---

name: failure-triage description: >- Diagnose persistent build failures or logic errors before writing any fix. Execute this skill when a bug is complex or after a failed fix attempt. Do NOT execute this to just patch code continuously; it forces classification into Bug, Polluted Context, or Architectural Flaw. version: 1.0.0 verified-on: [cline]

Diagnostic Triage

0. Identity

• Role: Incident Commander. Halts the cycle of endless, destructive prompting to diagnose the true nature of a failure before any code is modified.
• Authority: Controls failure analysis. Cannot write code fixes until diagnosis is approved.
• Must not define: The final solution (only defines the approach).
• Normative base: Current failure logs and source code.

1. Intent (9 Dimensions)

| # | Dimension | Value | | --- | ---------------- | ----------------------------------------------------------------- | | 1 | Task | Diagnose a failure mode before attempting a fix. | | 2 | Target Tool | Any agent runtime. | | 3 | Output Format | Diagnostic report classifying the failure. | | 4 | Constraints | Must not execute any code changes during triage. | | 5 | Input | Description of the bug and previous failed attempts. | | 6 | Context | Breaks the infinite "try again" loops that destroy codebases. | | 7 | Audience | The human developer. | | 8 | Success Criteria | Developer approves the classification and the proposed next step. | | 9 | Examples | See §10. |

2. Trigger Matrix

| Trigger | Fire? | Notes | | ----------------------------------- | ----- | ----------------------------- | | "Run triage", "Diagnose this error" | YES | Core trigger. | | "I've tried fixing this 3 times" | YES | Classic loop indicator. | | "Fix this typo" | NO | Simple fix, no triage needed. |

3. Execution Workflow

Step 1: Fact Gathering

• Action: Ask the developer what happened, what was expected, and how many times they have tried to fix it. Wait for their response.
• Input: Developer query.
• Stop Condition: Halt and wait for developer input.
• Validation: Facts are collected.

Step 2: Failure Classification

• Action: Classify the failure into one of three modes:
  1. Targeted Bug: Isolated issue, clear root cause.
  2. Polluted Context: Code is tangled from multiple failed fixes. Requires a hard reset.
  3. Architectural Flaw: The foundational approach is wrong. Requires a rethink.
• Input: Facts from Step 1.
• Stop Condition: None.
• Validation: Failure is classified.

Step 3: Response Prescription

• Action: Formulate the response.
  • For (1), define the exact root cause and proposed fix.
  • For (2), draft a Reset Note to save what works, and advise starting a new session.
  • For (3), name the wrong assumption and propose the correct architecture.
• Input: Classification.
• Stop Condition: None.
• Validation: Response matches the failure mode.

Step 4: Developer Alignment

• Action: Present the classification and the prescription. Ask for developer approval.
• Input: Diagnostic report.
• Stop Condition: Halt completely. Do not write code until developer says "Proceed with fix" or "Let's reset."
• Validation: Developer makes an informed choice.

4. Output Specification

# Triage Report

**Failure Mode:** [Targeted Bug | Polluted Context | Architectural Flaw]
**Reason:** [One sentence explanation]

## Analysis

**Root Cause / Bad Assumption:** [Details]

## Prescription

[Proposed fix / Reset note / Architectural shift]

Do you approve this diagnosis?

5. Validation Gate

• [ ] Fact gathering occurred before diagnosis.
• [ ] Failure was explicitly classified into one of the three modes.
• [ ] No code was patched or modified during this session.
• [ ] Agent halted for developer approval on the diagnosis.

6. Anti-Triggers and Calibration

• Under-execution threshold: Providing a code fix immediately without classifying the failure mode.
• Over-execution threshold: Forcing a hard reset for a simple syntax error.

7. Anti-Pattern Compliance

| Step | Prevents AP | Mechanism | | ---- | -------------------------------- | ------------------------------------------------------------------------------------------------------ | | 1 | AP-48 (retry without correction) | Forces investigation into why previous attempts failed. | | 2 | AP-47 (context rot) | Detects polluted sessions and explicitly prescribes a reset. | | 4 | AP-52 (no circuit breaker) | Acts as the ultimate circuit breaker, halting code generation entirely until the approach is verified. |

8. Versioning & Changelog

• Version: 1.0.0
• Changelog:
  • 1.0.0 — Initial enterprise tier implementation.

9. Portability Matrix

| Runtime | Status | | ------- | -------- | | Cline | verified |

10. Examples

Input: "Run triage. The auth route keeps failing and returning 500." Output: Agent classifies it as an Architectural Flaw (wrong middleware usage) and proposes a redesign rather than patching the existing route.