// the find
GammaLabTechnologies/harmonist
Portable AI agent orchestration with mechanical protocol enforcement. 186 agents, zero runtime dependencies.
Harmonist is a drop-in pack of 193 agent definition files for AI coding assistants (Cursor, Claude Code, Copilot, etc.) that pairs the catalogue with IDE hooks claiming to mechanically enforce protocol gates — requiring QA review, memory updates, and subagent dispatch before a turn can complete. It targets teams who've been burned by an LLM agreeing to follow rules and then quietly skipping them. Zero runtime dependencies, pure Python stdlib.
The stop-hook design is the genuinely interesting part: instead of trusting the model to self-police, the hook returns a followup_message to the AI enumerating what's missing and refuses to close the turn until the gaps are addressed. That's meaningfully stronger than prompt-only guidance, even if 'mechanical' overstates it. The repomap.py tool is a clean, stdlib-only symbol/import graph (ast + sqlite3) that lets the scout agent query structure instead of grepping files — real token savings for large codebases. Supply-chain integrity via MANIFEST.sha256 with upgrade-time tamper detection is rare in this space and handles a real threat: a modified security-reviewer.md that always returns 'approve'. Memory correlation IDs generated by hooks rather than by the LLM is a sound design — the audit linkage doesn't depend on the model being honest.
The 'mechanical enforcement' claim is marketing. The stop hook returns a followup_message asking the AI to fix things — the AI then decides whether to comply or emit PROTOCOL-SKIP. The readme itself says 'the AI literally cannot ship a code change that skipped review', which is false; it can, with a one-liner escape hatch. The skip-rate abuse detection is also self-reported via markers the model writes. The agent count (186 in the repo description, 193 in the README badge — a live discrepancy) signals the number is a marketing metric; many catalogue entries are thin markdown for extremely niche domains like Kuaishou strategist and academic narratologist, and quality at that breadth is hard to maintain. The integration path — 'paste integration-prompt.md into Cursor Agent mode and let the AI wire everything up' — is the weakest link for production use: when the LLM misreads an unusual project layout, the failure mode is a silently misconfigured enforcement layer, which is worse than no enforcement.