Model routing

Status: installation paused (2026-08-01).

The rule was removed from all eight repos that had it, and should not be reinstalled for now. It is 1,885 characters of always-loaded text asserting that down-tier delegation improves cost without hurting quality — a claim nothing has measured. Until there is evidence, installing it spends context budget on an unfalsified belief.

If invoked, say this and confirm before writing anything. The removal was deliberate and recent; a reinstall silently undoes it.

Removal still works normally and needs no confirmation.

To resume: delete this block and disable-model-invocation from the frontmatter. What would justify resuming is in docs/internals/model_routing_flywheel.md — the rule needs a definition of a good delegation outcome and a way to observe one before it earns its place in every session.

Install, update, or remove a per-project model-delegation rule. The rule is a plain .claude/rules/ file: it loads at session start in that project only, needs no plugin installed to keep working, and is removed by deleting the file.

The install has two independent layers. The base rule is standalone — no external tool, no CLI. One layer is opt-in: pre-shaped agents. The base is complete on its own.

A third layer, an agent-state feedback section, was removed in 0.5.0. It told Claude to record each delegation outcome via a CLI, into a table that never existed in the live database (the whole package was retired on 2026-08-02) — and the outcome it recorded was the orchestrator grading its own delegation. Delegation data is now recovered observationally from session transcripts instead. If a project still carries that section in its installed rule, delete it.

Design rationale: architecture.md "route to the cheapest capable model" — decomposition quality and model tiering are complements; well-scoped leaf tasks don't need the frontier model.

Install

1. Determine the project root: the git repository root, or the working directory if not in a git repo. If the user names a different target project, use that.
2. Base rule (always). Read references/model-delegation.md (relative to this skill) and write its content verbatim to <project-root>/.claude/rules/model-delegation.md, creating .claude/rules/ if needed. Verbatim copy keeps installs identical across projects — do not regenerate or paraphrase the rule text. This layer is fully standalone; nothing below is required for it to work.
3. If the target file already exists and differs, show the user the diff and ask before overwriting — it may carry local edits.
4. Agents layer (opt-in). Ask whether to also install the pre-shaped delegation agents (or install without asking if the user said "with agents"). If yes, copy verbatim references/agents/fast-executor.md and references/agents/task-coder.md to <project-root>/.claude/agents/, same diff-and-confirm treatment. These give delegation targets tailored execute-to-spec system prompts instead of a bare model override; the rule prefers them automatically when present.
5. Tell the user which layers were installed: the rule (and any agents) load automatically at the next session start. For the current session, adopt the rule's behavior immediately since you have just read it.

Update

Same as install; step 3's diff-and-confirm handles the existing files. If the installed rule still carries the removed agent-state feedback section (a ## Record outcomes block invoking agent-state delegation record), delete that section as part of the update.

Remove

Delete <project-root>/.claude/rules/model-delegation.md, and <project-root>/.claude/agents/fast-executor.md / task-coder.md if they were installed. Nothing else to clean up.

What the rule says

Delegation criteria in brief (full text in references/model-delegation.md): route tasks that are well-specified, mechanical, and verifiable to the cheapest capable model in a subagent; keep design, ambiguity, user interaction, and verification of returned work in the main loop on the strongest model. Tiers are named only as examples so the rule survives model-lineup changes. The base rule stops there — no external dependency.