Codex Deepsearch

Codex DeepSearch

Use this skill to make DeepSearch fit Codex's actual runtime: threads, workspaces, tools, skills, file edits, git state, progress updates, /goal, validation, and optional delegated workers. The goal is not a long report. The goal is decision-quality evidence that can survive review.

Minimum Contract

Every Codex DeepSearch run must:

1. Frame the decision question, audience, scope, exclusions, privacy boundaries, source priority, and stop conditions before broad searching.
2. Build a scenario matrix for cases where the answer can differ by role, permission, platform, version, region, plan, local/cloud execution, or implementation path.
3. Keep a source ledger and claim ledger before synthesis; read references/source-and-claim-ledgers.md when creating or reviewing them.
4. Label claims as retrieved_fact, source_summary, inference, needs_verification, or conflict.
5. Never treat search snippets, abstracts, tool summaries, model reasoning, or generated prose as retrieved_fact.
6. Separate web/open/browser evidence from local file evidence and user-provided evidence in both ledger and report.
7. Run the review gate in references/review-gate.md before final output or before accepting a worker report.
8. State whether the result can become spec, roadmap, product decision, or implementation input; if not, name the blocking gaps.

Codex Runtime Rules

Read references/codex-runtime-model.md when a task involves local files, /goal, resuming, long-running research, workspace boundaries, or a dirty worktree.

• Treat AGENTS.md and repo docs as maps and constraints, not as proof of external facts.
• Prefer rg, sed, git status, and git diff for local inspection.
• Use apply_patch for manual edits when the user explicitly asks to create research artifacts or update skill files.
• Do not default to writing inside a project repository. Ask or use an explicitly authorized research directory.
• Do not commit, stage, or push unless the user explicitly asks.
• If Python is needed, use uv or an existing project/skill validation command; do not modify system Python.
• In /goal mode, keep durable checkpoints: current phase, accepted evidence, unresolved gaps, next checks, and review status.

Research Workflow

Read references/research-workflow.md for the full flow:

frame -> scenario matrix -> source discovery -> source ledger -> claim ledger
-> counter-evidence -> synthesis -> review gate -> handoff/readiness decision

Use templates when a durable artifact is useful:

• templates/RESEARCH_DESIGN.md
• templates/SOURCE_LEDGER.md
• templates/CLAIM_LEDGER.md
• templates/DEEPSEARCH_REPORT.md
• templates/REVIEW_GATE.md

Tooling And Source Rules

Read references/codex-tooling-rules.md before using web/search/browser/MCP, local files, or delegated workers.

• Web search discovers candidates. Opened pages, fetched docs, rendered PDFs, or verified local files support retrieved_fact.
• Snippet-only sources stay source_summary or needs_verification.
• Local file reads must record path, access date, retrieval method, and whether the file is public/project/private/user-provided.
• Record unavailable tools explicitly: web disabled, browser unavailable, WebFetch unavailable, MiniMax MCP unavailable, login/paywall/region blocked, or local file denied.
• When facts may be current or unstable, browse or use official/current sources and record freshness limits.

Privacy And Repo Safety

Read references/privacy-and-repo-safety.md before inspecting user files or writing artifacts.

• Do not read private or ignored material unless the user explicitly authorizes it for this task.
• Treat workspace/, recordings/, transcripts/, tts/, audio/, secrets, private resumes, salary records, interview logs, .env, databases, and obvious personal exports as off limits by default.
• Check .gitignore and dirty worktree before writing inside any repo.
• Never use git add ..
• Do not upload private local material to remote tools or paste it into web services.

Optional Delegation

Read references/delegation-with-codex-with-cc.md when the user explicitly asks for subagents, worker execution, delegation, codex-with-cc, Claude Code, or when a broad research task needs parallel source scouting or independent review and the user permits it.

• codex-with-cc-plus and Claude Code are optional worker layers, not required dependencies.
• Main Codex owns the research design, task files, acceptance criteria, final synthesis, and review decision.
• Workers gather evidence or review ledgers; they do not decide final acceptance.
• If the user has codex-with-cc-plus, use it to widen search and analysis with bounded researcher/reviewer tasks and exact report requirements.

Output Status

Use:

• DONE: primary/high-quality sources were opened, material claims map to evidence, counter-evidence was checked, and the review gate accepts the result.
• DONE_WITH_CONCERNS: useful result, but tool limits, snippet-only claims, stale sources, unresolved assumptions, or manual checks remain.
• NEEDS_CONTEXT: the decision question, scenario matrix, private-data authorization, or target audience is too unclear.
• BLOCKED: no useful source discovery or permitted local evidence access is possible.

Final reports must make uncertainty visible and must not smooth over missing evidence.