orchestration

# Orchestration Protocol

This skill provides procedural knowledge for observable subagent orchestration.

## Core Principles

1. **You are the only decision-maker.** Subagents execute; they do not judge, interpret ambiguity, or make policy choices.

2. **Atomic tasks only.** Each subagent receives ONE well-defined task with explicit success criteria.

3. **Structured contracts.** Every dispatch includes:
   - Exact input
   - Expected output schema
   - Explicit constraints
   - Required evidence for claims

4. **Full observability.** Log and review every input/output. No hidden reasoning.

5. **Verification is separate from execution.** Parse and validate all subagent responses.

## Task Decomposition

When given a complex task:

1. Identify atomic units that can be executed independently
2. Determine dependencies between units
3. Sequence units respecting dependencies
4. For each unit, define:
   - Clear objective (single verb phrase)
   - Input data required
   - Expected output format
   - Success criteria

## Dispatch Pattern

For each atomic task, dispatch to the `executor` subagent:

```
Task: [Single atomic instruction]

Input:
[Structured data the executor needs]

Success Criteria:
- [Criterion 1]
- [Criterion 2]
```

## Response Verification

After each executor response, verify:

| Check | Pass Condition |
|-------|----------------|
| Schema compliance | All required YAML fields present |
| Understanding match | `understood` semantically matches intended task |
| Confidence threshold | `confidence >= 0.7` or flag for review |
| Evidence support | `evidence` logically supports `output` |
| No blockers | `blockers` is null or acceptable |

## Verification Actions

```
IF schema_invalid:
  → Log deviation: "schema_violation"
  → Retry with explicit schema reminder

IF understood != intended:
  → Log deviation: "misinterpretation"
  → Refine task description
  → Retry

IF confidence < 0.7:
  → Flag for manual review
  → OR dispatch verification subagent

IF