fix(translation): report a dropped thinking budget on the way to Responses

[harshitwandhare]

Follow-up to #792, which @afourniernv closed as superseded by #840 and asked for this part as a new PR.

What was broken. Since #840, an Anthropic request with "thinking": {"type": "enabled", "budget_tokens": 10000} translated to OpenAI Responses has no reasoning key. That part is right. But the budget is dropped with no diagnostic. A caller on the default policy gets no lossy_conversion warning, and a caller on LossyConversionPolicy::Reject gets the request translated anyway.

Why. The Anthropic branch of the Responses encoder builds reasoning from scratch and reads only thinking.type. The rest of thinking is discarded without going through push_lossy, which is how this codec reports the other content it cannot map, such as image and file blocks.

What the fix does. When the Anthropic thinking object has budget_tokens, the encoder calls push_lossy with a message that names it. The translated body does not change. adaptive and disabled report nothing, since #840 maps both onto effort.

I left thinking.display out. The request in #239 sends "display": "omitted" with adaptive thinking, so reporting it would make those requests fail under Reject. I can add it if you want it reported too.

How it was verified. I applied only the new tests to main at 64def564 first. anthropic_thinking_budget_drop_to_responses_is_reported fails there on the diagnostics count (left 0, right 1). The other new test passes on both, as a guard against reporting adaptive or disabled. With the fix, both pass, and #840's anthropic_thinking_to_responses_uses_normalized_effort passes unchanged.

Gates, run in WSL (Ubuntu 24.04, Python 3.12):

Gate	main at 64def564	This branch
cargo test --workspace --locked	875 passed, 0 failed (37 binaries)	877 passed, 0 failed
cargo test -p switchyard-runner --features prefill-router --locked	not run	66 passed, 0 failed
pytest tests/ -m "not integration"	139 passed	139 passed

Re-run 2026-10-11 after rebasing onto fc64565a: cargo test --workspace --locked 962 / 0 over 40 binaries (960 on main plus the two tests here), prefill-router 68 / 0, fmt and both clippy runs clean. pytest was not re-run, since the branch touches no Python.

cargo fmt --all --check, both cargo clippy ... -D warnings runs (with and without prefill-router), ruff check, mypy switchyard, actionlint and the SPDX header scan (226 files, 0 missing) are all clean on the branch.

Summary by CodeRabbit

• Bug Fixes
  • Requests that include an Anthropic thinking token budget now report that the budget is dropped during translation to OpenAI Responses.
  • When lossy conversions are set to be rejected, these requests now fail rather than proceeding without the budget.
  • Adaptive and disabled thinking settings without a token budget continue to translate without loss warnings, including under strict conversion settings.

[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: 6e40d0c5-49fc-4902-b37f-c133f3257d6e

---

📥 Commits

Reviewing files that changed from the base of the PR and between 64def56 and 6e8a83a.

---

📒 Files selected for processing (2)

• crates/switchyard-translation/src/codecs/responses/buffered.rs
• crates/switchyard-translation/tests/request_translation.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

Responses encoding now reports a lossy conversion when an Anthropic request includes thinking.budget_tokens. Tests check diagnostic reporting, rejection under the strict loss policy, and adaptive or disabled thinking controls without diagnostics.

Changes

Thinking budget handling

Layer / File(s)	Summary
Budget loss reporting and policy tests crates/switchyard-translation/src/codecs/responses/buffered.rs, crates/switchyard-translation/tests/request_translation.rs	Responses encoding reports when it drops budget_tokens. Tests verify the diagnostic, rejection under LossyConversionPolicy::Reject, and diagnostic-free translation for adaptive and disabled thinking controls.

Priority: ⬇️ Low

Estimated code review effort: 2 (Simple) | ~10 minutes

Merge Risk: ⚪ Minimal · up to 6e8a8

Responses translations now report or reject dropped Anthropic thinking budgets, while the tested budgetless adaptive and disabled controls remain quiet. The inspected changes show no concrete merge-blocking risk.

Pre-merge checks | 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage		Docstring coverage is 80.00% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 5 functions across 2 files.
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.
Description Check		Check skipped - CodeRabbit’s high-level summary is enabled.
Title check		The title clearly and concisely describes the main change: reporting a dropped Anthropic thinking budget during translation to OpenAI Responses.

---

---

_{A rabbit checks the budget line
Then marks the loss so all can see
Strict rules can stop the translation
Adaptive thoughts pass quietly
The rabbit hops through tests with glee}

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

[harshitwandhare]

Rebased onto 16cbe593. #892 renamed the _policy parameter of OpenAiResponsesCodec::encode_request to policy, so the old head still merged without a conflict but no longer compiled against main. That identifier is the only change.

On the new head f91516d7: cargo test --workspace --locked 905 passed / 0 failed against 903 / 0 on main (the +2 are this PR's tests), prefill-router 66 / 0, fmt and clippy -D warnings clean. The budget test still fails with main's encoder and passes here.

[harshitwandhare]

Rebased onto fc64565a, it was 24 behind. No conflicts and no code change. Gates at 67f8cf64: workspace 962 / 0 (960 on main plus the two tests here), prefill-router 68 / 0, fmt and clippy clean.