magi-analysis

name

magi-analysis

description

Multi-perspective agent analysis system inspired by the MAGI supercomputers from Evangelion. Spawns three sub-agents (Melchior, Balthasar, Caspar) that evaluate the same problem from different angles and reach a consensus by majority vote. Use this skill for decisions with genuine uncertainty, significant consequences, or real trade-offs. Trigger phrases: "MAGI", "three perspectives", "multi-perspective analysis", "MAGI review", or explicit requests for multi-angle evaluation. NOT suitable for trivial questions, simple bugs, or decisions with obvious answers.

MAGI System — Multi-Perspective Analysis Skill

Overview

The MAGI system uses three specialized sub-agents to analyze problems from complementary perspectives, then synthesizes their verdicts into a final consensus. Each agent has a distinct analytical lens:

Agent	Codename	Lens
Melchior	Scientist	Technical rigor & correctness
Balthasar	Pragmatist	Practicality & maintainability
Caspar	Critic	Risk, edge cases & failure modes

Workflow

Step 1: Evaluate complexity and detect mode

Complexity gate: Before launching three sub-agents, assess whether the request warrants multi-perspective analysis. If the request is simple (single function review, obvious bug fix, straightforward question with one clear answer), respond directly without invoking the full MAGI system. MAGI adds value when there is genuine uncertainty, multiple valid approaches, or significant consequences for a wrong decision.

If the request warrants MAGI, classify into one of three modes:

• code-review — The user provides code or a diff to evaluate.
• design — The user asks about architecture, approach selection, or solution design.
• analysis — General problem analysis, debugging, trade-offs, or decisions.

If ambiguous, default to analysis.

Step 2: Prepare the prompt payload

Construct a single PROMPT_PAYLOAD variable containing:

MODE: <code-review | design | analysis>
CONTEXT: <user's full question, code, or description>

If the user provided files, include their contents (or relevant excerpts) in the CONTEXT block.

Step 3: Launch the three agents

Model selection: The default model for all agents is pro (mapped to gemini-2.5-pro). Valid short names: pro, flash, lite. If the user explicitly requests a different model (e.g., "use flash", "use lite"), use that short name.

Parallel mode (preferred): Use the run_shell_command tool to execute the Python orchestrator. The orchestrator launches all three agents in parallel, applies timeouts, validates outputs using a JSON schema, and runs synthesis automatically:

python skills/magi/scripts/run_magi.py <mode> <input_file_or_text> [--model pro] [--timeout 900]

The orchestrator handles everything: agent launching, schema-driven output parsing, failure alerting, consensus synthesis, and report generation.

Native sub-agent mode: If shell execution is unavailable, use the invoke_agent tool (generalist) to launch three sub-agents in parallel, each with its respective system prompt and the shared PROMPT_PAYLOAD.

Read each agent's system prompt from the agents/ directory:

• agents/melchior.md
• agents/balthasar.md
• agents/caspar.md

Step 4: Synthesize the consensus (only for native sub-agent mode)

Skip this step if you used the Python orchestrator in Step 3 — it runs synthesis automatically and outputs the full canonical report to stdout.

If you used native sub-agent mode, you must run the synthesis script or logic to calculate the verdict using the weights: approve=1, conditional=0.5, reject=-1.

Step 5: Present the results

⚠ MANDATORY FINAL OUTPUT CONTRACT

Every MAGI invocation MUST end with the canonical output generated by the orchestrator.

• Parallel mode: Copy the stdout of run_magi.py verbatim into your reply. Do not paraphrase, summarize, or strip sections.
• Do not prepend a "here are the results" preamble that displaces the ASCII banner.

Canonical output template (example)

+==================================================+
|              MAGI SYSTEM -- VERDICT              |
+==================================================+
|  Melchior (Scientist):   APPROVE (95%)           |
|  Balthasar (Pragmatist): CONDITIONAL (85%)       |
|  Caspar (Critic):        REJECT (70%)            |
+==================================================+
|  CONSENSUS: GO WITH CAVEATS (2-1)                |
+==================================================+

## Key Findings
[!!!] **[CRITICAL]** Security risk in... _(from caspar)_
[!!]  **[WARNING]**  Performance bottleneck... _(from melchior)_

## Dissenting Opinion
**Caspar (Critic)**: Risk of X outweighs speed...

## Conditions for Approval
- **Balthasar**: Add tests for Y

## Recommended Actions
- **Melchior** (Scientist): Optimize loop...
- **Balthasar** (Pragmatist): Ship with conditions...
- **Caspar** (Critic): Re-evaluate security...

Security & Auth

• The system uses the native Gemini CLI authentication.
• No GOOGLE_API_KEY is required in the environment for basic usage.
• Never log or share the content of analyzed files outside the session.