@plumbus/ai-bedrock 0.3.0-beta.0 → 0.3.0-beta.1

package/CHANGELOG.md

@@ -1,11 +1,34 @@
 # @plumbus/ai-bedrock
 
+## 0.3.0-beta.1 — 2026-10-12
+
+### Added
+
+- Carry Converse/ConverseStream prompt caching from stable 0.2.2 into the beta family, preserving native structured-output schemas and per-call deadlines. Use with core 0.8.0-beta.10.
+
+### Fixed
+
+- **A request its timeout ends is reported as a timeout.** When `requestTimeout`, or the request's own `timeout` (a prompt's `model.timeoutMs` or a call's `timeoutMs` in core), ended a Converse call, it failed with the AWS SDK's bare `AbortError: Request aborted`, the same as a caller's cancellation. It now fails with core's `ProviderAPIError` shape (`name: 'ProviderAPIError'`, `providerName: 'bedrock'`, `retryable: false`, `timeoutMs`), message `Bedrock request timed out after <ms> ms`, which core's `isProviderAPIError` recognises; a stream ends with an `error` event carrying that message, also when the timeout cuts it off after some deltas — whether the SDK's event stream then breaks or just ends, which could otherwise be reported as some other error or as a finished answer. A caller's abort is still rethrown as raised.
+- **Embeddings honour a per-call timeout.** `embed()` gave every InvokeModel call the adapter's `requestTimeout`, ignoring the request, and a call it ended failed with a bare `AbortError`. `EmbeddingRequest.timeout` (core's RAG `ingest`/`retrieve` and `ctx.ai.retrieve` `timeoutMs`) now sets it, per InvokeModel call, and a call the timeout ends fails with the same `ProviderAPIError` shape as Converse (`Bedrock request timed out after <ms> ms`, `timeoutMs`), which core's `isProviderTimeoutError` recognises.
+- **An unusable timeout is refused before anything is sent.** A `requestTimeout` or request `timeout` that was not a whole number of milliseconds from 1 to 2 147 483 647 went straight to `AbortSignal.timeout`, which threw a bare `RangeError` for `-1`, `1.5` or `NaN` (an `AI_BEDROCK_REQUEST_TIMEOUT` that is not a number), ended the request at once for `0`, and fired after 1 ms above the maximum. `complete()`, `stream()` and `embed()` now fail with core's `AIInvalidRequestError` (`reason: 'timeout_invalid'`) before loading pricing or calling Bedrock, as core's OpenAI and Anthropic adapters do.
+- **Embeddings stop when the caller cancels.** `embed()` ignored the request's signal (core had none on `EmbeddingRequest`), so a cancelled `ctx.ai.retrieve` or RAG `ingest` kept every InvokeModel call running and sent the remaining texts. `embed()` now honours `EmbeddingRequest.signal` (core's new field): the calls in flight are aborted and fail as `complete()` does on a caller abort (the AWS SDK's `AbortError`, the caller's reason as its `cause`), no further text is sent once it fires between texts (rejecting with the caller's reason as raised), and an already-aborted signal rejects with its reason before pricing is loaded or Bedrock is called. Needs the `@plumbus/core` release that adds `EmbeddingRequest.signal`.
+
 ## 0.3.0-beta.0 — 2026-09-11 — core 0.8 beta family
 
 ### Upgrade boundary
 
 - Beta prerelease of the coordinated core 0.8 family (core 0.8.x, UI 0.9.x, MCP 0.7.x, voice 0.6.x, other add-ons 0.3.x), published under the branch-named npm dist-tag `plumbus-next` (`latest` stays on the 0.7 family). Previous caret ranges exclude it; install the whole family together and follow the [0.8 upgrade notes](../../docs/upgrading-core-0.8.md). Internal peers use prerelease-inclusive ranges (for example `>=0.8.0-beta.0 <0.9.0`) until the family goes stable.
 
