团队仓库能力补齐后,manage-teams 的原有说明仍要求用户在 UI 配置远程,并缺少安装、升级与发布重试入口。同步中英文动作和流程、修正 set_member_source 自动解析语义,将技能增版为 1.4.0;最低客户端 10.0.178,旧客户端不应加载新增动作。 Updates the bilingual manage-teams contract for all team repository actions, remote destination configuration, exact release tags and same-version retries, complete archiving and protected member updates. Bumps the Skill to 1.4.0 with a 10.0.178 client floor so older tool ABIs cannot load the new actions. Validation: i18n/schema/catalog checks passed with 105 pre-existing governance/provenance warnings; reviewed human translations and refreshed their source hash. New content uses generic roles and examples. No customer, tenant, credential or runtime fixture data is introduced. Related: desirecore/desirecore#1490. Client implementation: https://github.com/desirecore/desirecore/pull/4362. Do not infer release availability from this Skill PR. The authorized alternative pre-publication check passed: full working tree including hidden files, changed paths, branch and commit metadata, and this PR text were scanned for credential shapes and manually reviewed for customer-neutral content. No private token list was supplied or recorded.
14 KiB
license, name, description, version, type, risk_level, status, disable-model-invocation, tags, metadata, market
| license | name | description | version | type | risk_level | status | disable-model-invocation | tags | metadata | market | |||||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MIT | manage-teams | 创建和管理 Agent 团队,组织多 Agent 协作。Use when 需要多个 Agent 持续协作、建立组织架构,或发布、安装和同步团队仓库时。 | 1.4.0 | procedural | medium | enabled | true |
|
|
|
manage-teams Skill
L0: One-line Summary
Use ManageTeam to inspect, create, and govern Agent teams, with the required checks before organizational changes or remote synchronization.
L1: When to Use
Use a team when:
- multiple Agents need sustained collaboration around one task and a shared team workdir;
- a stable supervisor, membership, or parent-child organizational structure is required;
- a team repository must be published, installed, or synchronized.
Do not create a team when:
- one expert is needed once: use
Delegate(mode="sync" | "async"); - several experts only need to provide one-off opinions: use
Delegate(mode="fan-out"); - the work is temporary file exploration: use a Worker instead of creating a lasting organization.
A team defines organization, shared directories, and governance. Actual work is still dispatched to members with Delegate.
L2: Execution Specification
1. Inspect Before Mutating
- If
teamIdis unknown, callManageTeam(action="list")first. - Before modifying, disbanding, or synchronizing a team, call
ManageTeam(action="get", teamId=...)and verify its name, type, supervisor, members, local repository directory, and remote state. - Never guess a path under
~/.desirecore; use only the absolute repository path returned byget. listcan filter byparentTeamId. Withtree=true,teamIdselects the subtree root andparentTeamIdis ignored.
2. Action Reference
| action | Purpose | Key parameters and notes |
|---|---|---|
list |
List teams or an organization tree | parentTeamId?, tree?, teamId? |
get |
Inspect one team and its repository path | teamId |
create |
Create an ephemeral team | name or task; supervisor?, members?, memberRouting?, parentTeamId?, workdirMode? |
add_member |
Add one member | teamId, agentId |
add_members |
Add members in a batch | teamId, members |
remove_member |
Remove one member | teamId, agentId |
remove_members |
Remove members in a batch | teamId, members |
set_supervisor |
Replace the supervisor | teamId, agentId |
set_member_source |
Declare where a member Agent comes from | teamId, agentId, memberSource; git needs url (https, will be cloned) and ref (defaults to main), registry needs id+version, core/local need nothing else |
update |
Partially update team configuration | teamId; supports name/type/isolation/parentTeamId/description/avatar/avatarImage |
promote |
Promote an ephemeral team to persistent | teamId; one-way and never implicit |
disband |
Disband a team | teamId; enters approval; preserves a full archive, requiring workdir binding and content review on restore |
fork_team |
Install a team from a remote repository | url; name?, installMembers?; enters approval |
push |
Push a local team to its connected remote | teamId; enters approval |
pull |
Pull and validate a team from its connected remote | teamId; enters approval |
add_remote |
Configure or replace a publishing destination | teamId, url; optional connectionId, branch, isPrivate; does not create a repository |
remove_remote |
Remove remote configuration and the local connection binding | teamId; leaves the remote repository intact |
install_members |
Install or repair members from the existing lock | teamId; optional forceUpdate, default false |
upgrade_members |
Resolve a new lock and attempt safe updates | teamId; optional forceUpdate, default true; preserves local edits and commits |
release |
Create a local version or publish one exact version | teamId; push defaults to false, optional changelogEntry, expectedCommit; retry with releaseTag |
3. Create a Team
Before creation:
- Every Agent in
supervisorandmembersmust already exist. Verify IDs withManageAgent(action="list" | "get"); create or install a missing Agent through its corresponding Agent Skill before creating the team. - The DesireCore core Agent,
desirecore, cannot be a supervisor. When the core Agent initiates creation, it must explicitly choose a regular Agent assupervisor. - Normally do not add
desirecoreas a member. Reach core capabilities throughDelegateinstead. - One Agent may supervise only one team. If the intended supervisor already leads another team, assign a successor there first.
Choose the workdir mode deliberately:
merged(default): the shared team directory is primary while member and global workdirs remain available;team_only: exposes only the shared team directory, for high-reliability work where every member must operate on the same project. It does not delete member workdir configuration.
