Pre Push

ARCHIVED. Out of date, do not use.

This skill no longer loads: it lives in archive/, not skills/. Its stack detection is stale, and the work it gated is now covered by /factory (which adversarially reviews every diff before merge) and /remediate (which owns the test-verified fix loop). See the archive README for the why.

Everything below this line is the original skill, kept for reference.

Pre-Push

Everything you run before pushing to main, once you like the result. Each stage is a gate: it passes quietly, or trips on a genuine failure and stops for a fix. Never pushes — stops at the commit and hands back.

Setup

• Resolve the target repo root: git rev-parse --show-toplevel. Fails → not a git repo, stop. Run every stage from this path.
• git -C <root> status --short empty → nothing to ship, stop.
• Polish-and-ship, not build. Don't change behavior or add features.

Detect the stack

• Node: package.json exists.
• Python: any of pyproject.toml, setup.py, setup.cfg, requirements*.txt, Pipfile, tox.ini.

Both can be true — run both toolchains. Neither → skip stack commands, ask the user. Concrete commands per stack: stacks.md.

Gate: fixing what surfaces

A gate trips only on a genuine failure — a security defect, a failing test, a functional bug — never a style nit or theoretical note.

1. Show the user the problem, location, and error/repro. Ask: diagnosing-bugs, fix another way, or skip? Wait.
2. On go, invoke diagnosing-bugs with the specifics.
3. Re-run the tripped stage, then re-run test. Report pass/fail counts.
4. If it won't resolve in a reasonable pass, stop and hand back. Never commit over a blocker.

Style nits: fix directly or leave; don't pause for them.

Stages, in order

Run each from the repo root; let it finish before the next.

1. simplify — invoke simplify.
2. review — invoke code-review (Standards + Spec; reports, doesn't fix). Functional bug → gate.
3. security — cd <root> first, else security-review silently narrows scope; then invoke it. Real issue → gate.
4. test — from stacks.md: lint/format check, pre-commit hooks if configured, then the full suite. Report pass/fail. Genuine failure → gate. Retry flaky network: a run dying on ETIMEDOUT/ECONNRESET/ENOTFOUND/EAI_AGAIN/getaddrinfo/Could not resolve host/Read timed out/Retrying/registry 50[234] is infra, not a test failure — re-run up to 3×. Any other non-zero exit is genuine; take it to the gate, don't retry. 3 network failures → stop, tell the user, don't commit on unrun tests.
5. commit:
   • chmod +x scripts. A push ships the whole tree, so any tracked .sh-or-shebang file at mode 100644 breaks a Linux install. git ls-files, and for each script still 100644: git update-index --chmod=+x <path>. Report which you flipped.
   • Scan for secrets in changed/staged files (SSH logs, credentials, tokens). Suspicious → stop.
   • Invoke commit, or inline: match git log --oneline -20 convention; commit staged-only if anything's staged else git add -A; message states intent, not a file list; git commit with native hooks — never --no-verify/--no-gpg-sign/amend.

Done

Summarize: stack, simplify changes, review findings, debug-loop fixes, scripts flipped, test counts (and retries), commit hash. State it's committed, ready to push, and not pushed.