ems-dev-harness

Improvements to review

• Extend with quality agent
  • ponytail
  • sonar/similar
• Extend scout with architect companion agent
  • understand-everything
  • treesitter-mcp (maybe overlap with understand-everything)
  • backstage mcp
  • aimee (per: Stacking NeuralMind + Headroom + Ponytail)

The reliability layer for agent-authored pull requests/local develompent. It turns a clearly scoped td task into a reviewed PR with evidence, and makes the next run better when this one fails.

Humans steer. Agents execute. The harness is the product; target-repo code is the output. When agents struggle, fix the harness.

The stack is modern: Effect for types/errors/IO, the Claude Agent SDK as the default agent runtime, LangGraph for the pipeline graph, and @effect/cli for the CLI. See CONTEXT.md for the full design and vocabulary.

Core Loop

ems run --task td-1234 --repo-path ../target-repo

The pipeline runs deterministic phases — the LLM does the work inside each phase, TypeScript decides which phase runs next:

scout → implement → verify → review → quality → close → retrospective

Every run is event-sourced: an append-only JSONL log under <target-repo>/.ems/<task-slug>/ is the source of truth, and reduceEvents reconstructs PipelineState, so runs resume after a crash. Evidence gates (tested, manual-tested, reviewed) must pass before the closer opens a PR.

Architecture

Every provider adapter implements one interface — AgentRuntime.spawn — and provides it as an Effect Layer for the Runtime service tag. The pipeline asks for Runtime; --runtime claude|pi|openrouter|copilot picks the adapter.

CLI

ems run --task <td-id> [--repo-path <path>] [--runtime <r>] [--model <m>] [--merge]
                               # --merge: closer merges into the base branch
                               #   locally (git) instead of opening a PR
ems status <td-id>             # projected pipeline state
ems watch <task-slug>          # tail the event log
ems mark-tested <td-id>
ems mark-reviewed <td-id> --critical 0
ems init                       # write ~/.config/ems/config.json

Surfaces

The harness is exposed through five editor/agent integrations, each delegating to the ems CLI: a pi tool set, an opencode plugin, a VS Code extension (command palette), a copilot-cli MCP bridge, and a Claude Code plugin (/ems-run, /ems-status, /ems-watch slash commands).

Development

This is a Vite+ (vp) + Bun workspace.

vp install        # after pulling
vp check          # fmt + lint + typecheck (whole repo)
vp run -r test    # all package tests
vp run -r build   # build all packages

Per package: bun run check and bun run test.