clawmem-team（爪友团队）

ClawMem Team

Use this skill as the single entry point for ClawMem Team requests.

Keep the product boundary explicit:

• ClawMem plugin provides memory and atomic collaboration tools.
• This skill defines how those primitives are composed into a Team workflow.

Available templates

• main-worker-summary-queue.md

- Best for one coordinating agent plus worker agents. - Recommended starting shape is 1 main agent + 2 worker agents. - Use when work should move through a shared queue repo with explicit task labels and status changes. - If the user names this template directly, read references/templates/main-worker-summary-queue.md.

• reviewing.md

- Best for code review, PR review, design review, architecture review, or document inspection. - Use when review work needs a clear requester, reviewer set, artifact location, and completion rule. - If the user names this template directly, read references/templates/reviewing.md.

Users may refer to a template by filename, for example:

• "Based on main-worker-summary-queue.md, set up my team"
• "Based on reviewing.md, set up my team"

Select the right path

1. Confirm the user wants Team design, Team setup, Team verification, or a Team template.
2. If this is only an ordinary memory request, stay on the bundled clawmem skill instead.
3. Choose one of three paths:

- explicit template: the user directly names a file under references/templates/ - matched template: the user's request clearly fits one of the available templates - custom: no template clearly fits, so design a Team from scratch

Template readiness

Before producing a blueprint for a multi-agent template, confirm these readiness layers:

• environment readiness: the current OpenClaw host has ClawMem installed and enabled, and the bundled clawmem runtime skill is available
• participant inventory: which OpenClaw agents already exist and which ones still must be prepared
• ClawMem readiness: whether each selected agent is already configured, will bootstrap on first use, or is blocked
• runtime delegation readiness: whether the current control session can still dispatch real work to the required workers, or whether session refresh / pairing repair is still required

For main-worker-summary-queue.md:

• treat 1 main + 2 workers as the default minimum ready shape for this template
• do not treat a selected participant as ready until that agent can use the ClawMem runtime in the current OpenClaw environment
• if only the current agent is available, prepare two worker agents first when the runtime exposes agent-creation capability and the user approves it
• if the runtime cannot list or create agents, stop at a readiness plan and ask the user to confirm or prepare the missing agents
• if multiple agents already exist, ask whether to use the default 3-agent shape or to choose specific existing agents
• if the user has no stronger preference, keep the current agent as main and assign two workers
• if bootstrap changes the OpenClaw agent topology, gateway state, or pairing state, assume the current session may need refresh before worker dispatch can be verified
• do not report this template as ready until one real worker handoff succeeds through a working dispatch path instead of only by config inspection or main-authored proxy writes

Read the right reference

• For any Team blueprint, read references/blueprint.md.
• For Team bootstrap and mutation sequencing, read references/bootstrap.md.
• For readiness checks and end-to-end proof, read references/verification.md.
• For user-facing tone, progressive disclosure, and plain-language substitutions, read references/communication.md.
• For the main/worker/summary queue template, read references/templates/main-worker-summary-queue.md.
• For the reviewing template, read references/templates/reviewing.md.

Workflow label schema

Every Team produces an explicit Workflow Label Schema, even when starting from a template. It lives inside the canonical Team artifact so workers can read it at runtime. ClawMem deliberately stays label-agnostic, so the Team is the only place this schema can live.

A schema is valid when it names:

• task-kind labels: which kinds of work exist in this Team (e.g., queue:task, review:request)
• status labels: the finite set of status values a task can move through
• exactly one terminal status label: the label that means the task is finished
• assignment label format: how a task records its current owner (e.g., assignee:<agent-id>)
• terminal closure rule: when a task reaches the terminal status, the issue MUST also be set to state: closed via issue_update in the same mutation
• canonical polling filter: the issue_list filter participants use to pick up work (state: open plus the task-kind label), so closed tasks never reappear in the queue and participants on different hosts can converge through the same filter

Templates under references/templates/ provide ready-made schemas. Custom Teams define their own, but the structure above is required. Do not let a Team ship without the schema written down.

Team contract binding

The canonical Team artifact only works if participants actually read it before acting. The default binding path reuses the bundled clawmem skill's existing turn loop — no per-host install is needed, and the mechanism is the same whether the team lives on one machine or across many.

