runtime

Use a reasoning circuit breaker that detects repeated conclusions, recursive planning, and unproductive analysis. When triggered, the agent consolidates evidence, picks the best available decision, executes, and verifies instead of continuing to loop.

Enforce tool and context budgets: require every tool call to produce meaningful progress, prefer targeted file reads over directory scans, and reuse already-obtained information. Stop searching once sufficient evidence is available.

Search with a specific question or hypothesis and prefer sources in order: official documentation, specifications, source repositories, then project code. Avoid inspecting node_modules or browsing directories without a concrete reason.

Escalate only for architecture changes, security-sensitive operations, data integrity risks, destructive actions, complex cross-module dependencies, or ambiguous requirements. Routine tasks should use lightweight reasoning and targeted execution.

Yes, for localized tasks repository-wide exploration is prohibited unless justified by evidence. Agents must start from files directly related to the request and expand scope only when necessary.