Codex Readonly Analyze

Answer repository questions through grounded read-only analysis. This is the native Codex adaptation of OMX $analyze: keep ranked synthesis, explicit confidence, concrete file references, and evidence-vs-inference separation; omit OMX runtime routing.

Contract

• Do not edit files.
• Do not turn the answer into an implementation plan.
• Do not silently switch to fixing.
• Do not overclaim certainty.
• Do not ask the user for facts that are discoverable from the repo.
• If the answer needs current external facts, say so and use the appropriate research path.

Workflow

1. Restate the question in one sentence.
2. Identify the smallest likely evidence surface: files, configs, tests, generated artifacts, docs, task cards, or logs.
3. Read direct evidence first with rg, rg --files, and focused file reads.
4. Expand only when competing explanations remain plausible.
5. Rank explanations by support.
6. Separate evidence, inference, and unknowns.
7. Return a concise synthesis with file references and confidence.

Parallel Exploration

Use native subagents only when the question is broad enough to benefit from bounded independent lanes. Each lane must be read-only and answer one concrete sub-question. The main thread remains responsible for synthesis.

Output Shape

• Question
• Ranked synthesis
• Evidence
• Inference
• Unknowns / limits
• Next discriminating read-only probe, if useful

For simple one-file questions, answer directly after enough reading.

References

Read references/analysis-quality-checklist.html for the review checklist.