[joepetjr]

Summary

Ports the reviewed installer pattern from keycut (PR #2/#3). Pushing a v* tag builds a Windows
installer on windows-latest — PyInstaller one-file exe, wrapped by Inno Setup, installed to
Program Files, appended to the machine PATH, registered in Add/Remove Programs — and attaches it to
that tag's release. workflow_dispatch runs the identical build and stops short of publishing,
leaving the installer as an artifact, so the Windows half can be proven without a public release.

The external requirement, established from source

Claude Code (claude) must be installed separately and be on PATH. This is not incidental:
driver.py:85-96 builds ["claude", "-p", prompt, "--agent", …] and driver.py:100 runs it via
subprocess.run. Every unit of work is one such process, so with no claude binary every unit
returns FLAGGED. Corroborated at pyproject.toml:23-24 and README.md:44-50. The installer's
finish page and the README both say so, with the install command.

Authentication is inherited and never handled here — driver.py:9-10 states this module "never
reads, sets, or forwards credentials of its own", and a grep for API_KEY|api_key|getenv|requests| urllib|http across the package returns no credential handling. The README says only: run claude
once and sign in. No mechanism is invented, and nothing about pricing, quota or plan entitlements
appears anywhere.

The install command is documentation-verified, not recalled: winget install Anthropic.ClaudeCode
on the finish page, with the native installer and claude --version in the README, taken from
code.claude.com/docs/en/setup.

The __main__ hazard, resolved by execution

The entry point is claude_batch_runner.__main__:main, and that module carries its own
if __name__ == "__main__": sys.exit(main()) at lines 60-61 — a launcher that re-triggered it would
run argparse twice. Proved by running it that importing as a submodule binds __name__ to
claude_batch_runner.__main__, so the guard does not fire:

import OK; module __name__ = claude_batch_runner.__main__
PROBE PASSED: no premature exit

Had it fired, argparse would have exited 2 on the missing --spec. The launcher docstring records why.

Test plan

• python -m PyInstaller --onefile --console --name claude-batch-runner … → build complete
• dist/claude-batch-runner --help → exit 0; --spec examples/demo-campaign.json --dry-run →
  exit 0 with a full valid plan, so real logic is reached inside the freeze
• Frozen size 8.1 MB (Linux ELF; the Windows PE will differ)
• PyInstaller warn file: 13 missing-module lines, all Windows-only stdlib, macOS-only, or
  PyInstaller-internal. Zero package modules missing → no --hidden-import
• pytest -q → 65 passed, re-run unpiped by the dispatcher
• .iss pure ASCII, no BOM, [Registry] PATH on one line with zero \ continuations
• Workflow YAML parses; two mutually-exclusive guards, no build step guarded

Lint is pre-existing, not a regression: baseline at 625a44b is identically Found 9 errors /
3 files would be reformatted, all in files this PR does not touch. The launcher alone is clean.
ci.yml runs pytest only, so no lint job gates this.

Unverified until a real Windows run

The entire Windows half: the ISCC compile, the HKLM PATH add/remove Pascal (needs a real
install/uninstall, not just a build), the Windows freeze, elevation/SmartScreen, and the
gh release create-vs-upload branch. Merge this after keycut's dry run has gone green — the
same chain is now in four repos and a flaw found once should be fixed once.

Minor and pre-existing: --help prints usage: claude_batch_runner, not the exe name, from prog=
at __main__.py:18. Fixing it means touching src/, which is out of scope here.

Diff 285 sum (cap 300).