fix(codex-crew): bypass codex:codex-rescue for coordinator dispatch

[lklimek]

TL;DR: Codex dispatches from a coordinator now call the Codex runtime directly instead of routing through a Claude subagent that added no value and caused most of the recurring dispatch bugs.

User story

As a developer using Claudius to orchestrate Codex Sol coding work, I want a Codex dispatch to bind reliably to the worktree I intended and never silently collide with another dispatch, so that multi-worktree coding work completes predictably instead of stalling or getting duplicated.

Scenario

Base flow

A coordinator wants to hand a coding task to Codex Sol against a specific pre-created worktree, previously by spawning the codex:codex-rescue Claude subagent, which forwards the request to the codex-companion.mjs runtime as a single shell command.

Actual behavior

codex:codex-rescue's forwarding contract never exposes the runtime's own --cwd/--prompt-file flags. As a result: the broker/workspace binds to the coordinator's own working directory, not the target worktree, so every dispatch from one session shares one broker regardless of which worktree it's meant for; two dispatches close together collide (one strands at status=running forever, or silently returns a generic canned reply instead of doing the work); long prompts with embedded quotes or Rust Debug dumps can arrive corrupted; and the wrapper agent's own completion signal (idle_notification) is unreliable in both directions, so jobs finish 40–85 minutes before anyone finds out. Working around the cwd problem required a separate EnterWorktree/ExitWorktree dance that also blocked genuinely concurrent multi-worktree dispatch.

Expected behavior

The coordinator calls the codex-companion.mjs runtime directly, passing the target worktree and the prompt (as a file) as explicit arguments. Each dispatch is self-contained — distinct worktrees never share a broker, prompts never go through shell-quoting, and there is no wrapper-agent lifecycle to produce a false completion signal. Monitoring goes straight to the job's on-disk state, which now reliably reflects the intended worktree.

Detailed discussion

What was done

• skills/codex-crew/SKILL.md: new § Direct Dispatch documents calling codex-companion.mjs task directly with --cwd <worktree-abs-path> and --prompt-file <abs-path> for all coordinator-orchestrated dispatch; codex:codex-rescue is now scoped to the upstream, user-typed /codex:rescue interactive command only.
• § Sandbox & Workdir rule 3 and § Plan-Approval Gate: drop the EnterWorktree/ExitWorktree requirement, since --cwd binds the worktree directly and allows real concurrent multi-worktree dispatch.
• § Never Dispatch Concurrently to the Same --cwd (renamed from § Never Dispatch Back-to-Back from the Same cwd): collision risk is now scoped to literal same---cwd dispatches, not every dispatch in a session.
• § Monitoring a Codex Job: reframed around "no agent lifecycle to watch" instead of "don't trust the wrapper's unreliable signal."
• skills/codex-crew/references/sandbox-and-recovery.md: on-disk job state's workspaceRoot now reliably matches the intended worktree for direct dispatch; the old dispatching-session-cwd caveat is scoped to the interactive path only.
• Closed 3 stale memcan:todo items (project=claudius) this change resolves at the root: d8f1a08b (--write omission / self-reinvocation collision), 4d1f3d2f (quote/Rust-Debug-dump corruption), b6f79473 (same-cwd collision manifesting as an instant generic reply).
• .claude-plugin/plugin.json bumped 5.15.0 → 5.16.0; CHANGELOG.md updated.

No source code changed — this is a skill-doctrine change only, no new scripts.

Testing

claude plugin validate . passes (one pre-existing, unrelated warning about CLAUDE.md at plugin root).

Breaking changes

None. This changes agent guidance, not a public interface; any coordinator following the old EnterWorktree-based doctrine still works, it's just no longer necessary.

Checklist

• plugin.json version bumped
• CHANGELOG.md updated
• claude plugin validate . clean
• Related memcan:todo items closed

Prior work

None — first attempt at removing the wrapper-agent layer; prior PRs (#71–#75) hardened the wrapper-based approach instead.

Attribution

_{🤖 Co-authored by Claudius the Magnificent AI Agent}

[Copilot]

Pull request overview

Updates Claudius’ coordination doctrine to make Codex dispatches more reliable by removing the codex:codex-rescue wrapper from coordinator-orchestrated flows, and aligns several workflow/delegation skills around “goal-first, plan-gated” implementation.

Changes:

• Document direct coordinator dispatch to codex-companion.mjs task using explicit --cwd + --prompt-file, and adjust Codex monitoring/recovery guidance accordingly.
• Introduce/propagate “Development-Work Delegation (WHAT, not HOW)” guidance across workflow skills, delegate, and developer agent prompts (plan first, coordinator approval, then implement).
• Refactor watchdog doctrine: prefer MCP watchdog when available, keep built-in Monitor as fallback, and extract detailed fallback mechanics into a reference file; bump plugin version and update changelog.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
skills/codex-crew/SKILL.md	Switches coordinator doctrine to direct Codex runtime dispatch; updates plan/monitoring rules.
skills/codex-crew/references/sandbox-and-recovery.md	Updates job-state mapping notes for direct dispatch vs interactive wrapper path.
skills/grand-admiral/SKILL.md	Adds “Development-Work Delegation” doctrine and MCP-watchdog preference; trims fallback details.
skills/grand-admiral/references/stall-watchdog.md	New extracted reference for built-in watchdog mechanics and playbooks.
skills/delegate/SKILL.md	Updates pre-delegation checklist for watchdog choice + plan gate.
skills/track-minions/SKILL.md	Reframes durable tracking: in-session file primary; memcan for cross-session/project.
skills/workflow-trivial/SKILL.md	Aligns trivial workflow implementation phase guidance with new delegation doctrine.
skills/workflow-simplified/SKILL.md	Requires implementer-authored plan + coordinator sign-off per task before coding.
skills/workflow-feature/SKILL.md	Same plan gate + acceptance-criteria briefing pattern for feature workflow.
agents/developer-bilby.md	Adds an explicit “Implementation Plan Gate” before coding.
agents/architect-nagatha.md	Adjusts architecture guidance to leave file-level placement to implementer plan.
CHANGELOG.md	Adds release notes for new doctrine/version.
.claude-plugin/plugin.json	Bumps plugin version to 5.16.0.

Comments suppressed due to low confidence (1)

CHANGELOG.md:22

• The changelog introduces both 5.15.0 and 5.16.0 sections, but .claude-plugin/plugin.json is bumped directly to 5.16.0 in this PR. That makes the release notes ambiguous, and the 5.15.0 “Fixed” bullet about requiring EnterWorktree conflicts with the new 5.16.0 direct-dispatch doctrine (--cwd removes that requirement). Consider collapsing these notes into a single 5.16.0 entry so the changelog matches the single version bump and the final guidance.

## [5.15.0] - 2026-07-22

### Added

- **`skills/grand-admiral/references/stall-watchdog.md`**: new reference file — the built-in Monitor's discovery sources, event grammar, Multi-Session Hygiene traps, and STALL/GONE playbooks, extracted from `grand-admiral` § Recovery so they load only when the fallback path is actually used.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.