fix(claudeHelper): preserve latest assistant thinking blocks verbatim

[NomenAK]

What

Preserve the LATEST assistant message's thinking blocks verbatim instead of rewriting them to redacted_thinking. Apply the same invariant to non-Anthropic Claude-shape upstreams (kimi-coding, glmt, zai…) where the latest message's plain thinking.thinking text must be sent as-is.

Why

Anthropic's 2026 thinking-mode contract distinguishes:

• Older assistant messages: can be replayed as redacted_thinking { data } — Anthropic accepts the synthetic blob.
• Latest assistant message: must be sent verbatim (same type, signature, text). Modifying it triggers HTTP 400 "thinking or redacted_thinking blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.".

On our fork, we observed 49 occurrences/hour of this error on claude-opus-4-7 with conversational replay before this fix.

Distinction from already-merged 5c11b575: that commit added the data field on synthetic redacted_thinking (preventing messages.N.content.M.redacted_thinking.data: Field required 400s). It is a different invariant — about the field shape of rewritten blocks. This PR is about which assistant message is rewriteable at all: the index-based "skip rewrite on the latest assistant" guard. The two fixes are complementary and both required.

How

• Compute latestAssistantIndex once before the message loop in prepareClaudeRequest:

  let latestAssistantIndex = -1;
  for (let i = body.messages.length - 1; i >= 0; i--) {
    if (body.messages[i]?.role === "assistant") { latestAssistantIndex = i; break; }
  }

• Add a guard inside the rewrite loop:

  if (messageIndex === latestAssistantIndex && supportsRedactedThinking) {
    continue;  // never modify the latest Anthropic assistant message
  }

• Symmetric guard for non-Anthropic Claude-shape: preserve the latest message's plain thinking.thinking text when it's non-empty (only fall through to reasoningCache/placeholder for older messages or empty-text latest)
• 2 new tests in tests/unit/translator-claude-helper-thinking.test.ts covering Anthropic preservation and non-Anthropic symmetry; updates to 2 existing tests that previously encoded the all-rewrites behavior
• Test count: baseline 17 → 19 (+2 new)

Notes

• Live validation on our fork after this fix: error rate dropped from 49/h → 0/10m (no positive STREAM traffic during the validation window, but the error pattern was eliminated)
• The non-Anthropic symmetry is conservative: we still strip signature/data field cleanup and normalize type to "thinking" — only the text content is preserved verbatim on the latest message
• Happy to revise the test count expectation in CI if the coverage gate is strict; the pre-existing functions-coverage gap (59.26%) is not introduced by this commit

[gemini-code-assist]

Code Review

This pull request introduces provider-aware handling for Claude's thinking blocks, specifically addressing compatibility issues with non-Anthropic providers (like kimi-coding or glmt) that do not support redacted thinking blobs. It implements a reasoning cache lookup to restore original thinking text for these providers and adds a guard to preserve thinking blocks verbatim in the latest assistant message to comply with Anthropic's API requirements. Feedback focuses on improving maintainability by exporting the new placeholder constant from the helper module instead of duplicating it in test files.

[gemini-code-assist]

To improve maintainability and avoid duplication across the codebase, consider exporting this constant so it can be reused directly in the unit tests.

Suggested change

	const NON_ANTHROPIC_THINKING_PLACEHOLDER = "(prior reasoning summary unavailable)";
	export const NON_ANTHROPIC_THINKING_PLACEHOLDER = "(prior reasoning summary unavailable)";

[NomenAK]

Done in ddf1ba2 — exported the constant from claudeHelper.ts (now export const NON_ANTHROPIC_THINKING_PLACEHOLDER). Thanks.

[gemini-code-assist]

Instead of hardcoding the placeholder string again, you can import it from the claudeHelper module (assuming it is exported as suggested). This ensures that the tests stay in sync with the implementation and reduces the risk of regression if the placeholder text changes.

const { prepareClaudeRequest, NON_ANTHROPIC_THINKING_PLACEHOLDER: PLACEHOLDER } = await import("../../open-sse/translator/helpers/claudeHelper.ts");
const { DEFAULT_THINKING_CLAUDE_SIGNATURE } = 
  await import("../../open-sse/config/defaultThinkingSignature.ts");
const reasoningCache = await import("../../open-sse/services/reasoningCache.ts");

[NomenAK]

Done in ddf1ba2 — test now imports it via { NON_ANTHROPIC_THINKING_PLACEHOLDER: PLACEHOLDER } from the destructure, dropped the duplicated literal. Thanks.

[diegosouzapw]

Hey @NomenAK! Thanks for the rigorous claudeHelper rework.

Heads-up: this PR significantly overlaps with two PRs already merged into release/v3.8.0:

• d995713e (PR fix(translator): inject thinking placeholder for all Claude-shape upstreams #2161) — inject thinking placeholder for all Claude-shape upstreams
• 5c11b575 (PR fix(claudeHelper): emit data field on redacted_thinking, drop bogus signature #2191) — emit data field on redacted_thinking, drop bogus signature

Your approach (per-provider switch between redacted_thinking vs plain thinking) is structurally different from the upstream approach (provider-agnostic placeholder + data field). Mergeable shows UNKNOWN — likely heavy conflicts.

Could you rebase against release/v3.8.0 and identify the minimum delta that's still needed after the upstream work? Most likely:

• The "preserve latest assistant thinking blocks verbatim" logic (Anthropic 400 on modification)
• Any kimi-coding-specific shape that's still different from generic Claude-shape

A smaller, focused follow-up PR would be much easier to merge than the current full-rewrite. Leaving this open.

[NomenAK]

Thanks for the catch — already addressed in ddf1ba21b3 (pushed a couple of minutes after the review landed, so the threads weren't auto-resolved):

• NON_ANTHROPIC_THINKING_PLACEHOLDER exported from open-sse/translator/helpers/claudeHelper.ts:17
• Consumed via destructured import in tests/unit/translator-claude-helper-thinking.test.ts:229 and tests/unit/translator-helper-branches.test.ts

Happy to consolidate further if it's worth a follow-up commit.