feat(deepseek): implement interleaved thinking mode for deepseek-reasoner #9969
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
dosubot
bot
added
size:XL
Contributor
Re-review complete for commit range b84b5b5..31def3a. No new issues found.
Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues. |
![ellipsis-dev[bot]](https://avatars.githubusercontent.com/in/64358?s=60&v=4){kind=small}
![roomote[bot]](https://avatars.githubusercontent.com/in/1546624?s=60&v=4){kind=small}
hannesrudolph
added
the
Issue/PR - Triage
|
so needed too for kimi-k2-thinking here is some more details on that tech # Your tool implementation
def get_weather(city: str) -> dict:
return {"weather": "Sunny"}
# Tool schema definition
tools = [{
"type": "function",
"function": {
"name": "get_weather",
"description": "Retrieve current weather information. Call this when the user asks about the weather.",
"parameters": {
"type": "object",
"required": ["city"],
"properties": {
"city": {
"type": "string",
"description": "Name of the city"
}
}
}
}
}]
# Map tool names to their implementations
tool_map = {
"get_weather": get_weather
}
def tool_call_with_client(client: OpenAI, model_name: str):
messages = [
{"role": "system", "content": "You are Kimi, an AI assistant created by Moonshot AI."},
{"role": "user", "content": "What's the weather like in Beijing today? Use the tool to check."}
]
finish_reason = None
while finish_reason is None or finish_reason == "tool_calls":
completion = client.chat.completions.create(
model=model_name,
messages=messages,
temperature=1.0,
tools=tools, # tool list defined above
tool_choice="auto"
)
choice = completion.choices[0]
finish_reason = choice.finish_reason
if finish_reason == "tool_calls":
messages.append(choice.message)
for tool_call in choice.message.tool_calls:
tool_call_name = tool_call.function.name
tool_call_arguments = json.loads(tool_call.function.arguments)
tool_function = tool_map[tool_call_name]
tool_result = tool_function(**tool_call_arguments)
print("tool_result:", tool_result)
messages.append({
"role": "tool",
"tool_call_id": tool_call.id,
"name": tool_call_name,
"content": json.dumps(tool_result)
})
print("-" * 100)
print(choice.message.content) |
8 tasks
hannesrudolph
added
PR - Draft / In Progress
and removed
Issue/PR - Triage
hannesrudolph
force-pushed
the
feat/deepseek-interleaved-thinking
branch
2 times, most recently
from
9b50c0e to
5b34edb
Compare
hannesrudolph
moved this from PR [Draft / In Progress]
to PR [Needs Prelim Review]
in Roo Code Roadmap
Contributor
Review in progress.
Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues. |
mini2s
added a commit
to zgsm-ai/costrict
that referenced
this pull request
Dec 18, 2025
* Add vendor confidentiality section to the system prompt for stealth models (RooCodeInc#9742) Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com> * chore: add changeset for v3.35.3 (RooCodeInc#9743) * Changeset version bump (RooCodeInc#9745) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Matt Rubens <[email protected]> * Refactor: Remove line_count parameter from write_to_file tool (RooCodeInc#9667) * fix: remove reasoning toggles for GLM-4.5 and GLM-4.6 on z.ai provider (RooCodeInc#9752) Co-authored-by: Roo Code <[email protected]> * fix: handle malformed native tool calls to prevent hanging (RooCodeInc#9758) Co-authored-by: Matt Rubens <[email protected]> * chore: add changeset for v3.35.4 (RooCodeInc#9763) * Changeset version bump (RooCodeInc#9764) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Matt Rubens <[email protected]> * Convert the Roo provider tools for OpenAI (RooCodeInc#9769) * Update the evals keygen command (RooCodeInc#9754) * feat: Add provider routing selection for OpenRouter embeddings (RooCodeInc#9144) (RooCodeInc#9693) Co-authored-by: Sannidhya <[email protected]> * ux: Updates to CloudView (RooCodeInc#9776) * refactor: remove TabHeader and onDone callback from CloudView - Removed TabHeader component from CloudView as it is no longer needed - Removed onDone prop from CloudView component definition and usage - Updated all test files to reflect the removal of onDone prop - Kept Button import that was accidentally removed initially * Updates upsell copy to reflect today's product * Update webview-ui/src/components/cloud/CloudView.tsx Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com> * Update webview-ui/src/i18n/locales/ko/cloud.json Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com> * Update webview-ui/src/i18n/locales/zh-CN/cloud.json Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com> * Test fixes --------- Co-authored-by: Roo Code <[email protected]> Co-authored-by: Bruno Bergher <[email protected]> Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com> * Update model key