fix: preserve Anthropic thinking block signature for multi-turn conversations

[wnxd]

Fixes #217

Problem

When using Anthropic models with thinking enabled, follow-up requests fail with thinking.signature: Field required because the extension discards signature_delta from streaming responses.

Solution

Uses a static cache keyed by the assistant text content (which VS Code always preserves) to store thinking+signature pairs.

• During streaming: accumulate assistant text + thinking content + signature
• On stream end: cache {thinking, signature}[] keyed by assistant text
• In convertMessages(): look up cached thinking blocks by assistant text
• If no cache hit: skip thinking block entirely (avoiding 400 error)

Files Changed

• src/anthropic/anthropicApi.ts (+85 lines)
• src/commonApi.ts (+13 lines)

Testing

Tested with Claude models using thinking: {type: adaptive} — multi-turn conversations now work correctly.

[wnxd]

Thanks for the thorough review! All three points have been addressed in commit 5c5e8e1:

1. _pendingThinkingEntries not reset — Fixed. Added this._pendingThinkingEntries = undefined; right after the cache write, and captured entryCount beforehand for the logger.

2. Dead code branch (else if (thinkingSignatures.some(...))) — Removed. Since LanguageModelThinkingPart has no signature field, the thinkingSignatures array was always empty, making this branch unreachable.

3. thinkingSignatures always empty — Removed the variable entirely, along with thinkingParts and joinedThinking which were only consumed by the deleted branch.

[JohnnyZ93]

There are two issues to consider:

1. If some API providers do not return the signature, then the thinking part content will be lost during message conversion.
2. The cache uses _accumulatedAssistantText as the key. If one request does not return text, only thinking and tool_use, will it become invalid?