adrozdenko/soleri-orchestrate

Orchestrate — Full Work Loop

Run the full work loop with vault intelligence: plan the task, execute it step by step, and close with knowledge capture. Use this for any non-trivial task where missing context or skipping steps would cost time.

When to Use

• User gives a concrete work task ("implement auth", "refactor the parser", "fix the flaky test")
• Task spans multiple files or has cross-cutting concerns
• You want vault patterns and brain recommendations surfaced automatically

Subagent Model Selection (when dispatching)

If Step 2 fans out to subagents, pick the model per subagent — don't use one size for the whole loop:

• simple (exploration, lookup, classification) → haiku
• standard (code impl, refactors, test writing, routine review) → sonnet
• complex (architecture, grading, critical debug, plan creation) → opus

Fallback flows downward only (Opus→Sonnet with warning; Sonnet→Haiku only if task is clearly simple). Never silently escalate upward. Explicit user pin always overrides the rubric.

Before every Agent call, state one line:

Dispatching <description> on <model> (tier=<simple|standard|complex>, reason=<pattern match>).

Full rubric and examples live in the soleri-subagent-driven-development skill (canonical source) and soleri-parallel-execute skill.

Orchestration

Step 1: Plan

YOUR_AGENT_orchestrate op:orchestrate_plan
  params: {
    prompt: "<user task description>",
    context: { domain: "<inferred domain>" }
  }

Present the plan to the user. If grade < A-, ask for approval before proceeding.

Step 2: Execute

Once approved, begin execution. For each task in the plan:

YOUR_AGENT_orchestrate op:orchestrate_execute
  params: {
    planId: "<planId from step 1>",
    taskId: "<current task id>",
    output: "<what you did>"
  }

Log progress after each task: show completed count / total and current task title.

Step 3: Complete

After all tasks are done:

YOUR_AGENT_orchestrate op:orchestrate_complete
  params: {
    planId: "<planId>",
    summary: "<one-line outcome>",
    operatorSignals: {
      qualityNotes: "<any quality observations>",
      unexpectedComplexity: <true|false>
    }
  }

Report the completion summary and any knowledge captured.

Exit Criteria

All plan tasks executed, orchestrate_complete called with outcome, knowledge persisted to vault.