codex/skills/<role>/SKILL.md missing for any wave's task role: abort with /codex-init --theme=floor required to seed skill roster.
• shell_command timeout (default 30s) extended to 300s for git push.

When done

Print summary: phases completed, iters, cost, PR URL, dispatch strategy used. Exit 0/1/2/3 per spec.

References

• references/porting-notes.md — agent hook research status + sequential fallback rationale
• plugins/harness-floor/skills/agent-all/SKILL.md — source-of-truth pipeline