Addresses Codex review: concurrent callers could both pass the
_mcp_connected guard and race through _connect_mcp(). Added
_mcp_connecting flag set immediately to serialize attempts.
The flag was set before the connection attempt, so if any MCP server
was temporarily unavailable, the flag stayed True and MCP tools were
permanently lost for the session.
Closes#889
Address review feedback:
- Remove on_progress call for interim text to prevent duplicate
messages when the model simply answers a direct question
- Reset final_content to None before continue to avoid stale
interim text leaking as the final response on empty retry
Closes#705
Some LLM providers (MiniMax, Gemini Flash, GPT-4.1, etc.) send an
initial text-only response like "Let me investigate..." before actually
making tool calls. The agent loop previously broke immediately on any
text response without tool calls, preventing these models from ever
using tools.
Now, when the model responds with text but hasn't used any tools yet,
the loop forwards the text as progress to the user and gives the model
one additional iteration to make tool calls. This is limited to a
single retry to prevent infinite loops.
Closes#705
Add a `_consolidating` set to track which sessions have an active
consolidation task. Skip creating a new task if one is already in
progress for the same session key, and clean up the flag when done.
This prevents the excessive API calls reported when messages exceed
the memory_window threshold — previously every single message after
the threshold triggered a new background consolidation.
Closes#751
File tools now resolve relative paths (e.g., "test.txt") against the
workspace directory instead of the current working directory. This fixes
failures when models use simple filenames instead of full paths.
- Add workspace parameter to _resolve_path() in filesystem.py
- Update all file tools to accept workspace in constructor
- Pass workspace when registering tools in AgentLoop
Fix TypeError when LLM returns JSON objects instead of strings for
history_entry or memory_update.
Changes:
- Update prompt to explicitly require string values with example
- Add type checking and conversion for non-string values
- Use json.dumps() for consistent JSON formatting
Fixes potential memory consolidation failures when LLM interprets
the prompt loosely and returns structured objects instead of strings.
