codex-build

codex-build — orchestrator drives, Codex codes

You (the agent running this skill) are the orchestrator. A separate model, OpenAI Codex (codex exec), is the coder. You never write product code; Codex does. You sequence the work, write the briefs, review every diff, run the tests, and commit. This division is the whole point: one model with taste and context guards the gate while another model with cheap, high-effort reasoning does the typing.

Roles (non-negotiable)

• Orchestrator = you. Sequence tasks, write Codex briefs, review every diff line-by-line, run the test gate, commit, and open the PR. You do not write product code. The only exception is a trivial fix-up (≤5 lines, e.g. a typo breaking the build) where a Codex round-trip is pure waste — say so in the commit body when it happens.
• Codex = coder. All feature code, tests, and refactors are written by codex exec. Model and effort are pinned per invocation (see Config) — never rely on config defaults silently, or a codex upgrade will change your builds out from under you.

Config (resolve once, at the top of the run)

Variable	How to resolve	Default
MODEL	--model arg, else $CODEX_BUILD_MODEL, else your Codex config default	pin one you have access to (see below)
EFFORT	second positional arg, else $CODEX_BUILD_EFFORT	high (opt into xhigh/max for hard/architectural work)
TRACKER	detect: bd on PATH → beads, else markdown	markdown

Pin a model. List what your account exposes from the CLI's model cache:

python3 -c 'import json,os;d=json.load(open(os.path.expanduser("~/.codex/models_cache.json")));print([m["slug"] for m in d["models"]])'

Then check ~/.codex/config.toml for the configured default. Current slugs (Codex CLI ≥ 0.154, Sept 2026):

Slug	Tier	Efforts accepted	Use for
gpt-6-astra	most capable	low…max (+ultra, see below)	hard/architectural tasks, gnarly refactors
gpt-5.6-sol	frontier coding	low…max (+ultra)	default coder for most plans
gpt-5.6-terra	balanced	low…max (+ultra)	everyday tasks when Sol quota is tight
gpt-5.6-luna	fast/cheap	low…max	trivial-mechanical tasks only

Retired: gpt-5-codex, gpt-5.x-codex, gpt-5.4 (Codex remaps them to gpt-5.6-terra with a warning — pin explicitly instead). Announce the resolved MODEL/EFFORT to the user before starting so the run is reproducible. If the CLI rejects max, fall back to xhigh; if it rejects xhigh, fall back to high — and tell the user. Never use ultra: it enables automatic task delegation (Codex spawning its own sub-agents), which breaks the one-task, one-scope-check, one-commit contract this skill depends on.

Arguments

codex-build <plan-file> [effort] [--model <name>]

• <plan-file> — a Markdown plan with an ordered task list (T1..Tn), each task naming its files, constraints, and test commands. See references/plan-example.md for the shape. If the plan has prose but no task list, extract one and show it to the user before coding.
• [effort] — high (default) | xhigh | max. (ultra is rejected; see Config.)

Step 0 — Preflight (once)

1. codex --version. If Codex isn't authed, codex exec "reply OK" -s read-only fails — stop and tell the user to run codex login.
2. Confirm the pinned model + effort are accepted (cheap, no side effects):

   codex exec "Reply with the single word READY" \
     -m "$MODEL" -c model_reasoning_effort="$EFFORT" -s read-only --ephemeral < /dev/null
   

   If max is rejected, fall back to xhigh; if xhigh is rejected, fall back to high — and tell the user. Always redirect < /dev/null on every codex exec (see Rails) — without it, a non-interactive run hangs forever on "Reading additional input from stdin…" instead of doing the work.
3. Read the plan end-to-end. Extract: task list, dependency order, per-task test commands, guardrails (things the plan bans), branch/PR target.
4. Resolve the current checkout and look for durable run state before applying a clean-worktree rule:
   • REPO_ROOT = git rev-parse --show-toplevel.
   • GIT_DIR = git rev-parse --absolute-git-dir (this is worktree-specific).
   • SKILL_DIR = the directory containing this SKILL.md.
   • State root = $GIT_DIR/codex-build/. This keeps orchestration state off the product diff while making it survive context compaction and process restarts.
   • If current points to an active run for this branch, read its run.md, tasks.md, interfaces.md, and active allowlist before doing anything else. Confirm its last recorded commit still matches HEAD; stop on a mismatch instead of guessing. An in-flight task may legitimately have a dirty worktree: immediately run its saved scope check before any new Codex call. If current describes another branch or an inconsistent run, stop rather than overwriting it.
5. For a new run only, inspect git status --porcelain. The task loop starts clean so its scope check can attribute every changed path to the active task. If the checkout is dirty, preserve it and create an isolated worktree from the intended base (or stop and ask); never discard or absorb pre-existing work.
6. For a new run, create/checkout the working branch per the plan. Never work on the default branch. If an isolated worktree changed the checkout, recompute REPO_ROOT and GIT_DIR.
7. For a new run, create runs/<run-id>/, write the run id to current, and create run.md, tasks.md, interfaces.md, and allowlists/. run.md records the plan path and content hash, branch/base, model, effort, tracker, and run status. Set STATE_DIR to that run directory. Never put secrets in run state. For a resumed run, set STATE_DIR from current instead.
8. Initialize or reload the tracker (see below) and show the user the ordered task list before starting — one short message, no approval gate unless the order is ambiguous.