+## 0.2.2 — 2026-10-08
+
+### Added
+
+- Prompt caching on Converse / ConverseStream (#63). When a request sets `cache` (core 0.7.9+ passes it from `ctx.ai`, `runToolLoop`, and `AI_PROMPT_CACHE`), the adapter appends `cachePoint` after the system text, at the end of `toolConfig.tools`, and on the last message when requested. Only Claude model ids get marks, with or without an inference-profile prefix (`us.`, `jp.`, `us-gov.`, …); a model ARN gets none. Claude 3 Haiku, Claude 3 Sonnet, Claude 3 Opus, and Claude 3.5 Sonnet `20240620` get no marks because Bedrock rejects `cachePoint` for them with a ValidationException, and neither does Claude 3.5 Sonnet v2 (`20241022`), whose explicit caching AWS lists as Preview; other model families get none either, so those requests still succeed.
+
+### Compatibility
+
+- Loads and runs on any core 0.7.x. The `cache` option is resolved inside this package, which imports nothing from `@plumbus/core` that 0.7.0 lacks. On cores before 0.7.9, `ctx.ai` never sets `cache`, so requests are unchanged. The peer range stays `0.7.x`.
+
 ## 0.2.1 — 2026-09-10
 
 ### Fixed

package/README.md

@@ -46,7 +46,7 @@ An opt-in peer keeps core lean while giving AWS shops a first-class `provider: '
 
 ## Status
 
-Optional peer of `@plumbus/core` (version-locked **`0.2.x`**; required peer `@plumbus/core` **`0.8.x`**). **Runtime floor:** `@plumbus/core` **≥ 0.6.16** (Bedrock provider slot, env discovery, adapter-supplied `cost`, agent wiring v13). Install alone is not enough until the adapter is registered (env discovery or `createBedrockAdapter`).
+Optional peer of `@plumbus/core` (beta family version **`0.3.0-beta.1`**, npm dist-tag `plumbus-next`; required peer `@plumbus/core` **`>=0.8.0-beta.0 <0.9.0`**). The beta includes prompt caching through `ctx.ai`; on the stable 0.7 line caching needs core **≥ 0.7.9** and Bedrock **≥ 0.2.2**. Install alone is not enough until the adapter is registered (env discovery or `createBedrockAdapter`).
 
 ## Install
 
@@ -129,11 +129,12 @@ Full curl, normalize script, ConfigMap sketch: **[instructions/pricing.md](./ins
 - **Model access** — Converse fails with `AccessDenied` until the account enables the model (and Anthropic use-case forms where required).
 - **Embeddings** — Converse does not embed; use InvokeModel (Titan) via `createRAGPipeline({ provider })` + `ctx.ai.retrieve` / `ragPipeline.ingest`. There is no `ctx.ai.embed`. `plumbus rag ingest` does not auto-select Bedrock. Mantle typically has no embeddings.
 - **Tools** — use `runToolLoop` / `generateWithUsage({ tools })`; do not call the Bedrock SDK. Stream accumulates partial tool JSON. Parallel tool results are coalesced into the single user turn Converse requires.
+- **Prompt caching** — set `cache` on `ctx.ai.*` / `runToolLoop`, or `AI_PROMPT_CACHE=true` (core ≥ 0.7.9). Claude models get Converse `cachePoint` marks, except Claude 3 Haiku / Sonnet / Opus and Claude 3.5 Sonnet `20240620`, which reject them, and Claude 3.5 Sonnet v2 (Preview-only on Bedrock); other families get none.
 - **`toolChoice: 'none'`** — Bedrock has no true none; Plumbus omits `toolConfig` entirely. `'auto'` omits just the `toolChoice` field (model-dependent support).
 - **Missing pricing file** — file mode throws on first call; auto-download failure warns, leaves cost unknown, and backs off 5 minutes before retrying.
 - **Unmapped / unkeyed models** — inference works; **no `cost` field** is set until the pricing file has an explicit family-key row (auto-download is best-effort, not a complete AWS catalog). Core then falls back to its own catalog rather than recording the call as free.
 - **IAM** — Converse itself requires `bedrock:InvokeModel`; do not scope that action to embedding models only.
-- **Wiring** — after install, `plumbus init --patch` (core ≥ 0.6.16 / wiring v13) so agents see `instructions/`.
+- **Wiring** — after install, `plumbus init --patch` (wiring v13+, any core 0.7.x) so agents see `instructions/`.
 
 ## Documentation / Agent recipes