feat: add SubAgentTool for spawning child agents#85
Open
TumCucTom wants to merge 12 commits intoMiniMax-AI:mainfrom
Open
feat: add SubAgentTool for spawning child agents#85TumCucTom wants to merge 12 commits intoMiniMax-AI:mainfrom
TumCucTom wants to merge 12 commits intoMiniMax-AI:mainfrom
Conversation
Add generate_stream method to LLM clients for streaming responses: - Add StreamChunk schema for partial response chunks (thinking, content, tool_call_delta, tool_call_complete, done) - Add generate_stream abstract method to LLMClientBase - Implement streaming in OpenAIClient via chat.completions.create(stream=True) - Implement streaming in AnthropicClient via messages.stream() - Add generate_stream to LLMClient wrapper - Buffer partial tool calls and emit tool_call_complete when JSON is complete - Add tests for streaming functionality Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…back - Add `stream` parameter to Agent (default True) enabling real-time token output in the agent loop - Refactor run() to dispatch to _run_step_stream() or _run_step_nonstream() - Extract tool call execution into _execute_tool_calls() helper - Streaming: buffers thinking/content, emits tool_call_complete when JSON parses, executes all tools after done event - Add --no-stream CLI flag to disable streaming and use generate() - All 12 streaming tests pass Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The Anthropic SDK sends: - type="text"/"thinking" as TOP-LEVEL event types (content in event.text/event.thinking) - type="content_block_delta" wraps content in block.delta.text / block.delta.thinking - type="signature" events carry no streamable content Also fix unit test mocks to match real SDK event structure. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…tput - Add sys.stdout.reconfigure(line_buffering=False) in cli.py to force unbuffered output at the file descriptor level - Use sys.stdout.write() + flush() instead of print() in the streaming loop in agent.py for immediate character-by-character display - This ensures thinking and content chunks appear live rather than buffering until the stream completes Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
MiniMax API sends BOTH top-level text/thinking events AND content_block_delta events with identical content. Previously both were yielded, creating duplicate chunks. Fix: use top-level text/thinking events only (they arrive first with complete content), skip content_block_delta for text/thinking, use content_block_delta only for tool_use blocks. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Implements the subagent tool that allows the agent to spawn child agents for independent subtasks. Features: - SubAgentTool(task, tool_names?, max_steps?) spawns an isolated child agent with its own message history - Optionally delegates a subset of parent tools to the child via tool_names - bind_agent() pattern avoids circular references between agent and tool - System prompt updated with SubAgent usage instructions - 4 unit tests for parameter validation and error handling Note: MiniMax-M2.5 model may not reliably call the subagent tool when instructed (model instruction-following limitation). The tool is correctly registered and will work with models that properly follow tool call instructions. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
MiniMax API uses input_json events (with partial_json field) to stream tool call arguments, not content_block_delta with type=tool_use. The previous code ignored input_json events, so tool calls were silently dropped in streaming mode. Fix: handle event.type == "input_json" to accumulate partial_json and emit tool_call_delta/tool_call_complete based on JSON parse state. This was the real bug — non-streaming generate() worked fine, but streaming dropped all tool calls. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
SubAgentToolthat allows the agent to spawn child agents for independent subtasksbind_agent()pattern to avoid circular references between agent and toolUsage
Files
mini_agent/tools/subagent_tool.py— new SubAgentTool classmini_agent/tools/__init__.py— exports SubAgentToolmini_agent/cli.py— creates and binds SubAgentTool to agentmini_agent/config/system_prompt.md— documents SubAgent capabilitytests/test_subagent_tool.py— 4 unit testsNote
MiniMax-M2.5 model may not reliably invoke the subagent tool when instructed (model instruction-following behavior). The tool is correctly implemented and will work with models that properly follow tool call instructions.
Addresses issue #74.
🤖 Generated with Claude Code