Model-specific repo instructions

[enyst]

Update: Split per-directory AGENTS.md concept into a dedicated issue (#1283) and added a reference to GPT-5 prompt-driven behavior.

Scope of this issue now focuses on model-specific repo instructions only:

• Load CLAUDE.md only when using Claude models
• Load GEMINI.md only when using Gemini models

Related

• Prior discussion in PR Add support for loading CLAUDE.md and GEMINI.md as repo skills #1280 (adding CLAUDE.md/GEMINI.md handling)
• Per-directory AGENTS.md options (SDK vs prompt): AGENTS.md per-directory discovery: prompt-driven approach #1283
• GPT-5 specific system prompt (draft): Add GPT-5 specific system prompt template #1236

[enyst]

@OpenHands Fix this issue. Read it carefully and investigate the code, so that the solution you propose is inline with codebase patterns. Send a PR.

[openhands-ai]

I'm on it! enyst can track my progress at all-hands.dev

[openhands-ai]

Here’s a concise summary of what I implemented and shipped for issue #1282 (Model-specific repo instructions):

What I changed

• AgentContext
  • get_system_message_suffix now optionally accepts llm_model and llm_model_canonical.
  • When model info is provided, repo skills are filtered so:
    • CLAUDE.md (skill name "claude") is only included for Anthropic/Claude models.
    • GEMINI.md (skill name "gemini") is only included for Gemini models.
  • If model info is not provided, behavior is unchanged (backward-compatible).
• AgentBase
  • Passes self.llm.model and self.llm.model_canonical_name to AgentContext.get_system_message_suffix(), enabling model-specific filtering automatically at system prompt construction time.
• Tests
  • Added tests/sdk/context/test_agent_context_model_specific.py to verify correct inclusion/exclusion for Claude, Gemini, other providers, and canonical-name routing.

Quality checks

• Pre-commit hooks (ruff format/lint, pyright, etc.) passed.
• Unit tests:
  • New file: 4 tests passed.
  • Existing AgentContext tests also passed.

Why this matches the request

• CLAUDE.md is used only for Claude/Anthropic models; GEMINI.md only for Gemini models.
• No change to how skills are loaded; filtering happens at render time, minimizing disruption and preserving current architecture patterns.
• Backward compatible: existing callers of AgentContext require no changes.

PR details

• Branch: openhands/model-specific-repo-instructions-1282
• PR: Load model-specific repo instructions (CLAUDE.md for Claude only, GEMINI.md for Gemini only) #1328
• The PR description explains the approach and includes “Closes Model-specific repo instructions #1282”.

Checklist

• Implemented model-specific repo instruction filtering: done
• Ensured backward compatibility: done
• Added focused tests: done
• Ran hooks and tests locally: done
• Opened a PR against main and referenced the issue: done

If you’d like me to extend vendor detection to additional families or adjust logic for other third-party files, I can follow the same pattern.

_{View full conversation}