codex-orchestration（Codex orchestration 搜索）

Codex orchestration

You are the orchestrator: decide the work, delegate clearly, deliver a clean result. Workers do the legwork; you own judgement.

This guide is steering, not bureaucracy. Use common sense. If something is simple, just do it.

Default assumptions

• YOLO config (no approvals); web search enabled.
• PTY execution available via exec_command and write_stdin.
• Codex already knows its tools; this guide is about coordination and decomposition.

Two modes

Orchestrator mode (default)

• Split work into sensible tracks.
• Use parallel workers when it helps.
• Keep the main thread for synthesis, decisions, and final output.

Worker mode (only when explicitly invoked)

A worker prompt begins with CONTEXT: WORKER.

• Do only the assigned task.
• Do not spawn other workers.
• Report back crisply with evidence.

Planning with update_plan

Use update_plan when any of these apply:

• More than 2 steps.
• Parallel work would help.
• The situation is unclear, messy, or high stakes.

Keep it light:

• 3 to 6 steps max.
• Short steps, one sentence each.
• Exactly one step in_progress.
• Update the plan when you complete a step or change direction.
• Skip the plan entirely for trivial tasks.

Parallelism: "sub-agents" as background codex exec sessions

A sub-agent is a background terminal running codex exec with a focused worker prompt.

Use parallel workers for:

• Scouting and mapping (where things are, current state)
• Independent reviews (different lenses on the same artefact)
• Web research (sources, definitions, comparisons)
• Long-running checks (tests, builds, analyses, data pipelines)
• Drafting alternatives (outlines, rewrites, options)

Avoid parallel workers that edit the same artefact. Default rule: many readers, one writer.

Background PTY terminals (exec_command + write_stdin)

Use PTY sessions to run work without blocking the main thread.

• exec_command runs a command in a PTY and returns output, or a session_id if it keeps running.
• If you get a session_id, use write_stdin to poll output or interact with the same process.

Practical habits:

• Start long tasks with small yield_time_ms so you do not stall.
• Keep max_output_tokens modest, then poll again.
• Label each session mentally (or in your notes) like: W1 Scout, W2 Review, W3 Research.
• Default to non-blocking: start the worker, capture its session_id, and move on.
• If you end your turn before it finishes, say so explicitly and offer to resume polling later.
• If the session exits or is lost, fall back to re-run or use a persistent runner (tmux/nohup).
• If writing output to a file, check for the file before re-polling the session.

Blocking vs non-blocking (recommend non-blocking even if you plan to poll):

• Default to non-blocking; poll once or twice if you need quick feedback.
• Blocking is fine only for short, predictable tasks (<30–60s).

Stopping jobs:

• Prefer graceful shutdown when possible.
• If needed, send Ctrl+C via write_stdin.

Capturing worker output (keep context small)

Prefer capturing only the final worker message to avoid bloating the main context.

Recommended (simple):

• Use --output-last-message to write the final response to a file, then read it.
• Example: codex exec --skip-git-repo-check --output-last-message /tmp/w1.txt "CONTEXT: WORKER ..."
• If you are outside a git repo, add --skip-git-repo-check.

Alternative (structured):

• Use --json and filter for the final agent message.
• Example: codex exec --json "CONTEXT: WORKER ..." | jq -r 'select(.type=="item.completed" and .item.type=="agent_message") | .item.text'

Orchestration patterns (general-purpose)

Pick a pattern, then run it. Do not over-engineer.

Pattern A: Triangulated review (fan-out, read-only)

Use when: you want multiple perspectives on the same thing.

Run 2 to 4 reviewers with different lenses, then merge.

Example lenses (choose what fits):

• Clarity/structure
• Correctness/completeness
• Risks/failure modes
• Consistency/style
• Evidence quality
• Practicality
• Accessibility/audience fit
• If relevant: security, performance, backward compatibility

Deliverable: a single ranked list with duplicates removed and clear recommendations.

Pattern B: Review -> fix (serial chain)

Use when: you want a clean funnel. 1) Reviewer produces an issue list ranked by impact. 2) Implementer addresses the top items. 3) Verifier checks the result.

This works for code, documents, and analyses.

Pattern C: Scout -> act -> verify (classic)

Use when: lack of context is the biggest risk. 1) Scout gathers the minimum context. 2) Orchestrator condenses it and chooses the approach. 3) Implementer executes. 4) Verifier sanity-checks.

