#1415 Release v0.83.0

Release v0.83.0

Bumps version from 0.82.0 to 0.83.0. Promotes dev → main.

Changelog

Added

• LLMCompletionTruncatedError and LLMCompletionRefusedError: two subclasses of LLMCompletionError, the first for a text completion cut at its output limit or its context window, the second for one the model declined to write or a safety filter stopped. Both are content errors in the input domain, not retryable, answered with HTTP 422 and caller-facing under STRICT disclosure, and carry model_handle, pipe_code and the provider's stop_reason, a truncation also carrying the limit it hit as truncation_limit, max_tokens or context_window, the output_tokens used and the max_tokens sent; the message names the model, the pipe and those values, never the partial text. Each ends with a change_input next step fitted to what stopped the text: revise the prompt or the input after a refusal, raise the pipe's max_tokens or shorten its input at the output limit, set the pipe's max_tokens up to the model's limit when the request sent none, and shorten the input, lower the reasoning effort or choose a model with a larger window, never a higher max_tokens, at the context window.

Changed

• The Bedrock client's chat returns a BedrockChatResult (Breaking): BedrockClientProtocol.chat returns a named result holding the text, the token usage and Converse's stopReason, in place of a (text, usage) tuple, so a custom Bedrock client returns read_converse_response(response=…) or builds the result itself.
• A lowered Anthropic structured-output limit is said: when a structured generation on an Anthropic model lowers max_tokens to fit its structured-output timeout, the call logs it, at warning level when the pipe set the limit and at debug level when it is the model's default, and an error the lowered limit can explain, a structured output that failed validation or came back incomplete, or a timeout, ends with a sentence saying the limit was lowered and to what, never a rate limit, an authentication failure, an overload, a server error or a lost connection.

Fixed

• A truncated, refused or filtered text no longer passes for a result (Breaking): every LLM worker, OpenAI chat completions whether direct or behind a gateway such as Portkey or OpenRouter, the OpenAI Responses API, Anthropic, Google, Bedrock and Mistral, now reads its response's stop value before it returns a text, so a text that stopped short fails its pipe with LLMCompletionTruncatedError or LLMCompletionRefusedError where it used to come back as a successful output, empty or partial, its billed usage still reported and recorded on its span. Each provider's vocabulary is read, including a gateway's pass-through of another provider's value such as max_tokens or content_filtered, an unknown value being logged at warning level and taken as normal, except that an incomplete Responses API answer is a truncation whatever its reason; a Gemini prompt blocked before any candidate is a refusal, and structured generation keeps its configured re-asks. A Bedrock answer is read for every text block it holds, so one holding none fails with an LLMCompletionError where it used to raise an IndexError when its message was empty, or a KeyError when its first block was not text.

🤖 Generated with Claude Code

---

Summary by cubic

Promotes dev to main as v0.83.0. Every LLM worker now reads its provider's stop signal before returning text, so a completion cut at its output limit or context window, refused by the model, or stopped by a safety filter fails the pipe with LLMCompletionTruncatedError or LLMCompletionRefusedError instead of passing empty or partial text as a successful output. Billed usage is still reported once, and normal stops and configured retries behave as before.

Breaking changes

• BedrockClientProtocol.chat now returns a BedrockChatResult in place of a (text, usage) tuple. Custom Bedrock clients must return read_converse_response(response=…) or build the result themselves.

Also in this release

• Anthropic logs when it lowers max_tokens to fit a structured output's timeout — at warning level when the pipe set the limit, at debug when it is the model's default — and a failure that lowering can explain names the lowered limit.
• Version bumped from 0.82.0 to 0.83.0; changelog, error docs, and OpenAPI spec updated.