Git Push

Git Push

Purpose

Close the loop from dirty working tree → clean push → green GitHub Actions: commit everything pending when needed, integrate upstream with merge (not silent rebase unless the user prefers otherwise), push, watch workflows, and iterate on real failures until passes.

Authority

• git push allowed when the user invokes $git-push, names git-push, or clearly asks you to sync, push, and fix CI until green—or a checked-in workflow authorizes the same (for example execute-plan step 5 after all phases are committed). execute-plan does not authorize running git-push after each phase—only the final tail.
• This authorizes git commit, git pull/git merge (integrating origin), git push, reading CI logs, and code/test fixes needed to get workflows green.
• If push or CI remediation would be a surprise, stop and confirm.

Use When

• The user wants the branch on the remote and CI passing.
• You need the full “commit if needed → sync → push → babysit CI” routine.

Do Not Use When

• The user only wants a dry run or local-only commands.
• You are not allowed to touch origin or fix CI (no override).

Preconditions (agent)

1. Know the current branch and whether it has an upstream (git status -sb, git rev-parse --abbrev-ref @{u} when set).
2. git fetch origin before deciding you are in sync.

Procedure

1. Uncommitted changes

• If git status --porcelain is non-empty (including untracked you intend to keep), run git-commit first so the working tree is committed with a full what / why message.
• Do not bypass git-commit with a lazy one-liner unless the user explicitly overrides.

2. Integrate origin (merge-oriented)

• Preferred for this skill: after git fetch origin, integrate with merge, not rebase—unless the user explicitly asks for rebase.
• If an upstream exists: e.g. git pull --no-rebase (or git merge origin/<upstream-branch> after fetch) so local commits combine with remote updates.
• If there is no upstream yet, a reasonable first push is git push -u origin HEAD after the branch is ready—set upstream for future pulls.
• Merge conflicts: resolve carefully; if intent is ambiguous, stop and ask rather than guessing.

3. Push

• git push to the appropriate remote (usually origin) and branch.
• If push is rejected because the remote advanced, fetch, merge (per above), resolve conflicts, then push again—do not force-push unless the user explicitly requests it and it is safe for the branch.

4. Monitor GitHub Actions

• Identify the run(s) for the pushed commit / branch (for example gh run list --branch <branch> --limit 5 then gh run watch <run-id> --exit-status, or the Actions UI).
• Wait until the relevant workflow(s) finish.

5. On failure: fix root cause, repeat

• Read logs; reproduce locally when possible (pnpm lint, pnpm check, targeted tests, smoke, etc., matching what failed).
• Fix the underlying issue—not symptoms only—then commit (again via git-commit when there are new changes) and push, then re-watch CI.
• Loop until green or until you hit a blocker (permissions, flaky external, ambiguous product intent).

6. Coverage and quality gates

• If CI fails because test coverage (or a similar required threshold) is not met: Never satisfy the gate by lowering the threshold or disabling checks as a shortcut.
• Correct response: add or strengthen tests, exercise untested paths, or refactor for testability so coverage earns the bar.

Anti-patterns

• Pushing with a dirty tree (except when every pending change is intentionally left out—and then say so explicitly; default is commit first).
• Force-push to shared branches without explicit approval.
• “Fixing” CI by weakening lint, coverage, or type gates.
• Re-running CI repeatedly without changing what failed.

Related skills

• git-commit: message shape and default whole-tree staging (unless a narrower parent workflow applies).
• execute-plan: may narrow what gets committed per phase; this skill still applies to the push / merge / CI loop once commits are ready.

Source: mattmireles/kokoro-coreml — distributed by TomeVault.