Search tool now supports 4 backends with unified output (bocha,
qianfan, zhipu, linkai) and a routing layer:
- strategy 'auto' (default): pick first configured in canonical order
bocha > qianfan > zhipu > linkai
- strategy 'fixed': pin a specific provider
- agent may pass `provider` to override per-call (only exposed when
≥2 providers configured + auto strategy)
Browser sessions now reuse a Chromium user profile across runs by default
(`~/.cow/browser_profile`), so users only log in to a site once.
Three launch modes are selectable via `tools.browser` in config.json:
- persistent (default): Playwright Chromium with a persistent user_data_dir
- cdp: attach to an externally launched real Chrome via `cdp_endpoint`
(full fingerprints, ideal for sites with strict bot detection)
- fresh: clean context every run, set `persistent: false`
Also:
- Self-heal when the user closes the browser window mid-session: detect
closed page/context/browser via close listeners and exception scanning,
then transparently relaunch on the next request.
- Graceful CDP shutdown: disconnect only, never kill the user's Chrome.
- Friendly errors when the CDP endpoint is unreachable or the persistent
profile is locked, so the LLM can guide the user instead of looping.
- Fix tool config being silently overwritten by workspace config in
AgentInitializer; per-tool user settings (e.g. browser.cdp_endpoint)
are now merged instead of replaced.
- Update zh / en / ja docs with the new login-persistence section,
including the Chrome 137+ requirement to pair --remote-debugging-port
with a dedicated --user-data-dir.
Boot MCP servers (npx/uvx) on a background thread instead of blocking
agent init. Built-in tools serve traffic immediately while MCP comes
online; each new agent reads whatever is ready at creation time.
Idempotent via _mcp_loaded flag — concurrent sessions never re-fork
subprocesses. Per-server failures are isolated and warmup is triggered
in app.py so loading overlaps with channel startup.
Stability fixes in mcp_client.py:
- Fix stderr buffer overflow: start daemon thread to continuously drain
stderr pipe, preventing 64KB buffer fill that blocks child process
- Fix notification interference: loop readline and skip JSON-RPC messages
without 'id' field (notifications) instead of treating them as responses
- Fix concurrent race condition: wrap send+receive in _call_lock so
multiple sessions cannot interleave reads/writes on the same client
- Fix missing timeout: use select.select() with 30s timeout in
_readline_with_timeout() to prevent infinite block on dead MCP server
Config improvements in tool_manager.py:
- Add _normalize_mcp_configs() to support both list format (mcp_servers)
and dict format (mcpServers used by Claude Desktop / Cursor)
- Add _load_mcp_configs() to load from ~/cow/mcp.json first, falling back
to config.json mcp_servers field for backward compatibility
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two bugs found during end-to-end validation with Amap and Chrome DevTools
MCP servers:
1. MCP tools were loaded into ToolManager._mcp_tool_instances but never
added to the agent's tool list. AgentInitializer._load_tools() only
iterated tool_classes (built-in tools). Added a second pass to append
all MCP tool instances.
2. When a MCP server config contains an "env" dict, it was passed directly
to subprocess.Popen, replacing the entire process environment. This
caused npx to fail because PATH and other inherited vars were missing.
Fixed by merging config env on top of os.environ.
Validated with:
- @amap/amap-maps-mcp-server (12 tools, stdio + API key env var)
- chrome-devtools-mcp (29 tools, stdio + remote debugging port)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Allows CowAgent to dynamically load tools from any MCP server at startup,
extending the agent from a fixed toolset to an open, extensible tool ecosystem.
## What's added
- `agent/tools/mcp/mcp_client.py`: lightweight JSON-RPC client supporting both
stdio (subprocess) and SSE (HTTP) transports — zero extra dependencies
- `agent/tools/mcp/mcp_tool.py`: `McpTool` wraps a single MCP tool as a
`BaseTool`, with dynamic name/description/params set at instance level
- `agent/tools/tool_manager.py`: new `_load_mcp_tools()` loads MCP servers at
startup via `McpClientRegistry`; falls back gracefully on any error; no-op
when `mcp_servers` is not configured
- `config.py`: registers `mcp_servers` in `available_setting` with inline docs
## Design
- No new dependencies — JSON-RPC implemented from scratch using stdlib only
- MCP clients are long-lived (initialized once, shared across tool calls)
- `McpClientRegistry` holds all subprocess handles and shuts them down cleanly
- Server init failures are non-fatal: logged as warnings, agent continues normally
- Zero overhead when `mcp_servers` is absent from config
## Config example
```json
"mcp_servers": [
{
"name": "filesystem",
"type": "stdio",
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
}
]
```
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Further refinements on top of #2795:
- persist real session_id (notify_session_id) at task creation so group chats
correctly map back to the user's actual conversation
- mark scheduler turns with [SCHEDULED] (recognise legacy "Scheduled task"
prefix too for backward-compatible pruning)
- prune both DB and in-memory to scheduler_inject_max_per_session (default 3),
only marker-tagged pairs are touched; regular user turns never deleted
- send_message type gated by scheduler_inject_send_message (default false) —
fixed reminder text rarely benefits follow-up Q&A
Co-authored-by: huangrichao2020 <grdomai43881@gmail.com>
Address review feedback from #2794:
1. Use notify_session_id instead of receiver for correct group chat mapping
- Task creation should store the real session_id in action.notify_session_id
- Falls back to receiver for backward compatibility with old tasks
2. Add injection to all four execution branches:
- _execute_agent_task
- _execute_send_message
- _execute_tool_call
- _execute_skill_call (also fixed missing channel.send)
3. Add config switch and content truncation:
- scheduler_inject_to_session (default: true) to toggle the feature
- 2000 char limit prevents high-frequency tasks from bloating sessions
Fixes#2793
- Add call_vision method to all bot implementations (DashScope, Claude,
Gemini, ZhipuAI, MiniMax, Doubao, Moonshot, OpenAICompatibleBot)
using each vendor's native multimodal API format
- Remove call_with_tools/call_vision from Bot base class to fix MRO
shadowing issue with OpenAICompatibleBot mixin
- Refactor vision tool provider resolution: MainModel → other configured
models (auto-discovered) → OpenAI → LinkAI, with automatic fallback
- Return actual model name used in call_vision responses
- Sync config.json API keys to .env bidirectionally on startup
- Fix bot instance cache to detect bot_type/use_linkai config changes
- Add SSE reconnection support for web console
- Preserve image path hints in Gemini text for correct vision tool calls
- Update docs/tools/vision.mdx
Browser tool enhancements:
- Navigate action now auto-includes snapshot result, saving one LLM round-trip
- Wait for networkidle + 800ms after navigation for SPA/JS-rendered pages
- Prompt guides agent to screenshot key results and ask user for login/CAPTCHA help
- Fixed playwright version pinned to 1.52.0; mirror fallback to official CDN on failure
Web console file/image support:
- SSE real-time push for images and files via on_event (file_to_send)
- Added /api/file endpoint to serve local files for web preview
- Frontend renders images in media-content container (survives delta/done overwrites)
- File attachment cards with download links; RFC 5987 encoding for non-ASCII filenames
Tool workspace fix:
- Inject workspace_dir as cwd into send and browser tools (previously only file tools)
- Screenshots now save to ~/cow/tmp/ instead of project directory
