LLM Codex Readiness Unit Test

LLM Codex Readiness Unit Test

Instruction-first, in-session "readiness" for evaluating AGENTS/​PLANS documentation quality without any external APIs or SDKs. All checks run against the current working directory (cwd), with no monorepo discovery. Each run writes to .codex-readiness-unit-test/<timestamp>/ and updates .codex-readiness-unit-test/​latest.json. Keep execution deterministic (filesystem scanning + local command execution only). All LLM evaluation happens in-session and must output strict JSON via the provided references.

Quick Start

1. Collect evidence:
   • python skills/​codex-readiness-unit-test/​bin/​collect_evidence.py
2. Run deterministic checks:
   • python skills/​codex-readiness-unit-test/​bin/​deterministic_rules.py
3. Run LLM checks using references in references/ and store .codex-readiness-unit-test/<timestamp>/​llm_results.json.
4. If execute mode is requested, build a plan, get confirmation, run:
   • python skills/​codex-readiness-unit-test/​bin/​run_plan.py --plan .codex-readiness-unit-test/<timestamp>/​plan.json
5. Generate the report:
   • python skills/​codex-readiness-unit-test/​bin/​scoring.py --mode read-only|execute

Outputs (per run, under .codex-readiness-unit-test/<timestamp>/):

• report.json
• report.html
• summary.json
• logs/​* (execute mode)

Runbook

This skill produces a deterministic evidence file plus an in-session LLM evaluation, then compiles a JSON report and HTML scorecard. It requires no OpenAI API key and makes no external HTTP calls.

Minimal Inputs

• mode: read-only or execute (required)
• soft_timeout_seconds: optional (default 600)

Modes (Read-only vs Execute)

• Read-only: Collect evidence, run deterministic rules, and run LLM checks #3–#5. No commands are executed, check #6 is marked NOT_RUN, and no execution logs/​summary are produced.
• Execute: Everything in read-only plus a confirmed plan.json is executed via run_plan.py. This enables check #6 and produces execution logs + execution_summary.json for scoring.

Always ask the user which mode to run (read-only vs. execute) before proceeding.

Check Types

• Deterministic: filesystem-only checks (#1 AGENTS.md exists, #2 PLANS.md exists, #3 AGENTS.md <= 300 lines, #4 config.toml exists at repo root, repo .codex/, or user .codex/)
• LLM: in-session Codex evaluation (#3 project context, #4 commands, #5 loops; commands may live in AGENTS or referenced skills)
• Hybrid: deterministic execution + LLM rationale (#6 execution)

Skill references are discovered from AGENTS.md via $SkillName or .codex/​skills/<name> patterns; their SKILL.md files are added to evidence for the LLM checks.

All checks run relative to the current working directory and are defined in skills/​codex-readiness-unit-test/​references/​checks/​checks.json, weighted equally by default. Each run writes outputs to .codex-readiness-unit-test/<timestamp>/ and updates .codex-readiness-unit-test/​latest.json. The helper scripts read .codex-readiness-unit-test/​latest.json by default to locate the latest run directory.

Strict JSON + Retry Loop (Required)

For each LLM/​HYBRID check:

1. Run the specialized prompt expecting strict JSON.
2. If JSON is invalid or missing keys, run skills/​codex-readiness-unit-test/​references/​json_fix.md with the raw output.
3. Retry up to 2 additional attempts (max 3 total).
4. If still invalid: mark the check as WARN with rationale: "Invalid JSON from evaluator after retries".

The JSON schema is:

{
  "status": "PASS|WARN|FAIL|NOT_RUN",
  "rationale": "string",
  "evidence_quotes": [{"path":"...","quote":"..."}],
  "recommendations": ["..."],
  "confidence": 0.0
}

Single Confirmation (Required)

Combine the command summary and execute plan into one concise confirmation step. Present:

• The extracted build/​test/​dev loop commands (human-readable, labeled).
• The planned execute details (cwd, ordered commands, soft timeout policy, env). Ask for a single confirmation to proceed. Do not paste raw JSON, full evidence, or the full plan.json. If declined, mark execute-required checks as NOT_RUN.

Required Files

• .codex-readiness-unit-test/<timestamp>/​evidence.json (from collect_evidence.py)
• .codex-readiness-unit-test/<timestamp>/​deterministic_results.json (from deterministic_rules.py)
• .codex-readiness-unit-test/<timestamp>/​llm_results.json (from in-session references)
• .codex-readiness-unit-test/<timestamp>/​execution_summary.json (execute mode only)
• .codex-readiness-unit-test/<timestamp>/​report.json and .codex-readiness-unit-test/<timestamp>/​report.html (from scoring.py)
• .codex-readiness-unit-test/<timestamp>/​summary.json (structured pass/​fail summary from scoring.py)
• .codex-readiness-unit-test/​latest.json (stable pointer to the latest run directory)

Prompt Mapping

• #3 project_context_specified → skills/​codex-readiness-unit-test/​references/​project_context.md
• #4 build_test_commands_exist → skills/​codex-readiness-unit-test/​references/​commands.md
• #5 dev_build_test_loops_documented → skills/​codex-readiness-unit-test/​references/​loop_quality.md
• #6 dev_build_test_loop_execution → skills/​codex-readiness-unit-test/​references/​execution_explanation.md

plan.json schema (execute mode)

{
  "project_dir": "relative/​or/​absolute/​path (optional)",
  "cwd": "optional/​absolute/​path (defaults to current directory)",
  "commands": [
    {"label": "setup", "cmd": "npm install"},
    {"label": "build", "cmd": "npm run build"},
    {"label": "test", "cmd": "npm test"}
  ],
  "env": {
    "EXAMPLE": "value"
  }
}

Place plan.json inside the run directory (e.g., .codex-readiness-unit-test/<timestamp>/​plan.json).

llm_results.json schema

{
  "project_context_specified": {"status":"PASS","rationale":"...","evidence_quotes":[],"recommendations":[],"confidence":0.7},
  "build_test_commands_exist": {"status":"PASS","rationale":"...","evidence_quotes":[],"recommendations":[],"confidence":0.7},
  "dev_build_test_loops_documented": {"status":"WARN","rationale":"...","evidence_quotes":[],"recommendations":[],"confidence":0.6},
  "dev_build_test_loop_execution": {"status":"PASS","rationale":"...","evidence_quotes":[],"recommendations":[],"confidence":0.6}
}

Scoring Rules

• PASS = 100% of weight
• WARN = 50% of weight
• FAIL/​NOT_RUN = 0%
• Overall status: FAIL if any FAIL; else WARN if any WARN or NOT_RUN; else PASS.

Safety + Timeouts

• Denylisted commands are not executed and marked FAIL.
• Soft timeout defaults to 600s; hard cap defaults to 3x soft timeout.
• Execution logs are written to .codex-readiness-unit-test/<timestamp>/​logs/.