@vymalo/opencode-ratelimit 0.16.1 → 0.18.0

package/README.md

@@ -2,7 +2,9 @@
 
 OpenCode plugin that makes OpenAI-compatible providers **rate-limit aware**. It reads the IETF draft-03 rate-limit response headers emitted by [Envoy Gateway](https://gateway.envoyproxy.io/docs/tasks/traffic/global-rate-limit/)'s global rate limiting (`x-ratelimit-limit` / `-remaining` / `-reset`), proactively pauses new requests when the quota is exhausted, and backs off + retries on HTTP `429` — so your OpenCode client cooperates with the gateway instead of hammering it.
 
-Auth-agnostic by design: it runs as an OpenCode `config` hook and only wraps the provider's `fetch`. It never reads or sets `Authorization`, so it composes with `@vymalo/opencode-oauth2`, static API keys, or no auth at all.
+Auth-agnostic by design: it observes OpenCode's HTTP request/response hooks and never reads or sets `Authorization`, so it composes with `@vymalo/opencode-oauth2`, static API keys, or no auth at all.
+
+> **Requires OpenCode v2.** This release targets the v2 plugin API and does not run on v1. Migrating from `0.17.x`? See [Migrating from v1](#migrating-from-v1).
 
 ## Why use this
 
@@ -10,14 +12,14 @@ OpenCode has no built-in client-side rate-limit handling. If your inference sits
 
 ## How it works
 
-OpenCode's plugin API has **no post-response hook**, so the only way to observe response status and headers is to inject a custom `fetch` onto the provider's `options.fetch` during the `config` hook (OpenCode forwards it to the AI SDK provider). This plugin wraps that fetch:
+Four registrations on OpenCode's native request lifecycle:
 
-1. **Pre-request gate** — if the last response said `remaining: 0`, hold new requests until `x-ratelimit-reset` elapses.
-2. **Send** the request through the underlying fetch (or any fetch a prior plugin already installed).
-3. **Read** the rate-limit headers from the response and update per-provider state.
-4. **On `429`** — wait the reset window (`x-ratelimit-reset`, or `Retry-After` as a fallback) and retry, up to `maxRetries`.
+1. **`provider.transform`** — read which providers opted in via `settings.rateLimit`. Reads only; it never mutates a provider.
+2. **`http.request`** — pre-request gate. If the last response said `remaining: 0`, hold new requests until `x-ratelimit-reset` elapses.
+3. **`http.response`** — read the rate-limit headers and update per-bucket state.
+4. **`retry`** — on a `429`, return the reset window (`x-ratelimit-reset`, or `Retry-After` as a fallback) as the host's retry delay, up to `maxRetries`; or decline the retry so the `429` surfaces immediately.
 
-The `Response` is returned untouched (no `.clone()`), so the streaming body reaches OpenCode intact.
+The response body is never read — only `status` and `headers` — so the streaming body reaches OpenCode intact. Because `retry` hands the delay to OpenCode, the plugin never blocks a request thread on a `429`; the host sleeps, cancellably, and counts the attempts.
 
 > **Where the headers live.** Envoy attaches its rate-limit `BackendTrafficPolicy` to a specific route — in practice `/v1/chat/completions`, **not** `/v1/models`. So the plugin sees the quota on the inference calls that matter, and a `curl` of `/v1/models` may show no `x-ratelimit-*` at all. The `limit` header is also commonly *multi-policy*, e.g. `x-ratelimit-limit: 200, 200;w=60, 200000;w=60, 50000000;w=2592000` — the parser takes the first token as `limit`, while `remaining`/`reset` already track whichever bucket is closest to its cap (which is all the throttle logic needs).
 
@@ -29,23 +31,21 @@ npm install @vymalo/opencode-ratelimit
 
 ## Usage
 
-A provider opts in via `options.meta.rateLimit`. All fields are optional:
+A provider opts in via `settings.rateLimit`. All fields are optional:
 
 ```json
 {
-  "plugin": ["@vymalo/opencode-ratelimit"],
-  "provider": {
+  "plugins": ["@vymalo/opencode-ratelimit"],
+  "providers": {
     "my-provider": {
-      "npm": "@ai-sdk/openai-compatible",
-      "options": {
+      "package": "aisdk:@ai-sdk/openai-compatible",
+      "settings": {
         "baseURL": "https://api.example.com/v1",
-        "meta": {
-          "rateLimit": {
-            "enabled": true,
-            "maxWaitMs": 0,
-            "maxRetries": 5,
-            "headerPrefix": "x-ratelimit"
-          }
+        "rateLimit": {
+          "enabled": true,
+          "maxWaitMs": 0,
+          "maxRetries": 5,
+          "headerPrefix": "x-ratelimit"
         }
       }
     }
@@ -53,6 +53,21 @@ A provider opts in via `options.meta.rateLimit`. All fields are optional:
 }
 ```
 
+The plugin's own options (currently just `logLevel`) go on its `plugins` entry:
+
+```json
+{
+  "plugins": [
+    { "package": "@vymalo/opencode-ratelimit", "options": { "logLevel": "debug" } }
+  ]
+}
+```
+
+> **`logLevel` is a gate, not an override.** Records go to OpenCode's own logger, which applies the
+> host's level too — and Effect's default minimum is `Info`. So `"logLevel": "debug"` alone will not
+> surface `debug`/`trace` records; the host has to permit them as well
+> (`OPENCODE_LOG_LEVEL=DEBUG`). Both gates must be open.
+
 | Field | Default | Meaning |
 | --- | --- | --- |
 | `enabled` | `true` | Set `false` to opt a provider out while keeping the block. Omitting the whole `rateLimit` block also opts out. |
@@ -62,14 +77,15 @@ A provider opts in via `options.meta.rateLimit`. All fields are optional:
 | `headerPrefix` | `x-ratelimit` | Lowercased prefix of the draft-03 triple. Remap for gateways that use e.g. `ratelimit-*`. |
 | `tiers` | — | Optional array of policy bands keyed by reset magnitude. Replaces flat `maxWaitMs`/`maxRetries` with per-window behavior. See [Tiered policies](#tiered-policies). |
 
-If you also use `@vymalo/opencode-oauth2`, list it **before** this plugin in `plugin` (config hooks run in registration order). It is only a soft recommendation — this plugin's fetch wrapping is auth-independent.
+Ordering in `plugins` does not matter for this plugin. Each v2 plugin registers against a different domain, and this one touches neither auth nor the provider definition.
 
 ### Tiered policies
 
 A single `maxWaitMs` can't tell a **per-minute burst reset** (≤60s — worth waiting through) from a **monthly-budget reset** (days away — should error fast): both arrive as the same `x-ratelimit-reset`. The discriminator is the reset's *magnitude*, and `tiers` lets you act on it. Each tier matches resets up to `maxResetSeconds` (first match wins; ascending, catch-all last):
 
 ```jsonc
-"meta": { "rateLimit": {
+// providers.<id>.settings
+"rateLimit": {
   "scope": "model",
   "tiers": [
     { "maxResetSeconds": 120, "action": "wait", "maxWaitMs": 0, "maxRetries": 3 }, // ≤2min: wait it out
@@ -92,18 +108,18 @@ Notes:
 
 ### Per-model scope
 
-If your gateway buckets per model (e.g. Envoy `BackendTrafficPolicy` selectors that include the model), `scope: "model"` (the default) keys cooldown state per `(provider, model)` — so exhausting one model's bucket never blocks requests to a sibling model that still has quota. The model is read from the request body (`@ai-sdk/openai-compatible` sends a JSON body with `model`); when it can't be parsed the bucket falls back to provider-wide. Use `scope: "provider"` if your gateway buckets per provider/account.
+If your gateway buckets per model (e.g. Envoy `BackendTrafficPolicy` selectors that include the model), `scope: "model"` (the default) keys cooldown state per `(provider, model)` — so exhausting one model's bucket never blocks requests to a sibling model that still has quota. The model comes straight off the hook event, so there is no body parsing and no fallback case. Use `scope: "provider"` if your gateway buckets per provider/account.
 
 ## Caveats
 
 ### Long waits can be cut short by OpenCode's own timeouts
 
-OpenCode wraps the injected `fetch` with its own `headerTimeout` / `chunkTimeout` logic, and our pre-request gate and `429` backoff wait **inside** that fetch. With the default `maxWaitMs: 0`, a 55-second reset window means a 55-second wait — which OpenCode may abort if it exceeds the header timeout. When that happens the wait is cancelled cleanly (a `ratelimit_wait_aborted` event is logged and the abort is propagated as a normal cancellation).
+The pre-request gate holds the request inside OpenCode's own request path, so it competes with the provider's `headerTimeout` / `chunkTimeout`. With the default `maxWaitMs: 0`, a 55-second reset window means a 55-second hold — which OpenCode may abort if it exceeds the header timeout. When that happens the wait is cancelled cleanly (a `ratelimit_wait_aborted` event is logged and the abort is propagated as a normal cancellation). The `429` backoff is the host's own retry delay and is not subject to this.
 
 If you expect long waits, raise the provider's `headerTimeout` (and `chunkTimeout`) above your largest reset window:
 
