Fix Slow Endpoint

Fix a slow endpoint with Profyle traces

Work from measurements, not intuition: every claim you make about where time goes must come from a trace, and every fix must be verified with a new trace of the same request.

The profyle MCP server provides: doctor, slowest_endpoints, list_traces, analyze_trace, get_call_details, get_function_source, replay_request, compare_traces. All durations are in milliseconds.

0. Check the setup

Call doctor first. It reports which trace database is read, whether traces exist, which app wrote them and how (profyle run or middleware), and whether the app is running. If everything is ✓, go to step 1.

If there are no traces or the app is not running, get the app running under Profyle without editing the user's code:

1. Make sure Profyle is installed in the project's environment (pip install profyle, uv add --dev profyle, ...). Ask before installing anything.
2. Find how the project starts its dev server: README, Makefile, pyproject.toml scripts, Procfile, docker-compose, manage.py, or an app object such as main:app. Prefer the command with auto-reload: uvicorn main:app --reload, flask --app app run --debug, python manage.py runserver.
3. Propose the command prefixed with profyle run, for example profyle run uvicorn main:app --reload, and ask the user to confirm it (and the port). profyle run adds tracing to FastAPI, Starlette, Flask, Django, Tornado (also under gunicorn's tornado worker) and any ASGI app served by uvicorn. If the app already uses ProfyleMiddleware, run the command as is.
4. With the user's agreement, start it as a background process so it keeps running. It prints profyle ▸ tracing requests ... at start-up, then one line per request with the trace id and its main finding (profyle ▸ GET /orders 245 ms · #12 · repeated: list_orders → get_customer ×100 (83%)). Read those lines: they confirm tracing works.
5. Trigger the slow request: with curl if it is a GET that needs no login, otherwise ask the user to do it in the app. Then call doctor again.

If the app runs where you cannot start it (a container, a remote machine), explain the setup to the user instead: profyle run <their command>, or ProfyleMiddleware from profyle.asgi / profyle.wsgi (or "profyle.django.ProfyleMiddleware" in MIDDLEWARE), with the traces database shared with this project (<project>/.profyle/profile.db, or PROFYLE_DB).

1. Find the endpoint and its traces

• If the user named the endpoint, call list_traces(name_contains=...). Otherwise call slowest_endpoints and confirm with the user which one to work on.
• Listings show each trace's main finding (for example repeated: a → b ×100 (83%)). It is a starting point, not the diagnosis: confirm it with analyze_trace. A first, much slower trace usually shows warm-up (imports, regex compiling) instead.

2. Get a trustworthy baseline

• The first request after the server starts includes warm-up (imports, regex compiling, connection setup). Do not diagnose that trace if a later one exists.
• For a GET request, call replay_request(trace_id, times=3) before changing anything. This gives warm traces and a median baseline, and confirms that replay works (the app is reachable and credentials are not missing). If the app is not reachable, start it as in step 0.
• Never replay POST, PUT, PATCH or DELETE without the user's explicit permission. These requests may create or delete data. Ask first; only then pass allow_unsafe_method=true.
• Auth headers and cookies are redacted when recording. If the replay returns 401 or 403, ask the user for a token or session for local testing and pass it in headers. Never print credentials back.