Diagnose root cause and fix through four phases: observe → pattern → hypothesis → harden.

Phases

1. Observe and Capture

• Reproduce deterministically. Capture exact error, steps, environment, and variance.
• Record what is known vs unknown before editing any code.

2. Pattern Analysis

• Compare failing path with a known-good reference (code, commit, or contract).
• Identify mismatches in assumptions, data shape, config, ordering, boundaries.
• Narrow to the smallest collision zone that explains the failure.

3. Hypothesis Testing

• One hypothesis at a time: I expect X because Y proves Z.
• Change one variable per test. Log: hypothesis, test, result, next move.
• Failed hypothesis → return to Observe/Orient, not forward. This is a loop.

4. Implement and Harden

• Apply the smallest fix that resolves the confirmed root cause.
• Harden to make the bug class impossible, not just fix the symptom:
  • Entry validation
  • Domain invariants
  • Environment guardrails
  • Regression test coverage
• Verification claims require current-run command evidence (E3).
• Semantic regression from the fix → re-enter Phase 1. Non-semantic failure (lint, format) → fix inline.

Escalation into Root-Cause Chain

When root cause remains unclear after Phase 2, or the same failure pattern spans multiple modules: chain "why" questions, each grounded in observed evidence. Stop when the cause is actionable. Branching unknowns → return to Phase 1.

Unsticking Patterns

• Simplification cascade: strip non-essential parts until failure is obvious.
• Collision-zone narrowing: reduce to smallest interacting component set that still fails.
• Inversion check: validate assumptions by forcing opposite conditions.
• Boundary stress: test near limits to expose hidden conditions.

Escalation

• After 3 failed hypotheses: escalate with concrete evidence and attempted alternatives.
• Architectural uncertainty or repeated cross-module failures: re-enter planning (/do-plan) before broader changes.

Anti-Patterns

• Skipping reproduction because the fix "seems obvious"
• Bundling speculative edits across multiple subsystems
• Suppressing errors to force green checks
• Stopping at symptom relief when the source trigger is unknown
• Claiming completion without current-run verification evidence