for minimax in MODEL_DEFAULTS (RooCodeInc#9778) Co-authored-by: Roo Code <[email protected]> * Release v3.35.5 (RooCodeInc#9781) * Changeset version bump (RooCodeInc#9783) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Matt Rubens <[email protected]> * Use search_and_replace for minimax (RooCodeInc#9780) * fix: restore context when rewinding after condense (RooCodeInc#8295) (RooCodeInc#9665) * fix: remove omission detection logic to fix false positives (RooCodeInc#9787) Co-authored-by: Roo Code <[email protected]> * Fix Vercel AI Gateway model fetching (RooCodeInc#9791) Co-authored-by: Roo Code <[email protected]> * refactor: remove insert_content tool (RooCodeInc#9751) Co-authored-by: Roo Code <[email protected]> * feat: add reasoning_details support to Roo provider (RooCodeInc#9796) - Add currentReasoningDetails accumulator to track reasoning details - Add getReasoningDetails() method to expose accumulated details - Handle reasoning_details array format in streaming responses - Accumulate reasoning details by type-index key - Support reasoning.text, reasoning.summary, and reasoning.encrypted types - Maintain backward compatibility with legacy reasoning format - Follows same pattern as OpenRouter provider Co-authored-by: Roo Code <[email protected]> * chore: hide parallel tool calls experiment and disable feature (RooCodeInc#9798) * Update next.js (RooCodeInc#9799) * Fix the download count on the homepage (RooCodeInc#9807) * Default to native tools for all models in the Roo provider (RooCodeInc#9811) Co-authored-by: Roo Code <[email protected]> * Fix/cerebras conservative max tokens (RooCodeInc#9804) Co-authored-by: Matt Rubens <[email protected]> * Release v3.36.0 (RooCodeInc#9814) * Changeset version bump (RooCodeInc#9828) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Matt Rubens <[email protected]> * Merge remote-tracking branch 'upstream/main' into roo-to-main * ux: improved error messages and documentation links (RooCodeInc#9777) * Minor ui tweaks * Basic setup for richer API request errors * Better errors messages and contact link * i18n * Update webview-ui/src/i18n/locales/en/chat.json Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com> * Update webview-ui/src/i18n/locales/en/chat.json Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com> * Empty better than null * Update webview-ui/src/i18n/locales/nl/chat.json Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com> * i18n * Start retryAttempt at 1 * Reverse retryAttempt number, just ommit it from the message --------- Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com> Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com> * web: New Pricing Page (RooCodeInc#9821) * Removes Pro, restructures pricing page * Solves provider/credits * Update apps/web-roo-code/src/app/pricing/page.tsx Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com> * Updates agent landing pages to not mention a trial that doesn't exist * Updates agent-specific landing pages to reflect new home and trial * Indicate the agent landing page the user came from * Clean up the carousel --------- Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com> * Ignore input to the execa terminal process (RooCodeInc#9827) * fix: Overly round follow-up question suggestions (RooCodeInc#9829) Not that rounded * Always enabled reasoning for models that require it (RooCodeInc#9836) * ChatView: smoother stick-to-bottom during streaming (RooCodeInc#8999) * feat: add symlink support for slash commands in .roo/commands folder (RooCodeInc#9838) Co-authored-by: Roo Code <[email protected]> * fix: sanitize reasoning_details IDs to remove invalid characters (RooCodeInc#9839) * feat(evals-ui): Add filtering, bulk delete, tool consolidation, and run notes (RooCodeInc#9837) * Be safer about large file reads (RooCodeInc#9843) validateFileTokenBudget wasn't being called considering the output budget. * Revert "fix: sanitize reasoning_details IDs to remove invalid characters" (RooCodeInc#9846) * Merge remote-tracking branch 'upstream/main' into roo-to-main * Exclude the ID from Roo reasoning details (RooCodeInc#9847) * fix: prevent cascading truncation loop by only truncating visible messages (RooCodeInc#9844) * FIX + feat: add MessageManager layer for centralized history coordination (RooCodeInc#9842) * feat(web-evals): add multi-model launch and UI improvements (RooCodeInc#9845) Co-authored-by: Roo Code <[email protected]> * Revert "Exclude the ID from Roo reasoning details" (RooCodeInc#9850) * fix: handle unknown/invalid native tool calls to prevent extension freeze (RooCodeInc#9834) * feat: add gpt-5.1-codex-max model to OpenAI provider (RooCodeInc#9848) * Delete .changeset/symlink-commands.md * Release v3.36.1 (RooCodeInc#9851) * Changeset version