Name one end state that output shows: a test summary line, an exit code, a count, an empty queue. "The code is clean" and "tests pass" name no output, so the agent's own claim meets them.
Name the repository's own check command (just check, bun test), read from its justfile, package.json, or README, and require its output after the last edit. An earlier green run is stale evidence.
Bound every /goal with a turn count: or stop after 15 turns. /goal has no built-in cap.
Add a constraint that shows the check was not gamed, tied to a command: git diff --stat -- tests shows no changes. Otherwise deleting or skipping the failing test meets the goal. Never weaken the check to reach the end state; replace the goal or /goal clear instead.
Use a Stop hook when the check must really run (a /goal evaluator runs nothing and reads only the transcript), when the host is Codex, or when the gate must persist across sessions.
Bound every Stop hook's retries. stop_gate.py blocks at most --max-retries times in a row (default 3), counted per session while stop_hook_active is true, and starts over on a new stop. A gate that blocks on every stop_hook_active: true payload loops forever on Codex, which has no cap; Claude Code caps continuations at 8, so a higher value has no effect there.
Stop output on both hosts: {"decision": "block", "reason": "..."} on stdout with exit 0 to keep the agent working, {} to let it stop. Codex rejects plain text from a Stop hook as a hook error and ends the turn, so every path prints JSON.
Put the check's failing output in reason. The agent acts on the reason, so "tests failed" alone sends it back blind.
Set the hook timeout above the check's own timeout. A timed-out hook does not block, so the agent stops unchecked.
Merge one Stop entry into the existing hooks file and keep every other entry. Never rewrite the file from a template. Read the file first and show the user the diff.
Codex runs a new or edited hook only after the user reviews it in /hooks, in a trusted project. Say so, because you cannot run that step.

Workflow

Find the check command and run it once to learn its output and exit code.
Choose the form. Claude Code with an interactive user and no need to run the check: /goal. Otherwise a Stop hook.

For /goal, write one condition:

/goal `just check` exits 0 in the transcript after the last edit, `git diff --stat -- tests`
shows no changes, or stop after 15 turns

Read it as the evaluator: could a false claim in the transcript satisfy it?

For a Stop hook, copy scripts/stop_gate.py into the project (.claude/hooks/ or .codex/hooks/) so the entry does not depend on where the skill is installed, then add one entry to hooks.Stop in the existing file. Claude Code, .claude/settings.local.json (or .claude/settings.json if the user wants it shared), exec form so paths with spaces work:
```
{"hooks": [{"type": "command", "command": "python3",
  "args": ["${CLAUDE_PROJECT_DIR}/.claude/hooks/stop_gate.py", "--check", "just check",
           "--timeout", "300"],
  "timeout": 330}]}
```
Codex, .codex/hooks.json:
```
{"hooks": [{"type": "command",
  "command":
  "python3 \"$(git rev-parse --show-toplevel)/.codex/hooks/stop_gate.py\" --check 'just check'",
  "timeout": 150}]}
```
The hook timeout stays above the gate's --timeout (default 120 seconds).
Prove the hook with payloads piped to the exact entry command, with TMPDIR set to a scratch directory so the retry counter stays out of the real one. A Stop payload needs hook_event_name, session_id, cwd, and stop_hook_active. A failing check with stop_hook_active false prints a block with the output tail and Retry 1 of 3; a passing check prints {}; a failing check sent three more times with stop_hook_active true blocks twice, then prints {}.
Report the condition or entry, each payload result, how to remove it (/goal clear, or delete the one entry), and each step not run in the host.

Scripts

python3 scripts/stop_gate.py --check CMD [--timeout S] [--max-retries N] blocks Stop with the check's output tail until CMD exits 0, at most N times in a row (default 3), then lets the agent stop. It also lets the agent stop when its counter file in the temp directory cannot be written. Exit 0, or 2 for a non-Stop event, an empty or unclosed-quote CMD, or N below 1. On Windows, use py -3. Test: test_stop_gate.py.

References

Read references/goal-conditions.md when a /goal needs session behavior (resume, pause on limits, permission mode), more example conditions, or a repair for a goal that never ends.

write-agent-stop-condition

Write Agent Stop Condition

Rules