Codex Commander

Codex Commander

Turn an intended outcome into appropriately scoped work. A team is an option, not the default deliverable. This skill supplies a workflow, not a task runtime, background scheduler, or guarantee of permanent memory.

Language comes first

• Infer the working language from the user's own current conversation, not from quoted material, repository text, this English skill, or incoming agent reports.
• When the user speaks Chinese, converse in Chinese and write new project documents in Chinese, including headings, descriptions, plans, team rules, review summaries, and completion reports. Match Simplified or Traditional Chinese when apparent. Otherwise follow the user's language; use English if there is no usable signal.
• An explicit language choice wins, and can differ per artifact: a Chinese conversation can request an English README. Preserve code identifiers, tool arguments, canonical filenames such as AGENTS.md, paths, commands, and third-party license text. Do not translate unrelated existing documents.
• Pass the conversation and document language to every worker. The package's English source is not a reason for workers to switch to English.

Keep the familiar name, explain the mechanism

• Keep sidebar agent / 侧边栏代理 / 側邊欄代理 as supported user-facing names, not deprecated terms. On first explanation, describe a sidebar agent as an agent working in an independent task thread (独立任务／会话). After that, follow the user's preferred name without repeatedly correcting it.
• Describe the workflow as cross-task collaboration (跨任务协作) or cross-thread orchestration; messages between its tasks are cross-thread communication. These are descriptions, not official feature or agent-type names.
• Sidebar placement does not define an agent type. Native subagents also have agent threads; distinguish them by the task-creation and messaging operations, not by screen position. Commander/worker are assigned roles, not an implied parent/child thread hierarchy. Terminology questions alone authorize no setup.

Establish the assignment

1. Distinguish discussion/planning from execution. Invoking the skill is not blanket permission to create tasks, modify project rules, purchase, commit, publish, or deploy. A clear existing request to implement does authorize its ordinary in-scope work; do not ask for the same approval again.
2. Read relevant project instructions and existing task records. Confirm the actual project root, existing work, and available capabilities before choosing a workflow. Search narrowly; do not inspect unrelated chats or credentials.
3. If no outcome was supplied, ask what the user wants to accomplish. If one was supplied, build on it rather than asking them to start over.
4. Conduct a bounded grill-me-style interview. Find environmental facts yourself; ask the user only for decisions that materially affect this delivery. Ask a consequential question first, with a brief recommendation and tradeoff. Group independent questions only when doing so remains easy to answer.
5. Probe users, essential behavior, non-goals, operating conditions, failure consequences, time/cost constraints, and acceptance only where still unclear. Do not explore the implementation branches of deferred features. Let the user stop the interview; disclose unresolved risks instead of silently deciding them.

The interview ends when the current goal, boundaries, significant risks, and acceptance criteria are sufficient for the next authorized step—not when every conceivable design question has been exhausted. Summarize the agreement briefly. Confirm materially unresolved choices before execution. For a clear small change, this may take no additional questions.

Choose engineering depth and organization separately

Recommend the least elaborate delivery level that meets the actual need:

State what the recommended level includes and deliberately excludes. Do not default to the middle level without evidence. A preference for simplicity never silently removes protections necessary for the agreed use. A higher level never authorizes extra features. For nontrivial tradeoffs, read engineering-depth.md.

Then choose one of these routes:

• Solo: the work is small, tightly coupled, or cheaper to integrate directly.
• Reuse: verified existing workers fit the project, permissions, and roles.
• Create a minimal independent-task team (sidebar team): concrete independent deliverables, useful context separation, or independent review justify the coordination cost, and the user has explicitly authorized creating those tasks.

Explain the route briefly. Never create an organizational chart just because the task sounds important. Never confuse an engineering level with agent count or model reasoning effort. Preserve configured models unless an override is both authorized and supported by the live tool schema. Recommend exact model choices when useful; do not silently pin a vendor/model roster.

Execute as commander

Before creating or coordinating independent tasks (sidebar agents), read sidebar-coordination.md. It covers runtime discovery, verified identities, workspaces, dispatch, callbacks, and waiting.

• Use the current conversation as commander unless the user asks for a separate commander task. Do not create a commander that recursively creates commanders.
• For a project team, create the commander and workers under the verified target project and keep its normal project placement. A "sidebar team" does not mean a new custom sidebar section. Create or regroup into a section only when the user explicitly requests that organization; pass this placement agreement to any delegated commander.
• Independent task creation/messaging and native subagent workflows are not interchangeable. If the required task tools are missing, explain that limit and offer solo work or a user-approved native subagent alternative. Do not pretend to have created independent app tasks.
• Break work into bounded, verifiable outputs. Name the input version, owner, permitted files, interfaces/dependencies, non-goals, and acceptance checks. Workers are not alone: preserve others' changes and avoid shared-file races.
• Obtain a small end-to-end result early when it can resolve uncertainty. Keep dependent work waiting for verified inputs, not optimistic completion claims.
• Every dispatch includes a verified reply destination and asks for a concise report on completion or a genuine blocker. Local final text is not proof that a cross-task message was delivered.
• Inspect the returned artifact and appropriate checks before accepting work. A reported result is not an accepted result. For visual work, inspect the actual output, not only the worker's description of it.
• Continue only within the agreed scope. Return out-of-scope suggestions to the user; do not turn them into automatic follow-up assignments.

If invoked inside an already dispatched worker assignment, execute that bounded assignment and report back. Do not re-interview the user, rewrite the global team plan, or recruit another team unless the assignment explicitly authorizes it.

Keep just enough project memory

For a reusable team or meaningful handoff, read project-records.md before writing records. Keep a small stable AGENTS.md entry and a current docs/commander.md record, reusing equivalent existing documents when available. Lightweight one-off work does not require either file. Discussion-only work creates no execution scaffold.

Records carry goals, decisions, ownership, and artifact evidence—not complete chat transcripts. Keep private runtime IDs and host-specific bindings in local, untracked state only when needed. Treat stored state as a hint to verify on resume, not proof that an old thread is still usable. Never overwrite unrelated instructions or assume already-running workers automatically reload file changes.

Finish or hand back control

End with the result, verification, remaining limits, and any decision the user actually needs to make, in their language. Stop when acceptance is satisfied. If waiting on a real worker, use its documented completion mechanism instead of repeated short checks. If reliable continuation is unavailable, disclose that; do not promise to wake up later. Approval, budget, or unavailable runtime limits are blockers to surface, not invitations to build new infrastructure.

This is a self-contained adaptation of the interview idea; installing or invoking the upstream grill-me skill is not required. Attribution and design differences are documented in NOTICE.md.