Imported from tae2089/agent-team (
skills/agent-team-shared/SKILL.md). Install upstream withnpx skills add tae2089/agent-team --skill agent-team-shared. Copyright stays with the author.
Agent Team Shared Runtime
Use agent-team for daemonless coordination between orchestrators and workers. The CLI stores state only; it does not run agents, watch files, or require a background daemon.
Installation
agent-team must be available on $PATH.
agent-team --help
Initialize state once per project or custom state directory:
agent-team init
State
- Default state:
.agent-team/agent-team.db - Override state directory:
AGENT_TEAM_STATE_DIR=/path/to/state - Artifact convention:
_workspace/{run_id}/ - SQLite is the source of truth for runs, tasks, messages, inbox, sync status, and event history.
- Do not use direct peer-to-peer agent communication. Send compact messages with
agent-team message send; recipients check inboxes at checkpoints.
CLI Syntax
agent-team <resource> <command> [flags]
Examples:
agent-team run create --title "docs refactor"
agent-team task create --run run_docs --agent writer --title "draft docs"
agent-team inbox list --agent writer --run run_docs --unread
agent-team task complete --task task_docs --agent writer --evidence "Verified locally." --artifact "_workspace/run_docs/task_docs.md"
Global Flags
| Flag | Default | Description |
|---|---|---|
--params |
{} |
Strict JSON command arguments. Accepts inline JSON, @file.json, or - for stdin. |
--json |
{} |
Strict JSON payload. Use for richer content such as task bodies, metadata, evidence, artifact paths, message body, and reasons. |
Unknown JSON fields are rejected. If a value is supplied by both a named flag and JSON, the CLI returns input_conflict.
Shell Tips
- Boolean flags are presence-based in shell usage: use
--unreador--force. - Use JSON for explicit false values:
--params '{"force":false}'. - Quote JSON with single quotes in POSIX shells.
- Use
@file.jsonfor large JSON objects. - Use
-to read a JSON object from stdin. - Do not pass positional arguments; use flags,
--params, or--json.
Common Aliases
| Flag | JSON key | Meaning |
|---|---|---|
--run |
run_id |
Workflow scope. |
--task |
task_id |
Task scope. |
--msg |
msg_id |
Inbox message scope. |
--agent |
agent |
Logical agent name. |
--id |
id |
Optional caller-provided stable identifier. |
Common Value Semantics
| Value | Meaning |
|---|---|
metadata |
JSON object for compact machine-readable context. Put large content in artifacts. |
body |
Human-readable task or message content. |
reason |
Operational explanation persisted in state or event payloads. |
evidence |
Completion proof in text. State what was checked, changed, or verified. |
artifact |
Path to produced work, normally under _workspace/{run_id}/. |
JSON Output
Success:
{ "ok": true, "state_version": 12, "data": {}, "warnings": [] }
Failure:
{
"ok": false,
"state_version": 12,
"error": {
"code": "sync_conflict",
"message": "...",
"details": {},
"recovery": {
"summary": "...",
"actions": [],
"commands": [],
"docs": [],
"skills": []
}
}
}
Agents should branch on ok and error.code. Use error.recovery for operator-facing next steps and related helper skills. Human-readable messages are not stable interfaces.
Error Codes
| Code | Meaning | Typical action |
|---|---|---|
validation_error |
Required or invalid input. | Fix flags or JSON. |
input_conflict |
Same field supplied by flag and JSON. | Provide the value once. |
invalid_json |
JSON payload is malformed or has unknown fields. | Fix JSON shape. |
invalid_json_source |
@file or stdin JSON source is invalid. |
Fix the source argument. |
not_found |
Run, task, or message does not exist. | Check IDs. |
agent_mismatch |
Agent does not own the task/message. | Reassign or use the correct agent. |
sync_conflict |
Unread messages or incomplete dependencies. | Run sync check, resolve, retry. |
run_not_ready |
Run has unfinished tasks. | Finish or terminally resolve tasks before closing. |
invalid_run_state |
Run status does not allow operation. | Inspect run status. |
invalid_task_state |
Task status does not allow operation. | Inspect task and choose an allowed transition. |
internal_error |
Local filesystem, SQLite, state, install, or sandbox failure. | Check state dir permissions, run init, or inspect install/sandbox docs. |
Full recovery guidance lives in docs/errors.md. agent-team schema export exposes the same error_recovery catalog for agents and CI drift checks.
Skill Routing
| Need | Load |
|---|---|
| Run lifecycle | agent-team-run |
| Task lifecycle | agent-team-task |
| Messages and inbox | agent-team-inbox |
| Drift checks before completion | agent-team-sync |
| Audit, logs, schema, operational cleanup | agent-team-ops |
Use agent-team schema export to inspect the current machine-readable command contract.