fix(anthropic): drop thinking blocks with empty thinking text, not just missing signature

[shoemoney]

TLDR

Problem this solves:

• A thinking block with a valid-looking signature but empty thinking text is not dropped
• Anthropic rejects it with a 400 even though the signature is present
• This crashes completion()/acompletion() for anthropic/* and vertex_ai/claude-*

How it solves it:

• _is_unsignable_thinking_block() now also returns True when thinking is missing, not a string, or empty after .strip()
• Signature check runs first and is unchanged; redacted_thinking blocks are untouched

The sequential-mode branch of anthropic_messages_pt() already guards this case with len(thinking_block) > 0 and the comment "don't pass empty text blocks. anthropic api raises errors." But _drop_unsignable_thinking_blocks(), the filter used to build thinking_blocks earlier in the same function and the only guard on the anthropic_messages_pt() -> AnthropicConfig.transform_request() path, calls _is_unsignable_thinking_block() alone. The primary path was left unguarded. #36033 (Responses streaming adapter) and #27850 (Bedrock Converse) fixed the same class of bug elsewhere; this is the same fix on _is_unsignable_thinking_block().

User Flow

Before: a developer whose app replays assistant history through anthropic/claude-* or vertex_ai/claude-* gets a 400 whenever that history contains a thinking_blocks entry with an empty thinking field (for example, forwarded from a non-Anthropic reasoning provider's turn that had no summary text)

1. Client sends a multi-turn completion() call where messages[1]["thinking_blocks"] = [{"type": "thinking", "thinking": "", "signature": "sig_abc123"}]
2. The block is kept because it has a non-empty signature; only the signature was ever checked
3. The outbound Anthropic request contains {"type": "thinking", "thinking": "", "signature": "sig_abc123"}
4. Anthropic returns 400 messages.N.content.M.thinking: each thinking block must contain thinking and the call raises

After: the same history no longer reaches Anthropic with an empty thinking block

1. Same client call, same history
2. The block is dropped because thinking is empty, independent of the signature
3. The outbound request has no thinking block for that turn
4. The call proceeds instead of raising a 400

Relevant issues

None filed. Found by code reading in prompt_templates/factory.py and reproduced with a unit test (below).

Pre-Submission checklist

• I have added meaningful tests
• The handful of test files covering my change pass locally: uv run pytest tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py -v, 102 passed (97 pre-existing + 5 new)
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

This is a pure request-transform bug, so the proof is a unit test calling anthropic_messages_pt() directly with the malformed history shape and asserting the outbound content list. No live API key and no mocks; the function does no I/O.

Before (f005afa)

Empty-but-signed thinking block

1. anthropic_messages_pt() called with messages[1]["thinking_blocks"] = [{"type": "thinking", "thinking": "", "signature": "sig_abc123_looks_valid"}]
2. pytest -k test_anthropic_messages_pt_drops_empty_but_signed_thinking_block -> FAILED: assert 'thinking' not in ['thinking', 'text'] (the empty block was kept)

Whitespace-only thinking text

1. _is_unsignable_thinking_block({"type": "thinking", "thinking": " \n\t ", "signature": "sig_abc123_looks_valid"})
2. pytest -k test_is_unsignable_thinking_block_treats_whitespace_only_as_empty -> FAILED: assert False is True (the function said the block was signable)

After (9fa8d06)

Empty-but-signed thinking block

1. Same call as above
2. pytest -k test_anthropic_messages_pt_drops_empty_but_signed_thinking_block -> PASSED, thinking type is absent from the outbound content list

Whitespace-only thinking text

1. Same call as above
2. pytest -k test_is_unsignable_thinking_block_treats_whitespace_only_as_empty -> PASSED, function returns True

Full file: uv run pytest tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py -v -> 102 passed, 0 failed

Also ran ruff format --check and ruff check on the two changed files. Did not run the full make test-unit suite (left to CI per CONTRIBUTING.md) or any live Anthropic/Vertex calls.

Type

Bug Fix
Test

Caveats (if any)

Low

• The emptiness check uses .strip(), so whitespace-only text counts as empty. The sibling len(thinking_block) > 0 check is always combined with and not _is_unsignable_thinking_block(m), so the two agree in practice; this one is the stricter of the two because _drop_unsignable_thinking_blocks() calls it standalone.
• Related, not a duplicate: open PR fix(utils): keep nested thinking when dropping top-level thinking #37624 fixes a 400 with the same error string via apply_additional_drop_params(). It does not touch this function.

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

The PR extends Anthropic request shaping to discard signed thinking blocks whose text is missing, non-string, empty, or whitespace-only, preventing malformed replayed history from reaching Anthropic-compatible endpoints.

• Preserves non-empty signed and redacted thinking blocks.
• Adds focused regression coverage for empty, whitespace-only, unsigned, valid signed, and redacted blocks.
• The functional change appears sound, with only excessive explanatory comments requiring cleanup.

Confidence Score: 4/5

The PR appears safe to merge after the non-blocking excess-commentary issue is cleaned up.

The stricter predicate consistently drops malformed empty thinking blocks while preserving valid signed and redacted blocks, and the added tests exercise the relevant request transformation without network access.

Files Needing Attention: litellm/litellm_core_utils/prompt_templates/factory.py; tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/prompt_templates/factory.py	Correctly strengthens thinking-block validation, but adds substantially more explanatory commentary than the straightforward logic needs.
tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py	Adds meaningful regression cases in the mapped test file, though the tests contain unnecessarily extensive docstrings.

_{Reviews (1): Last reviewed commit: "fix(anthropic): drop thinking blocks wit..." | Re-trigger Greptile}

[greptile-apps]

Excessive predicate commentary

The extended request-history explanation around this straightforward signature-and-text predicate, along with similarly extensive test docstrings, duplicates implementation rationale and makes the behavior harder to scan and maintain. Keep the commentary focused on the non-obvious provider constraint.

Context Used: CLAUDE.md (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing shoemoney:fix/drop-empty-thinking-blocks (dfba491) with litellm_internal_staging (e5da593)}

[shoemoney]

code-quality is failing repo-wide right now, not on this diff. This branch touches only litellm/litellm_core_utils/prompt_templates/factory.py and its test file, no workflow files.

The job exits on a workflow startup invariant check:

ERROR: Workflow startup invariants violated:
  - .github/workflows/test-unit.yml: job `unit` gives pytest 20m but caps the job at 55m.
    Setup can use up to 35m plus 5m of runner overhead, so the job deadline would preempt
    pytest; raise job-timeout-minutes to at least 60.

Same failure on the latest runs for #38055, #38053 and #38051. Leaving it to whoever owns that workflow config rather than touching it from here.

[shoemoney]

CI note: code-quality / check_workflow_startup_safety failure is pre-existing on main and unrelated to this PR's changed files. All 3 shards cap timeout at 55m but need 60m (35m setup + 20m pytest + 5m overhead) on both base f005afa and head; staging already fixed to 60m and passes locally, and this PR only touches factory.py. No code fix required from this PR; rebase/label will clear it.