hiring-agent

In fetch_all_github_repos in github.py, _fetch_github_api() returns (status_code, data), so do not access response.status_code or response.text. Replace the error-handling block with:

elif status_code == 404:
    print(f"GitHub user not found: {username}")
    return []
else:
    print(f"GitHub API error: {status_code} - {repos_data}")
    return []

This prevents the application from crashing with NameError when GitHub returns 404, rate-limit, network, or other API errors.

Close each PyMuPDF Document explicitly. In pdf.py, update extract_text_from_pdf() to use a context manager:

import fitz

def extract_text_from_pdf(path):
    with fitz.open(path) as doc:
        return "".join(page.get_text() for page in doc)

The with statement automatically closes the PDF after extraction, releasing memory. If you cannot use a context manager, call doc.close() in a finally block. This prevents memory from growing continuously when processing batches of 50+ PDFs with PyMuPDF 1.24.x on Python 3.11.

Explicitly set UTF-8 encoding in every Path.write_text() call that writes cache files. On Windows, the default encoding may be cp1252, which cannot encode emojis, accented names, or other Unicode characters.

Change:

Path(github_cache_filename).write_text(
    json.dumps(github_data, indent=2, ensure_ascii=False)
)

To:

Path(github_cache_filename).write_text(
    json.dumps(github_data, indent=2, ensure_ascii=False),
    encoding="utf-8"
)

Apply the same encoding="utf-8" argument to all cache-writing calls in github.py and score.py, including the GitHub API cache around github.py line 49. For additional robustness, cache reads can handle empty or corrupted JSON by rebuilding the cache. The fix was verified on Windows 11 with Python 3.11.

Add the following before importing Protobuf or LLM libraries:

import os
import sys

# Use UTF-8 for Windows console output
if hasattr(sys.stdout, "reconfigure"):
    sys.stdout.reconfigure(encoding="utf-8", errors="replace")
if hasattr(sys.stderr, "reconfigure"):
    sys.stderr.reconfigure(encoding="utf-8", errors="replace")

# Force the pure-Python Protobuf implementation
os.environ.setdefault("PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION", "python")

The environment variable must be set before protobuf is imported. Alternatively, run:

$env:PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION="python"
python score.py

You can also enable UTF-8 mode with set PYTHONUTF8=1 or run chcp 65001 before starting the script. A fix is provided in PR #340.

A failed LLM extraction can write incomplete data, such as empty lists or null fields, to cache/resumecache_<resume>.json. On later runs, the corrupted cache is loaded instead of reprocessing the PDF, causing the Pydantic ValidationError.

Immediate workaround:

rm cache/resumecache_resume.json
python score.py data/resume.pdf

Delete the cache file matching the affected PDF if its name differs. You can also temporarily set DEVELOPMENT_MODE = False to bypass cache loading.

Permanent fix: validate the extracted JSONResume object before writing it to disk, never cache failed extractions, and validate cached data on load. If cache validation fails, delete or ignore the file and run a fresh PDF extraction. This behavior is addressed in PR #198; update to a release containing that PR if available.