[bug] OpenAI content_filter is silently translated to Anthropic end_turn

[Atharva-Kanherkar]

Symptom

When an Anthropic Messages request is served by an OpenAI-compatible backend, an upstream finish_reason of content_filter is translated to an Anthropic stop_reason of end_turn.

A response blocked or truncated by content filtering is therefore presented as a normal completion. The translation also emits no diagnostic indicating that the stop reason was lost.

Reproduction

Translate an OpenAI Chat response to Anthropic Messages with the following values:

• Response ID: chatcmpl-probe
• Model: gpt-4o
• Assistant content: partial answer before the block
• Finish reason: content_filter
• Prompt tokens: 10
• Completion tokens: 5
• Total tokens: 15

Call translate_response with WireFormat::OpenAiChat as the source, WireFormat::AnthropicMessages as the target, and the default translation policy.

The translated response reports stop_reason: "end_turn".

A control response with finish_reason: "stop" also reports stop_reason: "end_turn", making the filtered response indistinguishable from a normal completion.

Expected vs. actual

• Expected: Translation should not silently make a content-filter stop indistinguishable from a normal completion. It should preserve the distinction, emit a lossy-conversion diagnostic and honor the configured lossy-conversion policy, or translate the event as a failure.
• Actual: The response contains stop_reason: "end_turn" and no translation diagnostic.

Environment

• Switchyard commit SHA: fb3fc30
• Python: python3 --version reports 3.9.6, though this path is Rust-only
• OS / arch: macOS 26.2, Darwin 25.2.0, arm64
• Install path: source build
• Inbound format: Anthropic Messages
• Backend: OpenAI or another OpenAI-compatible upstream

Additional context

The buffered OpenAI decoder preserves content_filter as StopReason::ContentFilter in crates/switchyard-translation/src/codecs/openai_chat/buffered.rs:1133. The OpenAI encoder maps that normalized variant back to content_filter at line 1148.

The loss occurs in the Anthropic buffered encoder at crates/switchyard-translation/src/codecs/anthropic/buffered.rs:986, where StopReason::ContentFilter is mapped to end_turn.

The streaming path has the same behavior. The OpenAI stream decoder copies the upstream finish reason into the normalized stop event at crates/switchyard-translation/src/codecs/openai_chat/stream.rs:142. The Anthropic stream encoder stores that reason at crates/switchyard-translation/src/codecs/anthropic/stream.rs:206, then maps unrecognized values, including content_filter, to end_turn at line 571.

The stream accumulator also recognizes content_filter as StopReason::ContentFilter at crates/protocol/src/stream.rs:432.

Because the Anthropic vocabulary handled by this codec has no direct content-filter equivalent, the wire behavior needs a design decision. Protocol-safe options include:

1. Emit a lossy-conversion diagnostic and honor LossyConversionPolicy.
2. Translate a filtered generation into an error response or event.
3. Preserve the original reason in an explicitly supported extension, if one exists or is introduced.

I can provide the local probe as a regression test once the desired behavior is chosen.