@axlsdk/axl 0.22.1 → 0.22.3

package/README.md

@@ -67,7 +67,7 @@ agent({ model: 'deepseek:deepseek-reasoner' });           // DEEPSEEK_API_KEY
 agent({ model: 'ollama:llama3' });                        // local — no key, $0
 ```
 
-Presets: `openrouter`, `azure`, `xai`, `deepseek`, `mistral`, `groq`, `bedrock`, and self-hosted `ollama` / `vllm` / `lmstudio` / `llamacpp` / `sglang`. The unified `effort` knob and per-call cost tracking work across them. OpenRouter passes image input through to the selected model/route, but Axl does not claim every catalog model supports images, tools, or transcription. Build your own by cloning a `ProviderProfile`; see [provider details](../../docs/providers.md#openai-compatible-providers--presets) and [multimodal input](../../docs/multimodal-input.md).
+Presets: `openrouter`, `azure`, `xai`, `deepseek`, `mistral`, `groq`, `bedrock`, and self-hosted `ollama` / `vllm` / `lmstudio` / `llamacpp` / `sglang`. The unified `effort` knob and per-call cost tracking work across them. OpenAI Responses, Anthropic, Google, and OpenRouter accept image input without hard-coding a model catalog; the selected upstream model remains authoritative. Completed-file transcription is a separate, narrower contract. Build your own compatible provider by cloning a `ProviderProfile`; see [provider details](../../docs/providers.md) and [multimodal input](../../docs/multimodal-input.md).
 
 Each provider also accepts an opt-in `rateLimit` (`{ maxConcurrent?, minIntervalMs?, acquireTimeoutMs? }`) for proactive client-side pacing on top of the automatic 429/503/529 backoff — useful when a large fan-out (e.g. an eval) shares one API key. It caps in-flight request concurrency (not token throughput) for that provider's chat calls. See [Providers → Rate limiting](../../docs/providers.md#rate-limiting-opt-in).
 
@@ -604,7 +604,7 @@ const safe = agent({
 });
 ```
 
-When `onBlock` is `'retry'`, the LLM's blocked output is appended to the conversation (as an assistant message) along with a system message containing the block reason, then the LLM is re-called so it can self-correct. These messages **accumulate** across retries — if the guardrail blocks multiple times, the LLM sees all prior failed attempts and corrections before its next try. All retry messages are ephemeral — they are **not** persisted to session history, so subsequent session turns never see the blocked attempts. Schema retries and validate retries use the same accumulating pattern. Input guardrails always throw since the prompt is user-supplied. Throws `GuardrailError` if retries are exhausted or `onBlock` is `'throw'`.
+When `onBlock` is `'retry'`, the LLM's blocked output is appended to the conversation (as an assistant message) along with a user message containing the block reason, then the LLM is re-called so it can self-correct. These messages **accumulate** across retries — if the guardrail blocks multiple times, the LLM sees all prior failed attempts and corrections before its next try. All retry messages are ephemeral — they are **not** persisted to session history, so subsequent session turns never see the blocked attempts. Schema retries and validate retries use the same accumulating pattern. Input guardrails always throw since the prompt is user-supplied. Throws `GuardrailError` if retries are exhausted or `onBlock` is `'throw'`.
 
 For **business rule validation** on the parsed typed object (not raw text), use `validate` on `ctx.ask()`: