| name | manciple-agents |
| description | Coordinate Manciple task-runner workers efficiently across actionable Manciple work, with each worker using the manciple-mcp-task-runner skill and Manciple MCP tools. Use when the user asks to spawn Manciple workers, process multiple Manciple tasks in parallel, or coordinate Manciple task execution across sub-agents. |
Manciple Agents
Coordinate worker agents for Manciple task work. Use local CPU capacity as a hard ceiling, but let Manciple's deterministic dispatch plan choose the actual assignments. Each worker must use manciple-mcp-task-runner as its operating procedure.
Use this skill as the parent coordinator. Do not do task implementation yourself unless a small coordinator-side change is needed to integrate worker results.
Required Tools
Use multi_agent_v1.spawn_agent to spawn workers. If the tool is not available, stop and report that this skill requires sub-agent support.
Use the repo-local manciple-mcp-task-runner skill when available:
/Users/danielpatton/Documents/GitHub/promptops/.codex/skills/manciple-mcp-task-runner/SKILL.md
If that path does not exist, use the discovered manciple-mcp-task-runner skill path from the active skill list.
Coordinator Workflow
-
Detect the worker cap.
- Determine the available cores before spawning workers. Prefer reliable local mechanisms available in the runtime, such as Node
os.availableParallelism(), nproc, or sysctl -n hw.ncpu.
- Use the detected core count only as the maximum safe worker count. If no safe core count can be determined, stop and report the blocker.
-
Get the deterministic dispatch plan.
- Call
manciple_dispatch_plan first when MCP tools are available. If MCP is unavailable, run manciple dispatch-plan.
- Treat the returned assignments, deferrals, stop conditions, and verification commands as the scheduling source of truth.
- Spawn only assignments returned by the dispatch plan. Do not hand-pick additional tasks, infer extra workers from CPU capacity, or override deferrals in the model.
- If the user named task ids, still use the dispatch plan and spawn only matching returned assignments. Report any named task that is absent, deferred, or stopped.
-
Spawn the returned batch of worker agents in parallel.
- Cap the batch at the smaller of the detected worker cap and the number of assignments returned by the dispatch plan.
- Use
worker role. Do NOT use a full-history fork; embed all required context directly in each worker prompt.
- When Assignr MCP tools are available, include only the repo absolute path, skill file path, assigned task id, ownership boundary from the dispatch plan, and any dispatch-plan notes. Tell the worker to load compact task context with
manciple_get_task_packet before editing.
- When MCP tools are unavailable, include the compact packet from
manciple task-packet <task-id> plus the dispatch-plan notes.
- Pass the
manciple-mcp-task-runner skill as a skill item to each worker.
- Tell workers they are not alone in the codebase, not to revert edits made by others, and to respect
allowed_paths, forbidden_paths, and task spec write rules.
- Tell workers to create their own Assignr run logs and status transitions.
-
Continue coordinator work while workers run.
- Inspect broad repo state, likely conflict zones, or verification setup.
- Do not duplicate assigned worker implementation.
-
Integrate each completed batch.
- Wait for all workers in the batch before refreshing task state unless a worker failure blocks the run.
- Review each worker's final report, changed files, status, run log evidence, and verification receipt.
- Resolve conflicts only within the relevant task scope.
- Verify the integrated batch with
manciple_verify using profile coordinator when MCP tools are available, or manciple verify --profile coordinator from the CLI. Report the returned receipt.
- Run
manciple_check_lifecycle when MCP tools are available, or manciple check-lifecycle from the CLI. If lifecycle placement issues are reported, move only misplaced task YAML files to the directory matching their status and re-run the lifecycle check.
- Validate Assignr metadata with
manciple_validate when MCP tools are available.
- Get a fresh dispatch plan before starting any later batch. Stop when the plan returns no assignments or reports a stop condition.
-
Report the run.
- Include each worker's assigned task or sidecar responsibility.
- Include files changed, verification receipts, final task statuses, lifecycle check result, blockers, and follow-up work.
- Explicitly call out any worker that could not complete its task or verify its changes.
Use manciple_format_task or manciple format-task <task-id> --check only when a specific task's YAML formatting is in scope or a receipt asks for scoped formatting evidence. Do not run routine whole-repo YAML formatting loops as coordinator housekeeping.
Worker Prompt Template
Use this shape for every worker. Fill in ALL bracketed values before spawning; workers have no access to conversation history.
Use the manciple-mcp-task-runner skill at [ABSOLUTE skill path, e.g. /Users/danielpatton/Documents/GitHub/promptops/.codex/skills/manciple-mcp-task-runner/SKILL.md] for this assignment.
Repo root: [ABSOLUTE path to repo root, e.g. /Users/danielpatton/Documents/GitHub/promptops]
You are Worker [NUMBER] of [TOTAL_WORKERS_THIS_BATCH] in an Assignr dispatch-plan batch. The coordinator used CPU capacity as a ceiling, but this assignment comes from manciple_dispatch_plan / manciple dispatch-plan. You are not alone in the codebase. Do not revert edits made by others; adapt to existing changes and keep your edits narrowly scoped.
Assigned task or responsibility:
[task id or sidecar responsibility from the dispatch plan]
Ownership boundary:
[assignment ownership boundary from the dispatch plan, or read-only verification scope]
Task context:
- MCP tools available: [yes/no]
- Dispatch-plan notes: [minimal notes from the assignment, deferral, or stop-condition context]
- If yes: task id is the contract; load compact task context with manciple_get_task_packet before editing.
- If no: run manciple task-packet <task-id> before editing, or use this embedded compact packet: [packet]
Instructions:
- Use Assignr MCP tools or CLI commands as the source of truth.
- If assigned a task id, load compact context first, set it in_progress if needed, and obey allowed_paths, forbidden_paths, acceptance_criteria, verification.commands, and outputs_required.
- Compile the full prompt only when the compact packet is insufficient and domain context is explicitly needed.
- Do not edit files under .manciple/specs/tasks directly. Only status updates through manciple_set_status are allowed.
- Verify with manciple_verify profile worker or manciple verify --profile worker and report the returned receipt as the primary verification evidence.
- Use manciple_format_task or manciple format-task <task-id> --check only when scoped task YAML formatting evidence is needed.
- Create the Assignr run log with final task status, files changed, non-test commands, tests, acceptance evidence, risks, and the deterministic verify receipt; then set final task status according to manciple-mcp-task-runner.
- For completed implementation work that changed behavior, include Decisions Made; omit it only when blocked before meaningful changes.
- In your final response, report task id, final status, files changed, verification receipt, risks, blockers, and follow-ups.