Agent Conductor

AGENTS.md in, governed agent team out.

Agent Conductor is an MCP server that turns the two conventions the coding-agent ecosystem has converged on — AGENTS.md operating manuals and SKILL.md skills — from passive documentation into an active orchestration layer, with a consensus-hardened decision engine gating high-stakes changes.

• Mirrors: codeberg.org/cubiczan/agent-conductor · github.com/icohangar-ops/agent-conductor
• License: MIT
• Status: v0.1 — working scaffold; see Roadmap

The problem

Every serious agent tool — Claude Code, Cursor, Copilot, Codex, Gemini CLI — now reads an AGENTS.md at the repo root and a catalog of SKILL.md files. But both conventions are honor-system prose:

• Nothing compiles the contract. The non-negotiable rules, layer boundaries, and verification checklists live as markdown the agent may or may not internalize.
• Nothing gates the decision. An agent that's about to rewrite your scoring model proceeds with the same confidence as one renaming a variable.
• Nothing verifies the checklist ran. "Run npm test before handing off" is a suggestion, not a gate.

Conductor makes the conventions executable — without asking any agent tool to change. It ships as a standard MCP server, so anything that speaks MCP gets contract compilation, skill discovery, and decision gating for free.

How it works

MCP client (Claude Code / Cursor / Copilot / ...)
        │  stdio (JSON-RPC, MCP)
        ▼
┌────────────────────────────────────────────────┐
│ TypeScript front end (src/)                    │
│   contract/parser.ts   AGENTS.md → contract    │
│   contract/workspace.ts Multi-root merge       │
│   skills/loader.ts     SKILL.md discovery      │
│   server.ts            7 MCP tools             │
└────────────────┬───────────────────────────────┘
                 │  newline-delimited JSON, child stdio
                 ▼
┌────────────────────────────────────────────────┐
│ Python decision engine (engine/)               │
│   vendored Consensus Hardening Protocol core   │
│   R0 gates · foundation attacks · lifecycle    │
└────────────────────────────────────────────────┘

Three capability groups:

1. Contract — compile an AGENTS.md into structured mission, non-negotiable rules, layer do/don't boundaries, verification gates, skill recommendations, and an out-of-scope list.
2. Skills — discover SKILL.md skills across project and personal scopes with progressive disclosure: metadata costs ~100 tokens, bodies load only on demand.
3. Decision — gate work through the Consensus Hardening Protocol: a cheap R0 sanity gate before work starts, and an adversarial foundation-attack pass before a high-stakes change locks.

Quick start

Requirements: Node 23+ (runs TypeScript natively) and Python 3.9+ (stdlib only — the engine needs no pip installs).

git clone https://codeberg.org/cubiczan/agent-conductor.git
cd agent-conductor
npm install
npm test            # TypeScript tests (parser, skills, multi-root, live engine bridge)
npm run test:engine # Python bridge protocol tests
npm run build

Register with Claude Code:

claude mcp add agent-conductor -- node /path/to/agent-conductor/dist/index.js

Or in any MCP client's JSON config:

{
  "mcpServers": {
    "agent-conductor": {
      "command": "node",
      "args": ["/path/to/agent-conductor/dist/index.js"]
    }
  }
}

Set CONDUCTOR_PYTHON if your Python 3 lives somewhere other than python3.

Then, from any project that has an AGENTS.md:

"Load this project's agent contract, list its verification gates, and run a decision_adversary pass on the change I'm about to make."

Tool reference

contract_load

Compile an AGENTS.md (or CLAUDE.md) into a structured contract. Accepts a file path, a project directory, an explicit roots list, or a rootsFile map; defaults to the current working directory. Multi-root workspaces emit one contract whose layer table and verification commands are merged. Section bodies stay off this tool so callers remain inside progressive- disclosure budgets (metadata + structured fields only).

