Add trial harness and MCP server for automated multi-agent coordination

[noelsaw1]

Summary

This PR adds two major new capabilities to the Trinity coordination layer:

1. Trial harness (experiments/coordination-layer/harness/) — a complete system for running automated, repeatable multi-agent coordination trials from the CLI, replacing manual chat-panel babysitting with headless agent execution.
2. MCP server (experiments/coordination-layer/mcp/tick-mcp.js) — exposes all coordination verbs as typed MCP tools, allowing agents to coordinate via MCP instead of shelling out to the CLI.

Both are thin adapters over the existing src/ coordination engine, so behavior is identical to the CLI — they're protocol frontends, not reimplementations.

Key Changes

Trial Harness (harness/)

• Spec parser (src/spec.js) — deterministic, zero-LLM parsing of human-authored project specs (markdown) into structured task lists with validation (duplicate IDs, cycles, empty scopes all hard-fail).
• Orchestrator (src/run.js) — builds isolated per-run workspaces (throwaway git repos with their own .tick/ state), seeds backlogs, spawns agent CLIs concurrently, captures transcripts, runs verification, and generates reports.
• Driver abstraction (src/drivers.js) — unified interface for invoking Gemini, Codex, Claude, and mock agents headlessly (prompt on stdin, no chat UI).
• Preflight phase (src/preflight.js) — agents ask clarifying questions before work begins; questions are collected and a human gate pauses the run for review.
• Observability (src/observe.js) — structured JSONL spine + per-agent transcripts + final analysis report.
• CLI (bin/trial) — list, doctor, validate, preflight, run subcommands with options for agent override, transport selection, timeouts, and circuit-breaker tuning.
• Trial specs — three example projects (build-todo-api, build-url-shortener, debug-calc-bugs, debug-poisoned-task) demonstrating build, debug, and circuit-breaker scenarios.
• Fixtures — seeded-bug code (calc-bugs, poison) for debug trials.
• Test harness (test/smoke.sh) — validates the entire battery with the deterministic mock driver (no API keys needed).

MCP Server (mcp/tick-mcp.js)

• Minimal JSON-RPC 2.0 over newline-delimited stdio (the MCP stdio transport).
• Zero dependencies (per spike rules).
• Implements initialize, tools/list, tools/call, ping.
• Tools map 1:1 to CLI verbs: tick_init, tick_log, tick_project, tick_take, tick_next, tick_claim, tick_scope, tick_release, tick_break, tick_done, tick_reap, tick_analyze.
• Smoke test (mcp/test/mcp-smoke.js) validates the protocol end-to-end.
• Example .mcp.json wiring for Claude Code.

Integration

• Agent prompts for both CLI (prompts/agent-loop.md) and MCP (prompts/agent-loop-mcp.md) modes.
• Preflight prompt (prompts/preflight.md).
• Prompt builder (src/prompts.js) with template rendering.
• Process runner (src/proc.js) — spawn, feed stdin, capture stdout/stderr, enforce timeouts.

Notable Implementation Details

• Isolation by design: Each trial run gets its own workspace with its own .tick/ state; the real repo is never touched. Trials are safe to run repeatedly and in parallel.
• Deterministic mock driver (harness/src/mock-agent.js and test/fake-cli/) — stands in for real Gemini/Codex CLIs, exercises the full coordination protocol and harness observability without API keys.
• Spec validation is strict — duplicate task IDs, empty scopes, non-numeric priority, dependency cycles, and

https://claude.ai/code/session_01WnzAdCRGrrhukvW1etFLyB