/execute

name: claudikins-kernel:execute
description: Execute validated plans with isolated agents and two-stage review
argument-hint: <plan-path> | --resume | --status [--model opus|sonnet]
model: opus
agent_outputs:
  - agent: babyclaude
    capture_to: .claude/task-outputs/
    merge_strategy: none
  - agent: spec-reviewer
    capture_to: .claude/reviews/spec/
    merge_strategy: none
  - agent: code-reviewer
    capture_to: .claude/reviews/code/
    merge_strategy: none
  - agent: conflict-resolver
    capture_to: .claude/conflict-resolutions/
    merge_strategy: none
allowed-tools:
  - Read
  - Grep
  - Glob
  - Task
  - Bash
  - AskUserQuestion
  - TodoWrite
  - Skill
skills:
  - git-workflow
output-schema:
  type: object
  properties:
    session_id:
      type: string
    status:
      type: string
      enum: [completed, paused, aborted]
    plan_source:
      type: string
    tasks_completed:
      type: integer
    tasks_total:
      type: integer
    batches_completed:
      type: integer
    batches_total:
      type: integer
    branches_merged:
      type: array
      items:
        type: string
    branches_remaining:
      type: array
      items:
        type: string
  required: [session_id, status, tasks_completed, tasks_total]

claudikins-kernel:execute Command

You are orchestrating a task execution workflow with isolated agents and human checkpoints between batches.

Flags

| Flag | Effect | | --------------- | --------------------------------------------------- | | `--resume` | Resume from last checkpoint | | `--status` | Show current execution status | | `--abort` | Abort current execution (saves checkpoint) | | `--batch N` | Override batch size (default: from plan) | | `--model M` | Model for agents: `opus` or `sonnet` (default: opus)| | `--skip-review` | Skip code review (spec review still runs) | | `--dry-run` | Parse plan and show execution order without running | | `--timing` | Show task and batch durations | | `--trace` | Show execution trace at completion |

Merge Strategy

None - task outputs are saved per-task, not merged.

Philosophy

> "5-7 agents per SESSION, not 30 per batch. Features are the unit of work." - Boris

• One task = one branch (isolation prevents pollution)
• Fresh context per task (context: fork)
• Two-stage review (spec compliance, then code quality)
• Human checkpoints between batches (not individual tasks)
• Commands own git (agents never checkout/merge/push)

Load Skill

First, load the git-workflow skill for methodology:

Skill(git-workflow)

This provides:

• Task decomposition patterns
• Review criteria and thresholds
• Batch checkpoint decision trees
• Circuit breaker and tracing patterns

State Management

State file: `.claude/execute-state.json`

{
  "session_id": "exec-YYYY-MM-DD-HHMM",
  "plan_source": "path/to/plan.md",
  "started_at": "ISO timestamp",
  "status": "initialising|executing|paused|completed|aborted",
  "current_batch": 1,
  "current_task": null,
  "tasks": [...],
  "batches": [...],
  "last_checkpoint": null
}

Phase 0: Initialisation

Flag Handling

Check for flags first:

--status → Run execute-status.sh hook, display status, exit
--resume → Load checkpoint, resume from saved state
--abort → Save checkpoint, mark aborted, exit
--dry-run → Parse and display, don't execute
--model → Set agent model (opus|sonnet), stored in execute-state.json

Model Selection

If `--model` flag is provided, use the specified model. Otherwise, prompt the user:

AskUserQuestion({
  question: "Which model should agents use for this execution?",
  header: "Model Selection",
  options: [
    { label: "Opus", description: "Most capable — best for complex tasks (uses more session quota)" },
    { label: "Sonnet", description: "Fast and capable — good for straightforward tasks (lighter on quota)" }
  ]
})

Store the selected model in `execute-state.json` as `"model": "opus"|"sonnet"`. All `Task()` calls for babyclaude, spec-reviewer, and code-reviewer must use this value.

Plan Loading

1. Get plan path from argument or find most recent in `.claude/plans/` 2. Validate EXECUTION_TASKS markers exist (hook: validate-plan-format.sh) 3. Parse task table between markers 4. Build dependency graph

**On validation failure:**

Plan missing EXECUTION_TASKS markers.

The plan must include:
  <!-- EXECUTION_TASKS_START -->
  | # | Task | Files | Deps | Batch |
  ...
  <!-- EXECUTION_TASKS_END -->

Run claudikins-kernel:outline to generate a properly formatted plan.

Dependency Graph

Build from parsed table:

{
  "tasks": [
    {
      "id": "1",
      "name": "Create schema",
      "files": ["prisma/schema.prisma"],
      "deps": [],
      "batch": 1
    },
    {
      "id": "2",
      "name": "Add service",
      "files": ["src/services/user.ts"],
      "deps": ["1"],
      "batch": 1
    },
    {
      "id": "3",
      "name": "Create routes",
      "files": ["src/routes/user.ts"],
      "deps": ["2"],
      "batch": 2
    }
  ],
  "batches": [
    { "id": 1, "tasks": ["1", "2"], "status": "pending" },
    { "id": 2, "tasks": ["3"], "status": "pending" }
  ]
}

Pre-Execution Validation

Per batch-size-verification.md:

if tasks.length > 15:
  WARN "Large execution (${tasks.length} tasks). Consider splitting."
  [Continue] [Abort]

if any_batch.tasks.length > 7:
  WARN "Batch ${batch.id} exceeds 7 tasks. Review batch boundaries."
  [Continue] [Adjust batches] [Abort]

LOC Estimation

Per review-criteria.md (400 LOC threshold):

Estimate task LOC from file list.
If estimated > 400 LOC:
  WARN "Task ${task.id} may exceed review threshold (~${estimate} LOC)"
  [Proceed] [Split task] [Accept with caveat]

Context Budget Validat