| name | power-mode |
| description | Orchestrator for parallel agent teams. When invoked with a roadmap file path, drives every item through the full tackle → uat-generate → uat-auto pipeline until complete. When invoked with a task file path, runs that pipeline once and stops. |
| category | executing |
| model | claude-sonnet-4-6 |
| argument-hint | [roadmap-file-path | task-file-path] |
| disable-model-invocation | false |
| user-invocable | true |
Power Mode — Agent Team Orchestrator
Use this skill to drive tasks or roadmaps to completion using a parallel agent team. The primary goal is efficiency: collision-safe tasks must always run in parallel. You are the primary orchestrator. Use agent teams for parallel work.
Dispatch — Roadmap vs Task
Trigger: /power-mode <path> with no additional instructions.
Inspect the path argument:
- Path contains
roadmap or matches ROADMAP-NNN → Roadmap-to-Completion Orchestrator (see below)
- Path contains
task or matches TASK-NNN → Single-Task Executor (see below)
Parallelization Mandate
Spawning a parallel agent team is MANDATORY whenever tasks are collision-safe. Collision-safe means the tasks do not write to the same files or shared state. When in doubt, read the task steps — if file paths do not overlap, they are collision-safe.
- Minimum team size: 5 agents when 5 or more collision-safe tasks are available in the current wave.
- If the wave has fewer than 5 tasks, spawn one agent per task — still parallel, never sequential.
- Running collision-safe tasks one at a time is a performance failure. It is forbidden.
- The mandate applies equally to UAT generation (Phase 2) and UAT execution (Phase 3).
Collision check protocol before each wave:
- List the files each task is expected to touch (read each task file's Steps section).
- Group tasks that share no file paths — each group is safe to run in parallel.
- Spawn all agents in the safe group simultaneously in a single message with multiple Agent tool calls.
- Only serialize tasks within a group if a direct dependency is declared in the task file.
Team Member Requirements
Every agent spawned in power-mode MUST include mode: "bypassPermissions" and the following footer appended to its prompt:
IMPORTANT: Follow .docs/guides/mcp-tools.md for all file and code operations. Use Serena (mcp__serena__*) for ALL file/directory exploration and ALL code editing. Never use ls, cat, find, grep, sed, awk, or any shell command to inspect or modify files. Use Read/Edit/Write only for markdown and config files.
IMPORTANT: If at any point you are not 100% sure how to fix an issue or which implementation strategy to use, invoke /research first and act on its recommendation before proceeding.
All agent call examples below omit this footer for brevity — it is always required.
Single-Task Executor
Trigger: /power-mode <path/to/TASK-NNN.md> with no additional instructions.
Goal: Drive one task through the full tackle → UAT-generate → UAT-auto pipeline exactly once, then stop.
AUTONOMOUS EXECUTION MANDATE: Execute all three phases without pausing, asking questions, or awaiting approval. The only valid stopping point is after /uat-auto completes and the file-move housekeeping is done.
Phase 1 — Tackle
Agent({
description: "Tackle <TASK-NNN>",
prompt: "Run /tackle <task-path>. [append team member footer]",
mode: "bypassPermissions"
})
Wait for the agent to complete.
Phase 2 — Generate UAT tests
Agent({
description: "UAT generate <TASK-NNN>",
prompt: "Run /uat-generate <task-path>. [append team member footer]",
mode: "bypassPermissions"
})
Wait for the agent to complete.
Phase 3 — Run UATs
Agent({
description: "UAT auto <TASK-NNN>",
prompt: "Run /uat-auto <uat-path> (infer the UAT file path from the task path — same NNN slug, under wiki/work/uat/).
CRITICAL: After the automated UAT run completes you MUST complete Step 7 in full before stopping:
1. git mv the UAT file to wiki/work/uat/archive/
2. git mv the task file to wiki/work/tasks/archive/
3. Remove the task row from wiki/work/tasks/README.md
4. Flip the matching roadmap checkbox if any, and update the task path in the roadmap
UAT tests recorded as [FAIL: auto-judge: UI test requires human verification] do NOT block task completion — /uat-auto does not run UI tests. Proceed with Step 7 even if they remain pending.
