What problem does it solve? Multi-agent orchestration systems waste tokens on verbose inter-agent messages. This Skill defines a governed protocol for compressing agent-to-agent and agent-to-tool payloads using the Caveman compression provider, reducing token consumption while protecting critical channels from information loss. ## Core Features & Use Cases - Five-Phase Compression Pipeline: Bootstrap configuration, identify the message channel, apply channel-appropriate compression levels, run drift detection on responses, and log token statistics to wiki/log.md. - Channel-Aware Policy Enforcement: Maps sender/receiver/kind tuples to channels (orchestrator_to_subagent, tool_to_subagent, etc.) with conservative, aggressive, or custom policy profiles, enforcing invariants R.C1–R.C6 that forbid compression of user-facing output, artifacts, and cross-factory messages. - Drift Detection and Recovery: Scans receiver responses for ambiguity markers (AMBIGUOUS_HANDOFF, REQUEST_CLARIFY), retries once in normal mode, and escalates to human review after repeated drift. - Use Case: An orchestrator dispatches a task package to a dev sub-agent; the protocol compresses the payload per the configured profile, logs the token ratio, and automatically falls back to uncompressed mode if the sub-agent signals confusion. ## Quick Start Ask the agent to run a compression dry-run on a sample payload for the orchestrator_to_subagent channel using the conservative profile.
