BB Auto Router

Auto Router is a standalone bb extension that creates new threads after choosing a concrete provider, model, and reasoning level from:

1. a model-rated 0–100 task difficulty;
2. live Codex, Claude Code, and Cursor quota remaining;
3. a bundled CursorBench 3.2 score and cost-per-task snapshot; and
4. a user-controlled $ to $$$ frugality preference.

No third-party benchmark API is called at runtime. Grok 4.5's published scores are reduced by 10% as part of the local policy.

What happens to your prompt

Every submission first runs a short hidden classification turn on one of your installed providers, which may not be the provider the thread ends up on. Before routing, Auto Router spawns a hidden thread titled "Autorouter classification" and sends it your prompt text (truncated to 20,000 characters) plus any custom rating instructions you configured. Its only job is to return a 0-100 difficulty score.

• Provider choice: with the default automatic decision agent this is Cursor gpt-5.6-sol-medium, else Codex gpt-5.6-luna, else the first usable model with quota remaining. Pick a fixed classifier under Extensions → Plugins → Auto Router if you want your prompts to go to one known vendor.
• The classification thread is archived and stopped as soon as the score is read.

Residual risk

The classifier is a full coding agent, not a sandboxed text endpoint. It is instructed not to solve, explain, or act on the task, but prompt text is untrusted input and a crafted task could try to make it take actions. Auto Router bounds this two ways:

• It always runs the classifier in a project-default environment, never the environment or worktree the routed thread is headed for, so it cannot touch the checkout you are working in.
• It requests the least-privileged permission preset the provider advertises. bb 0.39 has no read-only preset available to plugins, so that floor is accept-edits — a determined injection could still edit files inside the throwaway classification workspace.

Install

npm install
bb plugin install . --yes

Open Auto Router in bb's sidebar, compose with bb's native new-thread composer, and submit. The extension preserves project, environment, permission, prompt, mentions, and attachments. It classifies the task, shows the score and selected agent, creates the real thread with an explicit provider/model/reasoning tuple, and opens it.

Settings live under Extensions → Plugins → Auto Router and autosave. The same policy is available to agents and scripts:

bb autorouter status
bb autorouter config --frugality 25
bb autorouter config --instructions "Route CSS-only work to Composer 2.5."
bb autorouter route --prompt "Fix the header spacing"

Routing policy

The classifier is instructed to return only compact JSON containing a 0–100 difficulty score. automatic prefers a model with the lightest available reasoning. Cursor currently advertises none on individual model rows while its ACP launch contract accepts low as the minimum, so the extension reconciles to low. If classification fails, routing uses 50/100 and records a plugin warning.

At the frugality endpoints, the capability curves are anchored to:

Difficulty	$	$$$
1	GPT-5.6 Luna low	Grok 4.5 medium
50	GPT-5.6 Luna high	GPT-5.6 Sol high
75	Grok 4.5 medium	Fable 5 high
100	GPT-5.6 Sol medium	Fable 5 max

Intermediate values interpolate between those anchors. Provider quota adds a penalty that grows each time remaining quota halves. A 1/100 task cannot reach the strongest endpoint and a 100/100 task cannot reach the weakest endpoint.

The classifier may return a model override only when the user prompt requests that model or custom instructions explicitly route the matching task to it. The extension independently checks that the model exists and its name is grounded in one of those two sources before honoring the override.

Supported providers

The provider, model, and CursorBench snapshot tables are compiled into the extension (router.ts, benchmarks.ts) and cover Codex, Claude Code, and Cursor as of v0.2.0. Models outside that table are still routable as fallbacks but do not get a capability or cost score, so the table needs a new release whenever a provider ships new models.

Current bb extension boundary

bb 0.39 plugins cannot intercept the native root New Thread submit or insert synthetic thought rows into a core thread timeline. Auto Router therefore uses its own sidebar page containing bb's native NewThreadComposer; routing activity is shown on that page and in a toast. The created thread itself is a normal bb thread owned by the selected provider.

Development

npm test
npm run typecheck
npm run build
bb plugin dev