Skip to content

AGENTS.md

Latest commit

 

History

History
96 lines (66 loc) · 3.6 KB

AGENTS.md

File metadata and controls

96 lines (66 loc) · 3.6 KB

OpenBench — Agent Contract

Minimize mandatory reading. Expand context only when triggered.

0. Read Policy

  1. Read §1 on every task.
  2. Read §2 only when a trigger matches.
  3. No trigger? Follow §3 and stop expanding context.

1. Mandatory Core

1.1 Identity and priority

  • Greet ckl at conversation start.
  • Priority: safety > correctness > maintainability > speed.
  • User: senior engineer; values clean architecture, minimal complexity, direct answers.

1.2 Code style (non-negotiable)

  • Max function body: 20 lines. Extract if exceeded.
  • No comments that restate code. Only "why" comments for non-obvious decisions.
  • Every abstraction must justify itself: if used <2 places, inline it.
  • Delete dead code immediately. No TODOs in committed code.
  • Type annotations are documentation. Verbose names > comments.
  • Between two correct approaches, pick the one with fewer moving parts.
  • Trust caller and type invariants; no unnecessary defensive code.
  • No compatibility shims when requirements change; redesign cleanly.
  • Modify only files relevant to the task.

1.3 Experiment discipline

  • One diff: experiments must test exactly ONE variable between agent_a and agent_b.
  • SDK first: prefer SDK-reported metrics over estimates; fall back only when SDK returns zero.
  • Isolation: every trial runs in its own tempfile.TemporaryDirectory.
  • Reports: every completed experiment gets a reports/YYYY-MM-DD-<slug>.md committed to git.

1.4 Delivery

  • Run lint before delivery: ruff check src/.
  • Small, scoped commits. Warn before destructive operations.
  • Follow Code Review Guide; fix P0/P1 before commit.

2. Progressive Disclosure (trigger-gated)

Trigger Load
Non-trivial multi-step task docs/guides/engineering-workflow.md; create docs/plans/YYYY-MM-DD-slug.md
Memory / history retrieval docs/guides/memory-management.md; summaries first
Code quality / simplification question docs/guides/code-simplification.md
Architecture or module boundary question §5 Project Snapshot below
User correction Codify preventive rule in docs/memory/user-patterns.md before resuming
Error or regression Write entry to docs/error-experience/entries/
Notable win or insight Write entry to docs/good-experience/entries/

No trigger → do not load.

3. Default Route

  1. Read §1 only.
  2. Inspect target files and neighboring patterns.
  3. Implement minimal correct change.
  4. Proportionate verification.
  5. Commit and report.

4. Experience Records

Type Entry path Summary path
Error docs/error-experience/entries/YYYY-MM-DD-slug.md docs/error-experience/summary/entries/YYYY-MM-DD-slug.md
Win docs/good-experience/entries/YYYY-MM-DD-slug.md docs/good-experience/summary/entries/YYYY-MM-DD-slug.md

Index files stay index-only (no narrative content).

5. Project Snapshot

  • Purpose: automated A/B testing loop to optimize Claude agent configurations.
  • Loop: ExperimentPlanner → ExperimentRunner → AutoEvaluator → ResultStore → repeat.
  • Key dirs: src/openbench/ (core), experiments/ (manual experiments), programs/ (ResearchProgram configs), results/ (raw JSONL data), reports/ (human-readable reports).
  • Entry point: openbench CLI (src/openbench/cli.py).

6. Detail Sources (trigger-gated only)

docs/guides/engineering-workflow.md · docs/guides/code-simplification.md · docs/guides/code-review-guide.md · docs/guides/memory-management.md