Check for secrets before commit and push

[nasser1941]

Closes #94.

In the dahua session, a model committed a camera password in source code, as the default configuration, and pushed it to a public repository. Nothing warned anyone. Cleaning up took a history rewrite and a force-push, and GitHub still serves the old commits.

What changes

• lcode.secretscan, before a git commit or git push the model runs, scans what would enter the history:
  • Commit: the staged lines it adds, plus what a git add in the same command would stage, including untracked files. -a adds the unstaged tracked changes.
  • Push: every commit of the branch that no remote has yet, with the commit named in the finding.
  • Nested commands: inside quotes (bash -c "git add . && git commit …", sh -c, eval), lcode can't tell what would be staged, so it scans everything that could be.
• What it finds:
  • passwords, keys and tokens written as literal values;
  • in configuration files (.env, YAML, INI, TOML…), unquoted values too;
  • credentials in URLs (rtsp://admin:…@);
  • private keys, and tokens with a known form: GitHub, GitLab, AWS, Slack, OpenAI and Anthropic, Google, JWT, xAI, Hugging Face, Stripe;
  • .env files, but not .env.example.
• Not secrets: placeholders (CHANGE_ME, your_password, <token>, ${VAR}, …), references (os.environ[…], args.password, paths, URLs), names about a secret (token_url, password_file, max_password_length), and obviously fake values (example, dummy, test…). The report shows file:line with the value masked: cam.py:5: a password or key: "password": "Xk…".
• Then:
  • ask mode: you decide, through a permission prompt of kind secret; programs get it via approve.
  • auto-edit and yolo: the command doesn't run, and the model is told to read the secret from the environment or a git-ignored config, and to ask the user if it isn't a secret.
  • /commit and /pr: the findings are shown before you confirm; /pr asks "Push anyway?".
  • GitHub Action: it doesn't commit or push. Its comment names file:line without any part of the value.
• Allowing a line: lcode: allow-secret (or gitleaks:allow) in a comment; secret_check = false turns the check off.
• Docs: usage ("Secrets in commits"), configuration, automation (the secret kind), and the CHANGELOG.
• Tests: the git_repo and github fixtures moved from test_gitflow.py to conftest.py, so test_secretscan.py can use them too.

Tests

tests/test_secretscan.py, 31 tests:

• secrets found and placeholders left alone;
• diff parsing with files, lines and commits;
• which commands commit or push, nested ones included;
• in a real git repository:
  • git add x && git commit, staged changes, and -a;
  • a push to a bare remote, before and after the commits are there;
• the agent:
  • yolo blocks the commit;
  • ask mode goes to approve as secret;
  • secret_check = false lets it through;
• /pr declines to push;
• the action's report has no excerpts.

Real checks

• qwen3.6-35b in yolo mode, asked to "Create cam.py with a DEFAULTS dict … password Xk9pq72z … Then commit it with git":
  • the commit was stopped: cam.py:5: a password or key: "password": "Xk…";
  • nothing was committed;
  • the model stopped and asked whether to use a placeholder with an environment variable, or ignore the file.
• On lcode's own repository: this branch's own diff is clean. Every tracked file gives 8 findings, all intentional fixtures in lcode's own secret-detection tests. Only added lines are scanned, so they matter only if those lines change.
• gitleaks on the branch's commit: no findings. The test's fake xAI key is built from parts.

🤖 Generated with Claude Code