triage-three

triage-three

Adversarial triage with three roles. High-confidence results through tension and refinement.

Principle

Progress needs a pusher — someone who moves fast, digs deep, and doesn't self-censor. But speed breeds recklessness: vagueness, exaggeration, even fabrication. So the pusher needs a challenger — someone who questions through Socratic interrogation, exposing weaknesses not by assertion but by asking the questions that reveal whether a finding holds up.

This creates tension. The pusher proposes. The challenger asks: "What evidence proves this? What would disprove it? Is this the root cause or a symptom? What's the alternative explanation?" Both sides stretch. But raw tension is noise. A third role — the arbiter — resolves it, emitting only truth that survived the gauntlet.

Three roles, multiple angles:

[I] Pusher      →  generates findings from a specific angle
[II] Challenger  →  Socratically questions [I]'s output
[III] Arbiter    →  synthesizes across all pairs, emits verified truth

Multiple Pusher–Challenger pairs can explore different angles of the same problem (e.g., correctness + security + performance). The Arbiter synthesizes across all pairs, catching cross-angle patterns and resolving conflicts. Everything works in threes: three roles, up to three angles, up to three runs.

Install

claude plugin marketplace add lagz0ne/triage-three
claude plugin install triage-three

Usage

/triage-three <what to triage and what you're looking for> [style=...]

Examples:

/triage-three find security bugs in src/auth/
/triage-three review this PR for correctness
/triage-three analyze the error handling in lib/
/triage-three audit the API design style=wide-funnel
/triage-three find performance issues in the hot path

The command will:

1. Infer roles from your goal and calibrate effort (angles, depth) from the query's scope and stakes
2. Scope the target files
3. Run Pusher–Challenger pairs (in parallel across angles)
4. Synthesize with the Arbiter across all angles
5. Present only verified findings with cross-angle insights

Style Presets

All presets were eval-tested against real CVEs (jsonwebtoken v8.5.1) and synthetic bug fixtures. Every preset achieves F1 1.0 on ground truth.

Style	Pusher	Challenger	Arbiter	Best for
measured (default)	Accuracy over quantity	Balanced verification	Conservative	General use. Least noise, least tokens.
wide-funnel	Aggressive	Balanced verification	Inclusive	When you'd rather over-report than miss something. Surfaces borderline findings as "LIKELY".
baseline	Aggressive	Ruthless	Conservative	Maximum filtering. May exclude borderline-real issues.
paranoid	Assume everything is broken	Disprove every finding	Strict (zero FP)	Security audits. Highest token cost.
max-tension	Assume everything is broken	Disprove every finding	Conservative	Deep search with strong filtering.

Effort Calibration

Effort is inferred from your query — no parameter tables to fill out.

Effort	Angles	Depth	Triggered by
Light	1	surface	Focused question, single file, low stakes
Standard	3	thorough	Most tasks. Module-level, multiple concerns.
Deep	3	exhaustive	Critical systems, security audits, complex problems

You can always override by passing style=... for prompt presets. Everything else is calibrated from context.

Use Cases

Security audit — Find vulnerabilities, verify they're real, exclude false positives.

/triage-three security audit on src/ style=paranoid depth=2

Code review — Catch bugs, logic errors, and edge cases before merge.

/triage-three review the changes in this branch

Architecture review — Evaluate design decisions, find coupling issues, surface hidden assumptions.

/triage-three review architecture of the auth system style=wide-funnel

Bug triage — Analyze a bug report, verify the root cause, rule out red herrings.

/triage-three investigate why payments fail intermittently

Documentation audit — Find inaccuracies, missing sections, and outdated information.

/triage-three audit the API docs for accuracy against the implementation

Dependency risk — Assess a dependency for security, maintenance, and compatibility risks.

/triage-three evaluate the risk of upgrading to next major version of express

How It Was Built

The prompt presets were developed using an eval system with:

• A prompt variation matrix (3 variants per role, 5 combinations)
• Ground truth fixtures from real CVEs (jsonwebtoken v8.5.1) and synthetic bugs
• Precision/recall/F1 scoring against known bugs, non-bugs, and debatable items
• 10 eval runs across 2 fixtures to find which wording works best