Create tracked proposal entries for software repositories. Use when the user asks Codex to turn ideas, bugs, review notes, TODOs, errors, feature requests, implementation/documentation drift, or product discussions into proposals.p0.md, proposals.p1.md, proposals.p2.md, proposals.feat.md, docs.updates.md, or similar local proposal files; investigate when needed, deduplicate existing entries, classify severity or recommendation, assign the next id, and write the proposal without starting implementation.
Distill reusable experience (pitfalls, optimal task trajectories, dead-ends, emergent preferences) from the current conversation into the project memory store at `.agents/memory/`. Use when the user asks to 总结经验 / 记下这次的经验 / distill / capture lessons learned / 把刚才的踩坑或方法记下来 / 提炼经验沉淀; after finishing a task that involved real trial-and-error or non-obvious gotchas; or when the user wants to persist what was learned so a future agent does not pay the same exploration cost. Drafts candidates, adversarially self-audits each (truly non-obvious? correct? will it mislead a future agent?), drops the unqualified ones, presents survivors for user confirmation, and writes only the confirmed ones.
Plan and implement tracked proposal items in software repositories. Use when the user asks Codex to fix or implement items from proposals.p0.md, proposals.p1.md, proposals.p2.md, proposals.feat.md, docs.updates.md, or similar proposal and documentation-update documents; decide implementation scope, clarify ambiguous requirements, coordinate subagents when useful, modify code, validate changes, and mark proposals as fixed or completed.
Index and maintain local reference repository collections such as ref_repos/, references/repos/, third_party_repos/, or similar directories. Use when Codex needs to create or update INDEX.md, INDEX.REPO.{repo-slug}.md, and INDEX.TOPIC.{topic-slug}.md files that explain what each reference repository contains, how its architecture works, what files/directories/commands/agents/skills are worth reading, which topics it informs, and how future agents should explore the collection. Supports initial indexing, incremental indexing of newly added local repos, manual refresh of explicitly named repo indexes, and topic aggregation without Git synchronization or automatic stale checks.
**READING SKILL** — Generate a hierarchical outline index (with line-number ranges) for a long Markdown file, then use it to navigate precisely to sections of interest. USE FOR: reading/understanding large .md documents; answering questions about specific sections without loading the entire file; exploring document structure. DO NOT USE FOR: editing Markdown; non-.md files.
Facilitate a patient conversation to review, clarify, and align AI-generated proposal files with the user's real intent before editing them. Use when the user wants to review or revise proposals.p0.md, proposals.p1.md, proposals.p2.md, proposals.feat.md, docs.updates.md, or similar tracked proposal documents; when they say proposals feel wrong, too severe, too vague, not aligned, duplicated, missing context, or need discussion; or when proposal changes should be driven by explanation, questioning, user answers, and iterative alignment rather than direct implementation.
Systematic review and proposal workflow for software repositories. Use when the user asks Codex to review code against PRD/spec/docs, audit a newly written codebase, find logic errors, implementation drift, security or reliability risks, fake/simple/lazy implementations, coordinate parallel subagents, write findings to proposals.p0.md/proposals.p1.md/proposals.p2.md, write feature ideas to proposals.feat.md, and record richer implementation or documentation differences in docs.updates.md.