fix: drop thinking blocks without valid signatures instead of sending signature=None

[Dan Shapiro (danshapiro)]

Fixes microsoft/amplifier#330

Root cause

Sessions that run partly on non-Anthropic providers (e.g. OpenAI reasoning turns under a mixed-provider routing matrix) persist canonical ThinkingBlocks with signature: None (or no signature field at all — OpenAI reasoning blocks use encrypted content payloads instead of Anthropic's cryptographic signature). On resume, this provider's history serialization copied the signature through as-is, so the request was sent with signature: null and Anthropic rejected it with HTTP 400:

messages.N.content.0.thinking.signature.str: Input should be a valid string

Because the same persisted history is replayed on every resume, the session is permanently bricked.

Fix

Anthropic requires a valid non-empty signature to accept a thinking block as input — the field cannot simply be omitted — so unsigned thinking blocks are now dropped entirely:

• _clean_content_block(): returns None for a thinking block whose signature is not a non-empty string (valid blocks pass through unchanged).
• _convert_messages() (all three call sites — thinking_block + tool_calls path, thinking_block-only path, structured-content path): skips dropped blocks. If an assistant message loses ALL of its content this way, a minimal placeholder text block ((internal reasoning omitted)) is emitted instead of an empty content array, which Anthropic also rejects.
• _convert_to_chat_response(): no longer persists response thinking blocks without a valid signature (getattr(block, "signature", None) previously flowed None into canonical history), so unsigned blocks cannot re-enter the transcript from this side either.

Evidence / prevalence

• Second reproduction with full diagnosis: Cross-provider resume: thinking blocks with null signature passed to Anthropic API cause invalid_request_error amplifier#330 (comment) (session bricked at messages.137, repaired by stripping 7 unsigned thinking blocks).
• A scan of one machine found 167 of 12,705 transcripts contain signature-less thinking blocks — every one of them bricks on Anthropic resume without this guard.

Tests

New tests/test_thinking_signature_guard.py (10 tests) covering:

• valid string signature → passed through unchanged
• signature: None → dropped, sibling text/tool_use blocks preserved
• missing signature field → dropped
• empty-string signature → dropped
• assistant message where ALL content was unsigned thinking → placeholder emitted, never an empty content array (both thinking_block and structured-content paths)
• _convert_to_chat_response skips unsigned / keeps signed thinking blocks

Full suite: 585 passed (uv run --with amplifier-core pytest). ruff format --check clean; ruff check introduces no new findings vs. main.

---

Generated with Amplifier