weiping-lab

Weiping Lab Project Skill

Use this skill for any non-trivial change to Weiping Lab. The goal is an honest research workbench: public-safe config, explicit evidence gates, reproducible workspace state, and no hidden dependency on retired collaboration systems.

1. Plan

• Read AGENTS.md, README.md, CLAUDE.md, pyproject.toml, lab/core/config.py, and lab/runtime.py.
• Run git status --short --branch and do not disturb unrelated user changes.
• Define one version-sized outcome. Avoid tiny isolated tweaks unless the user explicitly asks for one.
• Identify any relation to WEIPING_WIKI, WEIPING_COUNCIL, AGENT_RESOURCE, AGENTIC_SCIENCE, or deepseek-cli; use current files as evidence, not memory alone. Treat historical vipin-* references as compatibility aliases only.

2. Design

• Keep runtime configuration env-driven. Never add tracked private keys, endpoints, account names, or secret-shaped placeholders.
• Keep active memory routed through agentmemory. Local fallback may be an outbox under the configured workspace only; do not restore retired Agent Hub mailboxes or old markdown session dumps.
• Preserve the research gate sequence: phenomenon, kill-first, refine, experiment plan, bridge, results, paper write, review, citation audit, claim audit.
• Preserve docs/RESULT_CONTRACT.md: result files are untrusted, evidence labels are computed locally, and empirical claims require persisted links to paper_result block IDs.
• Make every new check observable through CLI, API, tests, or release scan.
• Keep cross-project links low-coupling: route maps, optional context variables, validation commands, and artifact formats are acceptable; private .env files, local DBs, caches, generated reports, and active services are not.

3. Develop

• Prefer scoped edits in lab/, api/, cli/, ui/, tests/, scripts/, and docs.
• Use apply_patch for manual edits.
• If lab/workspace/ source files are touched, verify they are not gitignored.
• Redact provider errors and persisted diagnostics before writing logs or outbox records.
• Use atomic replacement for durable JSON checkpoints. Never treat a partial result set, a producer-supplied label, or a partial model rubric as a passing gate.
• Bind READY state to the exact plan, paper, and result artifact hashes; any later mutation must downgrade the checkpoint and force deterministic evidence validation again.