Systematic Debugging

Adapted from superpowers' 4-phase methodology for Rune's multi-agent orchestration context.

Iron Law

NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST (DBG-001)

>

This rule is absolute. "I think I know what's wrong" is not investigation. Evidence first, hypothesis second, fix third.

Activation Triggers

This skill auto-activates when:

• A rune-smith worker fails the same task 3+ times
• Ward check fails after a fix attempt
• test-failure-analyst returns LOW confidence classification
• A mend-fixer cannot reproduce the reported finding
• Build/test output contains "error" but the cause is ambiguous

Phase 1: Observe — Gather Evidence (NOT theories)

DO NOT hypothesize yet. Only collect facts.

1. Read the EXACT error message — full stack trace, not just the summary line
2. Identify the FIRST error — cascading failures mislead; find the root
3. Record what changed: git diff HEAD~3 or check recent task completions
4. Record the last known good state: When did this last work?
5. Check git log for recent changes to affected files: git log --oneline -10 -- <file>
6. Check if another worker modified a shared file: git blame <file> | head -20

Output format:

Observation Log:
- First error: {exact message at file:line}
- Last good state: {commit hash or "unknown"}
- Recent changes: {N files changed by {who}}
- Cascading: {yes/no — are subsequent errors caused by the first?}

Phase 2: Narrow — Bisect the Problem Space

Binary search through the problem space:

1. Code bisection:

• Comment out half the changes, re-test
• Narrow to the minimal reproducing change
• git stash to test with/without local changes

1. Layer isolation:

• Is it a data problem? (check inputs, fixtures, test data)
• Is it a logic problem? (check transformations, conditionals)
• Is it an integration problem? (check boundaries, APIs, imports)
• Is it an environment problem? (check versions, paths, permissions)

1. Time isolation:

• git bisect start HEAD <last-known-good> to find breaking commit
• Check if the failure is intermittent (run 3x to confirm determinism)

Output format:

Narrowing Log:
- Bisection result: {minimal reproducing change}
- Layer: {data|logic|integration|environment}
- Deterministic: {yes|no — does it fail every time?}

Phase 3: Hypothesize — Form ONE Testable Theory

Based on Phase 2 evidence:

1. Form a SINGLE hypothesis (not multiple)
2. Design a test that would DISPROVE the hypothesis
3. Run the test
4. If disproved: Return to Phase 2 with the new evidence
5. If confirmed: Proceed to Phase 4

Critical rule: Do NOT skip to Phase 4 based on "it's probably this." Your hypothesis must survive a disproof test.

Output format:

Hypothesis: {one sentence}
Disproof test: {what you ran}
Result: {confirmed|disproved}

Phase 4: Fix — Minimal Change to Address Root Cause

1. Fix the ROOT CAUSE, not the symptom
2. Write a test that would have caught this failure
3. Verify the fix doesn't break adjacent functionality
4. Apply Fresh Evidence Gate (Inner Flame Layer 1.5):

• Cite exact command + output proving the fix works
• No "should work" or "probably fixed"

1. Document what you learned (for Echo persistence)

Defense-in-Depth (Post-Fix)

After fixing, add defensive layers at the point of failure:

• Input validation at the boundary where bad data entered
• Assertion at the point where the invariant was violated
• Test covering the specific failure mode

For detailed multi-layer defense strategy, see defense-in-depth.md.

Escalation Rules

For full escalation templates and messaging formats, see escalation-matrix.md.

Multi-Agent Debugging Context

When debugging in a team workflow:

• Check file ownership: Did another worker modify a shared file? (git blame)
• Check task dependencies: Is the failure caused by an incomplete prerequisite task?
• Check inscription.json: File ownership conflicts between workers
• Use SendMessage: Ask the Tarnished about cross-worker state
• Check echoes: Has this failure pattern been seen before? (echo_search)

Rationalization Red Flags