cost-lean-orchestrationlisted

# Cost-Lean Orchestration ## The core rule The context window of the main model is the most expensive real estate in the session. Every raw file, log dump, or search result loaded into it is paid for on every subsequent turn, and it crowds out the reasoning the main model is actually there for. So: **The main model synthesizes. Sonnet subagents read.** - Anything whose primary cost is *reading* (exploration, search, bulk file reading, log scanning, doc digestion, call-site auditing) → delegate to a **Sonnet** subagent. - Anything whose primary cost is *thinking* (architecture, tradeoffs, subtle correctness, root-cause judgment, final implementation of hard code, the answer to the user) → keep in the **main model**. - **Never route to Haiku.** Exploration quality feeds directly into synthesis quality; a missed call site or a misread invariant poisons everything downstream. Sonnet is the floor. If a subagent task itself demands deep reasoning, it doesn't belong in a subagent at all — see "Escalation" below. This is not optional frugality. Delegating reads is what *preserves* quality: it keeps the main context clean, focused, and far from compaction, so the expensive reasoning happens over distilled signal instead of raw noise. ## Routing table | Task | Route | Why | |---|---|---| | Map an unfamiliar codebase / find where X is implemented | Sonnet subagent(s) | Pure reading; only the map matters | | Read 3+ files, long files, generated code, vendored deps |