/bugfix Skill

Root cause based one-shot bug fix. Diagnose → generate requirements.md → delegate to /execute for fix, verification, and commit.

The Iron Law

NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST
No completion claims without evidence
Must stop after 3 failed attempts

Architecture

/bugfix "error description"

Phase 1: DIAGNOSE ─────────────────────────────────
  debugger + verification-planner + gap-analyzer (all parallel)
  → Blast-radius grep scan + Triage recommendation
  → User confirmation (includes triage hint)

Phase 2: REQUIREMENTS GENERATION ──────────────────
  Diagnosis results → requirements.md (hoyeon-cli req init + Write)

Phase 3: EXECUTE ──────────────────────────────────
  Skill("execute", args=spec_dir)
  → Success: Phase 5
  → HALT: Phase 4

Phase 4: RESULT HANDLING (if HALT) ────────────────
  Retry (max 3) with stagnation detection → Phase 3
  Circuit breaker → .hoyeon/debug/{slug}.md → suggest /specify

Phase 5: CLEANUP & REPORT ─────────────────────────
  Save .hoyeon/debug/{slug}.md → final summary

Execution Mode

Always runs the full investigation + execution pipeline. No SIMPLE/COMPLEX branching.

Phase 1	Retry
debugger + verification-planner + gap-analyzer	bugfix-managed (max 3)

/execute will prompt for dispatch/work/verify via AskUserQuestion when invoked — bugfix does not pre-select these modes.

---

Phase 1: DIAGNOSE

Step 1.1: Parse Input

Extract from user input:

• Bug description: error message, symptoms, reproduction steps
• Error output: stack trace, test failure logs (if available)
• Context: related files, recent changes (if available)

Initialize Debug State:

SESSION_ID = [from hook — $CLAUDE_SESSION_ID]
slug = convert bug description to kebab-case (e.g. "null-pointer-in-auth")
DEBUG_STATE = "$HOME/.hoyeon/$SESSION_ID/debug-state.md"
hoyeon-cli session set --sid $SESSION_ID --json '{"skill":"bugfix","debug":"'"$DEBUG_STATE"'"}'

Write(DEBUG_STATE):
# Debug: {bug description}
status: investigating

attempt: 0
slug: {slug}

## Symptoms (IMMUTABLE after Phase 1)
- expected: {from user input}
- actual: {from user input}
- error: {from user input}

## Diagnosis
root_cause: pending
spec_dir: pending

## Attempts

Step 1.2: Parallel Investigation

Always dispatch 2 agents in parallel:

Task(debugger):
  "Bug Description: {user input}
   Error Output: {error logs, if available}
   Context: {related files/recent changes, if available}

   Investigate this bug following your Investigation Protocol.
   Classify Bug Type, trace backward to root cause, assess Severity."

Task(verification-planner):
  "User's Goal: Fix the bug described below
   Current Understanding: {user input}
   Work Breakdown:
   - Reproduce bug with test
   - Apply minimal fix at root cause
   - Verify fix resolves the issue

   Focus on Auto items only (what commands prove the fix works).
   Keep it minimal — this is a bug fix, not a feature.

   Note: /bugfix uses Tier 1-3 (Auto items) only. Do not inline VERIFICATION.md.
   Tier 4 (sandbox items) are not needed. Mark sandbox section as 'bugfix mode — Tier 1-3 only'."

After receiving debugger results, update debug-state.md:

Update DEBUG_STATE:
  ## Diagnosis section:
  root_cause: {debugger's Root Cause — 1 line}
  bug_type: {classification}
  proposed_fix: {proposed fix — 1 line}

Step 1.3: Gap Analysis

Always run gap-analyzer after debugger results:

Task(gap-analyzer):
  "User's Goal: Fix the bug below
   Current Understanding: {debugger's full Bug Analysis Report}
   Intent Type: Bug Fix

   Focus on:
   - Whether root cause vs symptom distinction is correct
   - Whether proposed fix could break other areas
   - Whether similar bugs exist with the same pattern"

Step 1.3b: Blast-radius Quick Scan

After debugger identifies root_cause.location (e.g. src/auth/token.ts:42), extract the function/module name and run grep-based structural scan. Always use wc -l so the LLM compares integers, not raw grep output.

