Add SubQ backend with per-backend request limits; release 0.10.0

CHANGELOG.md

@@ -2,6 +2,22 @@
 
 ## Unreleased
 
+## 0.10.0
+
+### Added
+
+- A `subq` backend for SubQ's decision model: `createTypeSafe({ backend: "subq" })` sends judgments to `api.subq.ai` under `/v1/systemone` with the key from `SUBQ_API_KEY`. The default model is `subq-snap-preview`, priced at $0.05 per million input tokens. Model ids go to SubQ unchanged. Its model list, under `/v1/models`, is public, so `listModels()` there does not verify the key.
+- Per-backend request limits. `BackendLimits` names the questions in one request, the options in one Choice, and the levels in one Score; `DEFAULT_BACKEND_LIMITS` holds the schema's bounds, which apply to a backend that declares none. A registry entry or a `BackendEndpoint` declares its own in `limits`, and no limit can be above the schema's bound. `subq` takes at most 16 questions, a Choice of at least 2 options, and a Score of at most 10 levels. `ResolvedBackend.limits` and `TypeSafe.limits` give the active backend's limits without a request.
+- `chunkEvaluationRequest` takes a `backend` option and chunks to that backend's question limit.
+
+### Changed
+
+- `evaluate`, `evaluateMany`, and `evaluateAll` refuse a request over the backend's limits before they send it, with a `validation` error. On `typesafe`, `openrouter`, `commandcode`, and `liquid` the limits are the schema's own, so nothing new is refused there.
+- `evaluateAll` chunks to the client's question limit, and `chunkEvaluationRequest` to the `backend` option's: 16 on `subq`, 32 elsewhere. A `maxQuestions` above the limit is now capped at it; before, `chunkEvaluationRequest` made chunks that larger, which the schema then refused.
+- `prepareEvaluationRequest` takes an optional `limits`.
+- A backend that names `modelsField` without `modelsPath` now has its model list renamed to the field the SDK reads; before, the field was ignored unless `modelsPath` was set too. No existing registry backend is affected.
+- `TypeSafeBackend` now includes `"subq"`; a consumer with an exhaustive `switch` over it sees a new member.
+
 ## 0.9.1
 
 ### Fixed

README.md

@@ -85,7 +85,7 @@ An enabled extension with no usable key looks exactly like a working one — one
 
 The environment may lower an explicit cap but never raise it. A reached cap raises a `budget` error naming the cap, the amount used, and the day.
 
-Need more than one request? `evaluateAll(request)` asks any number of questions about one state (over 32 are chunked and fanned out), and `evaluateMany(requests)` runs several requests at once. Both preserve order, bound concurrency, never throw, and stop submitting once the budget is gone. [`pi-typesafe/calibrate`](docs/api.md#calibration-pi-typesafecalibrate) turns labelled cases into thresholds with AUC, a sweep, and a replay runner.
+Need more than one request? `evaluateAll(request)` asks any number of questions about one state (over the backend's limit, 32 or 16 on SubQ, are chunked and fanned out), and `evaluateMany(requests)` runs several requests at once. Both preserve order, bound concurrency, never throw, and stop submitting once the budget is gone. [`pi-typesafe/calibrate`](docs/api.md#calibration-pi-typesafecalibrate) turns labelled cases into thresholds with AUC, a sweep, and a replay runner.
 
 ## Commands
 
@@ -117,7 +117,7 @@ The main model can answer any of these questions in prose. It is slower, costs m
 Only the state and questions you (or the agent, once enabled) submit, to `https://api.typesafe.ai` only. No files, conversation history, or telemetry. Error messages never include upstream response bodies, headers, keys, or your submitted state.
 
 **Limits?**
-Per request: 32 questions and 64 KiB of JSON. Per session: 20 attempts, 15-second timeout, no automatic retries. Per day: no cap unless you set one. Session limits reset when a session starts or reloads; daily counters live in `~/.pi/agent/pi-typesafe/usage.json` and roll over at local midnight. The SDK's `TYPESAFE_BASE_URL` and `TYPESAFE_LOG_LEVEL` overrides are ignored.
+Per request: 32 questions (16 on SubQ) and 64 KiB of JSON. Per session: 20 attempts, 15-second timeout, no automatic retries. Per day: no cap unless you set one. Session limits reset when a session starts or reloads; daily counters live in `~/.pi/agent/pi-typesafe/usage.json` and roll over at local midnight. The SDK's `TYPESAFE_BASE_URL` and `TYPESAFE_LOG_LEVEL` overrides are ignored.
 