 ```json
-"options": { "headerTimeout": 90000, "chunkTimeout": 90000 }
+"settings": { "headerTimeout": 90000, "chunkTimeout": 90000 }
 ```
 
 ### State is in-memory only
@@ -138,3 +154,22 @@ The `./lib` subpath exports the testable internals for embedders: `parseRateLimi
 ## License
 
 MIT
+
+## Migrating from v1
+
+`0.18.0` targets the OpenCode **v2** plugin API; `0.17.x` was the last v1 release. See
+[ADR-0019](../../docs/adr/0019-opencode-v2-cutover.md) for why there is no dual-support build.
+
+| v1 | v2 |
+| --- | --- |
+| `"plugin": [...]` | `"plugins": [...]` |
+| `"provider": { ... }` | `"providers": { ... }` |
+| `"npm": "@ai-sdk/openai-compatible"` | `"package": "aisdk:@ai-sdk/openai-compatible"` |
+| `options.baseURL` | `settings.baseURL` |
+| `options.meta.rateLimit` | `settings.rateLimit` |
+| `options.headerTimeout` / `chunkTimeout` | `settings.headerTimeout` / `chunkTimeout` |
+| `pluginConfig.ratelimit.logLevel` | the plugin's own `options.logLevel` |
+| host `config.logLevel` drove the log level | `options.logLevel`, else `OPENCODE_LOG_LEVEL` |
+
+The `rateLimit` block's own fields — `enabled`, `scope`, `maxWaitMs`, `maxRetries`,
+`headerPrefix`, `tiers` — are unchanged, as is every `ratelimit_*` event name.

package/dist/config.d.ts

@@ -14,7 +14,7 @@ export declare const DEFAULT_SCOPE: RateLimitScope;
 *  - **Tiered**: a `tiers` array, each entry `{ maxResetSeconds?, action?,
 *    maxWaitMs?, maxRetries? }`, matched by reset magnitude.
 */
-export declare function parseRateLimitOptions(providerOptions: Record<string, unknown> | undefined): RateLimitOptions | null;
+export declare function parseRateLimitOptions(providerSettings: Record<string, unknown> | undefined): RateLimitOptions | null;
 /**
 * Select the policy tier for an observed reset. `tiers` is pre-sorted ascending
 * with a catch-all last, so the first tier whose bound covers `resetSeconds`

package/dist/config.js

@@ -2,7 +2,6 @@ export const DEFAULT_MAX_WAIT_MS = 0;
 export const DEFAULT_MAX_RETRIES = 5;
 export const DEFAULT_HEADER_PREFIX = "x-ratelimit";
 export const DEFAULT_SCOPE = "model";
-const META_KEY = "meta";
 const RATE_LIMIT_KEY = "rateLimit";
 function asRecord(value) {
 	if (!value || typeof value !== "object" || Array.isArray(value)) {
@@ -84,15 +83,13 @@ function normalizeTiers(tiers) {
 *  - **Tiered**: a `tiers` array, each entry `{ maxResetSeconds?, action?,
 *    maxWaitMs?, maxRetries? }`, matched by reset magnitude.
 */
-export function parseRateLimitOptions(providerOptions) {
-	if (!providerOptions) {
+export function parseRateLimitOptions(providerSettings) {
+	if (!providerSettings) {
 		return null;
 	}
-	const meta = asRecord(providerOptions[META_KEY]);
-	if (!meta) {
-		return null;
-	}
-	const rateLimit = asRecord(meta[RATE_LIMIT_KEY]);
+	// v2 providers have no free-form `options.meta`; `Provider.Settings` is an
+	// open struct, so the opt-in lives directly at `settings.rateLimit`.
+	const rateLimit = asRecord(providerSettings[RATE_LIMIT_KEY]);
 	if (!rateLimit) {
 		return null;
 	}

package/dist/config.js.map

@@ -1 +1 @@
-{"mappings":"AAEA,OAAO,MAAM,sBAAsB;AACnC,OAAO,MAAM,sBAAsB;AACnC,OAAO,MAAM,wBAAwB;AACrC,OAAO,MAAM,gBAAgC;AAE7C,MAAM,WAAW;AACjB,MAAM,iBAAiB;AAEvB,SAAS,SAAS,OAAqD;CACrE,IAAI,CAAC,SAAS,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;EAC/D,OAAO;CACT;CACA,OAAO;AACT;AAEA,SAAS,SAAS,OAAoC;CACpD,OAAO,OAAO,UAAU,YAAY,MAAM,KAAK,CAAC,CAAC,SAAS,IAAI,MAAM,KAAK,IAAI;AAC/E;AAEA,SAAS,UAAU,OAAgB,UAA4B;CAC7D,OAAO,OAAO,UAAU,YAAY,QAAQ;AAC9C;AAEA,SAAS,cAAc,OAAgB,UAA0B;CAC/D,IAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,QAAQ,GAAG;EACpE,OAAO,KAAK,MAAM,KAAK;CACzB;CACA,OAAO;AACT;;AAGA,SAAS,iBAAiB,OAAgB,UAA0B;CAClE,IAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,SAAS,GAAG;EACrE,OAAO,KAAK,MAAM,KAAK;CACzB;CACA,OAAO;AACT;AAEA,SAAS,QAAQ,OAAgC;CAC/C,OAAO,UAAU,cAAc,UAAU,UAAU,QAAQ;AAC7D;AAEA,SAAS,SAAS,OAAiC;CACjD,OAAO,UAAU,UAAU,UAAU;AACvC;;AAGA,SAAS,kBAAkB,OAA+B;CACxD,IAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,QAAQ,GAAG;EACpE,OAAO,KAAK,MAAM,KAAK;CACzB;CACA,OAAO;AACT;AAEA,SAAS,UAAU,KAA6C;CAC9D,OAAO;EACL,iBAAiB,kBAAkB,IAAI,eAAe;EACtD,QAAQ,SAAS,IAAI,MAAM;EAC3B,WAAW,iBAAiB,IAAI,WAAW,mBAAmB;EAC9D,YAAY,cAAc,IAAI,YAAY,mBAAmB;CAC/D;AACF;;;;;;;AAQA,SAAS,eAAe,OAAyC;CAC/D,MAAM,SAAS,CAAC,GAAG,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM;EACvC,IAAI,EAAE,oBAAoB,EAAE,iBAAiB,OAAO;EACpD,IAAI,EAAE,oBAAoB,MAAM,OAAO;EACvC,IAAI,EAAE,oBAAoB,MAAM,OAAO,CAAC;EACxC,OAAO,EAAE,kBAAkB,EAAE;CAC/B,CAAC;CACD,IAAI,CAAC,OAAO,MAAM,MAAM,EAAE,oBAAoB,IAAI,GAAG;EACnD,OAAO,KAAK;GACV,iBAAiB;GACjB,QAAQ;GACR,WAAW;GACX,YAAY;EACd,CAAC;CACH;CACA,OAAO;AACT;;;;;;;;;;;;AAaA,OAAO,SAAS,sBACd,iBACyB;CACzB,IAAI,CAAC,iBAAiB;EACpB,OAAO;CACT;CAEA,MAAM,OAAO,SAAS,gBAAgB,SAAS;CAC/C,IAAI,CAAC,MAAM;EACT,OAAO;CACT;CAEA,MAAM,YAAY,SAAS,KAAK,eAAe;CAC/C,IAAI,CAAC,WAAW;EACd,OAAO;CACT;CAEA,MAAM,UAAU,UAAU,UAAU,SAAS,IAAI;CACjD,IAAI,CAAC,SAAS;EACZ,OAAO;CACT;CAEA,MAAM,WAAW,MAAM,QAAQ,UAAU,KAAK,IAAI,UAAU,QAAQ;CACpE,IAAI;CACJ,IAAI,YAAY,SAAS,SAAS,GAAG;EACnC,QAAQ,eAAe,SAAS,KAAK,MAAM,UAAU,SAAS,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;CAC1E,OAAO;;EAEL,QAAQ,CACN;GACE,iBAAiB;GACjB,QAAQ;GACR,WAAW,iBAAiB,UAAU,WAAW,mBAAmB;GACpE,YAAY,cAAc,UAAU,YAAY,mBAAmB;EACrE,CACF;CACF;CAEA,OAAO;EACL,SAAS;EACT,OAAO,QAAQ,UAAU,KAAK;EAC9B,eAAe,SAAS,UAAU,YAAY,KAAK,sBAAqB,AAAC,CAAC,YAAY;EACtF;CACF;AACF;;;;;;AAOA,OAAO,SAAS,WACd,OACA,cACe;CACf,MAAM,QAAQ,gBAAgB;CAC9B,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,oBAAoB,QAAQ,SAAS,KAAK,iBAAiB;GAClE,OAAO;EACT;CACF;CACA,OAAO,MAAM,MAAM,SAAS;AAC9B","names":[],"sources":["../src/config.ts"],"version":3,"file":"config.js","sourceRoot":""}
+{"mappings":"AAEA,OAAO,MAAM,sBAAsB;AACnC,OAAO,MAAM,sBAAsB;AACnC,OAAO,MAAM,wBAAwB;AACrC,OAAO,MAAM,gBAAgC;AAE7C,MAAM,iBAAiB;AAEvB,SAAS,SAAS,OAAqD;CACrE,IAAI,CAAC,SAAS,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;EAC/D,OAAO;CACT;CACA,OAAO;AACT;AAEA,SAAS,SAAS,OAAoC;CACpD,OAAO,OAAO,UAAU,YAAY,MAAM,KAAK,CAAC,CAAC,SAAS,IAAI,MAAM,KAAK,IAAI;AAC/E;AAEA,SAAS,UAAU,OAAgB,UAA4B;CAC7D,OAAO,OAAO,UAAU,YAAY,QAAQ;AAC9C;AAEA,SAAS,cAAc,OAAgB,UAA0B;CAC/D,IAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,QAAQ,GAAG;EACpE,OAAO,KAAK,MAAM,KAAK;CACzB;CACA,OAAO;AACT;;AAGA,SAAS,iBAAiB,OAAgB,UAA0B;CAClE,IAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,SAAS,GAAG;EACrE,OAAO,KAAK,MAAM,KAAK;CACzB;CACA,OAAO;AACT;AAEA,SAAS,QAAQ,OAAgC;CAC/C,OAAO,UAAU,cAAc,UAAU,UAAU,QAAQ;AAC7D;AAEA,SAAS,SAAS,OAAiC;CACjD,OAAO,UAAU,UAAU,UAAU;AACvC;;AAGA,SAAS,kBAAkB,OAA+B;CACxD,IAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,QAAQ,GAAG;EACpE,OAAO,KAAK,MAAM,KAAK;CACzB;CACA,OAAO;AACT;AAEA,SAAS,UAAU,KAA6C;CAC9D,OAAO;EACL,iBAAiB,kBAAkB,IAAI,eAAe;EACtD,QAAQ,SAAS,IAAI,MAAM;EAC3B,WAAW,iBAAiB,IAAI,WAAW,mBAAmB;EAC9D,YAAY,cAAc,IAAI,YAAY,mBAAmB;CAC/D;AACF;;;;;;;AAQA,SAAS,eAAe,OAAyC;CAC/D,MAAM,SAAS,CAAC,GAAG,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM;EACvC,IAAI,EAAE,oBAAoB,EAAE,iBAAiB,OAAO;EACpD,IAAI,EAAE,oBAAoB,MAAM,OAAO;EACvC,IAAI,EAAE,oBAAoB,MAAM,OAAO,CAAC;EACxC,OAAO,EAAE,kBAAkB,EAAE;CAC/B,CAAC;CACD,IAAI,CAAC,OAAO,MAAM,MAAM,EAAE,oBAAoB,IAAI,GAAG;EACnD,OAAO,KAAK;GACV,iBAAiB;GACjB,QAAQ;GACR,WAAW;GACX,YAAY;EACd,CAAC;CACH;CACA,OAAO;AACT;;;;;;;;;;;;AAaA,OAAO,SAAS,sBACd,kBACyB;CACzB,IAAI,CAAC,kBAAkB;EACrB,OAAO;CACT;;;CAIA,MAAM,YAAY,SAAS,iBAAiB,eAAe;CAC3D,IAAI,CAAC,WAAW;EACd,OAAO;CACT;CAEA,MAAM,UAAU,UAAU,UAAU,SAAS,IAAI;CACjD,IAAI,CAAC,SAAS;EACZ,OAAO;CACT;CAEA,MAAM,WAAW,MAAM,QAAQ,UAAU,KAAK,IAAI,UAAU,QAAQ;CACpE,IAAI;CACJ,IAAI,YAAY,SAAS,SAAS,GAAG;EACnC,QAAQ,eAAe,SAAS,KAAK,MAAM,UAAU,SAAS,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;CAC1E,OAAO;;EAEL,QAAQ,CACN;GACE,iBAAiB;GACjB,QAAQ;GACR,WAAW,iBAAiB,UAAU,WAAW,mBAAmB;GACpE,YAAY,cAAc,UAAU,YAAY,mBAAmB;EACrE,CACF;CACF;CAEA,OAAO;EACL,SAAS;EACT,OAAO,QAAQ,UAAU,KAAK;EAC9B,eAAe,SAAS,UAAU,YAAY,KAAK,sBAAqB,AAAC,CAAC,YAAY;EACtF;CACF;AACF;;;;;;AAOA,OAAO,SAAS,WACd,OACA,cACe;CACf,MAAM,QAAQ,gBAAgB;CAC9B,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,oBAAoB,QAAQ,SAAS,KAAK,iBAAiB;GAClE,OAAO;EACT;CACF;CACA,OAAO,MAAM,MAAM,SAAS;AAC9B","names":[],"sources":["../src/config.ts"],"version":3,"file":"config.js","sourceRoot":""}

package/dist/lib.d.ts

@@ -1,6 +1,6 @@
-export { createOpencodeRatelimitPlugin, OpencodeRatelimitPlugin, type OpenCodePluginFactoryOptions } from "./opencode.js";
+export { createOpencodeRatelimitPlugin, OpencodeRatelimitPlugin, type OpenCodePluginFactoryOptions, PLUGIN_ID } from "./opencode.js";
 export { DEFAULT_HEADER_PREFIX, DEFAULT_MAX_RETRIES, DEFAULT_MAX_WAIT_MS, DEFAULT_SCOPE, parseRateLimitOptions, selectTier } from "./config.js";
 export { parseRateLimit } from "./headers.js";
-export { createJsonConsoleLogger, DEFAULT_LOG_LEVEL, fromOpenCodeLogLevel, type LogFields, type Logger, type LogLevel } from "./logging.js";
-export { createProviderState, DEFAULT_BACKOFF_MS, type InstallConfigInput, installRateLimiter, makeRateLimitFetch, type ProviderConfigLike, type ProviderRateState, type RateLimitDeps, type RateStateStore } from "./plugin.js";
+export { captureHostContext, createEffectLogger, createJsonConsoleLogger, DEFAULT_LOG_LEVEL, type LogFields, type Logger, type LogLevel, resolveLogLevel } from "./logging.js";
+export { type BucketRef, collectOptIns, createProviderState, createRateLimiter, DEFAULT_BACKOFF_MS, type ProviderRateState, type RateLimitDeps, type RateLimiter, type RateStateStore, type RetryDecision } from "./plugin.js";
 export type { RateLimitAction, RateLimitOptions, RateLimitScope, RateLimitSnapshot, RateLimitTier } from "./types.js";

package/dist/lib.js

@@ -1,7 +1,7 @@
-export { createOpencodeRatelimitPlugin, OpencodeRatelimitPlugin } from "./opencode.js";
+export { createOpencodeRatelimitPlugin, OpencodeRatelimitPlugin, PLUGIN_ID } from "./opencode.js";
 export { DEFAULT_HEADER_PREFIX, DEFAULT_MAX_RETRIES, DEFAULT_MAX_WAIT_MS, DEFAULT_SCOPE, parseRateLimitOptions, selectTier } from "./config.js";
