Routing: layer, then lane

The layer question, first

Before picking a model, ask whether one is needed at all. Deterministic, repeatable work goes to code that runs without a model (Python/SQL/shell). Judgment, glue and variable inputs go to instructions (skills, agents, rules). The question that decides it: does a model have to DO this, or only ORCHESTRATE it?

Source: home/.claude/CLAUDE.md §Routing.

The lanes

Lane	Agent	Model	Effort	Tools	When
Trivial	mecanico	haiku	—	Read, Grep, Glob, Bash	Search, list, read logs, extract, count, rename
Mechanical	operario	sonnet	low	(all)	The decision is already made; it only needs writing
Prose	redactor	sonnet	medium	(all)	READMEs, commits, reports, TDRs, translations
Recon	investigador	sonnet	medium	(all)	Before touching a repo the session doesn’t know
Ambiguous	profundo	opus	xhigh	(all)	Root cause unclear, real architectural trade-offs

Source: home/.claude/CLAUDE.md §Routing table; frontmatter of home/.claude/agents/mecanico.md, operario.md, redactor.md, investigador.md, profundo.md. mecanico is the only one of the five whose frontmatter restricts tools:; the other four inherit the full toolset.

A sixth agent, reviewer (opus, high, tools restricted to Read, Grep, Glob, Bash(git *)), does not appear in this table because it is not a lane you route ordinary work to — see docs/reviewer.md.

The six rules

1. Recon first. More than 3-4 files to read: delegate to investigador instead of reading them in this thread. Only the summary comes back.
2. The main thread decides, subagents execute. The session settles the what and the how; the writing is delegated.
3. profundo is earned, not assumed. The trigger is ambiguity, not size. A 20-file refactor with the pattern already defined is operario.
4. Do not delegate what is already in context. Handing a file already read to a subagent makes it more expensive, not cheaper.
5. More reasoning in the same thread: ultrathink, rather than raising the session level.
6. Never switch the session model to get a review. Reviewer >= author, but switching mid-session invalidates the prompt cache; spawn reviewer instead.

Source: home/.claude/CLAUDE.md §Routing, rules 1-6.

The cheap path is the default

CLAUDE_CODE_SUBAGENT_MODEL=sonnet is set in home/.claude/settings.json. Every delegation, unless an agent’s own frontmatter overrides the model (as mecanico does with haiku and profundo/reviewer do with opus), already runs on the cheap model. A large session with zero subagent calls is not using this lever.

Source: home/.claude/settings.json (CLAUDE_CODE_SUBAGENT_MODEL).

/route

The route skill (home/.claude/skills/route/SKILL.md) recommends model, effort and execution mode (main session / subagent / new session / worktree / claude -p batch) for a described task, before work starts. It answers in a fixed table (type, model, effort, where, cache cost, needs SPEC.md, needs plan mode, external review after) plus one line of reasoning.

It remembers no model: it reads two files every time and cites them in the table.

• rules/models.json — the facts. Ids, tier (fast < balanced < flagship < frontier), context, max output, prices per million tokens and default effort for each model. Written by python tools/model-intel.py from Anthropic’s public docs, without spending a model token. Do not hand-edit it: regenerate it.
• rules/routing.json — the judgement. Each lane (mecanico, operario, redactor, investigador, profundo, reviewer, session-default) with its model family, effort and why. Edited by hand after every catalogue change; decisions is a dated log of those changes (the kit ships it empty).

The cheap-first cascade maps task type to lane: trivial → mecanico; mechanical → operario; prose → redactor; feature in known code → session-default (or operario if the pattern already exists); unclear-cause debugging → profundo, from a reproduction command; architecture or a long autonomous run → session-default with plan mode and a SPEC.md; exploration → investigador; review → reviewer, whose tier must be at least the author’s. The model and effort are whatever that lane says today. If either file is missing or does not parse, /route says so and stops. tests/route_test.sh resolves five cases from both files and compares them with tests/expected/route-cases.txt.

The skill names the prompt-cache cost explicitly: each model (and, on most models, each effort level) has its own prompt cache, so switching either mid-session forces the next request to re-read the whole conversation with no cache hits — cache reads bill at roughly 10% of standard input, making a switch deep into a long session the most expensive keystroke available. Its conclusions: never recommend switching the session model to get a review (spawn reviewer instead); pick model and effort at the top of a session, and recommend a new session rather than a switch if the task needs a different one; opusplan re-switches the model on every plan-mode toggle, so it’s not recommended for long sessions; and /rewind beats /compact when abandoning a path, because rewind truncates back to an already-cached prefix while compact builds a new one.

Source: home/.claude/skills/route/SKILL.md.