plan-loop

name	plan-loop
description	Use when the user asks to plan a task before coding (e.g. "plan-loop", "/plan-loop", "plan this out first"). Draft a detailed plan, cross-review it with Claude, triage skeptically, iterate, then gate for sign-off.
metadata	{"short-description":"Plan → cross-review with Claude → triage → gate"}

Plan Loop (Codex)

Settle a plan before any code, Codex-side. Same workflow as Claude's /plan-loop.

Canonical procedure: read ~/.codex/skills/skyhook-skills-commands/plan-loop.md and follow it (detailed plan with file:line evidence + Certain/Confident/Needs-input tiers, cross-review, triage, iterate, gate). Codex translations:

• Cross-review the plan with the OTHER model — Claude, not Codex. Hand the plan to Claude via the claude-review skill, or directly. Prefer writing the plan to a temp file and asking Claude to read it; that avoids argv quoting issues with backticks, $, newlines, and very large plans:

  # Plan text already written to /tmp/plan.md
  claude -p "Read /tmp/plan.md and critique it as a skeptical senior engineer. Find flawed assumptions, missing steps, hidden complexity, simpler alternatives, and risks. Challenge it; don't rewrite it. Return ranked findings only." --model opus --permission-mode default --allowedTools Read Grep Glob Bash --no-session-persistence --output-format stream-json --include-partial-messages --verbose
  

  (Trivial plans skip this.) claude -p needs Keychain access — run it escalated/unsandboxed (see the claude-review skill's auth note), or it falsely reports "Not logged in" from inside Codex's sandbox. Use the streamed JSON to monitor progress; plain claude -p buffers until the final answer and can look hung during a long critique. Do not interrupt just because no final prose has appeared while tool_use, status, or thinking_tokens events are still arriving. Triage the final result text, not the JSON event noise.
• Triage Claude's critique inline — never auto-accept it; verify each point against the actual code, integrate the valid ones, reject weak ones with a stated reason. Cap at 2 cross-review cycles.
• Gate: default mode stops for sign-off before coding; --auto resolves Needs-input with best judgment, logs each call to NOTES.md, and proceeds — except a ceiling decision (prod/deploy/data-migrations · external-publish/money · security/auth/secrets · breaking-API/destructive-file-ops) stops regardless.

Narrate phases with the glyph-tagged banners and close with the short summary ledger defined in the canonical file. When invoked by the autodev skill, the approved/assumed plan flows straight into implementation.