bump (RooCodeInc#9840) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Matt Rubens <[email protected]> * feat: add dynamic settings support for Roo models from API (RooCodeInc#9852) * chore: restrict gpt-5 tool set to apply_patch (RooCodeInc#9853) * Fix chutes model fetching (RooCodeInc#9854) * Release v3.36.2 (RooCodeInc#9855) * Changeset version bump (RooCodeInc#9856) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Matt Rubens <[email protected]> * Better error logs for parseToolCall exceptions (RooCodeInc#9857) * (update): Add DeepSeek V3-2 Support for Baseten Provider (RooCodeInc#9861) Co-authored-by: AlexKer <[email protected]> * web: Product pages (RooCodeInc#9865) Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com> Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com> * fix: sanitize removed/invalid API providers to prevent infinite loop (RooCodeInc#9869) * Update xAI models catalog (RooCodeInc#9872) * refactor: decouple tools from system prompt (RooCodeInc#9784) * Stop making count_tokens requests (RooCodeInc#9884) * Default to using native tools when supported on openrouter (RooCodeInc#9878) * feat: change defaultToolProtocol default from xml to native (RooCodeInc#9892) * feat: change defaultToolProtocol to default to native instead of xml * fix: add missing getMcpHub mock to Subtask Rate Limiting tests --------- Co-authored-by: Roo Code <[email protected]> * Refactor: Unified context-management architecture with improved UX (RooCodeInc#9795) Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com> * Make eval runs deleteable (RooCodeInc#9909) * fix: add Kimi, MiniMax, and Qwen model configurations for Bedrock (RooCodeInc#9905) * fix: add Kimi, MiniMax, and Qwen model configurations for Bedrock - Add moonshot.kimi-k2-thinking with 32K max tokens and 256K context - Add minimax.minimax-m2 with 16K max tokens and 230K context - Add qwen.qwen3-next-80b-a3b with 8K max tokens and 262K context - Add qwen.qwen3-coder-480b-a35b-v1:0 with 8K max tokens and 262K context All models configured with native tool support and appropriate pricing. Fixes RooCodeInc#9902 * fix: add preserveReasoning flag and update Kimi K2 context window - Added preserveReasoning: true to moonshot.kimi-k2-thinking model - Added preserveReasoning: true to minimax.minimax-m2 model - Updated Kimi K2 context window from 256_000 to 262_144 These changes ensure: 1. Reasoning traces are properly preserved for both models 2. Roo correctly recognizes task completion 3. Tool calls within reasoning traces are handled appropriately 4. Context window matches AWS Console specification * fix: update MiniMax M2 context window to 196_608 for Bedrock Based on AWS CLI testing, the actual context window limit for MiniMax M2 on Bedrock is 196,608 tokens, not 230,000 as initially configured. * Update packages/types/src/providers/bedrock.ts Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com> --------- Co-authored-by: Roo Code <[email protected]> Co-authored-by: Matt Rubens <[email protected]> Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com> * fix: use foreground color for context-management icons (RooCodeInc#9912) * feat: add xhigh reasoning effort for gpt-5.1-codex-max (RooCodeInc#9900) * feat: add xhigh reasoning effort for gpt-5.1-codex-max * fix: Address openai-native.spec.ts test failure * chore: Localisation of 'Extra high' * chore: revert unrelated CustomModesManager refactoring --------- Co-authored-by: Hannes Rudolph <[email protected]> * feat: add search_replace native tool for single-replacement operations (RooCodeInc#9918) Adds a new search_replace tool that performs a single search and replace operation on a file, requiring the old_string to uniquely identify the target text with 3-5 lines of context. Parameters: - file_path: Path to file (relative or absolute) - old_string: Text to find (must be unique in file) - new_string: Replacement text (must differ from old_string) * Improve cloud job error logging for RCC provider errors (RooCodeInc#9924) * feat: configure tool preferences for xAI models (RooCodeInc#9923) * fix: process finish_reason to emit tool_call_end events (RooCodeInc#9927) * fix: suppress 'ask promise was ignored' error in handleError (RooCodeInc#9914) * fix: exclude apply_diff from native tools when diffEnabled is false (RooCodeInc#9920) Co-authored-by: Roo Code <[email protected]> * Try to make OpenAI errors more useful (RooCodeInc#9639) * refactor: consolidate ThinkingBudget components and fix disable handling (RooCodeInc#9930) * Add timeout to OpenAI Compatible Provider Client (RooCodeInc#9898) * fix: add finish_reason processing to xai.ts provider (RooCodeInc#9929) * Remove defaultTemperature from Roo provider configuration (RooCodeInc#9932) Co-authored-by: Roo