deepthonk AGENTS.md

DeepThonk Build Instructions

DeepThonk is a provider-neutral, OpenDeepThink-style reasoning optimizer exposed as a TypeScript CLI and MCP server.

The algorithm is from Zhou et al. 2026 (OpenDeepThink: Parallel Reasoning via Bradley–Terry Aggregation, arXiv:2605.15177); the authors' Python reference implementation is at https://github.com/ZhouShang0817/open-deep-think (MIT). DeepThonk is an independent TypeScript reimplementation — not a fork, no code vendored from the reference.

Tagline: thonk harder, not richer.

Build and maintain the project around one shared execution engine in packages/core. The CLI and MCP server are wrappers over that engine; do not duplicate algorithm logic in those packages.

Design principle: agent-composable surface. Every algorithm dimension — population shape (n, k, t, m), regularization (lambda), temperatures, prompt style, and per-phase prompt templates — must be reachable inline through MCP tool arguments and CLI flags, not only through YAML files. Inspection of every intermediate artifact must remain available as MCP resources. See docs/customization.md for the variable contract.

Core requirements:

• Implement population-based candidate generation, randomized pairwise comparison, Bradley-Terry aggregation, elite preservation, critique-guided mutation, bottom-quartile discard, and final dense ranking.
• Keep all model access behind the provider-neutral ModelDriver contract in packages/providers.
• Support fake, openai-compatible, and deepseek providers.
• Do not request hidden chain-of-thought. Ask models for final artifacts, concise rationales, critiques, and strict JSON where needed.
• Never log API keys or env var values. Do not write prompts or raw model output unless explicitly configured.
• MCP is a protocol wrapper over core execution, not a separate execution engine.
• The Streamable HTTP transport must keep DNS rebinding protection on and validate Host against the loopback bind (CVE-2025-66414 class). Do not remove that guard when extending the transport.
• Background MCP jobs (deepthonk.start) must wrap both their success and failure handlers so a filesystem error cannot escape as an unhandled rejection.
• Use pnpm, TypeScript, Vitest, Zod 4, Commander, YAML, and the stable official MCP TypeScript SDK package @modelcontextprotocol/sdk. Do not add p-limit: it was removed in v0.3.0 because nothing imported it — phaseRunner.ts implements its own worker pool.
• zod is part of @deepthonk/core's published API (index.ts re-exports schemas.ts). A Zod major bump is a breaking release, not a dependency chore.
• Anything test/ imports must be declared in the root package.json. test/ sits at the workspace root, so a package reaching it only transitively resolves locally and then fails CI with ERR_MODULE_NOT_FOUND. This has bitten yaml, @modelcontextprotocol/sdk, and zod.

Acceptance checks:

pnpm install
pnpm run build
pnpm test
pnpm --silent --filter deepthonk deepthonk plan --profile paper
rm -rf runs/test-quick
pnpm --silent --filter deepthonk deepthonk run --provider fake --profile quick --task examples/tasks/toy-math.txt --out runs/test-quick
pnpm --silent --filter deepthonk deepthonk inspect runs/test-quick