 export { parseRateLimit } from "./headers.js";
-export { createJsonConsoleLogger, DEFAULT_LOG_LEVEL, fromOpenCodeLogLevel } from "./logging.js";
-export { createProviderState, DEFAULT_BACKOFF_MS, installRateLimiter, makeRateLimitFetch } from "./plugin.js";
+export { captureHostContext, createEffectLogger, createJsonConsoleLogger, DEFAULT_LOG_LEVEL, resolveLogLevel } from "./logging.js";
+export { collectOptIns, createProviderState, createRateLimiter, DEFAULT_BACKOFF_MS } from "./plugin.js";
 
 //# sourceMappingURL=lib.js.map

package/dist/lib.js.map

@@ -1 +1 @@
-{"mappings":"AAAA,SACE,+BACA,+BAEK;AAEP,SACE,uBACA,qBACA,qBACA,eACA,uBACA,kBACK;AAEP,SAAS,sBAAsB;AAE/B,SACE,yBACA,mBACA,4BAIK;AAEP,SACE,qBACA,oBAEA,oBACA,0BAKK","names":[],"sources":["../src/lib.ts"],"version":3,"file":"lib.js","sourceRoot":""}
+{"mappings":"AAAA,SACE,+BACA,yBAEA,iBACK;AAEP,SACE,uBACA,qBACA,qBACA,eACA,uBACA,kBACK;AAEP,SAAS,sBAAsB;AAE/B,SACE,oBACA,oBACA,yBACA,mBAIA,uBACK;AAEP,SAEE,eACA,qBACA,mBACA,0BAMK","names":[],"sources":["../src/lib.ts"],"version":3,"file":"lib.js","sourceRoot":""}

package/dist/logging.d.ts

@@ -1,3 +1,4 @@
+import { type Context as EffectContext, Effect } from "effect";
 import type { LogLevel } from "./types.js";
 export type { LogLevel } from "./types.js";
 export declare const LOG_LEVEL_PRIORITY: Record<LogLevel, number>;
@@ -13,4 +14,37 @@ export interface Logger {
 	error(event: string, fields?: LogFields): void;
 }
 export declare function createJsonConsoleLogger(minLevel?: LogLevel): Logger;
-export declare function fromOpenCodeLogLevel(value: unknown): LogLevel | undefined;
+/**
+* Capture the host's Effect context from inside the plugin's own fiber. The
+* returned value is what {@link createEffectLogger} needs to put a log record
+* on the host's logger rather than on a default one.
+*
+* OpenCode installs its file logger as a `Logger.layer(...)` on the runtime
+* that runs plugin setup, and Effect carries loggers as fiber-level references
+* — so the context captured here is the only route to that sink. A bare
+* `Effect.runFork` would log to Effect's default logger (the terminal) instead,
+* which is exactly what ADR-0014 forbids.
+*/
+export declare const captureHostContext: Effect.Effect<EffectContext.Context<never>>;
+/**
+* Structured logger that writes through the host's Effect logger — the v2
+* replacement for v1's `client.app.log`.
+*
+* ADR-0014 (no plugin writes to the user's TUI) is preserved: OpenCode's own
+* logger is a `Logger.toFile` sink over the shared `opencode-<channel>.log`,
+* so records land in the host log stream and never on the terminal. The event
+* name becomes the log message and `fields` become log annotations, which
+* OpenCode's logfmt formatter flattens to `key=value` — the same shape the v1
+* `extra` payload produced.
+*
+* Host-side severity filtering is Effect's job (`OPENCODE_LOG_LEVEL`); the
+* `getMinLevel` gate here is only this plugin's own `logLevel` option, applied
+* before the record is built so a filtered-out call costs nothing.
+*/
+export declare function createEffectLogger(context: EffectContext.Context<never>, getMinLevel: () => LogLevel): Logger;
+/**
+* Resolve this plugin's log level from its own `logLevel` option, falling back
+* to the host's `OPENCODE_LOG_LEVEL` env var. v2 gives plugins no `config`
+* object, so the v1 `config.logLevel` channel is gone.
+*/
+export declare function resolveLogLevel(value: unknown): LogLevel | undefined;

package/dist/logging.js

@@ -1,3 +1,4 @@
+import { Effect } from "effect";
 export const LOG_LEVEL_PRIORITY = {
 	trace: 5,
 	debug: 10,
@@ -51,13 +52,73 @@ export function createJsonConsoleLogger(minLevel = DEFAULT_LOG_LEVEL) {
 		error: (event, fields) => write("error", event, fields)
 	};
 }
-export function fromOpenCodeLogLevel(value) {
+/**
+* Capture the host's Effect context from inside the plugin's own fiber. The
+* returned value is what {@link createEffectLogger} needs to put a log record
+* on the host's logger rather than on a default one.
+*
+* OpenCode installs its file logger as a `Logger.layer(...)` on the runtime
+* that runs plugin setup, and Effect carries loggers as fiber-level references
+* — so the context captured here is the only route to that sink. A bare
+* `Effect.runFork` would log to Effect's default logger (the terminal) instead,
+* which is exactly what ADR-0014 forbids.
+*/
+export const captureHostContext = Effect.withFiber((fiber) => Effect.succeed(fiber.context));
+/**
+* Structured logger that writes through the host's Effect logger — the v2
+* replacement for v1's `client.app.log`.
+*
+* ADR-0014 (no plugin writes to the user's TUI) is preserved: OpenCode's own
+* logger is a `Logger.toFile` sink over the shared `opencode-<channel>.log`,
+* so records land in the host log stream and never on the terminal. The event
+* name becomes the log message and `fields` become log annotations, which
+* OpenCode's logfmt formatter flattens to `key=value` — the same shape the v1
+* `extra` payload produced.
+*
+* Host-side severity filtering is Effect's job (`OPENCODE_LOG_LEVEL`); the
+* `getMinLevel` gate here is only this plugin's own `logLevel` option, applied
+* before the record is built so a filtered-out call costs nothing.
+*/
+export function createEffectLogger(context, getMinLevel) {
+	// `VYMALO_PLUGIN_CONSOLE_LOG=1` mirrors to stdout for local debugging. Off by
+	// default precisely because stdout is the user's terminal (ADR-0014).
+	const consoleAll = /^(1|true|yes|on)$/i.test(process.env.VYMALO_PLUGIN_CONSOLE_LOG ?? "");
+	const fallback = createJsonConsoleLogger("trace");
+	const write = (level, event, fields) => {
+		if (LOG_LEVEL_PRIORITY[level] < LOG_LEVEL_PRIORITY[getMinLevel()]) {
+			return;
+		}
+		if (consoleAll) {
+			fallback[level](event, fields);
+		}
+		const redacted = redactFields(fields);
+		// Effect has no `trace` severity below `debug`; fold the most-verbose tier
+		// onto debug, as the v1 adapter did for the host log API.
+		const record = level === "trace" ? Effect.logDebug(event) : level === "debug" ? Effect.logDebug(event) : level === "info" ? Effect.logInfo(event) : level === "warn" ? Effect.logWarning(event) : Effect.logError(event);
+		const annotated = redacted ? Effect.annotateLogs(record, redacted) : record;
+		Effect.runFork(Effect.provideContext(annotated, context));
+	};
+	return {
+		trace: (event, fields) => write("trace", event, fields),
+		debug: (event, fields) => write("debug", event, fields),
+		info: (event, fields) => write("info", event, fields),
+		warn: (event, fields) => write("warn", event, fields),
+		error: (event, fields) => write("error", event, fields)
+	};
+}
+/**
+* Resolve this plugin's log level from its own `logLevel` option, falling back
+* to the host's `OPENCODE_LOG_LEVEL` env var. v2 gives plugins no `config`
+* object, so the v1 `config.logLevel` channel is gone.
+*/
+export function resolveLogLevel(value) {
 	if (typeof value !== "string") {
 		return undefined;
 	}
 	switch (value.toUpperCase()) {
 		// DEBUG unlocks the most-verbose `trace` tier (host has no finer level).
 		case "DEBUG": return "trace";
+		case "TRACE": return "trace";
 		case "INFO": return "info";
 		case "WARN": return "warn";
 		case "ERROR": return "error";

package/dist/logging.js.map

@@ -1 +1 @@
-{"mappings":"AAIA,OAAO,MAAM,qBAA+C;CAC1D,OAAO;CACP,OAAO;CACP,MAAM;CACN,MAAM;CACN,OAAO;AACT;AAEA,OAAO,MAAM,oBAA8B;AAc3C,SAAS,aAAa,QAA2C;CAC/D,IAAI,CAAC,QAAQ;EACX,OAAO;CACT;CACA,MAAM,WAAsB,CAAC;CAC7B,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,MAAM,GAAG;EACjD,IAAI,uCAAuC,KAAK,GAAG,GAAG;GACpD,SAAS,OAAO;GAChB;EACF;EACA,SAAS,OAAO;CAClB;CACA,OAAO;AACT;AAEA,OAAO,SAAS,wBAAwB,WAAqB,mBAA2B;CACtF,MAAM,cAAc,mBAAmB;CAEvC,MAAM,SAAS,OAAiB,OAAe,WAA6B;EAC1E,IAAI,mBAAmB,SAAS,aAAa;GAC3C;EACF;EACA,MAAM,UAAU;GACd,IAAI,IAAI,KAAK,CAAC,CAAC,YAAY;GAC3B;GACA;GACA,GAAI,aAAa,MAAM,KAAK,CAAC;EAC/B;EACA,MAAM,OAAO,KAAK,UAAU,OAAO;EACnC,IAAI,UAAU,SAAS;GACrB,QAAQ,MAAM,IAAI;GAClB;EACF;EACA,IAAI,UAAU,QAAQ;GACpB,QAAQ,KAAK,IAAI;GACjB;EACF;EACA,QAAQ,IAAI,IAAI;CAClB;CAEA,OAAO;EACL,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;EACtD,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;EACtD,OAAO,OAAO,WAAW,MAAM,QAAQ,OAAO,MAAM;EACpD,OAAO,OAAO,WAAW,MAAM,QAAQ,OAAO,MAAM;EACpD,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;CACxD;AACF;AAEA,OAAO,SAAS,qBAAqB,OAAsC;CACzE,IAAI,OAAO,UAAU,UAAU;EAC7B,OAAO;CACT;CACA,QAAQ,MAAM,YAAY,GAA1B;;EAEE,KAAK,SACH,OAAO;EACT,KAAK,QACH,OAAO;EACT,KAAK,QACH,OAAO;EACT,KAAK,SACH,OAAO;EACT,SACE,OAAO;CACX;AACF","names":[],"sources":["../src/logging.ts"],"version":3,"file":"logging.js","sourceRoot":""}
+{"mappings":"AAAA,SAAwC,cAAc;AAMtD,OAAO,MAAM,qBAA+C;CAC1D,OAAO;CACP,OAAO;CACP,MAAM;CACN,MAAM;CACN,OAAO;AACT;AAEA,OAAO,MAAM,oBAA8B;AAc3C,SAAS,aAAa,QAA2C;CAC/D,IAAI,CAAC,QAAQ;EACX,OAAO;CACT;CACA,MAAM,WAAsB,CAAC;CAC7B,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,MAAM,GAAG;EACjD,IAAI,uCAAuC,KAAK,GAAG,GAAG;GACpD,SAAS,OAAO;GAChB;EACF;EACA,SAAS,OAAO;CAClB;CACA,OAAO;AACT;AAEA,OAAO,SAAS,wBAAwB,WAAqB,mBAA2B;CACtF,MAAM,cAAc,mBAAmB;CAEvC,MAAM,SAAS,OAAiB,OAAe,WAA6B;EAC1E,IAAI,mBAAmB,SAAS,aAAa;GAC3C;EACF;EACA,MAAM,UAAU;GACd,IAAI,IAAI,KAAK,CAAC,CAAC,YAAY;GAC3B;GACA;GACA,GAAI,aAAa,MAAM,KAAK,CAAC;EAC/B;EACA,MAAM,OAAO,KAAK,UAAU,OAAO;EACnC,IAAI,UAAU,SAAS;GACrB,QAAQ,MAAM,IAAI;GAClB;EACF;EACA,IAAI,UAAU,QAAQ;GACpB,QAAQ,KAAK,IAAI;GACjB;EACF;EACA,QAAQ,IAAI,IAAI;CAClB;CAEA,OAAO;EACL,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;EACtD,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;EACtD,OAAO,OAAO,WAAW,MAAM,QAAQ,OAAO,MAAM;EACpD,OAAO,OAAO,WAAW,MAAM,QAAQ,OAAO,MAAM;EACpD,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;CACxD;AACF;;;;;;;;;;;;AAaA,OAAO,MAAM,qBAAkE,OAAO,WACnF,UAAU,OAAO,QAAQ,MAAM,OAAO,CACzC;;;;;;;;;;;;;;;;AAiBA,OAAO,SAAS,mBACd,SACA,aACQ;;;CAGR,MAAM,aAAa,qBAAqB,KAAK,QAAQ,IAAI,6BAA6B,EAAE;CACxF,MAAM,WAAW,wBAAwB,OAAO;CAEhD,MAAM,SAAS,OAAiB,OAAe,WAA6B;EAC1E,IAAI,mBAAmB,SAAS,mBAAmB,YAAY,IAAI;GACjE;EACF;EACA,IAAI,YAAY;GACd,SAAS,MAAM,CAAC,OAAO,MAAM;EAC/B;EAEA,MAAM,WAAW,aAAa,MAAM;;;EAGpC,MAAM,SACJ,UAAU,UACN,OAAO,SAAS,KAAK,IACrB,UAAU,UACR,OAAO,SAAS,KAAK,IACrB,UAAU,SACR,OAAO,QAAQ,KAAK,IACpB,UAAU,SACR,OAAO,WAAW,KAAK,IACvB,OAAO,SAAS,KAAK;EAEjC,MAAM,YAAY,WAAW,OAAO,aAAa,QAAQ,QAAQ,IAAI;EACrE,OAAO,QAAQ,OAAO,eAAe,WAAW,OAAO,CAAC;CAC1D;CAEA,OAAO;EACL,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;EACtD,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;EACtD,OAAO,OAAO,WAAW,MAAM,QAAQ,OAAO,MAAM;EACpD,OAAO,OAAO,WAAW,MAAM,QAAQ,OAAO,MAAM;EACpD,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;CACxD;AACF;;;;;;AAOA,OAAO,SAAS,gBAAgB,OAAsC;CACpE,IAAI,OAAO,UAAU,UAAU;EAC7B,OAAO;CACT;CACA,QAAQ,MAAM,YAAY,GAA1B;;EAEE,KAAK,SACH,OAAO;EACT,KAAK,SACH,OAAO;EACT,KAAK,QACH,OAAO;EACT,KAAK,QACH,OAAO;EACT,KAAK,SACH,OAAO;EACT,SACE,OAAO;CACX;AACF","names":[],"sources":["../src/logging.ts"],"version":3,"file":"logging.js","sourceRoot":""}

