Research
Use this Codex skill for prior-art research and assumption-retirement work.
The canonical workflow is the Claude research skill at
.claude/skills/research/SKILL.md; follow it, with the Codex-specific
adjustments below.
Codex Adjustments
- Treat
AGENTS.mdandCLAUDE.mdas binding project guidance. - Use
rg,rg --files,sed,cargo doc,cargo expand, and focused file reads instead of Claude-onlyRead/Grep/Exploretool names. - Use Codex subagents only when the user explicitly asks for delegated or parallel agent work. Otherwise perform the research locally.
- Place memos under
docs/research/. - Update
docs/research/index.mdif present; otherwise updatedocs/index.md.
Required Output
Produce exactly one memo for each research topic:
spike-<slug>.mdfor a single time-boxed question with a runnable artifact.study-<slug>.mdfor a deep dive into vendored source.survey-<slug>.mdfor web/docs research where vendoring is not warranted.
Each memo must state the question, evidence, decision, risks, and cited source locations. For vendored repositories, pin and record the exact commit SHA.
Quality Bar
A memo is done when another engineer can answer the question without opening the upstream repo and can verify every behavioral claim through cited file paths, line numbers, command output, or official source links.