Performance Problem

[IvanKablar]

Environment

• Serena: 0.1.4-fad35ef4-dirty
• Client: Google Gemini CLI
• OS: Ubuntu 24.04
• Language: Python 3.12.3 (Pyright 1.1.407)
• Test Project: https://github.com/tiangolo/fastapi (1,178 Python files, 80 MB)
• Health Check: ✅ Passes

Configuration:

~/.serena/serena_config.yml

tool_timeout: 1200 # 20 minutes (default: 240s)
default_max_tool_answer_chars: 500000 # (default: 150000)

// .gemini/settings.json MCP timeout
"timeout": 1200000 // 20 minutes

---

Issue

Querying central framework classes causes severe performance degradation:

Metric	Without Serena	With Serena (20min timeout)
Time	7m 13s	41m 14s (5.7x slower)
Timeouts	0	2
Tokens	~270K	~476K

Tool execution:
x find_referencing_symbols {"name_path":"FastAPI",...}
MCP error -32001: Request timed out (after 20 minutes)

x think_about_whether_you_are_done {}
MCP error -32001: Request timed out (after 20 minutes)

---

Reproduction

git clone https://github.com/tiangolo/fastapi
cd fastapi

Configure Gemini CLI with Serena MCP server

gemini

Where is the class FastAPI used?

Wait 41 minutes for 2 timeouts...

The FastAPI class has ~3,290 references across the project.

---

Problems

1. find_referencing_symbols: Runs for full 20-minute timeout on heavily-used symbols (expected but problematic)
2. think_about_whether_you_are_done: Also times out for 20 minutes (unexpected - this should be instant!)

The second timeout suggests a broader issue beyond just LSP performance - possibly deadlock or task queue problems after the first timeout.

---

Suggested Fixes

High Priority:

1. Fail-fast: Pre-check reference count, abort early if > 500 references with helpful error message
2. Fix meta-tool timeout: think_about_whether_you_are_done should never take > 5 seconds

Nice-to-have:
3. Streaming/incremental results for find_referencing_symbols
4. Documentation on when to use LSP tools vs. pattern search

---

Impact

Serena becomes unusable for exploring unfamiliar codebases - the primary use case. Users querying central classes (the natural first query) experience 40+ minute hangs and abandon the
tool.

[MischaPanch]

Thanks, we'll have a look into why the reference finding is so slow.

think_about_whether_you_are_done: Also times out for 20 minutes (unexpected - this should be instant!)

This is instantaneous, but something went wrong and the previous tool call blocked the execution queue. If the queue is blocked, everything will always time out.

This is an instance of something going wrong and getting stuck, not a general perforrmance problem (rather a stability issue). We have recently added a visualization of the state of the tool execution queue to the dashboard, where a user can also cancel tasks, including the running one. With this, one can get the queue unstuck and the execution of think_about_whether_you_are_done would be instantaneous.

Please send us the logs of what was going on and why the find_references got stuck (if you still have them).