From 301e7a66387f99ab76754d08bca42f4a9930d3b1 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Tue, 26 May 2026 07:11:59 +0000
Subject: feat(executor,llm): LocalRunner agent-channel via OpenAI tool-use
 (Phase 5)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

LocalRunner previously ignored the AgentChannel and produced a single
fire-and-forget completion. It now declares the four agent back-channel tools
(ask_user/report_summary/spawn_subtask/record_progress) as OpenAI
function-calling definitions and runs a tool-use loop: each turn feeds tool
results back as message history (re-feed) until the model stops calling tools,
bounded by maxLocalToolTurns. ask_user converts a buffered question into a
*BlockedError so the task blocks like the container runners.

Adds tool-use support to the llm client (Tool/ToolCall/ToolFunction types,
Tools on ChatRequest, ToolCalls on ChatResponse + wire request/response). The
loop uses non-streaming Chat (tool_calls don't stream cleanly); assistant text
is still written to stdout.log in the Claude stream-json envelope so summary/
changestats parsing is unchanged.

Fully tested against a mock OpenAI endpoint + storeChannel: spawn/summary/
progress dispatch, ask_user blocking, token accumulation, and the llm tools
round-trip. NOTE: local resume re-feeds conversation state (Decision #8) — not
yet wired, so a blocked local task resumes fresh for now.

https://claude.ai/code/session_01SESwn7kQ7oP62trWw6pc39
---
 internal/llm/client.go | 37 +++++++++++++++++++++++++++++++++++++
 1 file changed, 37 insertions(+)

(limited to 'internal/llm/client.go')

diff --git a/internal/llm/client.go b/internal/llm/client.go
index 613ebe5..e3a6102 100644
--- a/internal/llm/client.go
+++ b/internal/llm/client.go
@@ -34,6 +34,38 @@ type Client struct {
 type Message struct {
 	Role    string `json:"role"`
 	Content string `json:"content"`
+	// ToolCalls is set on assistant messages that request tool invocations.
+	ToolCalls []ToolCall `json:"tool_calls,omitempty"`
+	// ToolCallID links a role:"tool" result message to the call it answers.
+	ToolCallID string `json:"tool_call_id,omitempty"`
+	// Name carries the tool name on role:"tool" result messages.
+	Name string `json:"name,omitempty"`
+}
+
+// Tool declares a function the model may call (OpenAI function-calling format).
+type Tool struct {
+	Type     string       `json:"type"` // always "function"
+	Function ToolFunction `json:"function"`
+}
+
+// ToolFunction describes a callable function and its JSON-schema parameters.
+type ToolFunction struct {
+	Name        string         `json:"name"`
+	Description string         `json:"description"`
+	Parameters  map[string]any `json:"parameters,omitempty"`
+}
+
+// ToolCall is a function invocation the model emitted in its response.
+type ToolCall struct {
+	ID       string           `json:"id"`
+	Type     string           `json:"type"` // "function"
+	Function ToolCallFunction `json:"function"`
+}
+
+// ToolCallFunction holds the called function's name and raw JSON arguments.
+type ToolCallFunction struct {
+	Name      string `json:"name"`
+	Arguments string `json:"arguments"`
 }
 
 // ChatRequest captures the parameters of a single Chat or ChatStream call.
@@ -45,6 +77,7 @@ type ChatRequest struct {
 	Temperature  *float64
 	MaxTokens    int
 	ResponseJSON bool
+	Tools        []Tool
 }
 
 // ChatResponse is the aggregated result of a chat completion.
@@ -54,6 +87,7 @@ type ChatResponse struct {
 	OutputTokens int
 	Model        string
 	FinishReason string
+	ToolCalls    []ToolCall
 }
 
 // Chat performs a non-streaming chat completion. Rate-limit errors (HTTP 429,
@@ -87,6 +121,7 @@ func (c *Client) Chat(ctx context.Context, req ChatRequest) (*ChatResponse, erro
 			OutputTokens: oai.Usage.CompletionTokens,
 			Model:        oai.Model,
 			FinishReason: oai.Choices[0].FinishReason,
+			ToolCalls:    oai.Choices[0].Message.ToolCalls,
 		}
 		return nil
 	})
@@ -133,6 +168,7 @@ func (c *Client) buildRequestBody(req ChatRequest, stream bool) ([]byte, error)
 		Model:    model,
 		Messages: req.Messages,
 		Stream:   stream,
+		Tools:    req.Tools,
 	}
 	if req.Temperature != nil {
 		payload.Temperature = req.Temperature
@@ -301,6 +337,7 @@ type openAIRequest struct {
 	Stream         bool            `json:"stream,omitempty"`
 	StreamOptions  *streamOptions  `json:"stream_options,omitempty"`
 	ResponseFormat *responseFormat `json:"response_format,omitempty"`
+	Tools          []Tool          `json:"tools,omitempty"`
 }
 
 type streamOptions struct {
-- 
cgit v1.2.3