Skip to content

AGENTS.md

Latest commit

 

History

History
104 lines (74 loc) · 6 KB

AGENTS.md

File metadata and controls

104 lines (74 loc) · 6 KB
Raw

AGENTS

Purpose

agentcli is the agent-native control plane for workflow manifests. It is not the durable runtime.

Use this repo when the task is about:

  • manifest authoring
  • validation
  • compile target behavior
  • scheduler inspection surfaces
  • JSON-RPC / MCP style integration

Do not reimplement scheduler durability here. Runtime concerns belong in openclaw-scheduler.

Working Rules

  • Prefer machine-readable output and raw JSON payloads.
  • Pass --json to force structured JSON output from any command (including help), except serve which runs the JSON-RPC server on stdio.
  • Keep manifests backend-portable unless the task explicitly targets openclaw-scheduler.
  • Default compile target is standalone.
  • Treat openclaw-scheduler as an adapter/backend, not as the core abstraction.
  • Preserve agent-safe behavior: safe output paths, strict validation, field masks, and sanitization.

Error Handling

All CLI errors are written to stderr as JSON with the following shape:

{
  "ok": false,
  "error": "Human-readable error message",
  "error_type": "validation_error | unknown_command | invalid_argument | parse_error | internal_error"
}

Use error_type for structured dispatch rather than parsing the error string. Validation errors include an additional validation field with the full error and warning arrays.

Discovery Flow

When first interacting with agentcli, use this sequence:

  1. agentcli version -- confirm the package and manifest spec version
  2. agentcli describe commands -- enumerate all available CLI commands
  3. agentcli describe rpc -- enumerate JSON-RPC methods and notifications
  4. agentcli targets -- list available compile targets with capabilities
  5. agentcli schema manifest -- get the machine-readable manifest schema

For JSON-RPC integrations, use agentcli.version and agentcli.describe with target: "rpc" to discover the method surface programmatically.

Useful Commands

  • agentcli version -- show package and manifest spec version
  • agentcli init -- bootstrap a local home directory with a starter manifest
  • agentcli targets -- list available compilation targets
  • agentcli paths -- show resolved home directory layout
  • agentcli schema manifest -- emit machine-readable schema for the manifest format
  • agentcli describe commands -- list all CLI commands with summaries
  • agentcli validate examples/hello-world.json
  • agentcli compile examples/hello-world.json --target standalone --explain
  • agentcli compile examples/hello-world.json --target openclaw-scheduler
  • agentcli apply <manifest> --adopt-by name --dry-run -- preview migration of existing scheduler jobs
  • agentcli inspect jobs --db /path/to/scheduler.sqlite --fields id,status
  • agentcli approve <manifest> <task-id> --by <principal> --reason <text> -- grant a local single-use approval for a gated task
  • agentcli approvals list [--status pending|consumed|expired|revoked|all] -- list approval records
  • agentcli approvals revoke <approval-id> --by <principal> --reason <text> -- revoke a pending approval
  • agentcli exec <manifest> <task-id> --approval-id <id> -- target a specific pending grant
  • agentcli serve

Coordination protocol (mandatory)

Multi-agent file coordination via the coord MCP server. Required calls every session:

  1. list_claims at task start.
  2. claim_files before editing.
  3. release_claims (or release_session to drop everything this MCP session holds, v0.6+) when done.

If claim_files returns conflicts, stop and ask the user. No edits outside claimed scope; no opportunistic refactors; shared config edits only with explicit user approval.

Sub-file (symbol-level) claims (v0.14+)

When two agents need different parts of the same file, claim symbols instead of whole files. Pass symbols={"src/auth/login.ts": ["handleLogin"]} to claim_files so the claim scopes to just that function. Two agents on disjoint symbols of the same file auto-coexist with no 409.

  • Foo::handleA (v0.16): method-level notation. Sibling methods of the same class auto-coexist; the bare class blocks all of its methods.
  • Outer::Inner::method (v0.17): recursive nesting works to any depth.
  • Server-side validation (v0.17): when COORD_REPO_ROOT is set, the server rejects symbols that do not exist in the file with a hint listing what is parseable. The MCP wrapper pre-validates locally before POST so typos fail fast; disable with COORD_DISABLE_CLIENT_VALIDATION=1.

Queueing instead of bouncing (v0.21+)

When a hot file is contested and your work is blocking the team, pass wait_seconds=60 to claim_files. The request joins a FIFO queue behind the blocking holder; on release the next queued requester is auto-granted. Pair with urgency='low' | 'normal' | 'high' | 'blocking' (v0.25) to jump ahead of lower-priority waiters; long-waiting entries age-boost one level after COORD_QUEUE_AGE_BOOST_SECONDS (v0.26). Abandon a wait early via cancel_queue_request(queue_id, engineer=...) (v0.26). Inspect your own queue rows via my_requests(queued=True) (v0.22).

Asking the holder directly (v0.6+ / v0.11+ decisions)

If queueing will not work either:

  • request_release files an explicit ask against the holder's claim. The holder's TTL shortens; their decision lands back in your my_requests view.
  • respond_to_request decisions (v0.11+): approved (release whole claim), denied (keep it), narrowed (close the claim, open a tighter one -- pass narrowed_pattern), coexist (let the requester have a sibling claim on the same scope -- pass coexist_pattern). coexist is cooperative not enforced; agents on the same file still handle imports and module-level state themselves.

Operator visibility

Poll pending_requests between operations to see who is blocked on your scope and respond via respond_to_request. The dashboard surfaces hotspot files, an auto-resolution heatmap, and a pending-queue panel for ambient awareness.