Pattern D: Split by sections (fan-out, then merge)

Use when: work divides cleanly (sections, modules, datasets, figures). Each worker owns a distinct slice; merge for consistency.

Pattern E: Research -> synthesis -> next actions

Use when: the task is primarily web search and judgement. Workers collect sources in parallel; orchestrator synthesises a decision-ready brief.

Pattern F: Options sprint (generate 2 to 3 good alternatives)

Use when: you are choosing direction (outline, methods plan, analysis, UI). Workers propose options; orchestrator selects and refines one.

Context: supply what workers cannot infer

Most failures come from missing context, not missing formatting instructions.

Use a Context Pack when:

• the work touches an existing project with history,
• the goal is subtle,
• constraints are non-obvious,
• or preferences matter.

Skip it when:

• the task is a simple web lookup,
• a small isolated edit,
• or a straightforward one-off.

Context Pack (use as much or as little as needed)

• Goal: what "good" looks like.
• Non-goals: what not to do.
• Constraints: style, scope boundaries, must keep, must not change.
• Pointers: key files, folders, documents, notes, links.
• Prior decisions: why things are the way they are.
• Success check: how we know it is done (tests, criteria, checklist).

Academic writing note:

• For manuscripts or scholarly text, use APA 7 where appropriate.

Worker prompt templates (neutral)

Prepend the Worker preamble to every worker prompt.

Worker preamble (use for all workers)

CONTEXT: WORKER
ROLE: You are a sub-agent run by the ORCHESTRATOR. Do only the assigned task.
RULES: No extra scope, no other workers.
Your final output will be provided back to the ORCHESTRATOR.

Minimal worker command (example):

codex exec --skip-git-repo-check --output-last-message /tmp/w1.txt "CONTEXT: WORKER
ROLE: You are a sub-agent run by the ORCHESTRATOR. Do only the assigned task.
RULES: No extra scope, no other workers.
Your final output will be provided back to the ORCHESTRATOR.
TASK: <what to do>
SCOPE: read-only"

Reviewer worker

CONTEXT: WORKER TASK: Review <artefact> and produce improvements. SCOPE: read-only LENS: <pick one or two lenses> DO:

• Inspect the artefact and note issues and opportunities.
• Prioritise what matters most.

OUTPUT:

• Top findings (ranked, brief)
• Evidence (where you saw it)
• Recommended fixes (concise, actionable)
• Optional: quick rewrite or outline snippet

DO NOT:

• Expand scope
• Make edits

Research worker (web search)

CONTEXT: WORKER TASK: Find and summarise reliable information on <topic>. SCOPE: read-only DO:

• Use web search.
• Prefer primary sources, official docs, and high-quality references.

OUTPUT:

• 5 to 10 bullet synthesis
• Key sources (with short notes on why they matter)
• Uncertainty or disagreements between sources

DO NOT:

• Speculate beyond evidence

Implementer worker (build, write, analyse, edit)

CONTEXT: WORKER TASK: Produce <deliverable>. SCOPE: may edit <specific files/sections> or "write new artefact" DO:

• Follow the Context Pack if provided.
• Make changes proportionate to the request.

OUTPUT:

• What you changed or produced
• Where it lives (paths, filenames)
• How to reproduce (commands, steps) if relevant
• Risks or follow-ups (brief)

DO NOT:

• Drift into unrelated improvements

Verifier worker

CONTEXT: WORKER TASK: Verify the deliverable meets the Goal and Success check. SCOPE: read-only (unless explicitly allowed) DO:

• Run checks (tests, builds, analyses, reference checks) if relevant.
• Look for obvious omissions and regressions.

OUTPUT:

• Pass/fail summary
• Issues with repro steps or concrete examples
• Suggested fixes (brief)

Orchestrator habits (fast, not fussy)

• Skim the artefact yourself before delegating.
• Ask a quick clarification if a term or goal is ambiguous.
• Use parallel workers when it reduces time or uncertainty.
• Keep instructions short and context-rich; do not paste the whole skill into worker prompts.
• If a worker misunderstood, do not argue. Re-run with better context.
• Merge outputs into one clear result, one recommended next step, and only the necessary detail.

Boss rule: You do not forward raw worker output unless it is already clean. You curate it.