SuperAPI AI Directory

starworkMultiagent

Agents & models

View repository

Updated 10-15 22:41

About this skill

starworkMultiagent

Use this Skill to turn the user's natural-language requests about "common agents", "current session responsibilities", "multi-Agent division of labor", "cross-Agent output sharing", "cross-session instructions", "viewing other lane progress", and "creating Agent teams" into a safe StarWork collaboration workflow.

starworkMultiagent not starwork multiagent the command itself. The Skill is responsible for determining intent, reading the necessary reference, directly calling the host's standard session tools, and recording the real results back to the StarWork project source of truth.The CLI only serves as the StarWork project source of truth, not the executor of host session actions.

If the user is only asking what StarWork is, how to get started, the installation entry point, or which capability to use, return to the starwork main entry point. When the user explicitly mentions multi-Agent, Agent Lanes, lane, cross-session messages, development Agent, product Agent, acceptance Agent, Codex standard session tools, or Codex session control, continue using this Skill.

workflow is a next beta capability; the stable Skill does not guide ordinary users to test workflow. If the user asks about workflow, explain that the next Skill is required or that they should wait for the official release.

Read the context first

Before starting, prioritize reading the following files in the current workspace, reading as many as exist:

AGENTS.md
_系统/上下文/current-projects.md
_系统/上下文/decisions.md
_系统/上下文/product-principles.md
_系统/任务/current-work.md
_系统/协作/agent-lanes.md
_系统/协作/shared.md
_system/context/current-project.md
_system/context/decisions.md
_system/tasks/current-work.md
_system/collaboration/agent-lanes.md
_system/collaboration/shared.md

For Chinese projects use _系统/协作/; for English projects use _system/collaboration/. If the user specifies a target directory, add --target <path>to all CLI commands; otherwise default to the current workspace.

Reference loading rules

When a user request matches a scenario, first read that scenario's reference, then perform the action.

If a reference file does not exist or cannot be read, do not continue executing the corresponding high-risk action; first explain that the Skill installation is incomplete, and ask the user to reinstall StarWork Skills using the complete directory.

Scenario Required reference
Determine user intent references/intent-routing.md
Before any MultiAgent write references/context-and-compatibility.md
Bind / rename / pin / archive / create session references/session-tools.md
Deliver to another lane / Agent / session references/delivery-guarantee.md、references/message-templates.md、references/session-tools.md
Create Agent team references/team-onboarding.md、references/session-tools.md、references/message-templates.md
Read lane status references/session-tools.md、references/lane-workspace-output-promotion.md
Register shared output / promote output references/lane-workspace-output-promotion.md
Writes, outputs, security boundaries references/safety-output-rules.md

Pre-protection

starworkInit is responsible for onboarding ordinary projects into StarWork; this Skill is only responsible for team collaboration within an existing StarWork workspace.

Before starting any multi-Agent write:

  1. First confirm that the target directory is a StarWork workspace; when the target is not a StarWork workspace, stop the multiagent write and switch to starworkInit for safe onboarding.
  2. rules_entry_status: pending_merge stop writing and switch to starworkInit to integrate the final AGENTS.md / host rule entry point.
  3. multiagent.compatibility.status is not current , do not perform write-type MultiAgent operations; first go through the v0.10 upgrade preview.
  4. Preview write-type commands first or wait for user confirmation; do not overwrite the user's business files.

Current session ID

Before any session control or cross-session operation, you must confirmcurrent session ID。

  • <codex_delegation> in source_thread_id takes priority as the current source session ID.
  • When the host or runtime environment explicitly provides current thread / current session metadata, use that value.
  • Do not infer the current session from historical worklogs, old bindings, similar titles, most recent update times, or guessed thread ids.
  • If the current session ID is unknown, stop binding, renaming, pinning, archiving, releasing, and delivery records that use the current session as the source.
  • Before sending, you must check that the target lane session is not equal to the current session; otherwise, block self-delivery by default unless the user explicitly requests local execution or recording only.

Must deliver

Steps whose target is another lane, Agent, or session are allmust-deliver steps. Must-deliver steps cannot be replaced by an explanation in the current reply, and "the message is ready" cannot be treated as completion.

There are only three valid outcomes:

Outcome Requirement
Real automatic delivery succeeded Confirm the target lane / session / current session ID, assemble the complete message, successfully call the host standard thread tool, then record the StarWork request
Explicit manual handoff When the tool is not visible or fails, first perform tool discovery; if it still fails, output manual_handoff_required, the complete copyable message, and state that it has not yet been delivered automatically
Explicit blockage When the target lane, target session, current session ID, or user confirmation is missing, enter blocked / unbound / needs_confirmation

The status write order is fixed:

确认目标 lane / session / current session
  -> 组装 STARWORK:MULTIAGENT_MESSAGE
  -> 调用 send_message_to_thread 或对应宿主标准工具成功
  -> 再执行 starwork multiagent request record delivered...

Do not record delivered_via_codex_thread_tool or delivered_via_claude_code_session_toolunless delivery actually succeeded; do not say "notified", "handoff completed", or "target task completed".

If send_message_to_thread or the corresponding host standard tool is not in the current tool list, first use tool discovery capabilities to find it. When tool discovery is not visible, discovery fails, or the call fails, enter manual_handoff_required, display the complete STARWORK:MULTIAGENT_MESSAGE v1, and clearly state that it has not yet been delivered automatically.

CLI and host tool boundaries

In the normal Codex App path, creation, delivery, reading, naming, pinning, and archiving are performed by the Skill directly calling the standard thread tools:create_thread、send_message_to_thread、read_thread、list_threads、set_thread_title、set_thread_pinned、set_thread_archived。

The CLI only maintains the StarWork project source of truth, for example:

  • starwork doctor --target <path> --json
  • starwork multiagent status --target <path> --json
  • starwork multiagent init
  • starwork multiagent add
  • starwork multiagent bind
  • starwork multiagent release
  • starwork multiagent share
  • starwork multiagent request record

Do not restore the old CLI automatic delivery or creation path as the normal Codex App path; when host tools are not visible, use tool discovery or manual handoff, and do not use the CLI to simulate automatic delivery.

Success criteria

When reporting to the user, explain in layers:

Status Meaning
Position created A responsibility slot already exists in StarWork
Session bound The responsibility slot is bound to a real AI session
Message delivered The handoff message has reached the target session, and the request can be recorded
Target task completed Must come from target lane feedback, worklog, shared output, or explicit session observation

Message delivered does not equal target task completed. Agent created / session bound also does not mean the task has started or completed.

Skill-owned Message

In the normal Codex App path, the Skill assembles the StarWork message itself and does not call the CLI template generator. The message must be complete and copyable; for details, read references/message-templates.md。

Output and security

  • Do not automatically decide which lanes a project should have; lane IDs, responsibilities, and write scopes must come from the current project context.
  • Do not treat example lanes as default templates.
  • Do not treat the lane workspace as the project's official output directory.
  • Before modifying files outside the lane, first register a shared request or obtain explicit user authorization.
  • When the tool is invisible or fails, directly provide manual_handoff_required and the complete copyable message.
  • Completion status must come from the target lane's explicit report, worklog, shared output, or read_thread observation.

Source of truth references

Read when complete boundaries, acceptance criteria, and protocol details are needed:

../starworkMultiagent-spec.md
../../core/agent-lanes-spec.md

Do not repeatedly maintain Agent Lanes protocol details within the main Skill; use the Core SPEC, MultiAgent SPEC, and references as the source of truth.