Codex Implement

Codex Implement

Non-interactive implementation via Codex CLI in a workspace-write sandbox: Codex reads the plan, edits the working tree directly, runs the project's lint/build on its own work, and reports back. One persistent thread per target, so a plan can be delegated in successive batches (or phase by phase) with full context retained.

State persisted under .claude/skills/codex-implement/state/<sanitized-target>.{thread,review.txt,events.ndjson} (the .review.txt file holds Codex's implementation report — the naming comes from the shared helpers). All operations go through this skill's own scripts/ wrappers: start.sh is a dedicated workspace-write script, while resume/reset/show pin the correct STATE_DIR and delegate to the shared codex-plan-review implementation — so there is no export to remember (a forgotten export would silently operate on a review thread instead of the implementation thread).

Arguments

• <target> — auto: start if no thread, resume if one exists. Usually a plan path (docs/1-plans/F_*.plan.md); a free-form label for unplanned work.
• Optional trailing instructions — scope control appended to the prompt, e.g. "Implement only: <batch checkboxes>" or "Now implement: <next batch>".
• reset <target> — drop state, next call starts fresh.
• show <target> — display the latest report without calling Codex.

Execution

1. Parse $ARGUMENTS: extract action (reset/show/auto) and target.

2. Auto — try start.sh first (exit code 2 = thread exists → use resume.sh):

   • Start: bash .claude/skills/codex-implement/scripts/start.sh --prompt-file .claude/skills/codex-implement/prompts/implement.tpl <target> [instructions]
   • Resume (next batch / additional scope): bash .claude/skills/codex-implement/scripts/resume.sh --prompt-file .claude/skills/codex-implement/prompts/continue.tpl [--notes "review corrections"] <target> [instructions]

3. Reset: bash .claude/skills/codex-implement/scripts/reset.sh <target>

4. Show: bash .claude/skills/codex-implement/scripts/show.sh <target>

5. Parse trailing tag of the report:

   • IMPLEMENTATION_COMPLETE — hand control back to the requester's batch review (TRIP-2).
   • IMPLEMENTATION_PARTIAL — read the report; resume with instructions for the remainder, or let the requester finish small leftovers directly.

Notes

• Run Codex calls in a background shell. Invoke start.sh / resume.sh via the Bash tool with run_in_background: true — never as a foreground/inline command. Implementation runs are the longest Codex calls in the workflow and will outlast the foreground command timeout; the background task notifies on completion, then read its output. reset.sh / show.sh are instant and fine in the foreground.
• Set CODEX_TIMEOUT=7200 (2 h) when invoking start.sh / resume.sh — a circuit breaker against hung runs only, deliberately far above any normal batch; bump higher for unusually large batches rather than risk a mid-run kill. Script default is 0 = no timeout. A timeout here kills Codex while it is editing the tree, leaving a partially modified working state — inspect via git status / git diff before retrying.
• --sandbox workspace-write on start; codex exec resume inherits it. Codex edits files and runs repo commands (lint/build); no network, no commits.
• Fixes are the requester's job. After Codex reports, the requester (TRIP-2 batch review) fixes problems directly in the tree — do NOT ping-pong fixes back to Codex. Resume only for genuinely new scope (next batch, large remainder), passing what was fixed and why via --notes.
• Separate STATE_DIR from the review skills — the same plan path can hold an implementation thread and a review thread without collision.
• Codex is instructed not to write tests (testing gate owns that) and not to touch release ceremony.
• Network is blocked in the sandbox: if the plan requires installing a new dependency, Codex will report it as a leftover — install it yourself during the batch review.
• Model/effort/tier defaults live in codex-plan-review/scripts/_common.sh (implementation → gpt-5.6-luna at high effort on the fast service tier; reviews → gpt-5.6-sol at xhigh on standard routing; derived from STATE_DIR). Adjust that one file to your preferred models, or override per run via CODEX_MODEL / CODEX_EFFORT / CODEX_TIER env vars; the scripts echo the effective values.
• Effort escalation (per batch). high is the default because plan batches are well-scoped and every batch passes through the requester's delta review plus the final full code review. Escalate a single batch to xhigh (CODEX_EFFORT=xhigh on that batch's start.sh/resume.sh call) when it involves any of:
  • novel core logic designed from scratch — an algorithm, data structure, protocol, or state machine with no existing pattern in the codebase to follow;
  • changes the testing gate can't meaningfully verify — correctness only observable at runtime or by inspection, with no automated check covering it;
  • concurrency, security, or data-integrity-sensitive code — where a subtle slip is costly and hard to spot in review;
  • cross-cutting changes — one batch touching many files or layers whose interactions must stay coherent. When none apply, stay at high — a slipped batch is caught and fixed directly in the delta review.