fix(protocol): normalize Anthropic's refusal stop reason when streaming

[v0ropaev]

stop_reason_from_str in crates/protocol/src/stream.rs:505 is what the stream accumulator uses to fold a streamed response into a buffered one. It knows content_filter but not refusal, which is the spelling Anthropic sends, and its own doc comment says it covers "the common OpenAI and Anthropic spellings".

The buffered Anthropic codec does map it, at codecs/anthropic/buffered.rs:1284:

Some("refusal") => StopReason::ContentFilter,

so the same provider value normalizes differently depending on how the response arrived. Folding a single MessageStop through ResponseAccumulator:

refusal        -> Some(Unknown)
content_filter -> Some(ContentFilter)

Unknown is not inert from there. anthropic_stop_reason maps it to "end_turn" (buffered.rs:1297), openai_finish_reason to "stop" (openai_chat/buffered.rs:1270), and observability.rs:169 labels it unknown, so the gen_ai.response.finish_reasons attribute loses it too. A buffered-from-stream Anthropic turn therefore reaches the client as a normal finish with stop_reason: "end_turn", and nothing in the telemetry says a refusal happened.

Where that path is taken: algorithms/util/buffered_response.rs:60 builds BufferedResponse.agg from this accumulator, and libsy-llm-client/src/run.rs:823 uses it for remember_canonical_response.

One arm. The test covers the whole vocabulary the two codecs share rather than just the new value, since the asymmetry is the actual defect:

length, max_tokens      -> MaxTokens
tool_calls, tool_use    -> ToolUse
content_filter, refusal -> ContentFilter
stop, end_turn          -> EndTurn

On main that test fails on refusal with assertion left == right failed: stop reason "refusal"; the other seven pass either way.

cargo test --workspace 924 passed, 0 failed. cargo fmt --all --check and cargo clippy -p switchyard-protocol --all-targets -- -D warnings clean on toolchain 1.99.

Two notes for review.

#909 edits this same match arm, adding Bedrock's content_filtered and guardrail_intervened and model_context_window_exceeded, and does not add refusal. So the two will conflict textually on this line and both are still needed; whichever lands first, the other is a one-line rebase.

What this does not cover: a streamed refusal also loses stop_details, which the buffered codec emits alongside refusal. That is a separate hole in the accumulator, which carries no field for it, and it needs a decision about where to keep it rather than a one-line map, so I left it out. Say the word if you want it in the same PR.

Summary by CodeRabbit

• Bug Fixes
  • Anthropic’s refusal stop reason is now recognized as a content-filter stop reason.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Repository: NVIDIA-NeMo/Switchyard/.coderabbit.yaml
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 1ebadb96-8b2e-4031-8f47-9166a590e3d9

📥 Commits

Reviewing files that changed from the base of the PR and between a3cdc51 and 1c9c8ae.

📒 Files selected for processing (1)

• crates/protocol/src/stream.rs

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

---

Walkthrough

The stream parser now maps refusal to StopReason::ContentFilter, alongside content_filter. A test checks normalization for the listed stop-reason spellings.

Changes

Stop-reason normalization

Layer / File(s)	Summary
Map and test stop-reason spellings crates/protocol/src/stream.rs	stop_reason_from_str maps refusal to StopReason::ContentFilter. A test checks normalization for the listed stop-reason spellings.

Priority: ➖ Normal

Estimated code review effort: 2 (Simple) | ~5 minutes

Merge Risk: ⚪ Minimal · up to 1c9c8

Streamed refusals now retain the same normalized finish reason as buffered responses. No material merge-blocking risk was identified.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: normalizing Anthropic's streamed refusal stop reason in the protocol.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 4 functions across 1 files.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

_{I sniff the stream and find a name,
“refusal” joins the filtered frame.
The tests now check each spelling right,
I hop through values, neat and bright.
Then tuck my ears and end the night.}

_{Comment @coderabbitai help to get the list of available commands.}