Skip to content

05-codex-goal.md

Latest commit

 

History

History
84 lines (62 loc) · 4.68 KB

05-codex-goal.md

File metadata and controls

84 lines (62 loc) · 4.68 KB
Raw

05 — Codex: The Goal Loop Without /goal

Codex CLI has no /goal and no separate verifier model. You get the same machine in fifteen lines of bash: codex exec does the work, your verifier makes the done/not-done call.

When to use

You're on OpenAI, not Anthropic. Everything else tonight still applies — the loop is easy, the verifier is the game. This file gives you the loop; ../verifiers/ gives you the game.

The pattern

Install and authenticate:

curl -fsSL https://chatgpt.com/codex/install.sh | sh    # or: npm install -g @openai/codex
codex login                                             # sign in with a ChatGPT plan
# or with an API key:
printenv OPENAI_API_KEY | codex login --with-api-key
codex login status

The verifier-gated loop. Save as codex-goal.sh in your project:

#!/usr/bin/env bash
# codex-goal.sh — /goal, reimplemented in bash. Your verifier is the judge.
set -u
PROMPT="Make every test in this repo pass. Run 'npm test' yourself, fix real causes. Do not weaken or delete tests."
VERIFY="bash verify.sh"     # exits 0 = done. Copy from ../verifiers/templates/, edit.
MAX_ITER=25

for i in $(seq 1 "$MAX_ITER"); do
  if OUT=$($VERIFY 2>&1); then
    echo "verifier green after $((i-1)) iterations"; exit 0
  fi
  if [ "$i" -eq 1 ]; then
    codex exec -s workspace-write "$PROMPT"
  else
    codex exec resume --last -c 'sandbox_mode="workspace-write"' \
      "The verifier still fails. Last 20 lines of its output:
$(printf '%s\n' "$OUT" | tail -20)
Fix the cause, then stop."
  fi
done
$VERIFY && { echo "verifier green at the wire"; exit 0; }
echo "MAX_ITER=$MAX_ITER reached, verifier still red" >&2
exit 1

One flag trap, verified against codex 0.142: codex exec takes -s workspace-write, but codex exec resume rejects -s — sandbox mode goes through -c 'sandbox_mode="workspace-write"' on resume. The script above handles both. ../hypergraph/scripts/codex_exec.sh is the hypergraph's adapter for executor=codex and exists so this divergence is fixed in exactly one place — read it if you want thread IDs, token counts, and clean error handling.

Two variants worth knowing:

  • Fresh context per iteration (the Ralph pattern): replace the resume --last branch with a plain codex exec "$PROMPT ...". Costs more tokens, beats context rot on long runs.
  • Machine-readable output: add --json for a JSONL event stream, or -o last-message.txt to capture the final agent message.

Run it

bash verify.sh; echo $?      # must be non-zero NOW, for the right reason
chmod +x codex-goal.sh
tmux new -s codex 'bash codex-goal.sh'   # detach: Ctrl-b d

Watch the first iteration complete before you walk away. For overnight, run it on a machine that stays awake — see 04-background-agent.md, Option C; everything there applies here unchanged. codex cloud exec can submit server-side tasks, but it's marked experimental in the CLI — verify it works on your account before trusting it with the night.

Parity gaps

Honest accounting, one line each:

  • No /goal: there is no separate verifier model making the done call — your script is the only judge, which is why it must be mechanical.
  • No Stop hook: nothing stops Codex from ending a turn early with work remaining — the outer bash loop is your persistence layer.
  • No budget flag: codex exec has no --max-budget-usd equivalent — set the spend cap in the OpenAI platform dashboard (../safety/cost-caps.md) and wrap the loop in ../verifiers/watchdog.sh.
  • Cloud/background execution (codex cloud) is experimental; tmux plus codex exec is the reliable overnight path today.

Failure modes

  • The model grades its own work. The prompt tells Codex to run tests; fine. But the loop's exit is $VERIFY running outside the session. Never let the agent's "all tests pass now" be the gate.
  • resume --last grabs the wrong session. Run two loops from the same directory and each --last resumes whichever finished most recently. One loop per directory, or capture the thread ID (codex_exec.sh prints thread_id: for exactly this).
  • Sandbox blocks the writes. Without -s workspace-write (or the -c sandbox_mode form on resume), Codex may be unable to edit files — the loop burns iterations producing nothing.
  • 25 iterations of a broken verifier. If verify.sh can never exit 0 — wrong path, missing dependency — you pay for MAX_ITER attempts at an impossible goal. Run the verifier by hand first.

Verifier hookup

Verifiers are provider-neutral bash: every template in ../verifiers/templates/ drops into VERIFY= above unchanged. Wrap the whole loop in ../verifiers/watchdog.sh for spend, runtime, and stall kills — it doesn't care which CLI is inside.