dg

name	dg
description	Adversarial code review mit Dinesh vs Gilfoyle (Silicon Valley). Aufruf: /dg (git diff), /dg 3 (3 Runden), /dg src/foo.rs (Datei), /dg src/foo.rs 3 (Datei + Runden). Zwei Agenten debattieren ueber Code-Qualitaet: Gilfoyle greift an, Dinesh verteidigt. Bestaetigte Issues werden im Final Summary gesammelt.

Dinesh vs Gilfoyle — Adversarial Code Review

Inspiriert von HBO Silicon Valley. Gilfoyle (Angreifer) findet Bugs, Dinesh (Verteidiger) verteidigt oder gibt zu. Wenn Dinesh einen Punkt nicht verteidigen kann, ist das ein bestaetigtes Problem.

Invocation Parsing

Parse the user's /dg invocation:

Input	Target	Max Rounds
/dg	git diff HEAD + git diff --staged	5
/dg 3	git diff HEAD + git diff --staged	3
/dg src/auth.rs	Read that file	5
/dg src/auth.rs 3	Read that file	3

If the diff is empty and no file was given, tell the user there's nothing to review.

Step 1: Collect Code

• Git diff mode: Run git diff HEAD and git diff --staged. Combine both.
• File mode: Read the specified file(s).

Store the result as CODE_UNDER_REVIEW.

Step 2: Debate Loop

Initialize: round = 0, debate_history = []

Loop:

2a. Dispatch Gilfoyle (Agent tool, research only)

Send to Agent with the full content of gilfoyle-agent.md from this skill directory as system context. Include:

• CODE_UNDER_REVIEW
• debate_history (all previous rounds)
• Current round number
• Instruction: "research only, do NOT edit any files"

Display Gilfoyle's BANTER to the user immediately.

2b. Check Gilfoyle Convergence

If all findings are repeats from previous rounds: stop loop. Announce: "Gilfoyle has run out of things to hate. Unprecedented."

2c. Dispatch Dinesh (Agent tool, research only)

Send to Agent with the full content of dinesh-agent.md from this skill directory as system context. Include:

• CODE_UNDER_REVIEW
• Gilfoyle's response from this round
• debate_history
• Current round number
• Instruction: "research only, do NOT edit any files"

Display Dinesh's BANTER to the user immediately.

2d. Check Dinesh Convergence

If ALL responses are [concede] with zero [defend] or [dismiss]: stop loop. Announce: "Dinesh has conceded defeat. As expected."

If both agents are repeating themselves: stop loop. Announce: "These two are going in circles. Separating them before it gets physical."

2e. Continue

Append both responses to debate_history. Increment round.

If round >= max_rounds: ask the user if they want more rounds or stop.

Step 3: Final Summary

Generate this exact format:

## Dinesh vs Gilfoyle Review — [target]
### [N] rounds of mass destruction

### Best of the Banter
2-4 highlights from the debate (the funniest or most brutal exchanges)

### Verdict
Categorize each finding:
- **Critical** (Gilfoyle won, Dinesh conceded) — these need fixing
- **Important** (Gilfoyle won after debate) — should fix
- **Contested** (Dinesh held his ground) — judgment call
- **Dismissed** (Gilfoyle was nitpicking) — ignore

### Strengths
What both agreed was solid (if anything)

### Score: Gilfoyle X | Dinesh Y
Count wins per category above.

### Recommended Changes
- [ ] Actionable checklist of confirmed issues (Critical + Important)

Rules

• Both agents must NEVER edit files. Research and review only.
• The debate must stay technically grounded. Entertainment is the vehicle, not the goal.
• Findings sections are the structured interface. Banter is for the user.
• If the code is genuinely good: Gilfoyle should struggle, Dinesh should be smug, and the summary should reflect that.