Skip to content

AGENTS.md

Latest commit

 

History

History
32 lines (24 loc) · 2.69 KB

AGENTS.md

File metadata and controls

32 lines (24 loc) · 2.69 KB
Raw

Instructions for coding agents (and humans)

One person, with whatever agents they run, maintains this whole repository. There is no ownership split by path, no other computer to message, and no decision to wait for: if a change needs a file, edit the file.

Read, in this order:

  1. README.md — the product and working paths.
  2. docs/CONTRACT.md — the exact MCP tools, states, envelopes. Source of truth; change it together with the code it describes.
  3. docs/SECURITY.md — the implemented trust boundary.
  4. docs/CLAIMS.md — the official benchmark figures and their claim record.

Working rules

  • Branches. Work on the branch that is checked out. Do not create or switch branches unless asked.
  • Commits. Small, coherent, typechecked (pnpm -r typecheck) and, where tests exist, green (pnpm test). Conventional-commit subjects (feat(runner-wasm): …). Explain why in the body.
  • Shared pieces. packages/shared, docs/CONTRACT.md, db/schema.sql and packages/runner-core are used by the server, the seller worker, the judge and the dashboard. When you change one, update every consumer in the same change and rebuild what depends on it (pnpm --filter @atm/runner-wasm build for the sandbox, pnpm --filter @atm/dashboard build for the UI served by the daemon).
  • Secrets. .env is never committed. scripts/setup-env.sh pastes keys silently. Never print keys, prompts, answers or ciphertext in logs or messages.
  • Honesty in code and demo. Mocks and unsandboxed paths are labelled in the interface (runtime: "node-unsandboxed", kind: "MOCK", PENDING_ONBOARDING). Do not call the Wasm sandbox a TEE, do not call model digests attestation, do not call test-mode money real.
  • Benchmark wording. Use the officially tested results from docs/CLAIMS.md: 30.8% faster task completion and 53.2% cheaper inference. Do not conflate these results with the separate bounded-v1 run or alter raw run artifacts to reflect them.
  • Done means verified. Name the commit, the exact command you ran, what it printed, and what is still missing.

One-machine setup

git clone git@github.com:pranayD21/AgentTaskMarketplace.git && cd AgentTaskMarketplace
brew install ollama wassette && brew services start ollama && ollama pull qwen3:8b   # Node ≥ 22.22 + pnpm 10
pnpm install
# put the keys in .env (or run `pnpm setup:env` and paste them)
pnpm -r typecheck && pnpm test

Run the server with pnpm dev:server and point the daemon or client at it with MARKETPLACE_URL=http://localhost:3000/mcp in .env. The seller dashboard's mock mode (pnpm --filter @atm/dashboard dev) needs no server, keys or Ollama.