// input
{ "path": "examples/pipeline-pulse" }

// output (abridged — real output from the bundled example)
{
  "source": "examples/pipeline-pulse/AGENTS.md",
  "title": "AGENTS.md — Pipeline Pulse CRM",
  "mission": "Pipeline Pulse CRM is a lightweight, local-first pipeline review dashboard...",
  "rules": [
    "Deterministic logic — same inputs → same scores, labels, and summaries...",
    "Logic in crm.js — keep main.js thin (fetch, render, events).",
    "... (6 total)"
  ],
  "layers": [
    { "layer": "src/crm.js", "role": "Domain logic",
      "do": "Deterministic scoring, filtering, summaries", "dont": "DOM manipulation" }
  ],
  "gates": [
    { "name": "Code change checklist", "commands": ["npm test"], "notes": "" },
    { "name": "Before completion", "commands": [], "notes": "npm test — all green...\n..." }
  ],
  "skills": [
    { "task": "CRM scoring / forecast changes", "skill": "obra/test-driven-development",
      "url": "https://github.com/obra/superpowers/...", "why": "Tests-first changes to deterministic logic" }
  ],
  "outOfScope": ["External CRM integrations (Salesforce, HubSpot, etc.)", "..."],
  "sectionCount": 28
}

The parser is lossless: sections it doesn't recognize are preserved verbatim, so nothing in an unconventional AGENTS.md is dropped.

contract_verification

Returns only the verification gates — the named checklists and shell commands that must pass before work is handed off. Pair it with your agent's workflow: run the commands, confirm success, then declare done. Accepts the same single-root or multi-root inputs as contract_load.

skills_list

Discover SKILL.md skills visible from a project root or a declared multi-root workspace. Metadata only. Extra source roots outside a module directory are included when they appear in the roots list or map.

// input
{ "projectRoot": "examples/pipeline-pulse" }

// output
{
  "skills": [
    {
      "name": "pipeline-scoring",
      "description": "Explain and modify scoreDealRisk weights in src/crm.js with matching test updates...",
      "version": "0.1.0",
      "scope": "project"
    }
  ]
}

Search order (first hit per skill name wins). In a multi-root workspace the project rows run for each declared root, then personal scopes run once:

skill_load

Load the full SKILL.md body for one named skill — the on-demand half of progressive disclosure. Call it only when the task matches the skill's description.

decision_gate

The Consensus Hardening Protocol R0 gate: the cheapest, highest-leverage check, run before doing the work.

// input
{ "solvable": true, "scoped": false, "valid": true, "worth_it": true }

// output
{ "verdict": "HALT", "results": { "Solvable": "PASS", "Scoped": "FATAL", "Valid": "PASS", "Worth_it": "PASS" } }

Any FATAL answer halts: stop and reframe before burning tokens on a problem that isn't scoped, isn't understood, or isn't worth solving.

decision_adversary

A one-shot adversarial pass for high-stakes changes: CHP attacks the claim's foundations, scores them 0–100, and returns devil's-advocate findings plus a session status.

// input
{
  "claim": "Change scoreDealRisk stale-activity weight from 20 to 30",
  "context": "Tests updated; label distribution checked against fixture"
}

// output
{
  "status": "EXPLORING",          // or HALT / REFRAME_REQUIRED
  "foundation_score": 77,
  "findings": [
    "Treat every financial number as unverified until tied to source data.",
    "Require explicit flip criteria for any provisional recommendation."
  ],
  "verification_failures": ["PENDING third-party validation"],
  "report": "## TriangulationRunner Adversary Pass\n..."
}

Statuses map to the CHP decision lifecycle (EXPLORING → PROVISIONAL_LOCK → LOCKED, with HALT and REFRAME_REQUIRED exits): EXPLORING means the claim survived the attack and work may proceed toward a lock; HALT/REFRAME_REQUIRED mean the foundations failed.

engine_status

Health-check the Python engine subprocess. Returns { ok, engine: "chp", version }.

What the parser recognizes

contract_load is convention-based, not schema-based. It extracts the patterns AGENTS.md files in the wild actually use:

Headings inside code fences are ignored; tables tolerate emphasis in headers; markdown links and emphasis are stripped from extracted text.

Writing skills