package/dist/opencode.d.ts

@@ -1,11 +1,36 @@
-import type { Plugin } from "@opencode-ai/plugin";
+import { Plugin } from "@opencode/plugin/effect";
 import { type Logger } from "./logging.js";
+import { type RateLimiter } from "./plugin.js";
+/** Stable plugin id. v2 requires one and keys storage/registrations by it. */
+export declare const PLUGIN_ID = "vymalo.ratelimit";
 export interface OpenCodePluginFactoryOptions {
+	/** Override the logger (tests). Defaults to the host's Effect logger. */
 	logger?: Logger;
-	fetchImpl?: typeof fetch;
+	/** Defaults to `Date.now`. */
 	now?: () => number;
+	/** Defaults to a real `setTimeout`-based, abort-aware sleep. */
 	sleep?: (ms: number, signal?: AbortSignal) => Promise<void>;
+	/** Observe the limiter the plugin built (tests). */
+	onLimiter?: (limiter: RateLimiter) => void;
 }
-export declare function createOpencodeRatelimitPlugin(factoryOptions?: OpenCodePluginFactoryOptions): Plugin;
-export declare const OpencodeRatelimitPlugin: Plugin;
+/**
+* Build the OpenCode v2 plugin.
+*
+* Three registrations, one per lifecycle seam:
+*
+*  - `provider.transform` reads the `settings.rateLimit` opt-in set. It only
+*    *reads* — the limiter needs to know which providers it manages, and a
+*    transform is the seam that reruns on every provider reload. Transform
+*    callbacks must stay synchronous and cheap, which this is.
+*  - `session.hook("http.request")` is the pre-request gate.
+*  - `session.hook("http.response")` reads the `x-ratelimit-*` triple.
+*  - `session.hook("retry")` decides whether a 429 waits or surfaces.
+*
+* v1 had to do all of this inside one injected `provider.options.fetch`,
+* because OpenCode exposed no post-response hook. v2's native HTTP seams make
+* the fetch wrapper unnecessary — and the `retry` hook hands the backoff sleep
+* to the host, so the plugin no longer blocks a request thread on a 429.
+*/
+export declare function createOpencodeRatelimitPlugin(factoryOptions?: OpenCodePluginFactoryOptions): Plugin.Plugin;
+export declare const OpencodeRatelimitPlugin: Plugin.Plugin;
 export default OpencodeRatelimitPlugin;

package/dist/opencode.js

@@ -1,60 +1,82 @@
-import { createJsonConsoleLogger, DEFAULT_LOG_LEVEL, fromOpenCodeLogLevel, LOG_LEVEL_PRIORITY } from "./logging.js";
-import { installRateLimiter } from "./plugin.js";
-const PLUGIN_SERVICE_NAME = "opencode-ratelimit-plugin";
+import { Plugin } from "@opencode/plugin/effect";
+import { Effect } from "effect";
+import { captureHostContext, createEffectLogger, DEFAULT_LOG_LEVEL, resolveLogLevel } from "./logging.js";
+import { collectOptIns, createRateLimiter } from "./plugin.js";
+/** Stable plugin id. v2 requires one and keys storage/registrations by it. */
+export const PLUGIN_ID = "vymalo.ratelimit";
 /**
-* Pipe plugin logs through OpenCode's `client.app.log` so they show up in the
-* host's structured log stream, with the JSON console as a reliable fallback.
-* Mirrors the pattern used by `@vymalo/opencode-oauth2` and `-models-info`.
+* Build the OpenCode v2 plugin.
+*
+* Three registrations, one per lifecycle seam:
+*
+*  - `provider.transform` reads the `settings.rateLimit` opt-in set. It only
+*    *reads* — the limiter needs to know which providers it manages, and a
+*    transform is the seam that reruns on every provider reload. Transform
+*    callbacks must stay synchronous and cheap, which this is.
+*  - `session.hook("http.request")` is the pre-request gate.
+*  - `session.hook("http.response")` reads the `x-ratelimit-*` triple.
+*  - `session.hook("retry")` decides whether a 429 waits or surfaces.
+*
+* v1 had to do all of this inside one injected `provider.options.fetch`,
+* because OpenCode exposed no post-response hook. v2's native HTTP seams make
+* the fetch wrapper unnecessary — and the `retry` hook hands the backoff sleep
+* to the host, so the plugin no longer blocks a request thread on a 429.
 */
-function createOpenCodeLogger(client, getMinLevel) {
-	const fallback = createJsonConsoleLogger("debug");
-	// OpenCode already captures plugin logs via client.app.log (and filters them
-	// by its own log level). Mirroring every event to stdout on top of that just
-	// floods the terminal, so only mirror warn/error to the JSON console by
-	// default; set VYMALO_PLUGIN_CONSOLE_LOG=1 to restore full console output.
-	const consoleAll = /^(1|true|yes|on)$/i.test(process.env.VYMALO_PLUGIN_CONSOLE_LOG ?? "");
-	const write = (level, event, fields) => {
-		if (LOG_LEVEL_PRIORITY[level] < LOG_LEVEL_PRIORITY[getMinLevel()]) {
-			return;
-		}
-		if (consoleAll || level === "warn" || level === "error") {
-			fallback[level](event, fields);
-		}
-		// The host's log API has no `trace` level — fold it onto `debug` so the most
-		// verbose tier still reaches the structured stream (it's already gated above
-		// by `getMinLevel`, which DEBUG maps to `trace`).
-		const hostLevel = level === "trace" ? "debug" : level;
-		void client.app.log({ body: {
-			service: PLUGIN_SERVICE_NAME,
-			level: hostLevel,
-			message: event,
-			extra: fields
-		} }).catch(() => {
-			/* best-effort */
-		});
-	};
-	return {
-		trace: (event, fields) => write("trace", event, fields),
-		debug: (event, fields) => write("debug", event, fields),
-		info: (event, fields) => write("info", event, fields),
-		warn: (event, fields) => write("warn", event, fields),
-		error: (event, fields) => write("error", event, fields)
-	};
-}
 export function createOpencodeRatelimitPlugin(factoryOptions = {}) {
-	return async ({ client }) => {
-		let currentLogLevel = DEFAULT_LOG_LEVEL;
-		const logger = factoryOptions.logger ?? createOpenCodeLogger(client, () => currentLogLevel);
-		return { config: async (config) => {
-			currentLogLevel = fromOpenCodeLogLevel(config.logLevel) ?? DEFAULT_LOG_LEVEL;
-			installRateLimiter(config, {
+	return Plugin.define({
+		id: PLUGIN_ID,
+		effect: Effect.fn(function* (ctx) {
+			// v2 gives plugins no `config` object, so the v1 `config.logLevel`
+			// channel is gone: take this plugin's own `logLevel` option first, then
+			// the host's `OPENCODE_LOG_LEVEL`.
+			const logLevel = resolveLogLevel(ctx.options.logLevel) ?? resolveLogLevel(process.env.OPENCODE_LOG_LEVEL) ?? DEFAULT_LOG_LEVEL;
+			let logger = factoryOptions.logger;
+			if (!logger) {
+				const hostContext = yield* captureHostContext;
+				logger = createEffectLogger(hostContext, () => logLevel);
+			}
+			const limiter = createRateLimiter({
 				logger,
-				fetchImpl: factoryOptions.fetchImpl,
 				now: factoryOptions.now,
 				sleep: factoryOptions.sleep
 			});
-		} };
-	};
+			factoryOptions.onLimiter?.(limiter);
+			const log = logger;
+			yield* ctx.provider.transform((editor) => {
+				limiter.configure(collectOptIns(editor.list().map((record) => ({
+					id: record.provider.id,
+					settings: record.provider.settings
+				})), log));
+			});
+			yield* ctx.session.hook("http.request", (event) => Effect.promise(() => limiter.acquire({
+				providerId: event.model.providerID,
+				modelID: event.model.id
+			}, event.request.signal)));
+			yield* ctx.session.hook("http.response", (event) => Effect.sync(() => {
+				limiter.observe({
+					providerId: event.model.providerID,
+					modelID: event.model.id
+				}, {
+					status: event.response.status,
+					headers: event.response.headers
+				});
+			}));
+			yield* ctx.session.hook("retry", (event) => Effect.sync(() => {
+				const decision = limiter.decideRetry({
+					providerId: event.model.providerID,
+					modelID: event.model.id
+				}, {
+					status: event.error.status,
+					retryAfterMs: event.error.retryAfterMs
+				}, event.attempt);
+				// Only speak up about a 429 we manage; leave every other failure to
+				// the host's own retry policy.
+				if (decision.retry) {
+					event.decision = decision;
+				}
+			}));
+		})
+	});
 }
 export const OpencodeRatelimitPlugin = createOpencodeRatelimitPlugin();
 export default OpencodeRatelimitPlugin;

package/dist/opencode.js.map

@@ -1 +1 @@
-{"mappings":"AAEA,SACE,yBACA,mBACA,sBAEA,0BAGK;AACP,SAAkC,0BAA0B;AAE5D,MAAM,sBAAsB;;;;;;AAgB5B,SAAS,qBAAqB,QAA+B,aAAqC;CAChG,MAAM,WAAW,wBAAwB,OAAO;;;;;CAKhD,MAAM,aAAa,qBAAqB,KAAK,QAAQ,IAAI,6BAA6B,EAAE;CAExF,MAAM,SAAS,OAAiB,OAAe,WAAuB;EACpE,IAAI,mBAAmB,SAAS,mBAAmB,YAAY,IAAI;GACjE;EACF;EACA,IAAI,cAAc,UAAU,UAAU,UAAU,SAAS;GACvD,SAAS,MAAM,CAAC,OAAO,MAAM;EAC/B;;;;EAIA,MAAM,YAAY,UAAU,UAAU,UAAU;EAChD,KAAK,OAAO,IACT,IAAI,EACH,MAAM;GACJ,SAAS;GACT,OAAO;GACP,SAAS;GACT,OAAO;EACT,EACF,CAAC,CAAC,CACD,YAAY;;EAEb,CAAC;CACL;CAEA,OAAO;EACL,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;EACtD,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;EACtD,OAAO,OAAO,WAAW,MAAM,QAAQ,OAAO,MAAM;EACpD,OAAO,OAAO,WAAW,MAAM,QAAQ,OAAO,MAAM;EACpD,QAAQ,OAAO,WAAW,MAAM,SAAS,OAAO,MAAM;CACxD;AACF;AAEA,OAAO,SAAS,8BACd,iBAA+C,CAAC,GACxC;CACR,OAAO,OAAO,EAAE,aAAa;EAC3B,IAAI,kBAA4B;EAChC,MAAM,SAAS,eAAe,UAAU,qBAAqB,cAAc,eAAe;EAE1F,OAAO,EACL,QAAQ,OAAO,WAA2B;GACxC,kBAAkB,qBAAqB,OAAO,QAAQ,KAAK;GAC3D,mBAAmB,QAA8B;IAC/C;IACA,WAAW,eAAe;IAC1B,KAAK,eAAe;IACpB,OAAO,eAAe;GACxB,CAAC;EACH,EACF;CACF;AACF;AAEA,OAAO,MAAM,0BAAkC,8BAA8B;AAE7E,eAAe","names":[],"sources":["../src/opencode.ts"],"version":3,"file":"opencode.js","sourceRoot":""}