Role names are user-defined. Do not assume main / worker / reviewer unless the user or the chosen template declares those roles. Whatever role names the contract defines, every one of them uses the same binding path:

1. Store the contract as a stable ClawMem memory node, labeled with a stable pair such as kind:convention topic:team-contract. The orchestrating agent analyzes the user's setup to decide where it lives — a dedicated team / coordination repo, or the task repo itself. Either is valid; state the choice in the blueprint.
2. Every task issue this Team creates opens its body with an explicit citation, for example:

> Team contract: memory_get #<contract-id> (kind:convention topic:team-contract) before acting. Labels and terminal closure are defined there.

1. The bundled clawmem skill already instructs every ClawMem-compliant agent to call memory_get when a specific memory id is mentioned. The returned body comes back as regular tool output, not wrapped by auto-recall as "historical notes".

Defensive inline: also include the Workflow Label Schema's most critical rules directly in the task body (allowed labels plus the done ⇒ state:closed rule), so a participant that skips the fetch still sees the minimum.

Do not rely on auto-recall alone to deliver the Team contract to participants. The plugin frames auto-recall as background context, not as binding instructions — use the explicit memory_get citation above.

Workflow

1. Identify the user's goal, constraints, participants, and preferred collaboration style.
2. If the user directly names a template file, use that template first.
3. Otherwise, choose a template when it clearly fits. If no template clearly fits, design a custom Team.
4. Inspect or confirm the participant readiness required by the selected template or custom design.
5. If the chosen path needs multiple agents, decide whether to reuse existing agents or prepare missing ones before bootstrap.
6. Produce one explicit Team blueprint using references/blueprint.md, even when starting from a template.
7. Keep the Team contract explicit in the blueprint:

- roles - repos - access model - issue or comment protocol - canonical Team artifact - verification path - environment readiness state - participant readiness state - runtime delegation readiness state

1. If the user approves actual changes, bootstrap the Team using references/bootstrap.md.
2. If bootstrap changed agent config, restarted the gateway, or otherwise invalidated the current control path, refresh or re-establish the session before end-to-end worker verification.
3. When the Team is configured, verify it using references/verification.md.
4. If the user wants a demo, seed only one minimal workflow object.

Decision rules

• If the user explicitly names a template file, prefer that template unless it clearly conflicts with the request.
• Prefer a matched template when the user asks for a common Team shape.
• Prefer custom design when the user describes a unique workflow, org boundary, or access model.
• Do not assume the current host can list or create OpenClaw agents. Use those capabilities only when the runtime exposes them.
• For templates with a minimum participant shape, stop at readiness planning when the required agents are still missing.
• If several agents are already available for main-worker-summary-queue.md, ask whether to use the default 3-agent shape or a user-selected subset.
• If a selected agent lacks ClawMem configuration but can bootstrap on first use, record that state in the blueprint instead of treating it as a blocker.
• If the current session cannot dispatch to the required workers after bootstrap, report partial or blocked instead of ready.
• Do not treat main-authored proxy comments, config inference, or shared-identity authorship alone as proof that a worker actually executed the task.
• If bootstrap changes agent config or restarts the gateway, expect a fresh session or repaired pairing before claiming end-to-end success.
• Ask for clarification only when the goal, participants, or execution model is still ambiguous after the first pass.
• Do not invent a hidden Team protocol inside the plugin. If a protocol is needed, make it explicit in the blueprint or template output.
• Do not refer to other sibling skills. This single skill owns the full Team journey.
• Keep one canonical Team artifact. Do not scatter the Team contract across multiple uncoordinated files or issues.
• Use ClawMem collaboration tools first. Use shell or raw API only when the required operation is not exposed by tools.
• Require explicit approval before writes that create agents or change orgs, teams, invites, memberships, or permissions.
• Do not let a template bypass blueprint, bootstrap, or verification. Templates guide the design; the Team still must be made explicit and verified.

Required output

Every major phase ends with two layers (see references/communication.md):

1. Plain-language summary for the user

- what this Team will do, in one or two sentences and domain terms - who plays each role and what the user still has to decide - the single next concrete step, translated into the user's words

1. Technical detail, separated and only when helpful

- chosen path and why it fits - environment, participant, and runtime delegation readiness state - the blueprint or mutation plan - the Workflow Label Schema and terminal closure rule when relevant

Default to the user's current language for layer 1. Keep schema identifiers (kind:*, topic:*, queue:task, task-status:done, state:closed, etc.) machine-readable in both layers.