Improve tool calling response handling

[MichaelAnders]

Summary

• Additional logging for tool call parsing and execution
• Hard-coded shell commands for reliable tool execution
• Deduplication of tool calls within a single response
• Collect and return results from all called tools
• Ollama uses specified Ollama model
• Fix double-serialized JSON parameters from some providers
• Fix stripThinkingBlocks() destroying markdown bullet points in Ollama responses
• Handle Ollama offline gracefully with retry and 503 response
• Add optional TOPIC_DETECTION_MODEL env var to redirect topic detection to a lighter model
• Add optional SUGGESTION_MODE_MODEL env var to control suggestion mode LLM calls (none to disable, or redirect to other model)
• Remove tool_result_direct short-circuit and improve agentic loop
• Add external file read with tilde expansion and user approval flow
• Improve tool calling response handling for Ollama models

Problems

• Topic detection requests use the same large model as the main request. For users running local models, this adds unnecessary GPU load for a simple classification task.
• Several issues with the agentic loop and Ollama integration:

1. stripThinkingBlocks() was destroying valid response content - The regex matched standard markdown bullets as "thinking markers", causing most list-based responses to be truncated after the first heading.
2. Suggestion mode uses GPU resources - Every user message triggers a full agentic loop with tools for suggestion prediction, blocking responses on large models (70b+).
3. Tool calling response handling had edge cases with Ollama's response format.

Configuration for TOPIC_DETECTION_MODEL

# Use main model (default, unchanged behavior)
TOPIC_DETECTION_MODEL=default

# Redirect to a lighter model
TOPIC_DETECTION_MODEL=llama3.2:1b

Files Changed

File	Change
src/api/middleware/logging.js	Enhanced request/response logging
src/api/middleware/request-logging.js	Additional request logging
src/api/router.js	Router updates for tool handling
src/config/index.js	Ollama model configuration, added TOPIC_DETECTION_MODEL env var (default: "default"), wired into config object and reloadConfig() for hot reload support
src/orchestrator/index.js	Tool call deduplication, result collection, direct result return
src/tools/index.js	Deep-parse double-serialized JSON params, execution logging
src/tools/stubs.js	AskUserQuestion handler improvements
src/tools/workspace.js	Workspace tool updates

Test plan

• Verify tool calls are deduplicated when model returns duplicates
• Verify results from all called tools are collected and returned
• Verify Ollama uses the configured model
• Verify double-serialized JSON parameters (e.g. stringified arrays) are correctly parsed
• Verified markdown bullet points survive in Ollama responses (the original bug)
• Tested Ollama offline/online transitions
• Tested suggestion mode disable (SUGGESTION_MODE_MODEL=none)