v0.17.0

Added

• replay_thinking provider setting, for Claude behind LiteLLM. The OpenAI and OpenAI-compatible clients now keep the signatures LiteLLM streams in delta.thinking_blocks (and redacted thinking), so saved replies hold signed Thinking and RedactedThinking blocks as the Anthropic client's do. With replay_thinking = true on a provider (ProviderConfig::with_replay_thinking, AgentConfig::with_replay_thinking, OpenAiClient::with_replay_thinking, OpenAiCompatibleClient::with_replay_thinking in code), the replies of the current tool loop send that thinking back as thinking_blocks. Anthropic needs this to keep extended thinking on through tool calls, and LiteLLM's modify_params otherwise turns thinking off for those requests. Off by default. Unsigned thinking and older turns' thinking are never sent. See replay_thinking in docs/configuration.md.
• Error::from_response(response) builds the Error::HttpError for a non-2xx reqwest::Response: its status, body and Retry-After. Custom LlmClients can use it so their errors get the same retry handling as the built-in clients.
• HTTP errors are structured over ACP. A provider's HTTP error used to reach ACP clients as the string "HTTP 429: {...}". It's now an internal_error whose data is { "kind": "http_error", "status", "retry_after_secs", "body" }, with body parsed if it's JSON, so a client that knows its provider's error format can read it (see docs/api.md).

Changed

• Breaking: Error::HttpError has a retry_after: Option<Duration> field, the response's Retry-After header. Code that builds the variant needs to set it (or use Error::from_response), and patterns need .. or the field. Error::retry_after() returns it.
• Retry-After is honoured. On a 429 or 5xx, the LLM and embedding clients wait the provider's Retry-After instead of their own backoff. A Retry-After over 10 seconds isn't retried, and the error is returned at once. A limit that lifts in hours used to be retried three times over about 7 s before the user saw it.
• HTTP 529 is retried. Anthropic answers 529 (overloaded_error) when it's temporarily overloaded. The request failed at once; it's now retried like a 503.