DACP Assembler
Compose deterministic agent communication bundles from skill library artifacts.
When to Use
- Handing off work between agents (task assignment, verification, escalation)
- Packaging structured data with executable scripts for a receiving agent
- Upgrading prose-only messages to structured bundles for reliability
Quick Start
- Classify handoff -- identify the handoff type from the taxonomy
- Check drift -- look up historical drift rate for this pattern
- Determine fidelity -- run the decision model (Level 0-3)
- Search catalog -- find matching scripts and schemas
- Compose bundle -- assemble intent + data + code at the decided level
Fidelity Levels
| Level | Name | Contents | |-------|------|----------| | 0 | PROSE | Intent markdown only | | 1 | PROSE_DATA | Intent + structured JSON data | | 2 | PROSE_DATA_SCHEMA | Intent + data + JSON Schema validation | | 3 | PROSE_DATA_CODE | Intent + data + schemas + executable scripts |
For detailed criteria and decision tree, see @references/fidelity-levels.md
Assembly Process
- Receive handoff request with intent, data, and context
- Assess data complexity (none / simple / structured / complex)
- Count available skills from catalog
- Build fidelity decision input from request + computed values
- Run
determineFidelity()to get proposed level - Apply SAFE-02 clamping if changing from current level (max 1 step)
- Compose artifacts at the decided level
- Record assembly rationale (justification, skills, artifacts)
- Build manifest with provenance metadata
For worked examples at each level, see @references/assembly-patterns.md
Handoff Types
- task-assignment, verification-request, data-transformation
- configuration-update, research-handoff, status-report
- question-escalation, patch-delivery
Full taxonomy at @references/handoff-taxonomy.md
Safety Boundaries
- No code generation -- assembler selects from existing catalog entries, never generates code
- Size limits -- 50KB data payloads, 10KB scripts, enforced at bundle creation
- Bounded learning -- max 1 fidelity level change per cycle (SAFE-02)
- Mandatory provenance -- every script must have source skill + version