Repository navigation
[FEATURE]: Smart Rules - Context-aware instruction injection with Claude Code compatibility #10096
Description
Feature hasn't been suggested before.
- I have verified this feature I'm about to request hasn't been suggested before.
Describe the enhancement you want to request
Feature Request
Problem Statement
Currently, there's no way to automatically inject context-specific instructions based on the files being worked with. Users who want file-type-specific guidance (e.g., TypeScript conventions, React patterns, CSS formatting rules)
must either:
- Include all rules in a single
AGENTS.md/CLAUDE.mdfile (bloating every prompt) - Manually reference rules when needed (error-prone and tedious)
Additionally, users migrating from Claude Code have existing .claude/rules/ directories that aren't recognized by OpenCode.
Proposed Solution
Implement a Smart Rule system that:
- Automatically injects relevant rules based on files accessed during the session
- Supports both
.opencode/rules/and.claude/rules/directories for cross-compatibility - Uses glob patterns to match files (e.g.,
**/*.ts,src/components/**/*.tsx) - Caches rule discovery for performance (loaded once per project)
Rule File Format
---
description: "TypeScript formatting guidelines"
paths:
- "**/*.ts"
- "**/*.tsx"
alwaysApply: false
---
# TypeScript Guidelines
- Use strict mode
- Prefer interfaces over types Rule Discovery Locations
| Location | Scope | Precedence |
|---|---|---|
~/.config/opencode/rules/ |
Global | Lowest |
~/.claude/rules/ |
Global (Claude compat) | Low |
.claude/rules/ |
Project (Claude compat) | Medium |
.opencode/rules/ |
Project | Highest |
Configuration
Feature gated behind experimental flag:
export OPENCODE_EXPERIMENTAL_SMART_RULES=true
# or
export OPENCODE_EXPERIMENTAL=true Use Cases
- Language-specific conventions: TypeScript strict mode, Python PEP8, Go formatting
- Framework patterns: React hooks rules, Next.js best practices, Vue composition API
- Project-specific guidelines: Team coding standards, architecture decisions
- Migration from Claude Code: Seamless support for existing
.claude/rules/directories
Benefits
- Reduced prompt bloat: Only relevant rules are injected
- Better organization: Rules are modular and file-based
- Cross-tool compatibility: Works with existing Claude Code setups
- Zero config for basic use: Just create markdown files in the rules directory
Technical Considerations
- Piggybacks on existing
FileTime.read()for tracking (no new hooks) - Uses
Instance.state()for per-project caching - Pre-compiles
minimatchpatterns for performance - Early exit when feature is disabled (zero overhead)
Acceptance Criteria
- Rules in
.opencode/rules/are discovered and parsed - Rules in
.claude/rules/are discovered (Claude Code compatibility) - Global rules from
~/.config/opencode/rules/work - Glob patterns correctly match files
-
alwaysApply: truerules are always included - Higher precedence rules override lower ones (by name)
- Feature is disabled by default (experimental flag)
- No performance impact when feature is disabled
Activity
github-actions commentedon Jan 22, 2026
This issue might be a duplicate of existing issues. Please check:
Feel free to ignore if none of these address your specific case.
dkraemerwork commentedon Jan 22, 2026
You can just use pre or post turn hooks.
raedkit commentedon Jan 22, 2026
Have you read the full feature and checked the proposed pull request ? The implementation is made to not be triggered/verified for each turn, and it's path based activated.
Also this feature is proposed in Claude Code and it will help people migrate from Claude Code or work in parallel in the 2 tools like the already implemented skills and CLAUDE.md files.
dkraemerwork commentedon Jan 22, 2026
Yes, counter question:
When do you think this check of Inject in message or not will be done?
dkraemerwork commentedon Jan 22, 2026
Spoiler: In any turn, its possible to touch different files. So the check is made each turn..
dkraemerwork commentedon Jan 22, 2026
The solution for claude code also just is: pre turn, ls files, is there a CLAUDE.md, yes? read, then continue; else, no, continue
florian-trehaut commentedon Jan 23, 2026
Great feature! This would really help users who work with both Claude Code and OpenCode.
Question about scope expansion: Would it be possible to extend this compatibility to also support
.claude/agents/alongside.claude/rules/?I have custom subagents defined in
.claude/agents/(e.g., specialized code reviewers for Rust, architecture, security, etc.) that I'd like to use in both tools without maintaining duplicate configurations.The frontmatter format differs slightly between Claude Code and OpenCode agents:
name:,tools:(as list),model: haikutools:(as object with booleans),model: anthropic/claude-haiku-4-*A similar discovery + conversion layer for agents would enable seamless cross-tool workflows.
github-actions commentedon Mar 26, 2026
To stay organized issues are automatically closed after 90 days of no activity. If the issue is still relevant please open a new one.