# Extract from debugger's Root Cause section:
#   fn   = function or symbol name at the root cause (e.g. "parseToken")
#   mod  = module path without extension (e.g. "src/auth/token")
# If the function name cannot be reliably extracted, set fn="" and skip caller_count.

total_files=$(git ls-files | wc -l | tr -d ' ')
callers=$(git grep -l -F "$fn" 2>/dev/null | wc -l | tr -d ' ')
importers=$(git grep -l -F "$mod" 2>/dev/null | wc -l | tr -d ' ')
test_refs=$(git grep -l -F "$fn" -- '**/*test*' '**/*spec*' 2>/dev/null | wc -l | tr -d ' ')
hot_path=$(git grep -l -F "$fn" -- 'migrations/**' 'auth/**' 'payment/**' 'billing/**' 'schema/**' 2>/dev/null | wc -l | tr -d ' ')

Use -F (fixed string) to avoid regex injection from extracted symbols. Empty or malformed symbols → skip that signal, do NOT fail the phase.

Store the scan in DEBUG_STATE:

Update DEBUG_STATE:
  ## Diagnosis:
    blast:
      total_files: {total_files}
      callers: {callers}
      importers: {importers}
      tests: {test_refs}
      hot_path: {hot_path}

Blind-spot note: grep misses dynamic dispatch, reflection, and non-JS/Python import forms. Low counts = no data, not safety.

Step 1.3c: Triage Recommendation

Compute a routing hint from the scan. This is advisory only — actual routing decision happens at Step 1.4 with user input.

Rules (first match wins):

Condition	Hint
hot_path > 0	/specify recommended — touches critical path (auth/payment/migration/schema)
debugger.severity == COMPLEX AND callers > 10 AND callers > total_files × 0.05	/specify recommended — wide blast radius ({callers} callers, >5% of repo)
debugger.assumptions section non-empty AND root cause text contains uncertainty markers (?, 추정, possibly, may, unclear)	/discuss recommended — root cause unclear
none of the above	bugfix appropriate

Store the hint in DEBUG_STATE:

Update DEBUG_STATE:
  triage:
    hint: {specify|discuss|bugfix}
    reason: "{1-line reason}"

Step 1.4: User Confirmation

Present debugger results + blast-radius scan + triage hint for user confirmation.

The triage hint is advisory — the user always makes the final routing decision.

AskUserQuestion:
  header: "Root Cause & Triage"
  question: "Review diagnosis and choose how to proceed."

  Display:
  - Bug Type: [classification]
  - Root Cause: [file:line + 1-line description]
  - Proposed Fix: [change description — 1 line]
  - Verification: [verification commands from verification-planner]
  - Assumptions: [debugger's Assumptions section]
  - Key warnings from Gap Analysis

  - Blast Radius: callers={N}, importers={M}, tests={T}, hot_path={H}
  - Triage: {hint from Step 1.3c} — {reason}
    (e.g. "/specify recommended — touches auth/ (hot path)")

  options:
  - "Continue with bugfix" → Phase 2
      (if triage hint != bugfix, show as "Continue with bugfix (override triage)")
  - "Switch to /specify" → Hand off requirements.md + debug-state.md to /specify
  - "Switch to /discuss" → Exit bugfix, suggest running /discuss with the bug context
  - "Root cause is different" → Re-run Step 1.2 with user's additional info
  - "Stop" → Exit

Handoff on /specify selection:

# requirements.md hasn't been written yet (Phase 2 is skipped on handoff),
# but debug-state.md contains all Phase 1 findings. Save a skeleton so /specify
# has a starting point, then exit bugfix.

hoyeon-cli req init ${SPEC_DIR} --type bugfix --goal "Fix: {bug description}"
# Append debugger report + blast scan to ${SPEC_DIR}/requirements.md as context
# (sections: ## Debug Context, ## Blast Radius)

Update DEBUG_STATE: status: escalated, reason: "user_triage_specify"

print("Handed off to /specify. Run: /specify {SPEC_DIR}")

---

Phase 2: REQUIREMENTS GENERATION

Convert diagnosis results into requirements.md format. requirements.md is the standard format consumed by /execute, and serves as escalation context for /specify on failure.

Step 2.1: Initialize

SPEC_DIR = "$HOME/.hoyeon/$SESSION_ID"

hoyeon-cli req init ${SPEC_DIR} --type bugfix --goal "Fix: {bug description}"

This creates ${SPEC_DIR}/requirements.md with a stub template.

