LLM Codex Readiness Integration Test

This skill runs a multi-stage integration test to validate agentic execution quality. It always runs in execute mode (no read-only mode).

Outputs

Each run writes to .codex-readiness-integration-test/<timestamp>/ and updates .codex-readiness-integration-test/latest.json.

New outputs per run:

• agentic_summary.json and logs/agentic.log (agentic loop execution)
• llm_results.json (automatic LLM evaluation)
• summary.txt (human-readable summary)

Pre-conditions (Required)

• Authenticate with the Codex CLI using the repo-local HOME before running the test.
  Run these in your own terminal (not via the integration test):
  HOME=$PWD/.codex-home XDG_CACHE_HOME=$PWD/.codex-home/.cache codex login
  HOME=$PWD/.codex-home XDG_CACHE_HOME=$PWD/.codex-home/.cache codex login status
• The integration test creates {repo_root}/.codex-home and {repo_root}/.codex-home/.cache/codex as its first step.

Workflow

0. Ask the user how to source the task.
   • Offer two explicit options: (a) user provides a custom task/prompt, or (b) auto-generate a task.
   • Do not run the entry point until the user chooses one option.
1. Generate or load {out_dir}/prompt.pending.json.
   • Use the integration test's expected prompt path, not prompt.json at the repo root.
   • With the default out dir, this path is .codex-readiness-integration-test/prompt.pending.json.
   • If --seed-task is provided, it is used as the starting task.
   • If not provided, generate a task with skills/codex-readiness-integration-test/references/generate_prompt.md and save the JSON to {out_dir}/prompt.pending.json.
   • The user must approve the prompt before execution (no auto-approve mode). Make sure to output a summary of the prompt when asking the user to approve.
2. Execute the agentic loop via Codex CLI (uses AGENTS.md and change_prompt).
3. Run build/test commands from the prompt plan via skills/codex-readiness-integration-test/scripts/run_plan.py.
4. Collect evidence (evidence.json), deterministic checks, and run automatic LLM evals via Codex CLI.
5. Score and write the report + summary output.

Configuration

Optional fields in {out_dir}/prompt.pending.json:

• agentic_loop: configure Codex CLI invocation for the agentic loop.
• llm_eval: configure Codex CLI invocation for automatic evals.

If these fields are omitted, defaults are used.

Requirements

• The LLM evaluator must fail if evidence mentions the phrase Context compaction enabled.
• Use qualitative context-usage evaluation (no strict thresholds).

What this test covers well

• Runs Codex CLI against the real repo root, producing real filesystem edits and git diffs.
• Executes the approved change prompt and then runs the build/test plan in-repo.
• Captures evidence, deterministic checks, and LLM eval artifacts for review.

What this test does not represent

• The agentic loop may use non-default flags (e.g., bypass approvals/sandbox), so interactive guardrails differ.
• Uses a dedicated HOME (.codex-home), which can change auth/config/cache vs normal CLI use.
• Auto-generated prompts and one-shot execution do not simulate interactive guidance.
• MCP servers/tools are not exercised unless explicitly configured.

Notes

• The prompts in skills/codex-readiness-integration-test/references/ expect strict JSON.
• Use skills/codex-readiness-integration-test/references/json_fix.md to repair invalid JSON output.
• This skill calls the codex CLI. Ensure it is installed and available on PATH, or override the command in {out_dir}/prompt.pending.json.
• If the agentic loop detects sandbox-blocked tool access, it now writes requires_escalation: true to {run_dir}/agentic_summary.json and exits with code 3. Re-run the integration test with escalated permissions in that case.