Code <[email protected]> * feat: forbid time estimates in architect mode (RooCodeInc#9931) Co-authored-by: Roo Code <[email protected]> * feat: streaming tool stats + token usage throttling (RooCodeInc#9926) Co-authored-by: Matt Rubens <[email protected]> * feat: Make Architect save to `/plans` and gitignore it (RooCodeInc#9944) Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com> Co-authored-by: Roo Code <[email protected]> * feat: add announcement support CTA and social icons (RooCodeInc#9945) * fix: display actual API error message instead of generic text on retry (RooCodeInc#9954) * feat(roo): add versioned settings support with minPluginVersion gating (RooCodeInc#9934) * Revert "feat: change defaultToolProtocol default from xml to native" (RooCodeInc#9956) * fix: return undefined instead of 0 for disabled API timeout (RooCodeInc#9960) * feat(deepseek): update DeepSeek models to V3.2 with new pricing (RooCodeInc#9962) Co-authored-by: roomote[bot] <219738659+roomote[bot]@users.noreply.github.com> * Add a way to save screenshots from the browser tool (RooCodeInc#9963) * Add a way to save screenshots from the browser tool * fix: use cross-platform paths in BrowserSession screenshot tests * fix: validate screenshot paths to prevent filesystem escape --------- Co-authored-by: Roo Code <[email protected]> * Tweaks to baseten model definitions (RooCodeInc#9866) * fix: always show tool protocol selector for openai-compatible (RooCodeInc#9966) * feat: add API error telemetry to OpenRouter provider (RooCodeInc#9953) Co-authored-by: Roo Code <[email protected]> * fix: validate and fix tool_result IDs before API requests (RooCodeInc#9952) Co-authored-by: cte <[email protected]> Co-authored-by: Roo Code <[email protected]> Co-authored-by: Hannes Rudolph <[email protected]> * fix: respect explicit supportsReasoningEffort array values (RooCodeInc#9970) * v3.36.3 (RooCodeInc#9972) * fix(activate): unify webview panel identifier to use consistent tabPanelId * feat(gemini): add minimal and medium reasoning effort levels (RooCodeInc#9973) Co-authored-by: Roo Code <[email protected]> Co-authored-by: cte <[email protected]> * Delete changeset files (RooCodeInc#9977) * Add missing release notes for v3.36.3 (RooCodeInc#9979) * feat: add error details modal with on-demand display (RooCodeInc#9985) * feat: add error details modal with on-demand display - Add errorDetails prop to ErrorRow component - Show Info icon on hover in error header when errorDetails is provided - Display detailed error message in modal dialog on Info icon click - Add Copy to Clipboard button in error details modal - Update generic error case to show localized message with details on demand - Add i18n translations for error details UI * UI Tweaks * Properly handles error details * i18n * Lighter visual treatment for errors --------- Co-authored-by: Roo Code <[email protected]> Co-authored-by: Bruno Bergher <[email protected]> * Fix: Correct TODO list display order in chat view (ROO-107) (RooCodeInc#9991) Co-authored-by: Roo Code <[email protected]> * fix: prevent premature rawChunkTracker clearing for MCP tools (RooCodeInc#9993) * fix: filter out 429 rate limit errors from API error telemetry (RooCodeInc#9987) Co-authored-by: Roo Code <[email protected]> Co-authored-by: cte <[email protected]> * Release v3.36.4 (
Summary
Implements proper interleaved thinking support for DeepSeek V3's thinking mode (deepseek-reasoner model) following the DeepSeek API documentation at https://api-docs.deepseek.com/guides/thinking_mode
This PR managed to get the evals to 85.6% by extending the time of the runs from 5 to 10 min.
Changes
thinking: { type: "enabled" }parameter fordeepseek-reasonermodelreasoning_contentfrom DeepSeek API that yields reasoning chunkstool_use→ OpenAItool_callsformat) for thinking modetool_result→ OpenAItoolmessages)getReasoningContent()method to accumulate and expose reasoning contentHow It Works
Per DeepSeek's API documentation:
thinking: { type: "enabled" }fordeepseek-reasonermodelreasoning_contentis yielded as reasoning chunks and accumulatedreasoning_contentis preserved and passed back to the API for continuationImportant
Implements interleaved thinking mode for DeepSeek reasoner, handling reasoning content and tool calls, with updated message conversion and comprehensive tests.
deepseek-reasonermodel indeepseek.ts.thinking: { type: "enabled" }parameter fordeepseek-reasoner.reasoning_contentand tool calls increateMessage().convertToR1Format()to handlereasoning_contentand tool calls.tool_useto OpenAItool_callsandtool_resulttotoolmessages.deepseek.spec.ts.r1-format.spec.ts.deepseek.ts.This description was created by
for 31def3a. You can customize this summary. It will automatically update as commits are pushed.