code-review

name	code-review
description	Code review skill for the Ollama Agent Harness — enforces architectural conventions from the Claude Code paper
domain	code-review
confidence	medium
source	generated by CopilotForge
triggers	["review this","check my code","code review"]

Context

Reviews code changes against the architectural conventions established for the Ollama Agent Harness. The harness borrows patterns from the Claude Code paper: minimal scaffolding, deny-first safety, context-as-scarce-resource, append-only state, and subagent isolation.

Patterns

Review Checklist

1. Agent loop integrity — Does the change preserve the simple while-loop pattern? No explicit state graphs or planning frameworks added?
2. Permission enforcement — Are new tools gated through the permission system? Does deny-first ordering hold?
3. Context budget — Does the change increase context consumption? Are tool results capped? Are large outputs truncated or summarized?
4. Append-only persistence — Do session storage changes only append? No in-place mutations of transcript files?
5. Tool classification — Are new tools correctly classified as read-only (concurrent-safe) or exclusive (state-modifying)?
6. Subagent isolation — Do subagent interactions return summary-only? No full history leaking to parent?
7. Type safety — Explicit types on function signatures? No any in public APIs?
8. Error recovery — Are errors surfaced as tool results for the model to adapt, not silently swallowed?
9. Ollama API usage — Correct use of ollama-js streaming, tool calling, and message format?
10. Security — No command injection in tool dispatch? No unvalidated user input passed to shell?

TypeScript-Specific Checks

• Strict mode compliance
• No implicit any
• Proper async/await usage (no floating promises)
• Import organization (path aliases over relative paths)

Examples

Good: Tool with permission check

const result = await permissions.check(toolCall);
if (result.decision === 'deny') {
  return { error: result.reason };
}
return await tool.execute(toolCall.input);

Bad: Tool bypassing permissions

// Missing permission check
return await tool.execute(toolCall.input);

Anti-Patterns

• Approving code that adds explicit planning/state-graph frameworks without strong justification.
• Missing permission checks on new tool implementations.
• In-place mutation of session transcript files.
• Unbounded tool results that consume excessive context.
• Using any as a type escape hatch in core subsystem interfaces.