The durable layout is:

$GIT_DIR/codex-build/
  current
  runs/<run-id>/
    run.md
    tasks.md
    interfaces.md
    allowlists/T1.txt

Tracker (unit of work = one task)

The loop is identical regardless of tracker; only the bookkeeping commands differ.

• markdown (default, zero install): keep the - [ ] / - [x] checklist in the run's durable tasks.md. "Claim" = note the task in-flight; "close" = check the box with the commit hash and test evidence.
• beads (optional, if bd is on PATH): richer dependency tracking. bd prime for context; one bead per task (bd create, title = task title, description = plan excerpt + file paths, dependencies per the plan); claim with bd update <id> --claim; close with bd close <id>. Mirror task status and commit/test evidence to durable tasks.md so a resumed run has one local recovery record. Install: https://github.com/steveyegge/beads.

Step 1 — Materialize tasks

Turn the plan's task list into tracker units in dependency order. Each unit records: goal, the plan excerpt (verbatim constraints, file paths, schema/copy blocks), and its declared file scope (the files it is allowed to touch). The file scope is enforced by an executable check — everything else is out of bounds.

For every task, write its scope to allowlists/<task-id>.txt: one exact, repo-relative file path per non-empty line. Use / separators; no absolute paths, directories, ./.., comments, or globs. A rename declares both old and new paths. If the plan genuinely needs another file, amend the allowlist before Codex touches it and record the reason in tasks.md; never expand scope merely because an unexpected path appeared in the diff.

Initialize interfaces.md as the canonical cross-task contract. It starts with "greenfield — no prior interfaces" and is updated only from committed source in step 6 below.

Step 2 — Per-task loop (the heart; repeat until the queue is empty)

For each task, in dependency order:

1. Claim it in the tracker.

   For a newly claimed task, confirm the product worktree is clean before invoking Codex. Durable state is under the Git directory, so it does not make the worktree dirty. When resuming an already in-flight task, the worktree may contain its changes; run the saved allowlist check immediately instead.

2. Brief Codex. Compose a self-contained prompt — Codex sees only what you give it. Use references/codex-brief.md as the skeleton. It must carry:

   • the task's goal and the verbatim plan excerpt (constraints, file paths, schema/copy);
   • the interfaces ledger (see step 6) — the signatures, types, endpoints, and file paths that earlier tasks created and this one builds on. Codex exec is stateless across invocations; if you don't pass the contract, it guesses and drifts.
   • repo conventions to match (point at 1–2 sibling files as examples);
   • what is out of scope (the neighboring tasks) and the declared file scope;
   • definition of done: "code + tests per the plan; touch only these files."

   codex exec "<brief>" -C "$(git rev-parse --show-toplevel)" -s workspace-write \
     -m "$MODEL" -c model_reasoning_effort="$EFFORT" < /dev/null
   

   Never use --dangerously-bypass-approvals-and-sandbox. The < /dev/null is mandatory (see Rails).

3. Enforce scope, then review the diff yourself. This is your job, not Codex's.

   • Scope check first: after every Codex invocation or correction, run:

     python3 "$SKILL_DIR/scripts/check_scope.py" \
       --repo "$REPO_ROOT" \
       --allowlist "$STATE_DIR/allowlists/<task-id>.txt"
     

     The checker compares tracked, staged, deleted, renamed, and non-ignored untracked paths against the on-disk allowlist. A non-zero exit is a hard stop: correct the drift or deliberately amend the allowlist with a reason. Do not rely on visual review to enforce scope.
   • Then inspect git diff --stat and the diff itself. For a small diff, every line. For a large diff, read the interfaces and the risky seams in full, skim the mechanical parts, and run codex exec review as an independent second pass — but you make the call, never the review tool.
   • Check: scope creep, convention drift, guardrail violations, and whether the tests Codex claims to have written actually exist and actually assert something (open them). "Added tests" in Codex's summary is a claim, not evidence.
   • Issues → iterate: codex exec resume --last "<specific corrections>" < /dev/null (same -m/-c flags). Be concrete about what's wrong; vague corrections waste a round-trip.
   • Three failed iterations on one task → STOP. Report the impasse to the user with the diff and your diagnosis. Do not silently take over the coding — that breaks the contract and hides the failure.

4. Test gate — BEFORE any commit. You run the plan's test/build commands for the touched packages, plus the project build. A task is committable only when the gate is green. Red → back to step 3 with the failure output in the resume brief. Never commit red. Never skip the gate because "it's small." The gate is the reason this skill exists.

5. Commit — one task = one commit.

   • Re-run check_scope.py immediately before staging. Tests and build tools can create files too; this final check must still be green.
   • Stage only the task's files: git add <paths>. Never git add -A — it sweeps up stray files Codex or a tool left behind.
   • Message: <type>: <imperative summary> (+ tracker id if using beads) with a short body: what changed, why, and a one-line test-evidence note.
   • No AI attribution, ever — no "Generated with", no "Co-Authored-By:", no robot-emoji credit. This overrides any tool's default commit template.