+{"mappings":"AAAA,SAAS,cAAc;AACvB,SAAS,cAAc;AAEvB,SACE,oBACA,oBACA,mBAGA,uBACK;AACP,SAAS,eAAe,yBAA2C;;AAGnE,OAAO,MAAM,YAAY;;;;;;;;;;;;;;;;;;;AA+BzB,OAAO,SAAS,8BACd,iBAA+C,CAAC,GACjC;CACf,OAAO,OAAO,OAAO;EACnB,IAAI;EACJ,QAAQ,OAAO,GAAG,WAAW,KAAK;;;;GAIhC,MAAM,WACJ,gBAAgB,IAAI,QAAQ,QAAQ,KACpC,gBAAgB,QAAQ,IAAI,kBAAkB,KAC9C;GAEF,IAAI,SAAS,eAAe;GAC5B,IAAI,CAAC,QAAQ;IACX,MAAM,cAAc,OAAO;IAC3B,SAAS,mBAAmB,mBAAmB,QAAQ;GACzD;GAEA,MAAM,UAAU,kBAAkB;IAChC;IACA,KAAK,eAAe;IACpB,OAAO,eAAe;GACxB,CAAC;GACD,eAAe,YAAY,OAAO;GAElC,MAAM,MAAM;GAEZ,OAAO,IAAI,SAAS,WAAW,WAAW;IACxC,QAAQ,UACN,cACE,OAAO,KAAK,CAAC,CAAC,KAAK,YAAY;KAC7B,IAAI,OAAO,SAAS;KACpB,UAAU,OAAO,SAAS;IAC5B,EAAE,GACF,GACF,CACF;GACF,CAAC;GAED,OAAO,IAAI,QAAQ,KAAK,iBAAiB,UACvC,OAAO,cACL,QAAQ,QACN;IAAE,YAAY,MAAM,MAAM;IAAY,SAAS,MAAM,MAAM;GAAG,GAC9D,MAAM,QAAQ,MAChB,CACF,CACF;GAEA,OAAO,IAAI,QAAQ,KAAK,kBAAkB,UACxC,OAAO,WAAW;IAChB,QAAQ,QACN;KAAE,YAAY,MAAM,MAAM;KAAY,SAAS,MAAM,MAAM;IAAG,GAC9D;KAAE,QAAQ,MAAM,SAAS;KAAQ,SAAS,MAAM,SAAS;IAAQ,CACnE;GACF,CAAC,CACH;GAEA,OAAO,IAAI,QAAQ,KAAK,UAAU,UAChC,OAAO,WAAW;IAChB,MAAM,WAAW,QAAQ,YACvB;KAAE,YAAY,MAAM,MAAM;KAAY,SAAS,MAAM,MAAM;IAAG,GAC9D;KAAE,QAAQ,MAAM,MAAM;KAAQ,cAAc,MAAM,MAAM;IAAa,GACrE,MAAM,OACR;;;IAGA,IAAI,SAAS,OAAO;KAClB,MAAM,WAAW;IACnB;GACF,CAAC,CACH;EACF,CAAC;CACH,CAAC;AACH;AAEA,OAAO,MAAM,0BAAyC,8BAA8B;AAEpF,eAAe","names":[],"sources":["../src/opencode.ts"],"version":3,"file":"opencode.js","sourceRoot":""}

package/dist/plugin.d.ts

@@ -1,11 +1,11 @@
 import type { Logger } from "./logging.js";
-import type { RateLimitOptions } from "./types.js";
+import type { RateLimitOptions, RateLimitSnapshot } from "./types.js";
 /** Fallback backoff when a 429 carries no `x-ratelimit-reset` and no `Retry-After`. */
 export declare const DEFAULT_BACKOFF_MS = 1e3;
 /**
 * Mutable, in-memory rate-limit state for a single bucket (a provider, or a
 * `(provider, model)` pair when `scope: "model"`). Lives for the process
-* lifetime in the provider's state store. Never persisted — a reset window is
+* lifetime in the limiter's state store. Never persisted — a reset window is
 * measured in seconds, so survival across process boots is pointless.
 */
 export interface ProviderRateState {
@@ -25,6 +25,13 @@ export interface ProviderRateState {
 	* back to its own sleep, so it never waits past its required time / `maxWaitMs`.
 	*/
 	cooldownPromiseUntilMs?: number;
+	/**
+	* Last snapshot read from this bucket's responses. The v2 `retry` hook sees a
+	* `SessionError` (status + `retryAfterMs`) but no response headers, so the
+	* reset window it needs for tier selection and backoff is carried here from
+	* the `http.response` hook that observed it.
+	*/
+	lastSnapshot?: RateLimitSnapshot;
 }
 export declare function createProviderState(): ProviderRateState;
 export interface RateLimitDeps {
@@ -33,36 +40,59 @@ export interface RateLimitDeps {
 	now?: () => number;
 	/** Defaults to a real `setTimeout`-based, abort-aware sleep. */
 	sleep?: (ms: number, signal?: AbortSignal) => Promise<void>;
-	/** Underlying fetch when a provider has no pre-existing `options.fetch`. Defaults to `globalThis.fetch`. */
-	fetchImpl?: typeof fetch;
-}
-export interface ProviderConfigLike {
-	options?: Record<string, unknown>;
-}
-export interface InstallConfigInput {
-	provider?: Record<string, ProviderConfigLike | undefined>;
 }
 /** A per-bucket state store (key = provider id, or `${providerId}\0${model}`). */
 export type RateStateStore = Map<string, ProviderRateState>;
+/** What the `retry` hook should do with a failed request. */
+export type RetryDecision = {
+	retry: false;
+} | {
+	retry: true;
+	delay: number;
+};
+/** Identifies the bucket a hook event belongs to. */
+export interface BucketRef {
+	providerId: string;
+	modelID?: string;
+}
 /**
-* Walk every provider in the assembled OpenCode config; for each one that has
-* opted in via `options.meta.rateLimit`, wrap its `options.fetch` with a
-* rate-limit-aware fetch. Synchronous — all real work happens lazily inside the
-* wrapper at request time.
+* The rate-limit engine, shaped for OpenCode v2's three request-lifecycle
+* seams rather than v1's single `fetch` wrapper:
+*
+*  - {@link RateLimiter.acquire} is the pre-request gate (`http.request`): it
+*    waits out a cooldown a previous `"wait"` tier armed.
+*  - {@link RateLimiter.observe} reads the rate-limit headers off a response
+*    (`http.response`) and arms or clears the gate for the *next* callers.
+*  - {@link RateLimiter.decideRetry} answers the `retry` hook: a `"wait"` tier
+*    returns a delay for the host to sleep, an `"error"` tier declines the retry
+*    so the 429 surfaces immediately.
+*
+* The host owns the retry loop in v2, so the plugin never sleeps on a 429 — it
+* returns a delay and OpenCode waits (cancellably). The plugin still sleeps in
+* `acquire`, because there is no host hook for "hold this request back".
 */
-export declare function installRateLimiter(input: InstallConfigInput, deps: RateLimitDeps): void;
+export interface RateLimiter {
+	/** Replace the opted-in provider set. Called from the provider transform. */
+	configure(entries: ReadonlyMap<string, RateLimitOptions>): void;
+	/** Resolved options for a provider, or `undefined` when it has not opted in. */
+	options(providerId: string): RateLimitOptions | undefined;
+	acquire(ref: BucketRef, signal?: AbortSignal): Promise<void>;
+	observe(ref: BucketRef, response: {
+		status: number;
+		headers: Headers;
+	}): void;
+	decideRetry(ref: BucketRef, error: {
+		status?: number;
+		retryAfterMs?: number;
+	}, attempt: number): RetryDecision;
+}
 /**
-* Build the fetch wrapper for one provider. The wrapper:
-*  1. Resolves the bucket (provider, or `(provider, model)` under `scope: "model"`).
-*  2. Pre-request gate: if that bucket's cooldown is armed, waits until it clears.
-*  3. Sends the request via the underlying fetch.
-*  4. Reads the rate-limit headers; selects the policy tier by reset magnitude.
-*     On `remaining: 0` it arms the gate for the next callers (only when the
-*     matched tier is `"wait"`).
-*  5. On a 429: a `"wait"` tier waits the reset window and retries (up to the
-*     tier's `maxRetries`); an `"error"` tier surfaces the 429 immediately.
-*
-* The Response is returned untouched (no `.clone()`) — we only read `status`
-* and `headers`, so the body stream is delivered to OpenCode intact.
+* Read each provider's `settings.rateLimit` opt-in. v2 has no free-form
+* `options.meta` on a provider — `Provider.Settings` is an open struct, so the
+* opt-in moves from v1's `options.meta.rateLimit` to `settings.rateLimit`.
 */
-export declare function makeRateLimitFetch(providerId: string, opts: RateLimitOptions, store: RateStateStore, deps: RateLimitDeps, delegate?: typeof fetch): typeof fetch;
+export declare function collectOptIns(providers: readonly {
+	id: string;
+	settings?: Record<string, unknown>;
+}[], logger: Logger): Map<string, RateLimitOptions>;
+export declare function createRateLimiter(deps: RateLimitDeps): RateLimiter;

package/dist/plugin.js

@@ -9,143 +9,114 @@ export function createProviderState() {
 	};
 }
 /**
-* Walk every provider in the assembled OpenCode config; for each one that has
-* opted in via `options.meta.rateLimit`, wrap its `options.fetch` with a
-* rate-limit-aware fetch. Synchronous — all real work happens lazily inside the
-* wrapper at request time.
+* Read each provider's `settings.rateLimit` opt-in. v2 has no free-form
+* `options.meta` on a provider — `Provider.Settings` is an open struct, so the
+* opt-in moves from v1's `options.meta.rateLimit` to `settings.rateLimit`.
 */
-export function installRateLimiter(input, deps) {
-	const providers = input.provider;
-	deps.logger.trace("ratelimit_install_start", { providerCount: providers ? Object.keys(providers).length : 0 });
-	if (!providers) {
-		deps.logger.info("ratelimit_plugin_initialized", { providerCount: 0 });
-		return;
-	}
-	let enabledCount = 0;
-	for (const [providerId, providerConfig] of Object.entries(providers)) {
-		if (!providerConfig) {
-			continue;
-		}
-		deps.logger.trace("ratelimit_provider_optin_check", {
-			providerId,
-			hasOptions: Boolean(providerConfig.options)
-		});
-		const opts = parseRateLimitOptions(providerConfig.options);
+export function collectOptIns(providers, logger) {
+	const out = new Map();
+	for (const provider of providers) {
+		const opts = parseRateLimitOptions(provider.settings);
 		if (!opts) {
-			deps.logger.debug("ratelimit_provider_skipped", {
-				providerId,
+			logger.debug("ratelimit_provider_skipped", {
+				providerId: provider.id,
 				reason: "not_opted_in"
 			});
 			continue;
 		}
-		deps.logger.trace("ratelimit_provider_optin_resolved", {
-			providerId,
-			scope: opts.scope,
-			headerPrefix: opts.headerPrefix,
-			tiers: opts.tiers.length
-		});
-		const options = providerConfig.options ??= {};
-		// Compose with any fetch a prior plugin already installed (capture it now,
-		// not at call time, so we delegate to the original — not to ourselves).
-		const delegate = typeof options.fetch === "function" ? options.fetch : undefined;
-		deps.logger.trace("ratelimit_fetch_wrapped", {
-			providerId,
-			composedWithExistingFetch: Boolean(delegate)
-		});
-		const store = new Map();
-		options.fetch = makeRateLimitFetch(providerId, opts, store, deps, delegate);
-		enabledCount += 1;
-		deps.logger.info("ratelimit_provider_enabled", {
-			providerId,
+		out.set(provider.id, opts);
+		logger.info("ratelimit_provider_enabled", {