 ## For extension authors
 
@@ -142,7 +142,7 @@ if (!answer.ok) return { skipped: answer.errorCode === "budget" };  // never thr
 
 Your extension owns its own user consent and budget; `/typesafe enable` applies only to this package's tool. Check `authState()` rather than your own consent flag before you report that judgments are on.
 
-Judgments can also go through OpenRouter: `createTypeSafe({ backend: "openrouter" })` sends them to `openrouter.ai` with the key from `OPENROUTER_API_KEY`. Command Code serves the same protocol: `createTypeSafe({ backend: "commandcode" })` sends them to `api.commandcode.ai` with the key from `COMMANDCODE_API_KEY` and the model `typesafe/jev`. So does Liquid AI: `createTypeSafe({ backend: "liquid" })` sends them to `api.liquid.ai` with the key from `LIQUID_API_KEY` and the model `d1:free`, which is free; pass `model: "d1"` for the paid model, priced at $0.040 per million input tokens. Liquid counts the state once per question, where TypeSafe counts it once per request, so a request with N questions costs about N times the state on Liquid. Speed, measured on 2026-10-03 with one request at a time: `d1:free` took a median of about 14 seconds and sometimes failed with HTTP 504 after about 46 seconds, while `d1` took about 0.5 seconds. With `d1:free`, raise `timeoutMs` above its 15-second default, or pass `model: "d1"`. Or pass a `BackendEndpoint` object as `backend` to name any host that serves the protocol: it brings its own key variable, never reads `TYPESAFE_API_KEY`, and `backendHost()` reports its destination host for your consent text. These backends have no login store, so `/typesafe login` does not apply to them. Pass the same `backend` to `authState`, `keySituation`, and `ensureApiKey`, or the status you report describes the TypeSafe key while the requests use another one. The `/typesafe` commands and the `typesafe_evaluate` tool always use the TypeSafe backend. Every export — the client, `ask`, batching, the usage ledger, auth state, and the `pi-typesafe/calibrate` and `pi-typesafe/ui` entry points — is in [docs/api.md](docs/api.md).
+Judgments can also go through OpenRouter: `createTypeSafe({ backend: "openrouter" })` sends them to `openrouter.ai` with the key from `OPENROUTER_API_KEY`. Command Code serves the same protocol: `createTypeSafe({ backend: "commandcode" })` sends them to `api.commandcode.ai` with the key from `COMMANDCODE_API_KEY` and the model `typesafe/jev`. So does Liquid AI: `createTypeSafe({ backend: "liquid" })` sends them to `api.liquid.ai` with the key from `LIQUID_API_KEY` and the model `d1:free`, which is free; pass `model: "d1"` for the paid model, priced at $0.040 per million input tokens. Liquid counts the state once per question, where TypeSafe counts it once per request, so a request with N questions costs about N times the state on Liquid. Speed, measured on 2026-10-03 with one request at a time: `d1:free` took a median of about 14 seconds and sometimes failed with HTTP 504 after about 46 seconds, while `d1` took about 0.5 seconds. With `d1:free`, raise `timeoutMs` above its 15-second default, or pass `model: "d1"`. SubQ too: `createTypeSafe({ backend: "subq" })` sends them to `api.subq.ai` with the key from `SUBQ_API_KEY` and the model `subq-snap-preview`, priced at $0.05 per million input tokens. SubQ takes at most 16 questions in one request, a Choice of at least 2 options, and a Score of at most 10 levels: `evaluate` refuses a larger request before it sends it, `evaluateAll` splits to 16, and `client.limits` tells you the limit without a request. Or pass a `BackendEndpoint` object as `backend` to name any host that serves the protocol: it brings its own key variable, never reads `TYPESAFE_API_KEY`, and `backendHost()` reports its destination host for your consent text. These backends have no login store, so `/typesafe login` does not apply to them. Pass the same `backend` to `authState`, `keySituation`, and `ensureApiKey`, or the status you report describes the TypeSafe key while the requests use another one. The `/typesafe` commands and the `typesafe_evaluate` tool always use the TypeSafe backend. Every export — the client, `ask`, batching, the usage ledger, auth state, and the `pi-typesafe/calibrate` and `pi-typesafe/ui` entry points — is in [docs/api.md](docs/api.md).
 
 ## Development