What the agent does
spaCy 3 retired the one-word model shortcuts. On 3.8.18, spacy.load("en") raises OSError: [E941] Can't find model 'en'. It looks like you're trying to load a model from a shortcut, which is obsolete as of spaCy v3.0, and tells you to use spacy.load("en_core_web_sm"). The package carries the table it uses for that message, OLD_MODEL_SHORTCUTS, which maps en to en_core_web_sm, de to de_core_news_sm, es, pt, fr, it, nl, el, nb and lt to their _core_news_sm packages and xx to xx_ent_wiki_sm. python -m spacy download en still works, with a warning that shortcuts are deprecated, and installs en_core_web_sm; the code that loads en afterwards fails anyway, because the package is importable only under its full name.
web64/nlpserver is a Flask service that wraps several NLP libraries behind HTTP endpoints for callers in other languages. requirements.txt lists spacy with no version. The /spacy/entities handler in nlpserver.py reads a lang form field, defaults it to en on line 120, and calls spacy.load(lang) on line 124, so the model name comes straight from the request and the default is exactly the shortcut spaCy 3 refuses. wip/spacy_entities.py has spacy.load("en") on line 5. The README's Step 3 tells users to run python3 -m spacy download en, es and xx, and docker/Dockerfile does the same for en and xx after pip install -r requirements.txt, on a python:3.8.5-slim-buster base. With no pin, a fresh install gets spaCy 3 and every entity request answers a 500.
The agent reproduces it in a virtual environment: install the requirements, run python -m spacy download en, confirm the warning and the installed en_core_web_sm, then call spacy.load("en") and keep the E941 traceback. The fix has two parts. In nlpserver.py it resolves the request's lang through a small table before loading: a full name is used as given, a two-letter code is mapped to its _core_web_sm or _core_news_sm package, and an unknown code answers a JSON error instead of a traceback. The README and the Dockerfile change their download lines to the full package names, and requirements.txt gets a lower bound so the install and the docs describe the same major version.
To test the endpoint without the rest of the stack, the agent starts the Flask app in the sandbox with only spaCy installed, since the other libraries are imported inside their handlers and /status reports what is missing, and posts text=... with lang=en and with no lang to /spacy/entities, checking the JSON carries entity labels. The README's own example sentence serves as input. It also loads each language the Dockerfile installs once, to show xx resolves to xx_ent_wiki_sm. Nothing is pushed to the Docker image and the service is not deployed; the sandbox has no Docker, and the Dockerfile change is checked by reading, with the build left to CI.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "claude",
"input": "This service fails with OSError: [E941] Can't find model 'en' when spacy.load runs. Read requirements.txt, the README and the Dockerfile and report how spacy is specified and which models the docs tell users to download. In a clean virtual environment install the requirements, run python -m spacy download en, and reproduce the error by calling spacy.load('en'); keep the traceback and the installed spacy version. Find every spacy.load call with file and line, and note which take the model name from a request parameter. Fix the loading so a two-letter language code resolves to the full pipeline package spaCy 3 expects (read the mapping from the installed spacy package, not from memory), a full package name passes through unchanged, and an unknown code returns a JSON error. Update the download commands in the README and the Dockerfile to full package names and add a lower bound on spacy. Start the Flask app with only spacy installed and test the entity endpoint with lang=en and with no lang, using the README's example sentence, and include the responses. Do not build the Docker image and do not deploy. Open a pull request with the traceback, the list of load calls, and the endpoint responses.",
"resources": [
{
"url": "https://github.com/your-org/your-app"
}
]
}- Get an API key. In the console, open Settings, then API keys, and create a key for your org. Send it in the Authorization header as a Bearer token.
- Send the request. POST the body in request.json to https://api.phoenix.so/o/{org}/sessions, with your repository as the resources URL. The run starts in its own cloud sandbox.
- Read what came back. Follow the run on its live stream or wait for the run.completed webhook, then read the result and its trajectory.
