adaptive-routing

Default: direct execution

Complete the task directly when it is reasonably contained and can be validated through local inspection, compilation, tests, linting, or a focused runtime check.

Examples:

• fixing a localized bug;
• adding an endpoint;
• changing a CLI option;
• updating a query;
• adding tests;
• modifying configuration;
• implementing a small feature;
• refactoring a limited component;
• updating documentation.

Delegation rules

Before spawning any subagent, identify a narrow question or a concrete decision, risk, domain constraint, or failure mode. A full blocker is not required: use one focused specialist early when its output is likely to change the next implementation step.

• Delegate when a specialist can answer faster or with higher confidence than reading alone.
• Stop delegation when the next agent is unlikely to change the decision.
• Keep agent outputs concise; pass summaries, not full transcripts.
• Do not chain planner then explorer then worker then reviewer.
• Use one focused specialist for concrete uncertainty or a specialist domain; keep its scope narrow and read-only.

Allowed escalation paths

Use at most one specialist per concrete decision, risk, or failure mode:

• lean-planner: ambiguous requirements, architectural choices with multiple viable designs, destructive migrations, difficult rollback, or conflicting requirements after a short scan.
• lean-explorer: one narrow question about an unfamiliar or cross-cutting code path when local reading would duplicate broad context. Use Luna at high.
• lean-debugger: root-cause investigation with Luna at max after Luna reproduces or inspects a difficult, intermittent, cross-layer, previously failed, or multi-hypothesis failure.
• lean-native: native or unsafe systems code with Luna at max: C/C++, Rust, SIMD, CUDA, Metal, FFI, memory ownership, or atomics.
• lean-reviewer: when the user explicitly requests a review, or for security, concurrency, migrations, public API, unsafe code, large diffs, or changes where tests cannot provide reasonable confidence.
• lean-performance: performance work with a reproducible benchmark, profile, or identified hot path. Use Luna at high.
• lean-ai-ml: model inference, training, quantization, RAG, OCR, TTS, CUDA, Metal, MLX, llama.cpp, or vLLM. Use Luna at high.
• lean-infrastructure: deployment, CI/CD, containers, networking, cloud, observability, scaling, or reliability. Use Luna at high.