If any tests remain as - [ ] Pass because stub indicators were detected, do NOT mark the task as complete. Leave the UAT file in wiki/work/uat/ and report the stub-detected tests in your summary. [append team member footer]",
mode: "bypassPermissions"
})
Done. No further iteration.
Roadmap-to-Completion Orchestrator
Trigger: /power-mode <path/to/ROADMAP-NNN.md> with no additional instructions.
Goal: Drive the roadmap to completion — every item through the full tackle → UAT-generate → UAT-auto pipeline — looping through waves of parallelizable work until no implementation work remains.
AUTONOMOUS EXECUTION MANDATE: Once triggered, drive the roadmap to completion without stopping. Do NOT pause between waves to report progress or ask for permission to continue. The only valid stopping point is when /roadmap-next reports no items remaining. A roadmap item is complete once it has been through the full pipeline — UAT tests requiring human verification do NOT block loop progression.
Forbidden actions in orchestrator mode
- ❌ Stopping after a wave or phase to ask "shall I continue?"
- ❌ Asking clarifying questions about scope, naming, or strategy mid-run
- ❌ Treating a roadmap "Phase" boundary as a stopping point
- ❌ Blocking loop progression because some UAT tests require human verification
Orchestrator loop
Run until /roadmap-next <file> reports no items remaining:
Step 1 — Get the next wave
Invoke /roadmap-next <file> to identify the next group of parallelizable unchecked items.
Step 2 — Materialize inline placeholders
For any item that is an inline placeholder (no existing task file link), spawn a task-creation agent:
Agent({
description: "Create task for <item>",
prompt: "Run /task-add for this roadmap item: '<item description>'. Link the resulting task back into the roadmap file at wiki/work/roadmaps/<file>. Run /research before writing the task. Make all scope, naming, and structuring decisions autonomously — infer reasonable defaults from the roadmap context and existing task files in wiki/work/tasks/. [append team member footer]",
mode: "bypassPermissions"
})
Wait for all task-creation agents to complete.
Step 3 — Three-phase pipeline for the wave
Run tackle → UAT-generate → UAT-auto for all tasks in the current wave. Apply the Parallelization Mandate at every phase.
Phase 1 — Tackle (parallel, minimum 5 agents when 5+ collision-safe tasks available):
Agent({ description: "Tackle TASK-NNN", prompt: "Run /tackle <path>. [append team member footer]", mode: "bypassPermissions" })
Agent({ description: "Tackle TASK-MMM", prompt: "Run /tackle <path>. [append team member footer]", mode: "bypassPermissions" })
// ... one per task, all in the same message
Wait for all to complete.
Phase 2 — Generate UAT tests (parallel, read-only):
Agent({ description: "UAT generate TASK-NNN", prompt: "Run /uat-generate <task-path>. [append team member footer]", mode: "bypassPermissions" })
// ... one per task, all in the same message
Wait for all to complete.
Phase 3 — Run UATs (parallel when UAT files don't collide):
Agent({
description: "UAT auto TASK-NNN",
prompt: "Run /uat-auto <uat-path>.
CRITICAL: After the run completes, complete Step 7 before stopping:
1. git mv the UAT file to wiki/work/uat/archive/
2. git mv the task file to wiki/work/tasks/archive/
3. Remove the task row from wiki/work/tasks/README.md
4. Flip the matching roadmap checkbox and update the task path in the roadmap
UAT tests recorded as [FAIL: auto-judge: UI test requires human verification] do NOT block this step. If stub-detected tests remain, do NOT mark the task complete — leave the UAT file in place and report the stubs. [append team member footer]",
mode: "bypassPermissions"
})
Step 4 — Repeat
Immediately return to Step 1. No pausing, no reporting, no confirmation. Loop until /roadmap-next reports no items remaining.
