ai-agent-thinkroom AGENTS.md

Repository instructions

Authority

Implement docs/specification.md. Preserve the product semantics in thinkroom_ai_think_tank_product_concept.md. Do not silently reduce P0 scope or release gates.

Engineering method

• Use strict RED → GREEN → REFACTOR for behavior changes.
• Keep domain and application layers independent of FastAPI, SQLite, subprocess, provider SDKs, CLI, and MCP.
• Prefer small cohesive modules and explicit typed ports over framework magic.
• Treat model output as untrusted input: validate schemas, bound retries, and preserve evidence verification status.
• Never use shell=True or interpolate user input into commands.
• Do not add distributed infrastructure without an observed need.
• Keep production defaults loopback-only and fail closed on unsafe public binding while authentication is absent.

Required gates

The implementation is not complete until these pass from a clean checkout:

uv lock --check
uv sync --locked --all-extras --dev
uv run ruff format --check .
uv run ruff check .
uv run mypy src
uv run pytest
python3 scripts/build_release.py --out-dir /absolute/external/empty/dist

Add package-install and native production-process smoke tests/scripts. A production wheel release must include requirements-production.txt, uv.lock, and the flat asset verify_locked_runtime.py; install hash-locked dependencies first and the wheel with --no-deps. Docker is an operator-owned reference integration, not part of the v0.2 native release claim. Docker availability or smoke status does not gate the native release.

Claim discipline

A scripted backend proves orchestration, not model quality. A real provider smoke proves integration, not correctness of the research. SQLite release support is single-instance only.