+			providerId: provider.id,
 			scope: opts.scope,
 			tiers: opts.tiers.length,
 			headerPrefix: opts.headerPrefix
 		});
 	}
-	deps.logger.info("ratelimit_plugin_initialized", { providerCount: enabledCount });
+	logger.info("ratelimit_plugin_initialized", { providerCount: out.size });
+	return out;
 }
-/**
-* Build the fetch wrapper for one provider. The wrapper:
-*  1. Resolves the bucket (provider, or `(provider, model)` under `scope: "model"`).
-*  2. Pre-request gate: if that bucket's cooldown is armed, waits until it clears.
-*  3. Sends the request via the underlying fetch.
-*  4. Reads the rate-limit headers; selects the policy tier by reset magnitude.
-*     On `remaining: 0` it arms the gate for the next callers (only when the
-*     matched tier is `"wait"`).
-*  5. On a 429: a `"wait"` tier waits the reset window and retries (up to the
-*     tier's `maxRetries`); an `"error"` tier surfaces the 429 immediately.
-*
-* The Response is returned untouched (no `.clone()`) — we only read `status`
-* and `headers`, so the body stream is delivered to OpenCode intact.
-*/
-export function makeRateLimitFetch(providerId, opts, store, deps, delegate) {
+export function createRateLimiter(deps) {
 	const now = deps.now ?? Date.now;
 	const sleep = deps.sleep ?? defaultSleep;
-	const underlying = delegate ?? deps.fetchImpl ?? globalThis.fetch;
 	const { logger } = deps;
-	const isRequest = (value) => typeof Request !== "undefined" && value instanceof Request;
-	const wrapped = async (input, init) => {
-		// Honor an abort signal whether it rides on `init` or on a `Request` input.
-		const signal = init?.signal ?? (isRequest(input) ? input.signal : undefined) ?? undefined;
-		const model = opts.scope === "model" ? await modelFromRequest(input, init) : undefined;
-		if (opts.scope === "model") {
-			logger.trace("ratelimit_model_resolved", {
-				providerId,
-				model,
-				matched: model !== undefined
-			});
-		}
-		const key = model ? `${providerId}\u0000${model}` : providerId;
+	let configured = new Map();
+	const store = new Map();
+	const optionsFor = (providerId) => configured.get(providerId);
+	const bucket = (ref, opts) => {
+		const model = opts.scope === "model" ? ref.modelID : undefined;
+		const key = model ? `${ref.providerId}\u0000${model}` : ref.providerId;
 		let state = store.get(key);
-		const bucketExisted = state !== undefined;
+		const existed = state !== undefined;
 		if (!state) {
 			state = createProviderState();
 			store.set(key, state);
 		}
-		logger.trace("ratelimit_fetch_invoked", {
-			providerId,
-			scope: opts.scope,
+		return {
+			state,
+			key,
 			model,
-			bucketKey: key,
-			bucketExisted,
-			cooldownUntilMs: state.cooldownUntilMs
-		});
-		// 2. Pre-request gate — wait out a cooldown a previous "wait" tier armed.
-		if (state.cooldownUntilMs > now()) {
+			existed
+		};
+	};
+	return {
+		configure(entries) {
+			configured = new Map(entries);
+			// Drop state for providers that are no longer opted in so a reload does
+			// not leave an orphaned cooldown gating a provider nobody manages.
+			for (const key of [...store.keys()]) {
+				const providerId = key.split("\0", 1)[0] ?? key;
+				if (!configured.has(providerId)) {
+					store.delete(key);
+				}
+			}
+		},
+		options: optionsFor,
+		async acquire(ref, signal) {
+			const opts = optionsFor(ref.providerId);
+			if (!opts) {
+				return;
+			}
+			const { state, key, model, existed } = bucket(ref, opts);
+			logger.trace("ratelimit_acquire", {
+				providerId: ref.providerId,
+				scope: opts.scope,
+				model,
+				bucketKey: key,
+				bucketExisted: existed,
+				cooldownUntilMs: state.cooldownUntilMs
+			});
+			if (state.cooldownUntilMs <= now()) {
+				return;
+			}
 			const waitMs = clampWait(state.cooldownUntilMs - now(), state.cooldownMaxWaitMs);
 			logger.trace("ratelimit_cooldown_gate_armed", {
-				providerId,
+				providerId: ref.providerId,
 				model,
 				cooldownUntilMs: state.cooldownUntilMs,
 				remainingMs: state.cooldownUntilMs - now(),
 				clampedWaitMs: waitMs,
 				maxWaitMs: state.cooldownMaxWaitMs
 			});
-			if (waitMs > 0) {
-				logger.info("ratelimit_throttle_wait", {
-					providerId,
-					model,
-					waitMs
-				});
-				await waitGate(state, sleep, now, state.cooldownMaxWaitMs, signal, logger, providerId, model, "pre_request");
+			if (waitMs <= 0) {
+				return;
 			}
-		}
-		// 3-5. Attempt loop.
-		let attempt = 0;
-		for (;;) {
-			// A Request body is single-use; send a fresh clone each attempt so a
-			// wait-tier retry doesn't fail with "body already used". (The original is
-			// never sent directly, so each clone has an unconsumed body.)
-			const attemptInput = isRequest(input) ? input.clone() : input;
-			logger.trace("ratelimit_underlying_fetch", {
-				providerId,
+			logger.info("ratelimit_throttle_wait", {
+				providerId: ref.providerId,
 				model,
-				attempt
+				waitMs
 			});
-			const response = await underlying(attemptInput, init);
-			const snapshot = readSnapshot(response, opts, now(), logger, providerId);
+			await waitGate(state, sleep, now, state.cooldownMaxWaitMs, signal, logger, ref.providerId, model, "pre_request");
+		},
+		observe(ref, response) {
+			const opts = optionsFor(ref.providerId);
+			if (!opts) {
+				return;
+			}
+			const { state, model } = bucket(ref, opts);
+			const snapshot = readSnapshot(response.headers, opts, now(), logger, ref.providerId);
+			state.lastSnapshot = snapshot;
 			logger.trace("ratelimit_snapshot_parsed", {
-				providerId,
+				providerId: ref.providerId,
 				model,
 				status: response.status,
 				limit: snapshot.limit,
@@ -155,56 +126,89 @@ export function makeRateLimitFetch(providerId, opts, store, deps, delegate) {
 				retryAfterMs: snapshot.retryAfterMs
 			});
 			logger.debug("ratelimit_quota", {
-				providerId,
+				providerId: ref.providerId,
 				model,
 				remaining: snapshot.remaining,
 				limit: snapshot.limit,
 				resetSeconds: snapshot.resetSeconds
 			});
-			if (response.status !== 429) {
-				// Arm the gate for the NEXT callers when the window is exhausted — but
-				// only under a "wait" tier. An "error" tier wants the next request to
-				// hit a real 429 and be surfaced, so we leave the gate unarmed.
-				if (snapshot.remaining !== undefined && snapshot.remaining <= 0 && snapshot.resetAtMs !== undefined) {
-					const resetForTier = effectiveResetSeconds(snapshot);
-					const tier = selectTier(opts.tiers, resetForTier);
-					logger.trace("ratelimit_exhausted_tier_selected", {
-						providerId,
+			const resetForTier = effectiveResetSeconds(snapshot);
+			const tier = selectTier(opts.tiers, resetForTier);
+			if (response.status === 429) {
+				// The host will ask `decideRetry` what THIS caller does. Arm the gate so
+				// every OTHER concurrent caller throttles too (v1 got this from the
+				// shared wrapper state); an "error" tier clears it so the fail-fast is
+				// actually fast.
+				if (tier.action === "wait") {
+					const waitMs = clampWait(computeBackoff(snapshot, now()), tier.maxWaitMs);
+					state.cooldownUntilMs = now() + waitMs;
+					state.cooldownMaxWaitMs = tier.maxWaitMs;
+					logger.trace("ratelimit_cooldown_set", {
+						providerId: ref.providerId,
 						model,
-						resetSeconds: resetForTier,
-						maxResetSeconds: tier.maxResetSeconds,
-						action: tier.action,
+						cooldownUntilMs: state.cooldownUntilMs,
 						maxWaitMs: tier.maxWaitMs,
-						maxRetries: tier.maxRetries
+						source: "429_backoff"
+					});
+				} else {
+					state.cooldownUntilMs = 0;
+					state.cooldownMaxWaitMs = 0;
+					logger.trace("ratelimit_cooldown_cleared", {
+						providerId: ref.providerId,
+						model,
+						source: "429_error_tier"
 					});
-					if (tier.action === "wait") {
-						state.cooldownUntilMs = snapshot.resetAtMs;
-						state.cooldownMaxWaitMs = tier.maxWaitMs;
-						logger.trace("ratelimit_cooldown_set", {
-							providerId,
-							model,
-							cooldownUntilMs: state.cooldownUntilMs,
-							maxWaitMs: tier.maxWaitMs,
-							source: "exhausted"
-						});
-					} else {
-						// Error tier → don't arm, and drop any stale cooldown a prior
-						// "wait" window left so the next request hits a real 429 at once.
-						state.cooldownUntilMs = 0;
-						state.cooldownMaxWaitMs = 0;
-						logger.trace("ratelimit_cooldown_cleared", {
-							providerId,
-							model,
-							source: "exhausted_error_tier"
-						});
-					}
 				}
-				return response;
+				return;
+			}
+			// Arm the gate for the NEXT callers when the window is exhausted — but
+			// only under a "wait" tier. An "error" tier wants the next request to hit
+			// a real 429 and be surfaced, so we leave the gate unarmed.
+			if (snapshot.remaining === undefined || snapshot.remaining > 0 || snapshot.resetAtMs === undefined) {
+				return;
 			}
+			logger.trace("ratelimit_exhausted_tier_selected", {
+				providerId: ref.providerId,
+				model,
+				resetSeconds: resetForTier,
+				maxResetSeconds: tier.maxResetSeconds,
+				action: tier.action,
+				maxWaitMs: tier.maxWaitMs,
+				maxRetries: tier.maxRetries
+			});
+			if (tier.action === "wait") {
+				state.cooldownUntilMs = snapshot.resetAtMs;
+				state.cooldownMaxWaitMs = tier.maxWaitMs;
+				logger.trace("ratelimit_cooldown_set", {
+					providerId: ref.providerId,
+					model,
+					cooldownUntilMs: state.cooldownUntilMs,
+					maxWaitMs: tier.maxWaitMs,
+					source: "exhausted"
+				});
+				return;
+			}
+			state.cooldownUntilMs = 0;
+			state.cooldownMaxWaitMs = 0;
+			logger.trace("ratelimit_cooldown_cleared", {
+				providerId: ref.providerId,
+				model,
+				source: "exhausted_error_tier"
+			});
+		},
+		decideRetry(ref, error, attempt) {
+			const opts = optionsFor(ref.providerId);
+			if (!opts || error.status !== 429) {
+				return { retry: false };
+			}
