fix(agent): handle deepseek empty reasoning

[zerob13]

Summary by CodeRabbit

• Improvements

  • Enhanced reasoning content handling for DeepSeek models with refined preservation logic and clearer separation of concerns.
  • Improved model detection and configuration for more reliable compatibility behavior.

• Tests

  • Added test coverage for DeepSeek reasoning preservation with compatibility settings disabled.
  • Added test coverage validating empty reasoning content handling.

[coderabbitai]

📝 Walkthrough

Walkthrough

This change decouples the handling of non-empty and empty interleaved reasoning preservation in the context builder, introducing a new isDeepSeekSeriesModelId predicate to classify DeepSeek models, and refining how preserveReasoningContent and preserveEmptyReasoningContent flags are resolved during configuration initialization.

Changes

Cohort / File(s)	Summary
DeepSeek Model Classification src/shared/model.ts	Introduces new exported predicate isDeepSeekSeriesModelId that normalizes and checks for DeepSeek substring in model IDs.
Interleaved Reasoning Configuration src/main/presenter/agentRuntimePresenter/index.ts	Refines resolveInterleavedReasoningConfig to use the DeepSeek series predicate for both preserveReasoningContent (now model-driven) and preserveEmptyReasoningContent (now solely driven by DeepSeek classification).
Context Building Logic src/main/presenter/agentRuntimePresenter/contextBuilder.ts	Decouples "preserve non-empty reasoning" from "preserve empty reasoning" by introducing shouldPreserveEmptyReasoning flag, extending applyReasoningContent with allowEmptyReasoning parameter, and adding explicit early returns for no-tool-call and empty-tool-calls paths.
Test Coverage test/main/presenter/agentRuntimePresenter/agentRuntimePresenter.test.ts, test/main/presenter/agentRuntimePresenter/contextBuilder.test.ts	Validates that DeepSeek v4 preserves both reasoning flags when interleaved thinking compatibility is disabled (unlike GPT-4 behavior), and confirms empty reasoning content is not added when the prior assistant message lacks reasoning blocks.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~22 minutes

Possibly related PRs

• fix: interleaved reasoning persistence across tool calls #1543: Both PRs modify the contextBuilder.ts flow and the applyReasoningContent helper to refine assistant content assembly and interleaved reasoning preservation control.
• fix(deepchat): support interleaved thinking #1373: Both PRs build on interleaved-reasoning feature handling, with this PR further decoupling empty-reasoning cases and tying preservation behavior to DeepSeek model ID classification.

Poem

🐰 With whiskers twitching, we've untangled the knot,
DeepSeek now shines with reasoning thoughts,
Empty or full, each gets its own path—
No more mixed signals, just logic precise as math! 🎯

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change: fixing handling of DeepSeek empty reasoning in the agent, which is the core focus of the changeset.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/deepseek-interleved

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Review rate limit: 7/8 reviews remaining, refill in 7 minutes and 30 seconds.}

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@src/main/presenter/agentRuntimePresenter/contextBuilder.ts`:
- Around line 331-333: The empty-reasoning fallback is currently guarded only by
preserveEmptyInterleavedReasoning, causing inconsistent behavior; change the
guards in the contexts where the fallback is returned (the block using
assistantContent and the similar block later) to require both
preserveInterleavedReasoning && preserveEmptyInterleavedReasoning before
returning [{ role: 'assistant', content: assistantContent }], so the
empty-preservation path is only taken when interleaved reasoning preservation is
enabled too.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: aaa19f72-66f6-42f0-ad76-7a498f10f28c

📥 Commits

Reviewing files that changed from the base of the PR and between c2e8699 and 02c9376.

📒 Files selected for processing (5)

• src/main/presenter/agentRuntimePresenter/contextBuilder.ts
• src/main/presenter/agentRuntimePresenter/index.ts
• src/shared/model.ts
• test/main/presenter/agentRuntimePresenter/agentRuntimePresenter.test.ts
• test/main/presenter/agentRuntimePresenter/contextBuilder.test.ts

[coderabbitai]

⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Guard empty-reasoning fallback with the same composite condition.

Line 331 and Line 357 check preserveEmptyInterleavedReasoning directly, but empty-preservation elsewhere is gated by preserveInterleavedReasoning && preserveEmptyInterleavedReasoning. This creates inconsistent behavior when only the empty flag is true.

Proposed fix

-    if (preserveEmptyInterleavedReasoning) {
+    if (shouldPreserveEmptyReasoning) {
       return [{ role: 'assistant', content: assistantContent }]
     }
@@
-    if (preserveEmptyInterleavedReasoning) {
+    if (shouldPreserveEmptyReasoning) {
       return [{ role: 'assistant', content: assistantContent }]
     }

Also applies to: 357-359

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/main/presenter/agentRuntimePresenter/contextBuilder.ts` around lines 331
- 333, The empty-reasoning fallback is currently guarded only by
preserveEmptyInterleavedReasoning, causing inconsistent behavior; change the
guards in the contexts where the fallback is returned (the block using
assistantContent and the similar block later) to require both
preserveInterleavedReasoning && preserveEmptyInterleavedReasoning before
returning [{ role: 'assistant', content: assistantContent }], so the
empty-preservation path is only taken when interleaved reasoning preservation is
enabled too.