Use memberRouting to express routing intent without pinning a Provider or model:
{
"supervisor-agent": {
"tier": "flagship",
"requiredCapabilities": ["reasoning"],
"reasoning": "high"
},
"member-agent": {
"tier": "balanced"
}
}
- keys must belong to the selected
supervisorormembers; - Agents using a fixed model must not appear in
memberRouting; - omitted Smart members retain their current routing profile; the concrete Provider/model is resolved when that member executes work.
Example:
{
"action": "create",
"name": "Contract Review Project",
"supervisor": "legal-lead",
"members": ["contract-reviewer", "risk-analyst"],
"task": "Review contracts continuously and consolidate risks",
"workdirMode": "team_only"
}
4. Change Organization and Configuration
- Prefer batch member actions to avoid observable intermediate states.
set_supervisorusesagentId; first verify that the Agent does not already supervise another team.set_member_sourcedeclares provenance, it does not move files. A team whose roster still contains alocalmember is not distributable —members.lock.jsoncannot pin an ID that exists only on this machine, so a fork elsewhere would silently come up short a member. Switch each member togitorregistrybefore publishing. The action resolves and writes the lock automatically; inspect its receipt. It cannot change a member'srole; useset_supervisorfor that.updateis a patch: omitted fields remain unchanged.parentTeamId: nulldetaches the team and makes it top-level; an empty string is invalid.typeonly allowsephemeral → persistent. Repeating the current value is idempotent; usepromotefor an explicit upgrade.isolation:softuses shared session isolation;harduses independent Agent copies.descriptionis the marketplace-facing team description, not thetasksupplied at creation.
Use a declared avatar with:
{
"action": "update",
"teamId": "team-id",
"avatar": { "char": "CR", "color": "purple" }
}
For an image avatar, use avatarImage.source with dc-media://<mediaId>, a bare mediaId, or an image path inside the workdir. PNG/JPEG/WebP are supported. Do not pass an HTTP(S) URL or base64. Remove the image with { "remove": true }; remove and source are mutually exclusive.
5. Team Lifecycle
- Disband an ephemeral team after its project is complete so the organization does not accumulate stale teams.
- Use
promoteonly for an explicit long-term collaboration requirement; promotion is one-way. disbandarchives the complete team directory and ends its local workdir binding, even without Git initialization. Archive failure preserves the original data and reports failure. The action enters approval; the caller's approval mode determines whether a human card appears. To restore, first check that the team ID is free, then bind the workdir and approve its content again.
6. Team Repository and Remote Synchronization
The team directory is a Git repository containing governance data such as team.json, member locks, and shared/rules.md.
For local Git work:
- obtain the absolute repository path with
get; - run
status/log/diff/add/commit/tagwith Bash in that directory; - after the local commit is complete, call
ManageTeam(action="push").
Remote fork_team/push/pull must go through ManageTeam because it enforces team Schema validation, roster consistency, the core-Agent supervisor prohibition, workspace types, out-of-bounds symlink checks, and approval. Do not bypass those controls with raw git push/pull. This rule is not based on an assumption that Agents can never access credentials.
push/pullrequire a remote configuration. Useadd_remoteto configure the destination and an existing connection when needed. Client team settings are also available; the Agent chooses the configuration path.- Locally created and forked teams do not inherit a directly pushable remote configuration by default.
fork_teamdefaults toinstallMembers=true; a same-ID local Agent that has diverged from its lock is protected and skipped rather than overwritten.pullmay replace local team configuration. Inspect local state first and identify the target remote in the approval card.
Release and member versions:
releasedefaults to a local version. Setpush=truefor publication; approval identifies the target and version. Only the branch and this exact release tag are pushed, never all tags or private backup/archive refs.- Inspect the team with
get; constrain a new release HEAD withexpectedCommit. Inspecttag/version/commitHashandbranchPushed/tagPushed/pushedseparately. - After network failure or an unknown remote outcome, inspect the remote and retry with the returned
releaseTag; do not create another version. OmitchangelogEntryon retries. install_membersfollows the exact existing lock.upgrade_membersresolves declarations and attempts safe updates to new pins. Both preserve local edits and commits, avoid downgrades, and do not approve installed content.complete=falseoroutdated/diverged/failed/unreachablemeans members remain unaligned. Neithersuccess=true, skipped counts nor zero failures proves all members were upgraded. Inspect each receipt before choosing repairs.- Conflicting pins for an Agent shared by multiple teams do not authorize overwriting a local fork; preserve data and report the conflict.
- Remote actions, remote-source declarations and destination configuration enter approval.
ask-externalretains approval for them;allow-allrespects the user's full trust in the acting Agent.
7. Dispatch and Finish
After team creation, dispatch work to the supervisor or members with Delegate:
- one member:
Delegate(target=..., mode="sync" | "async", teamId=...); - several members:
Delegate(targets=[...], mode="fan-out", teamId=...); - prefer team members for sustained collaboration; a one-off outside opinion does not require membership.
Report the team name and ID, type, supervisor and members, workdir mode, organizational changes, and whether remote operations completed. Never expose credentials or a remote URL containing a token.
8. Failure Recovery
Agent does not exist: verify the ID; create or install the Agent, then retry.Core Agent cannot supervise: explicitly choose a regular Agent assupervisor.Supervisor already leads another team: runset_supervisoron the existing team before retrying.Remote not configured: useadd_remoteto configure the destination, or use client team settings; do not guess a hidden API.- Local content changed or conflicts exist: obtain the directory with
get, inspect Git state, preserve user changes, and only then decide whether to commit, pull, or retry. - If this Skill is missing or disabled, the minimal operation may still be executed from the
ManageTeamaction/parameter Schema and tool error messages. Never bypass the tool by editing AgentFS directly.