What comes back
A pull request that touches nlpserver.py, README.md, docker/Dockerfile and requirements.txt. The description leads with the E941 traceback from the sandbox and the installed version, then the mapping the code now applies, taken from the installed package's own OLD_MODEL_SHORTCUTS table so the names match what spacy download installs.
The record of the run holds the download command with its deprecation warning, the failing load, the Flask app starting, and the two curl calls to /spacy/entities with their JSON. A reviewer can read the entity labels for the README's sentence and compare them with the response example the README already shows.
Before merging, a reviewer rebuilds the Docker image in CI, since the sandbox did not, and decides whether the other language models the service advertises should be installed in the image or left to the README. The pull request lists the codes the mapping covers and the one it was tested with.
Run it on every 5xx on an endpoint
A service like this one fails on the first entity request after a rebuild, which an uptime check or an error tracker sees as a 500. A Grafana contact point can call a Rush webhook trigger directly with Authorization: Bearer <its secret> in a custom header, so an alert on the endpoint's error rate starts one run with the alert payload after the prompt. Set the trigger call's Idempotency-Key to the alert's fingerprint so a repeated notification starts no second run.
Sentry cannot send that header; a handler on your server receives its webhook and forwards it to the trigger's URL. Either way the run is the automation's, with its own max_runs_per_hour cap, and its pull request is read from the session.
The use case this job follows, with its automation and trigger: Fix a bug from an alert.
Why run it on Rush?
The fix is a name table; the proof is a running service. Showing that the endpoint answers again means installing spaCy and a language model, starting Flask, and posting real text, and a sandbox does that with nothing else on the machine, so the test shows what a fresh deployment would see rather than what a developer's laptop with three years of models happens to load.
The result is specific to this code, not to spaCy in general: it is the lang parameter on line 120 and the default on line 124 that turn a library change into a public 500, and the pull request says so with the lines. Other services built on the same pattern, a code search for spacy.load( finds them, get the same request as separate runs, each with its own sandbox and record.
Frequently asked questions
Why does spacy download en work but spacy.load('en') fail?
spacy download en prints a deprecation warning and installs en_core_web_sm, the package the shortcut used to point at. spacy.load no longer follows shortcuts, so loading en raises E941 while loading en_core_web_sm works. The agent changes the load, not just the download.
Where does the mapping of codes to models come from?
From the installed spaCy package: its OLD_MODEL_SHORTCUTS table maps en to en_core_web_sm, xx to xx_ent_wiki_sm and the other two-letter codes to their _core_news_sm packages. The agent reads it from the package rather than hard-coding names from memory.
Can I pin spacy below 3 instead?
A pin keeps the shortcuts, at the cost of staying on a line that the current models are not built for. The port is a few lines in one handler plus two documentation commands, and the pull request includes the pin bound in the other direction so the install and the docs agree.
Does the agent need polyglot and the other libraries?
No. The handlers import their libraries lazily and /status lists what is missing, so the agent installs spaCy alone, starts the app and tests the one endpoint. The rest of the stack is left to the image build.
What happens when a caller sends an unknown lang?
Today it raises inside spacy.load and the caller gets a 500. The fix answers a JSON error with the codes the service knows, which the test covers with a made-up code.
More jobs like this: Fix a bug from an alert
- Fix "AttributeError: module 'ssl' has no attribute 'wrap_socket'" on Python 3.12 (ssl)
- Fix "AttributeError: module 'streamlit' has no attribute 'beta_columns'" and the beta_expander Rename (Streamlit)
- Fix "AttributeError: module 'streamlit' has no attribute 'experimental_rerun'" (Streamlit)
- Fix "strings.Title Has Been Deprecated" (SA1019) in a Go CLI (Go strings)
- Fix "'Switch' is not exported from 'react-router-dom'" after upgrading to React Router 6 (React Router)
- Fix "It looks like you're trying to use `tailwindcss` directly as a PostCSS plugin" after Tailwind CSS 4 (Tailwind CSS)