+			const { state, model } = bucket(ref, opts);
+			// Prefer what the response hook saw; fall back to the error's own
+			// `retryAfterMs` when this failure never produced an observed response.
+			const snapshot = state.lastSnapshot ?? (error.retryAfterMs !== undefined ? { retryAfterMs: error.retryAfterMs } : {});
 			const resetForTier = effectiveResetSeconds(snapshot);
 			const tier = selectTier(opts.tiers, resetForTier);
 			logger.trace("ratelimit_429_tier_selected", {
-				providerId,
+				providerId: ref.providerId,
 				model,
 				attempt,
 				resetSeconds: resetForTier,
@@ -214,100 +218,50 @@ export function makeRateLimitFetch(providerId, opts, store, deps, delegate) {
 				maxRetries: tier.maxRetries
 			});
 			if (tier.action === "error") {
-				// Drop any cooldown a prior "wait" tier armed, so the fail-fast is
-				// actually fast — later requests must not sleep a stale window first.
-				state.cooldownUntilMs = 0;
-				state.cooldownMaxWaitMs = 0;
-				logger.trace("ratelimit_cooldown_cleared", {
-					providerId,
-					model,
-					source: "429_error_tier"
-				});
 				logger.warn("ratelimit_failfast", {
-					providerId,
+					providerId: ref.providerId,
 					model,
 					resetSeconds: snapshot.resetSeconds
 				});
-				return response;
+				return { retry: false };
 			}
 			if (attempt >= tier.maxRetries) {
 				logger.trace("ratelimit_retries_exhausted", {
-					providerId,
+					providerId: ref.providerId,
 					model,
 					attempt,
 					maxRetries: tier.maxRetries
 				});
 				logger.error("ratelimit_giveup", {
-					providerId,
+					providerId: ref.providerId,
 					model,
 					attempts: attempt
 				});
-				return response;
+				return { retry: false };
 			}
 			const backoffMs = computeBackoff(snapshot, now());
 			const waitMs = clampWait(backoffMs, tier.maxWaitMs);
 			logger.trace("ratelimit_429_backoff_computed", {
-				providerId,
+				providerId: ref.providerId,
 				model,
 				attempt: attempt + 1,
 				rawBackoffMs: backoffMs,
 				clampedWaitMs: waitMs,
 				maxWaitMs: tier.maxWaitMs
 			});
-			state.cooldownUntilMs = now() + waitMs;
-			state.cooldownMaxWaitMs = tier.maxWaitMs;
-			logger.trace("ratelimit_cooldown_set", {
-				providerId,
-				model,
-				cooldownUntilMs: state.cooldownUntilMs,
-				maxWaitMs: tier.maxWaitMs,
-				source: "429_backoff"
-			});
 			logger.warn("ratelimit_429_backoff", {
-				providerId,
+				providerId: ref.providerId,
 				model,
 				attempt: attempt + 1,
 				waitMs,
 				resetSeconds: snapshot.resetSeconds
 			});
-			await waitGate(state, sleep, now, tier.maxWaitMs, signal, logger, providerId, model, "backoff");
-			attempt += 1;
+			return {
+				retry: true,
+				delay: waitMs
+			};
 		}
 	};
-	return wrapped;
-}
-/**
-* Best-effort extraction of the `model` from an OpenAI-compatible request.
-* The AI SDK calls `fetch(url, { body: "<json>" })`, so the common path reads a
-* JSON string body synchronously (non-destructive). It also supports the
-* `fetch(new Request(...))` shape by reading a clone of the Request body, so a
-* Request-style caller doesn't silently collapse to the provider-wide bucket.
-* Anything unparseable → `undefined` → caller falls back to the provider bucket.
-*/
-async function modelFromRequest(input, init) {
-	const fromInit = modelFromBody(init?.body);
-	if (fromInit !== undefined) {
-		return fromInit;
-	}
-	if (typeof Request !== "undefined" && input instanceof Request) {
-		try {
-			return modelFromBody(await input.clone().text());
-		} catch {
-			return undefined;
-		}
-	}
-	return undefined;
-}
-function modelFromBody(body) {
-	if (typeof body !== "string") {
-		return undefined;
-	}
-	try {
-		const parsed = JSON.parse(body);
-		return typeof parsed.model === "string" && parsed.model.length > 0 ? parsed.model : undefined;
-	} catch {
-		return undefined;
-	}
 }
 /**
 * Reset magnitude (seconds) used for tier selection: prefer `x-ratelimit-reset`,
@@ -324,9 +278,9 @@ function effectiveResetSeconds(snapshot) {
 	}
 	return undefined;
 }
-function readSnapshot(response, opts, nowMs, logger, providerId) {
+function readSnapshot(headers, opts, nowMs, logger, providerId) {
 	try {
-		return parseRateLimit(response.headers, opts.headerPrefix, nowMs);
+		return parseRateLimit(headers, opts.headerPrefix, nowMs);
 	} catch (error) {
 		logger.debug("ratelimit_header_parse_failed", {
 			providerId,

package/dist/plugin.js.map

@@ -1 +1 @@
-{"mappings":"AAAA,SAAS,uBAAuB,kBAAkB;AAClD,SAAS,sBAAsB;;AAK/B,OAAO,MAAM,qBAAqB;AA2BlC,OAAO,SAAS,sBAAyC;CACvD,OAAO;EAAE,iBAAiB;EAAG,mBAAmB;CAAE;AACpD;;;;;;;AA6BA,OAAO,SAAS,mBAAmB,OAA2B,MAA2B;CACvF,MAAM,YAAY,MAAM;CACxB,KAAK,OAAO,MAAM,2BAA2B,EAC3C,eAAe,YAAY,OAAO,KAAK,SAAS,CAAC,CAAC,SAAS,EAC7D,CAAC;CACD,IAAI,CAAC,WAAW;EACd,KAAK,OAAO,KAAK,gCAAgC,EAAE,eAAe,EAAE,CAAC;EACrE;CACF;CAEA,IAAI,eAAe;CACnB,KAAK,MAAM,CAAC,YAAY,mBAAmB,OAAO,QAAQ,SAAS,GAAG;EACpE,IAAI,CAAC,gBAAgB;GACnB;EACF;EACA,KAAK,OAAO,MAAM,kCAAkC;GAClD;GACA,YAAY,QAAQ,eAAe,OAAO;EAC5C,CAAC;EACD,MAAM,OAAO,sBAAsB,eAAe,OAAO;EACzD,IAAI,CAAC,MAAM;GACT,KAAK,OAAO,MAAM,8BAA8B;IAAE;IAAY,QAAQ;GAAe,CAAC;GACtF;EACF;EACA,KAAK,OAAO,MAAM,qCAAqC;GACrD;GACA,OAAO,KAAK;GACZ,cAAc,KAAK;GACnB,OAAO,KAAK,MAAM;EACpB,CAAC;EAED,MAAM,UAAW,eAAe,YAAY,CAAC;;;EAG7C,MAAM,WACJ,OAAO,QAAQ,UAAU,aAAc,QAAQ,QAAyB;EAC1E,KAAK,OAAO,MAAM,2BAA2B;GAC3C;GACA,2BAA2B,QAAQ,QAAQ;EAC7C,CAAC;EACD,MAAM,QAAwB,IAAI,IAAI;EACtC,QAAQ,QAAQ,mBAAmB,YAAY,MAAM,OAAO,MAAM,QAAQ;EAC1E,gBAAgB;EAChB,KAAK,OAAO,KAAK,8BAA8B;GAC7C;GACA,OAAO,KAAK;GACZ,OAAO,KAAK,MAAM;GAClB,cAAc,KAAK;EACrB,CAAC;CACH;CAEA,KAAK,OAAO,KAAK,gCAAgC,EAAE,eAAe,aAAa,CAAC;AAClF;;;;;;;;;;;;;;;AAgBA,OAAO,SAAS,mBACd,YACA,MACA,OACA,MACA,UACc;CACd,MAAM,MAAM,KAAK,OAAO,KAAK;CAC7B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,aAAa,YAAY,KAAK,aAAa,WAAW;CAC5D,MAAM,EAAE,WAAW;CAEnB,MAAM,aAAa,UACjB,OAAO,YAAY,eAAe,iBAAiB;CAErD,MAAM,UAAwB,OAAO,OAAO,SAAS;;EAEnD,MAAM,SAAS,MAAM,WAAW,UAAU,KAAK,IAAI,MAAM,SAAS,cAAc;EAChF,MAAM,QAAQ,KAAK,UAAU,UAAU,MAAM,iBAAiB,OAAO,IAAI,IAAI;EAC7E,IAAI,KAAK,UAAU,SAAS;GAC1B,OAAO,MAAM,4BAA4B;IAAE;IAAY;IAAO,SAAS,UAAU;GAAU,CAAC;EAC9F;EACA,MAAM,MAAM,QAAQ,GAAG,WAAW,QAAQ,UAAU;EACpD,IAAI,QAAQ,MAAM,IAAI,GAAG;EACzB,MAAM,gBAAgB,UAAU;EAChC,IAAI,CAAC,OAAO;GACV,QAAQ,oBAAoB;GAC5B,MAAM,IAAI,KAAK,KAAK;EACtB;EACA,OAAO,MAAM,2BAA2B;GACtC;GACA,OAAO,KAAK;GACZ;GACA,WAAW;GACX;GACA,iBAAiB,MAAM;EACzB,CAAC;;EAGD,IAAI,MAAM,kBAAkB,IAAI,GAAG;GACjC,MAAM,SAAS,UAAU,MAAM,kBAAkB,IAAI,GAAG,MAAM,iBAAiB;GAC/E,OAAO,MAAM,iCAAiC;IAC5C;IACA;IACA,iBAAiB,MAAM;IACvB,aAAa,MAAM,kBAAkB,IAAI;IACzC,eAAe;IACf,WAAW,MAAM;GACnB,CAAC;GACD,IAAI,SAAS,GAAG;IACd,OAAO,KAAK,2BAA2B;KAAE;KAAY;KAAO;IAAO,CAAC;IACpE,MAAM,SACJ,OACA,OACA,KACA,MAAM,mBACN,QACA,QACA,YACA,OACA,aACF;GACF;EACF;;EAGA,IAAI,UAAU;EACd,SAAS;;;;GAIP,MAAM,eAAe,UAAU,KAAK,IAAI,MAAM,MAAM,IAAI;GACxD,OAAO,MAAM,8BAA8B;IAAE;IAAY;IAAO;GAAQ,CAAC;GACzE,MAAM,WAAW,MAAM,WAAW,cAAc,IAAI;GACpD,MAAM,WAAW,aAAa,UAAU,MAAM,IAAI,GAAG,QAAQ,UAAU;GACvE,OAAO,MAAM,6BAA6B;IACxC;IACA;IACA,QAAQ,SAAS;IACjB,OAAO,SAAS;IAChB,WAAW,SAAS;IACpB,cAAc,SAAS;IACvB,WAAW,SAAS;IACpB,cAAc,SAAS;GACzB,CAAC;GACD,OAAO,MAAM,mBAAmB;IAC9B;IACA;IACA,WAAW,SAAS;IACpB,OAAO,SAAS;IAChB,cAAc,SAAS;GACzB,CAAC;GAED,IAAI,SAAS,WAAW,KAAK;;;;IAI3B,IACE,SAAS,cAAc,aACvB,SAAS,aAAa,KACtB,SAAS,cAAc,WACvB;KACA,MAAM,eAAe,sBAAsB,QAAQ;KACnD,MAAM,OAAO,WAAW,KAAK,OAAO,YAAY;KAChD,OAAO,MAAM,qCAAqC;MAChD;MACA;MACA,cAAc;MACd,iBAAiB,KAAK;MACtB,QAAQ,KAAK;MACb,WAAW,KAAK;MAChB,YAAY,KAAK;KACnB,CAAC;KACD,IAAI,KAAK,WAAW,QAAQ;MAC1B,MAAM,kBAAkB,SAAS;MACjC,MAAM,oBAAoB,KAAK;MAC/B,OAAO,MAAM,0BAA0B;OACrC;OACA;OACA,iBAAiB,MAAM;OACvB,WAAW,KAAK;OAChB,QAAQ;MACV,CAAC;KACH,OAAO;;;MAGL,MAAM,kBAAkB;MACxB,MAAM,oBAAoB;MAC1B,OAAO,MAAM,8BAA8B;OACzC;OACA;OACA,QAAQ;MACV,CAAC;KACH;IACF;IACA,OAAO;GACT;GAEA,MAAM,eAAe,sBAAsB,QAAQ;GACnD,MAAM,OAAO,WAAW,KAAK,OAAO,YAAY;GAChD,OAAO,MAAM,+BAA+B;IAC1C;IACA;IACA;IACA,cAAc;IACd,iBAAiB,KAAK;IACtB,QAAQ,KAAK;IACb,WAAW,KAAK;IAChB,YAAY,KAAK;GACnB,CAAC;GACD,IAAI,KAAK,WAAW,SAAS;;;IAG3B,MAAM,kBAAkB;IACxB,MAAM,oBAAoB;IAC1B,OAAO,MAAM,8BAA8B;KACzC;KACA;KACA,QAAQ;IACV,CAAC;IACD,OAAO,KAAK,sBAAsB;KAChC;KACA;KACA,cAAc,SAAS;IACzB,CAAC;IACD,OAAO;GACT;GAEA,IAAI,WAAW,KAAK,YAAY;IAC9B,OAAO,MAAM,+BAA+B;KAC1C;KACA;KACA;KACA,YAAY,KAAK;IACnB,CAAC;IACD,OAAO,MAAM,oBAAoB;KAAE;KAAY;KAAO,UAAU;IAAQ,CAAC;IACzE,OAAO;GACT;GAEA,MAAM,YAAY,eAAe,UAAU,IAAI,CAAC;GAChD,MAAM,SAAS,UAAU,WAAW,KAAK,SAAS;GAClD,OAAO,MAAM,kCAAkC;IAC7C;IACA;IACA,SAAS,UAAU;IACnB,cAAc;IACd,eAAe;IACf,WAAW,KAAK;GAClB,CAAC;GACD,MAAM,kBAAkB,IAAI,IAAI;GAChC,MAAM,oBAAoB,KAAK;GAC/B,OAAO,MAAM,0BAA0B;IACrC;IACA;IACA,iBAAiB,MAAM;IACvB,WAAW,KAAK;IAChB,QAAQ;GACV,CAAC;GACD,OAAO,KAAK,yBAAyB;IACnC;IACA;IACA,SAAS,UAAU;IACnB;IACA,cAAc,SAAS;GACzB,CAAC;GACD,MAAM,SACJ,OACA,OACA,KACA,KAAK,WACL,QACA,QACA,YACA,OACA,SACF;GACA,WAAW;EACb;CACF;CAEA,OAAO;AACT;;;;;;;;;AAUA,eAAe,iBACb,OACA,MAC6B;CAC7B,MAAM,WAAW,cAAc,MAAM,IAAI;CACzC,IAAI,aAAa,WAAW;EAC1B,OAAO;CACT;CACA,IAAI,OAAO,YAAY,eAAe,iBAAiB,SAAS;EAC9D,IAAI;GACF,OAAO,cAAc,MAAM,MAAM,MAAM,CAAC,CAAC,KAAK,CAAC;EACjD,QAAQ;GACN,OAAO;EACT;CACF;CACA,OAAO;AACT;AAEA,SAAS,cAAc,MAAuD;CAC5E,IAAI,OAAO,SAAS,UAAU;EAC5B,OAAO;CACT;CACA,IAAI;EACF,MAAM,SAAS,KAAK,MAAM,IAAI;EAC9B,OAAO,OAAO,OAAO,UAAU,YAAY,OAAO,MAAM,SAAS,IAAI,OAAO,QAAQ;CACtF,QAAQ;EACN,OAAO;CACT;AACF;;;;;;;AAQA,SAAS,sBAAsB,UAAiD;CAC9E,IAAI,SAAS,iBAAiB,WAAW;EACvC,OAAO,SAAS;CAClB;CACA,IAAI,SAAS,iBAAiB,WAAW;EACvC,OAAO,KAAK,KAAK,SAAS,eAAe,GAAI;CAC/C;CACA,OAAO;AACT;AAEA,SAAS,aACP,UACA,MACA,OACA,QACA,YACmB;CACnB,IAAI;EACF,OAAO,eAAe,SAAS,SAAS,KAAK,cAAc,KAAK;CAClE,SAAS,OAAO;EACd,OAAO,MAAM,iCAAiC;GAC5C;GACA,OAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;EAC9D,CAAC;EACD,OAAO,CAAC;CACV;AACF;AAEA,SAAS,eAAe,UAA6B,OAAuB;CAC1E,IAAI,SAAS,cAAc,WAAW;EACpC,OAAO,KAAK,IAAI,GAAG,SAAS,YAAY,KAAK;CAC/C;CACA,IAAI,SAAS,iBAAiB,WAAW;EACvC,OAAO,SAAS;CAClB;CACA,OAAO;AACT;;AAGA,SAAS,UAAU,IAAY,WAA2B;CACxD,MAAM,cAAc,KAAK,IAAI,GAAG,EAAE;CAClC,OAAO,YAAY,IAAI,KAAK,IAAI,aAAa,SAAS,IAAI;AAC5D;;;;;;;;;;;;;;;;;;;;;AAsBA,eAAe,SACb,OACA,OACA,KACA,WACA,QACA,QACA,YACA,OACA,OACe;CACf,MAAM,aAAa,YAAY,IAAI,IAAI,IAAI,YAAY,OAAO;CAC9D,OAAO,MAAM,wBAAwB;EACnC;EACA;EACA;EACA,iBAAiB,MAAM;EACvB;CACF,CAAC;CACD,SAAS;EACP,MAAM,SAAS,KAAK,IAAI,MAAM,iBAAiB,UAAU;EACzD,MAAM,cAAc,SAAS,IAAI;EACjC,IAAI,eAAe,GAAG;GACpB,OAAO,MAAM,sBAAsB;IAAE;IAAY;IAAO;GAAM,CAAC;GAC/D;EACF;EACA,OAAO,MAAM,uBAAuB;GAClC;GACA;GACA;GACA;GACA,aAAa,MAAM,oBAAoB;EACzC,CAAC;;EAGD,IAAI,UAAU,MAAM;EACpB,IACE,WACA,MAAM,2BAA2B,aACjC,MAAM,yBAAyB,IAAI,IAAI,aACvC;GACA,UAAU;EACZ;EACA,IAAI,CAAC,SAAS;GACZ,MAAM,UAAU,MAAM,WAAW,CAAC,CAAC,cAAc;IAC/C,IAAI,MAAM,oBAAoB,SAAS;KACrC,MAAM,kBAAkB;KACxB,MAAM,yBAAyB;IACjC;GACF,CAAC;;;GAGD,IAAI,CAAC,MAAM,iBAAiB;IAC1B,MAAM,kBAAkB;IACxB,MAAM,yBAAyB,IAAI,IAAI;GACzC;GACA,UAAU;EACZ;EAEA,IAAI;GACF,MAAM,cAAc,SAAS,MAAM;EACrC,SAAS,OAAO;GACd,IAAI,aAAa,KAAK,GAAG;IACvB,OAAO,KAAK,0BAA0B;KAAE;KAAY;KAAO;IAAM,CAAC;GACpE;GACA,MAAM;EACR;CAEF;AACF;AAEA,SAAS,cAAc,SAAwB,QAAgD;CAC7F,IAAI,CAAC,QAAQ;EACX,OAAO;CACT;CACA,IAAI,OAAO,SAAS;EAClB,OAAO,QAAQ,OAAO,aAAa,MAAM,CAAC;CAC5C;CACA,OAAO,IAAI,SAAe,SAAS,WAAW;EAC5C,MAAM,gBAAgB,OAAO,aAAa,MAAM,CAAC;EACjD,OAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;EACxD,QAAQ,WACA;GACJ,OAAO,oBAAoB,SAAS,OAAO;GAC3C,QAAQ;EACV,IACC,UAAU;GACT,OAAO,oBAAoB,SAAS,OAAO;GAC3C,OAAO,KAAK;EACd,CACF;CACF,CAAC;AACH;AAEA,SAAS,aAAa,IAAY,QAAqC;CACrE,OAAO,IAAI,SAAe,SAAS,WAAW;EAC5C,IAAI,QAAQ,SAAS;GACnB,OAAO,aAAa,MAAM,CAAC;GAC3B;EACF;EACA,MAAM,QAAQ,iBAAiB;GAC7B,QAAQ,oBAAoB,SAAS,OAAO;GAC5C,QAAQ;EACV,GAAG,EAAE;EACL,MAAM,gBAAgB;GACpB,aAAa,KAAK;GAClB,OAAO,aAAa,MAAqB,CAAC;EAC5C;EACA,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;CAC3D,CAAC;AACH;AAEA,SAAS,aAAa,QAA4B;CAChD,MAAM,SAAS,OAAO;CACtB,IAAI,kBAAkB,OAAO;EAC3B,OAAO;CACT;CACA,OAAO,IAAI,aAAa,6BAA6B,YAAY;AACnE;AAEA,SAAS,aAAa,OAAyB;CAC7C,OAAO,iBAAiB,SAAS,MAAM,SAAS;AAClD","names":[],"sources":["../src/plugin.ts"],"version":3,"file":"plugin.js","sourceRoot":""}
+{"mappings":"AAAA,SAAS,uBAAuB,kBAAkB;AAClD,SAAS,sBAAsB;;AAK/B,OAAO,MAAM,qBAAqB;AAkClC,OAAO,SAAS,sBAAyC;CACvD,OAAO;EAAE,iBAAiB;EAAG,mBAAmB;CAAE;AACpD;;;;;;AAyDA,OAAO,SAAS,cACd,WACA,QAC+B;CAC/B,MAAM,MAAM,IAAI,IAA8B;CAC9C,KAAK,MAAM,YAAY,WAAW;EAChC,MAAM,OAAO,sBAAsB,SAAS,QAAQ;EACpD,IAAI,CAAC,MAAM;GACT,OAAO,MAAM,8BAA8B;IACzC,YAAY,SAAS;IACrB,QAAQ;GACV,CAAC;GACD;EACF;EACA,IAAI,IAAI,SAAS,IAAI,IAAI;EACzB,OAAO,KAAK,8BAA8B;GACxC,YAAY,SAAS;GACrB,OAAO,KAAK;GACZ,OAAO,KAAK,MAAM;GAClB,cAAc,KAAK;EACrB,CAAC;CACH;CACA,OAAO,KAAK,gCAAgC,EAAE,eAAe,IAAI,KAAK,CAAC;CACvE,OAAO;AACT;AAEA,OAAO,SAAS,kBAAkB,MAAkC;CAClE,MAAM,MAAM,KAAK,OAAO,KAAK;CAC7B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,EAAE,WAAW;CAEnB,IAAI,aAAa,IAAI,IAA8B;CACnD,MAAM,QAAwB,IAAI,IAAI;CAEtC,MAAM,cAAc,eAAuB,WAAW,IAAI,UAAU;CAEpE,MAAM,UAAU,KAAgB,SAA2B;EACzD,MAAM,QAAQ,KAAK,UAAU,UAAU,IAAI,UAAU;EACrD,MAAM,MAAM,QAAQ,GAAG,IAAI,WAAW,QAAQ,UAAU,IAAI;EAC5D,IAAI,QAAQ,MAAM,IAAI,GAAG;EACzB,MAAM,UAAU,UAAU;EAC1B,IAAI,CAAC,OAAO;GACV,QAAQ,oBAAoB;GAC5B,MAAM,IAAI,KAAK,KAAK;EACtB;EACA,OAAO;GAAE;GAAO;GAAK;GAAO;EAAQ;CACtC;CAEA,OAAO;EACL,UAAU,SAAS;GACjB,aAAa,IAAI,IAAI,OAAO;;;GAG5B,KAAK,MAAM,OAAO,CAAC,GAAG,MAAM,KAAK,CAAC,GAAG;IACnC,MAAM,aAAa,IAAI,MAAM,MAAU,CAAC,CAAC,CAAC,MAAM;IAChD,IAAI,CAAC,WAAW,IAAI,UAAU,GAAG;KAC/B,MAAM,OAAO,GAAG;IAClB;GACF;EACF;EAEA,SAAS;EAET,MAAM,QAAQ,KAAK,QAAQ;GACzB,MAAM,OAAO,WAAW,IAAI,UAAU;GACtC,IAAI,CAAC,MAAM;IACT;GACF;GACA,MAAM,EAAE,OAAO,KAAK,OAAO,YAAY,OAAO,KAAK,IAAI;GACvD,OAAO,MAAM,qBAAqB;IAChC,YAAY,IAAI;IAChB,OAAO,KAAK;IACZ;IACA,WAAW;IACX,eAAe;IACf,iBAAiB,MAAM;GACzB,CAAC;GAED,IAAI,MAAM,mBAAmB,IAAI,GAAG;IAClC;GACF;GAEA,MAAM,SAAS,UAAU,MAAM,kBAAkB,IAAI,GAAG,MAAM,iBAAiB;GAC/E,OAAO,MAAM,iCAAiC;IAC5C,YAAY,IAAI;IAChB;IACA,iBAAiB,MAAM;IACvB,aAAa,MAAM,kBAAkB,IAAI;IACzC,eAAe;IACf,WAAW,MAAM;GACnB,CAAC;GACD,IAAI,UAAU,GAAG;IACf;GACF;GACA,OAAO,KAAK,2BAA2B;IAAE,YAAY,IAAI;IAAY;IAAO;GAAO,CAAC;GACpF,MAAM,SACJ,OACA,OACA,KACA,MAAM,mBACN,QACA,QACA,IAAI,YACJ,OACA,aACF;EACF;EAEA,QAAQ,KAAK,UAAU;GACrB,MAAM,OAAO,WAAW,IAAI,UAAU;GACtC,IAAI,CAAC,MAAM;IACT;GACF;GACA,MAAM,EAAE,OAAO,UAAU,OAAO,KAAK,IAAI;GACzC,MAAM,WAAW,aAAa,SAAS,SAAS,MAAM,IAAI,GAAG,QAAQ,IAAI,UAAU;GACnF,MAAM,eAAe;GACrB,OAAO,MAAM,6BAA6B;IACxC,YAAY,IAAI;IAChB;IACA,QAAQ,SAAS;IACjB,OAAO,SAAS;IAChB,WAAW,SAAS;IACpB,cAAc,SAAS;IACvB,WAAW,SAAS;IACpB,cAAc,SAAS;GACzB,CAAC;GACD,OAAO,MAAM,mBAAmB;IAC9B,YAAY,IAAI;IAChB;IACA,WAAW,SAAS;IACpB,OAAO,SAAS;IAChB,cAAc,SAAS;GACzB,CAAC;GAED,MAAM,eAAe,sBAAsB,QAAQ;GACnD,MAAM,OAAO,WAAW,KAAK,OAAO,YAAY;GAEhD,IAAI,SAAS,WAAW,KAAK;;;;;IAK3B,IAAI,KAAK,WAAW,QAAQ;KAC1B,MAAM,SAAS,UAAU,eAAe,UAAU,IAAI,CAAC,GAAG,KAAK,SAAS;KACxE,MAAM,kBAAkB,IAAI,IAAI;KAChC,MAAM,oBAAoB,KAAK;KAC/B,OAAO,MAAM,0BAA0B;MACrC,YAAY,IAAI;MAChB;MACA,iBAAiB,MAAM;MACvB,WAAW,KAAK;MAChB,QAAQ;KACV,CAAC;IACH,OAAO;KACL,MAAM,kBAAkB;KACxB,MAAM,oBAAoB;KAC1B,OAAO,MAAM,8BAA8B;MACzC,YAAY,IAAI;MAChB;MACA,QAAQ;KACV,CAAC;IACH;IACA;GACF;;;;GAKA,IACE,SAAS,cAAc,aACvB,SAAS,YAAY,KACrB,SAAS,cAAc,WACvB;IACA;GACF;GACA,OAAO,MAAM,qCAAqC;IAChD,YAAY,IAAI;IAChB;IACA,cAAc;IACd,iBAAiB,KAAK;IACtB,QAAQ,KAAK;IACb,WAAW,KAAK;IAChB,YAAY,KAAK;GACnB,CAAC;GACD,IAAI,KAAK,WAAW,QAAQ;IAC1B,MAAM,kBAAkB,SAAS;IACjC,MAAM,oBAAoB,KAAK;IAC/B,OAAO,MAAM,0BAA0B;KACrC,YAAY,IAAI;KAChB;KACA,iBAAiB,MAAM;KACvB,WAAW,KAAK;KAChB,QAAQ;IACV,CAAC;IACD;GACF;GACA,MAAM,kBAAkB;GACxB,MAAM,oBAAoB;GAC1B,OAAO,MAAM,8BAA8B;IACzC,YAAY,IAAI;IAChB;IACA,QAAQ;GACV,CAAC;EACH;EAEA,YAAY,KAAK,OAAO,SAAS;GAC/B,MAAM,OAAO,WAAW,IAAI,UAAU;GACtC,IAAI,CAAC,QAAQ,MAAM,WAAW,KAAK;IACjC,OAAO,EAAE,OAAO,MAAM;GACxB;GACA,MAAM,EAAE,OAAO,UAAU,OAAO,KAAK,IAAI;;;GAIzC,MAAM,WACJ,MAAM,iBACL,MAAM,iBAAiB,YAAY,EAAE,cAAc,MAAM,aAAa,IAAI,CAAC;GAC9E,MAAM,eAAe,sBAAsB,QAAQ;GACnD,MAAM,OAAO,WAAW,KAAK,OAAO,YAAY;GAChD,OAAO,MAAM,+BAA+B;IAC1C,YAAY,IAAI;IAChB;IACA;IACA,cAAc;IACd,iBAAiB,KAAK;IACtB,QAAQ,KAAK;IACb,WAAW,KAAK;IAChB,YAAY,KAAK;GACnB,CAAC;GAED,IAAI,KAAK,WAAW,SAAS;IAC3B,OAAO,KAAK,sBAAsB;KAChC,YAAY,IAAI;KAChB;KACA,cAAc,SAAS;IACzB,CAAC;IACD,OAAO,EAAE,OAAO,MAAM;GACxB;GAEA,IAAI,WAAW,KAAK,YAAY;IAC9B,OAAO,MAAM,+BAA+B;KAC1C,YAAY,IAAI;KAChB;KACA;KACA,YAAY,KAAK;IACnB,CAAC;IACD,OAAO,MAAM,oBAAoB;KAAE,YAAY,IAAI;KAAY;KAAO,UAAU;IAAQ,CAAC;IACzF,OAAO,EAAE,OAAO,MAAM;GACxB;GAEA,MAAM,YAAY,eAAe,UAAU,IAAI,CAAC;GAChD,MAAM,SAAS,UAAU,WAAW,KAAK,SAAS;GAClD,OAAO,MAAM,kCAAkC;IAC7C,YAAY,IAAI;IAChB;IACA,SAAS,UAAU;IACnB,cAAc;IACd,eAAe;IACf,WAAW,KAAK;GAClB,CAAC;GACD,OAAO,KAAK,yBAAyB;IACnC,YAAY,IAAI;IAChB;IACA,SAAS,UAAU;IACnB;IACA,cAAc,SAAS;GACzB,CAAC;GACD,OAAO;IAAE,OAAO;IAAM,OAAO;GAAO;EACtC;CACF;AACF;;;;;;;AAQA,SAAS,sBAAsB,UAAiD;CAC9E,IAAI,SAAS,iBAAiB,WAAW;EACvC,OAAO,SAAS;CAClB;CACA,IAAI,SAAS,iBAAiB,WAAW;EACvC,OAAO,KAAK,KAAK,SAAS,eAAe,GAAI;CAC/C;CACA,OAAO;AACT;AAEA,SAAS,aACP,SACA,MACA,OACA,QACA,YACmB;CACnB,IAAI;EACF,OAAO,eAAe,SAAS,KAAK,cAAc,KAAK;CACzD,SAAS,OAAO;EACd,OAAO,MAAM,iCAAiC;GAC5C;GACA,OAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;EAC9D,CAAC;EACD,OAAO,CAAC;CACV;AACF;AAEA,SAAS,eAAe,UAA6B,OAAuB;CAC1E,IAAI,SAAS,cAAc,WAAW;EACpC,OAAO,KAAK,IAAI,GAAG,SAAS,YAAY,KAAK;CAC/C;CACA,IAAI,SAAS,iBAAiB,WAAW;EACvC,OAAO,SAAS;CAClB;CACA,OAAO;AACT;;AAGA,SAAS,UAAU,IAAY,WAA2B;CACxD,MAAM,cAAc,KAAK,IAAI,GAAG,EAAE;CAClC,OAAO,YAAY,IAAI,KAAK,IAAI,aAAa,SAAS,IAAI;AAC5D;;;;;;;;;;;;;;;;;;;;;AAsBA,eAAe,SACb,OACA,OACA,KACA,WACA,QACA,QACA,YACA,OACA,OACe;CACf,MAAM,aAAa,YAAY,IAAI,IAAI,IAAI,YAAY,OAAO;CAC9D,OAAO,MAAM,wBAAwB;EACnC;EACA;EACA;EACA,iBAAiB,MAAM;EACvB;CACF,CAAC;CACD,SAAS;EACP,MAAM,SAAS,KAAK,IAAI,MAAM,iBAAiB,UAAU;EACzD,MAAM,cAAc,SAAS,IAAI;EACjC,IAAI,eAAe,GAAG;GACpB,OAAO,MAAM,sBAAsB;IAAE;IAAY;IAAO;GAAM,CAAC;GAC/D;EACF;EACA,OAAO,MAAM,uBAAuB;GAClC;GACA;GACA;GACA;GACA,aAAa,MAAM,oBAAoB;EACzC,CAAC;;EAGD,IAAI,UAAU,MAAM;EACpB,IACE,WACA,MAAM,2BAA2B,aACjC,MAAM,yBAAyB,IAAI,IAAI,aACvC;GACA,UAAU;EACZ;EACA,IAAI,CAAC,SAAS;GACZ,MAAM,UAAU,MAAM,WAAW,CAAC,CAAC,cAAc;IAC/C,IAAI,MAAM,oBAAoB,SAAS;KACrC,MAAM,kBAAkB;KACxB,MAAM,yBAAyB;IACjC;GACF,CAAC;;;GAGD,IAAI,CAAC,MAAM,iBAAiB;IAC1B,MAAM,kBAAkB;IACxB,MAAM,yBAAyB,IAAI,IAAI;GACzC;GACA,UAAU;EACZ;EAEA,IAAI;GACF,MAAM,cAAc,SAAS,MAAM;EACrC,SAAS,OAAO;GACd,IAAI,aAAa,KAAK,GAAG;IACvB,OAAO,KAAK,0BAA0B;KAAE;KAAY;KAAO;IAAM,CAAC;GACpE;GACA,MAAM;EACR;CAEF;AACF;AAEA,SAAS,cAAc,SAAwB,QAAgD;CAC7F,IAAI,CAAC,QAAQ;EACX,OAAO;CACT;CACA,IAAI,OAAO,SAAS;EAClB,OAAO,QAAQ,OAAO,aAAa,MAAM,CAAC;CAC5C;CACA,OAAO,IAAI,SAAe,SAAS,WAAW;EAC5C,MAAM,gBAAgB,OAAO,aAAa,MAAM,CAAC;EACjD,OAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;EACxD,QAAQ,WACA;GACJ,OAAO,oBAAoB,SAAS,OAAO;GAC3C,QAAQ;EACV,IACC,UAAU;GACT,OAAO,oBAAoB,SAAS,OAAO;GAC3C,OAAO,KAAK;EACd,CACF;CACF,CAAC;AACH;AAEA,SAAS,aAAa,IAAY,QAAqC;CACrE,OAAO,IAAI,SAAe,SAAS,WAAW;EAC5C,IAAI,QAAQ,SAAS;GACnB,OAAO,aAAa,MAAM,CAAC;GAC3B;EACF;EACA,MAAM,QAAQ,iBAAiB;GAC7B,QAAQ,oBAAoB,SAAS,OAAO;GAC5C,QAAQ;EACV,GAAG,EAAE;EACL,MAAM,gBAAgB;GACpB,aAAa,KAAK;GAClB,OAAO,aAAa,MAAqB,CAAC;EAC5C;EACA,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;CAC3D,CAAC;AACH;AAEA,SAAS,aAAa,QAA4B;CAChD,MAAM,SAAS,OAAO;CACtB,IAAI,kBAAkB,OAAO;EAC3B,OAAO;CACT;CACA,OAAO,IAAI,aAAa,6BAA6B,YAAY;AACnE;AAEA,SAAS,aAAa,OAAyB;CAC7C,OAAO,iBAAiB,SAAS,MAAM,SAAS;AAClD","names":[],"sources":["../src/plugin.ts"],"version":3,"file":"plugin.js","sourceRoot":""}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vymalo/opencode-ratelimit",
-  "version": "0.16.1",
+  "version": "0.18.0",
   "description": "OpenCode plugin that makes OpenAI-compatible providers rate-limit aware: it reads Envoy Gateway / IETF draft-03 rate-limit response headers (x-ratelimit-limit/remaining/reset), proactively throttles requests when the quota is exhausted, and backs off and retries on HTTP 429.",
   "license": "MIT",
   "author": "vymalo contributors",
@@ -16,6 +16,7 @@
   "keywords": [
     "opencode",
     "opencode-plugin",
+    "opencode-v2",
     "rate-limit",
     "throttle",
     "429",
@@ -47,11 +48,15 @@
     "access": "public"
   },
   "dependencies": {
-    "@opencode-ai/plugin": "1.15.10"
+    "@opencode/plugin": "2.2.6"
   },
   "devDependencies": {
-    "vite": "^8.2.1",
-    "vitest": "^4.1.7"
+    "effect": "4.1.0",
+    "vite": "^8.3.0",
+    "vitest": "^5.0.1"
+  },
+  "peerDependencies": {
+    "effect": "^4.0.0"
   },
   "scripts": {
     "build": "node ../../scripts/build-package.mjs",