fix(deepseek): preserve reasoning_content through full pipeline for DeepSeek V4 models

[kang-heewon]

Summary

Rebased against upstream release/v3.8.0 (bbf3ba0). The 4 core DeepSeek V4 reasoning replay commits from this PR were already cherry-picked by upstream, so only the PR review feedback changes remain.

Remaining Changes (2 files, 4 insertions, 2 deletions)

1. responseSanitizer.ts — Extract inline regex to constant

• Extracted /deepseek[-/]v4/i inline regex to DEEPSEEK_V4_SANITIZER_MODEL_PATTERN constant
• Improves readability and makes the detection intent explicit

2. translator/index.ts — Boundary-aware DeepSeek detection

• Changed normalizedModel.includes("deepseek") → /(^|\/)deepseek/i.test(normalizedModel)
• Prevents false matches on model strings that happen to contain "deepseek" as a substring

Verification

• npm run typecheck:core — clean (1 pre-existing ioredis error unrelated)
• lsp_diagnostics — clean on both changed files
• 1 commit (e8268db) on top of upstream/release/v3.8.0

[gemini-code-assist]

Code Review

This pull request introduces support for DeepSeek V4 models, specifically focusing on their reasoning capabilities. Key changes include updating the provider registry, enhancing the response sanitizer to preserve reasoning content for DeepSeek V4, and extending the reasoning cache to support request-based keys when tool calls are absent. Feedback from the review suggests centralizing the model detection regex patterns to avoid inconsistencies and refining the model identification logic in the translator to prevent overly broad matching that could lead to false positives.

[diegosouzapw]

Hey @kang-heewon! Thanks for the deep work on DeepSeek V4 reasoning preservation.

This PR overlaps significantly with work already merged in release/v3.8.0 (FASE 5-9 consolidation, plus upstream cherry-picks):

• bc73fa24 — xiaomi-mimo reasoning replay
• 9e7cb90a — sanitizer preserves reasoning_content on tool_calls
• 28801557 — responses delta.reasoning_content

The new release/v3.8.0 HEAD has reconciliable changes around sanitizer + reasoning cache that probably overlap with parts of your diff (requireReasoningReplay, reasoning cache key change to requestId+messageIndex).

Could you rebase against the current release/v3.8.0 and:

1. Reconcile with the new cache-key signature
2. Verify requiresReasoningReplay callers are all updated (it changed shape — provider was kept positional in upstream, your PR changes it to an options object)
3. Keep only the delta that's still missing (e.g., the DeepSeek-V4-specific supportsReasoning: true for 14 models)

Happy to review the rebased version. Leaving open.

[kang-heewon]

Rebase complete. Here is what was done:

Rebased on upstream/release/v3.8.0 (HEAD: bbf3ba0). The 4 overlapping commits were already cherry-picked upstream, so they were dropped during rebase:

• preserve reasoning_content = upstream 18ef28e
• widen sanitizer regex + requestId = upstream 72dd7c9
• messageIndex 0 = upstream 4726dea
• fix cache key = upstream c2bf5c7

Regarding your 3 requests:

1. Cache-key signature — upstream and our PR now agree: buildAssistantMessageCacheKey(requestId, messageIndex) → request:${requestId}:message:${messageIndex}. No conflict.
2. requiresReasoningReplay callers — upstream already uses the object parameter signature ({ provider, model, thinkingEnabled?, supportsReasoning? }). Verified: reasoningCache.ts definition + translator/index.ts caller both match. No adjustment needed.
3. supportsReasoning for DeepSeek V4 models — already registered in upstream's providerRegistry.ts with supportsReasoning: true for 18 entries. No delta to add.

Remaining delta (2 files, 4 insertions, 2 deletions):

• responseSanitizer.ts: Extracted inline regex to DEEPSEEK_V4_SANITIZER_MODEL_PATTERN constant
• translator/index.ts: Changed includes("deepseek") to /(^|\/)deepseek/i boundary regex

Both changes are PR review feedback from gemini-code-assist. The PR description has been updated to reflect the current scope.