菜单
Interrogate code changes to surface decisions, alternatives, and confidence levels. Use when user asks to "interrogate", "what decisions", "hardest part", "alternatives", or wants to understand the reasoning behind code changes.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | dev:interrogate |
| description | Interrogate code changes to surface decisions, alternatives, and confidence levels. Use when user asks to "interrogate", "what decisions", "hardest part", "alternatives", or wants to understand the reasoning behind code changes. |
| allowed-tools | Bash(git diff:*), Bash(git show:*), Bash(gh pr diff:*), Bash(aaa session:*) |
| argument-hint | <unstaged|staged|changes|commit|pr> [--quick|--skeptical] |
Ask "why" instead of just reading code. Surfaces assumptions, decisions, and confidence levels.
@context/workflows/interrogate.md
Parse arguments to determine:
unstaged (default), staged, changes (staged+unstaged), commit <hash>, pr <number>, or range <ref1>..<ref2>--quick (minimal output) or --skeptical (extra validation)This skill operates in two distinct modes depending on the target:
| Mode | Targets | Behavior | Richness |
|---|---|---|---|
| Live | changes, staged, unstaged | Direct introspection - answer from current session memory | ★★★ Best |
| Forensic | commit <hash>, pr <N>, range | Load session transcript via aaa session cat | ★★ Good |
| Fallback | Forensic when no cc-session-id | Diff-only analysis (speculative) | ★ Basic |
When interrogating current changes (changes, staged, unstaged), Claude can answer directly from memory of the current session. This provides the richest answers because:
No file lookup needed. Just answer the three questions directly from memory.
When interrogating past commits (commit, pr, range), attempt to load the session transcript:
git log --format="%(trailers:key=cc-session-id,valueonly)" <hash>aaa session cat --commit <hash>Forensic mode is second-best to live mode - it has the conversation context but not the "lived experience" of the current session.
When no cc-session-id trailer exists (older commits or external contributions):
Based on target, gather the code diff:
git diff
git diff --cached
git diff HEAD
git show <hash>
# Also attempt session retrieval:
aaa session cat --commit <hash>
gh pr diff <number>
# For each commit in PR, attempt session retrieval
git diff <ref1>..<ref2>
# For each commit in range, attempt session retrieval
Apply the three core questions from the workflow to the gathered diff:
Live mode: Answer these from direct memory of the current session.
Forensic mode: Use loaded session transcript to inform answers. If no session available, derive answers from diff analysis only.
Format output per @context/workflows/interrogate.md#Output-Format based on selected mode.
Include a session context indicator in output:
[Live] - Answers from current session memory[Forensic: <session-id>] - Answers informed by loaded session transcript[Diff-Only] - No session context availableWalk through a milestone GUIDE.md with agent-browser, testing every step. Fixes code bugs and guide inaccuracies as they're found. Use when user asks to "test the guide", "run guide and fix", "validate guide", or "run-guide-and-fix".
Generate a comprehensive GUIDE.md for a project milestone. Use when user asks to "write guide", "generate guide", "create demo guide", or "write-guide". Analyzes milestone stories, git changes, and project setup to produce a step-by-step demo guide.
Complete reference for using the All Agents CLI (`aaa`). Use this skill whenever the user asks to run an aaa command, wants to plan/build/review with Ralph, needs to know the right CLI invocation, asks about available commands or flags, mentions 'aaa', 'ralph', 'subtasks', 'milestones', 'cascade', or needs help constructing any `aaa` CLI command. Also trigger when user says things like 'plan stories', 'build the queue', 'check status', 'review subtasks', 'run calibration', or any development workflow task that maps to an aaa command.
Automates browser interactions for web testing, form filling, screenshots, and data extraction. Use when the user needs to navigate websites, interact with web pages, fill forms, take screenshots, test web applications, or extract information from web pages.
Create Obsidian .canvas files programmatically with correct format. Use when user asks to "create canvas", "make canvas", "canvas file", "obsidian canvas", or needs to generate a .canvas for Obsidian.
Run Ralph calibration checks to analyze intention drift, technical quality, and self-improvement opportunities. Use when user asks to "ralph calibrate", "check drift", "analyze sessions", or needs to verify work alignment.