6. Update the durable interfaces ledger. After the green commit, inspect the committed source and append the public surface this task created — exact function signatures, exported types, routes/endpoints, config keys, and file paths — to interfaces.md. Tag the entry with task id and commit hash. If a later task changes a contract, append a superseding entry; do not silently leave stale guidance. Record "No public interface changes" when applicable. The relevant ledger slice feeds step 2 of the next task and is reloaded after compaction, keeping Codex from re-inventing or mismatching contracts.

7. Close the task in the tracker with a one-line note. Persist the status, tests, and commit hash to tasks.md before giving the user a brief progress line and moving on.

Step 3 — Finish (after the last task)

1. Full-suite gate one final time: all packages + build + any lint/screenshot scripts the plan names. Green or you don't ship.
2. git pull --rebase then git push. git status must show up-to-date with origin. On rebase conflicts, resolve minimally and re-run the gate — never force-push over shared history.
3. Exactly ONE PR for the whole feature, targeting the branch the plan specifies. Body = plan summary: context, what shipped, test evidence, out-of-scope/follow-ups. No AI attribution anywhere.
   • GitHub: gh pr create. GitLab: glab mr create. No forge CLI: push the branch and give the user the compare URL to open the PR themselves.
4. Close out the tracker (file follow-ups, close finished units).
5. Mark run.md complete and record the final commit, gate evidence, and PR URL.
6. Report: tasks completed/blocked, commit list, PR URL, anything deferred.

Rails

• One task per commit; one PR per plan. No batching, no splitting.
• Disk state is canonical. Chat progress is a rendering of run.md, tasks.md, task allowlists, and interfaces.md, never the only copy of run state.
• Every codex exec gets < /dev/null. With a prompt passed as an argument and an open stdin, codex exec blocks indefinitely on "Reading additional input from stdin…" and never runs. Redirecting stdin from /dev/null gives it immediate EOF so it proceeds. This is the single most common way to make the whole run silently hang.
• Tests before every commit — non-negotiable.
• Effort stays high/xhigh/max. A trivial-mechanical task may drop to high (never lower, never ultra); note it.
• If the plan and the user's live instructions conflict, the user wins — note the deviation in the PR body.
• Secrets never go into briefs, commits, or PR bodies.
• A commit that later tasks built on turned out wrong? Fix forward with a new task/commit. Don't rewrite pushed history.

Driving codex exec (field-tested constraints)

• Codex will run git commit AND git push unbidden. Observed mid-run: a correction round ended "complete and pushed" with nothing in the brief asking for it — one round after the same session reported .git was read-only. Capability varies between rounds; never infer from one round that git is unavailable to it. Put this block, verbatim, at the top of every brief:

  ## ABSOLUTE PROHIBITION — NO GIT WRITES, NO bd, NO xcodegen
  Do NOT run git commit, push, add, stash, checkout, reset, or any writing git command.
  Do NOT run `bd` (it writes to git). Leave your work as uncommitted changes.
  Read-only git (status, diff, log) is fine.
  

  Verify rather than trust: capture git rev-parse HEAD before each Codex call and compare after. A moved HEAD is a hard stop.

• The bd ban holds even on beads repos. Codex will obey the repo's own AGENTS.md and run bd prime, which internally attempts git add and rewrites .beads/issues.jsonl. Tracker writes belong to the orchestrator only.

• codex exec resume rejects -C and -s (error: unexpected argument '-C' found). It inherits cwd — and filters recorded sessions by it — and inherits the resumed session's sandbox. Working shape:

  cd "$REPO_ROOT" && codex exec resume --last "<corrections>" \
    -m "$MODEL" -c model_reasoning_effort="$EFFORT" < /dev/null
  

  The < /dev/null is still mandatory on resume.

• xcodebuild is impossible inside Codex's sandbox. Xcode's SwiftPM local-package resolution spawns a nested sandbox-exec the outer sandbox refuses (sandbox_apply: Operation not permitted); swift test works. Treat every Codex build claim as unverified — the orchestrator runs the build gate. Say so in the brief so Codex doesn't waste rounds retrying with cache redirection.

• Exit 144 means resume, don't restart. A codex exec run killed mid-stream (e.g. a usage-limit kill; log ends mid-diff with no error banner) exits 144 with the working tree intact. Grep the log head for session id: and codex exec resume <session-id> from the repo cwd — no -C/-s, < /dev/null still required.

• Codex's sandbox denies Go module cache writes. go test fails with open ~/go/pkg/mod/cache/download/....lock: operation not permitted — an environment failure that reads like a code failure and triggers identical patch retries. Brief Codex to set a workspace-local GOMODCACHE=$PWD/.gomodcache, or run the test gate yourself outside the sandbox.

See also

• scripts/check_scope.py — executable file-scope gate.
• references/codex-brief.md — the Codex brief skeleton (the highest-leverage part).
• references/plan-example.md — the shape of an input plan.
• references/walkthrough.md — a full trace of one task, start to finish.