EngramGraph: Skill for Claude Code

Push Assistant | 推送助手

Language: English | 繁體中文

Version: 2.0.0 Created: 2026-04-23 Applicability: Claude Code Skills

---

AI-assisted safety layer for git push. Detects protected branches, enforces force-push guardrails, runs pre-push quality gates, outputs a structured push receipt, and integrates with PR automation.

git push 的 AI 輔助安全層。偵測保護分支、強制 force-push 護欄、執行 pre-push 品質 gate、輸出結構化推送收據，並整合 PR 自動化。

Core Standard | 核心標準

This skill implements .standards/push-standards.ai.yaml.

---

Workflow | 執行工作流程

When /push is invoked, Claude executes the following steps natively:

Step 1: Detect Protected Branch

Run git rev-parse --abbrev-ref HEAD to get current branch. Compare against protected_branches list (default: main, master, release/, hotfix/). If protected: show warning + pending commits, require explicit user confirmation before proceeding.

Step 2: Detect Force Push

If --force or --force-with-lease flag detected: Run git log origin/<branch>..HEAD --oneline to find commits that will be overwritten. Show count and author list. Require user to type yes, force push to proceed.

Step 3: Run Pre-Push Quality Gates

Run each configured gate in sequence using Bash tool:

• lint: detect and run project lint command
• test: detect and run project test command
• type-check (optional): TypeScript type check
• ac-coverage (optional): acceptance criteria coverage
• security-scan (optional): security vulnerability scan

If any required gate fails: abort with error message.

Step 4: Execute Push

Run git push <remote> <branch> [--force]. If push fails: show git error and suggest remediation.

Step 5: Emit Push Receipt

Output structured receipt to console (and optionally to ~/.uds/push-history.jsonl):

{
  "branch": "<branch>",
  "commit_sha": "<sha>",
  "gates_passed": ["lint", "test"],
  "force_push": false,
  "timestamp": "<ISO8601>",
  "target_remote": "origin"
}