| name | monozukuri-codex-readiness |
| description | Verify whether Monozukuri can run with Codex in loop or full-auto workflows by checking agent availability, skill discovery, dry-run behavior, and the targeted Bats validation path. Use for Monozukuri Codex readiness, loop, agent doctor, or skill availability questions. Do not use for unrelated repositories or generic CI triage. |
Monozukuri Codex Readiness
Verify current behavior with commands before answering whether Codex, skills, agents, or loop/full-auto workflows are usable.
Preflight
Run from the Monozukuri repository root.
- Capture worktree state with
git status --short; do not stage broad paths in dirty checkouts.
- Check Codex availability:
./orchestrate.sh agent doctor codex
codex --version
- Check agent selection and setup state:
./orchestrate.sh agent list
./orchestrate.sh setup status --agent codex
- Check skill discovery if the question mentions skills:
scripts/skill-discovery.sh
If that script is unavailable or its interface has changed, use the repository's current discovery command and state the substitution.
Dry Run
Use a non-interactive dry run to verify orchestration without spending model calls:
SKILL_TIMEOUT_SECONDS=180 ./orchestrate.sh run --dry-run --non-interactive --no-ui
Report exact failures. Do not soften shell errors as "environment issues" until the blocking command and stderr are identified.
Known Failure Checks
When readiness fails, check these common regressions before proposing a new path:
- macOS Bash 3.2 incompatibility such as
${var^^}; prefer tr '[:lower:]' '[:upper:]'.
set -u failures from empty arrays such as ${flags[@]}.
- Adapter output piping that loses error-envelope writes; temp-file capture is safer for fallback
adapter_tee.
- Tests that accidentally read the user's global Claude configuration; isolate
CLAUDE_CONFIG_DIR when testing Claude adapter behavior.
make test wrappers that hide the true Bats exit code; prefer direct bats ... when a trustworthy result matters.
Validation
Choose the narrowest validation path that proves the claim:
- Syntax/static checks for shell edits:
bash -n cmd/run.sh cmd/setup.sh lib/agent/adapter-codex.sh lib/agent/thin-shell-base.sh scripts/agent-discovery.sh
shellcheck --severity=error <changed-shell-files>
git diff --check
- Loop command behavior:
bats test/integration/loop_command.bats
- Full local confidence:
bats test/unit test/integration test/conformance test/properties
Keep local validation separate from remote PR checks. If GitHub status is relevant, inspect it separately and report pending checks distinctly from local test results.
Output
Return:
- Readiness verdict:
ready, partially ready, or blocked.
- Evidence from each command, including versions, discovered skills count if available, and dry-run result.
- Fixes applied or the smallest next fix needed.
- Exact validation commands and outcomes.
- Any assumptions made for
MONOZUKURI_AUTONOMY=full_auto.
In full_auto, never ask for human input. Make a defensible default, record it under open questions or assumptions, and continue.