Step 2.2: Write requirements content

Overwrite ${SPEC_DIR}/requirements.md with the full requirements derived from diagnosis results. Use the Write tool directly.

What to include:

---
type: bugfix
goal: "Fix: {bug description}"
non_goals:
  - "No refactoring beyond the fix"
  - "No unrelated feature changes"
---

# Requirements

## R-B1: {Bug fix requirement title — describes the broken behavior}
- behavior: {one-sentence description of what should work correctly after fix}

#### R-B1.1: {sub-requirement — the core bug scenario}
- given: {the precondition/state that triggers the bug}
- when: {the action that exercises the broken path}
- then: {the expected (post-fix) outcome — what "fixed" looks like}

#### R-B1.2: {sub-requirement — edge case or similar issue, if any}
- given: {precondition for the edge case}
- when: {action that triggers it}
- then: {expected outcome}

## R-T1: Minimal diff constraint
- behavior: Fix targets root cause only with minimal code changes

#### R-T1.1: Root cause targeting
- given: {root cause identified at file:line}
- when: the fix is applied
- then: only the root cause location is modified, no unrelated changes

## Constraints
- Minimal diff (<5% of codebase)
- Fix root cause, not symptom
- No refactoring beyond what the fix requires

Mapping rules:

• Convert debugger's reproduction steps into GWT (given/when/then) for each sub-requirement
• given: the precondition/state that triggers the bug
• when: the action that exercises the broken path
• then: the expected (post-fix) outcome
• If debugger identified edge cases / similar issues, add sub-requirements for each (one sub per scenario)
• Include a constraints section with fix-scope guardrails

Step 2.3: Register

hoyeon-cli session set --sid $SESSION_ID --spec "$SPEC_DIR"

Update debug-state.md with spec_dir: ${SPEC_DIR}.

---

Phase 3: EXECUTE

Hand off requirements.md to /execute. Execute2 reads requirements.md from the spec dir and derives plan.json internally.

Skill("execute", args="${SPEC_DIR}")

What execute handles:

• Prompts the user for dispatch / work / verify via AskUserQuestion
• Worker dispatch according to the chosen dispatch mode
• Round-level commit
• Verify recipe according to the chosen verify depth
• Final report

Result judgment:

IF execute completed successfully (all tasks done, report output):
  → Phase 5

IF execute HALTED:
  → Phase 4

---

Phase 4: RESULT HANDLING

When execute HALTs.

Step 4.1: Read Failure Context

# Extract failure reason from execute's HALT output
# or read from context dir's audit.md, issues.json
CONTEXT_DIR = "${SPEC_DIR}/context"
failure_reason = {execute HALT output or last triage result from audit.md}

Step 4.2: Retry

# Read current attempt from debug-state.md
attempt = debug_state.attempt + 1
MAX_ATTEMPTS = 3

IF attempt >= MAX_ATTEMPTS:
  → Step 4.5 (Circuit Breaker)

Stagnation Detection (attempt >= 2):

# Compare with previous attempt failure info
previous = debug_state.Attempts[-1]
current_reason = failure_reason

SPINNING:    same file/component fails consecutively
OSCILLATION: A fails → B fails → A fails (circular)
NO_PROGRESS: different failures each time, previous fixes cause regressions

Pattern-specific retry_hint:
  SPINNING    → "Different root cause likely. Consider: previous root cause
                 was wrong — trace further back from the symptom."
  OSCILLATION → "Circular dependency. Fix both sides simultaneously."
  NO_PROGRESS → "Fundamental misunderstanding. Re-read error output.
                 Consider: multiple independent bugs? Missing dependency?"
  (no pattern) → "Different approach needed. Do NOT repeat previous attempt."

Step 4.4: Update Requirements & Re-execute

# 1. Record attempt in debug-state.md
Append to DEBUG_STATE ## Attempts section:
  ### Attempt {attempt}
  result: FAIL
  reason: {failure_reason}
  pattern: {detected pattern or "none"}
  hint: {retry_hint}

# 2. Update attempt counter
Update DEBUG_STATE: attempt: {attempt}

# 3. Add failure context to requirements.md
#    Read current requirements.md, append failure context to the Constraints
#    section or add a "## Known Issues" section with the failure info and
#    retry hint so the next worker has context.
#    Use Edit tool to append to requirements.md.

# 4. Re-invoke execute
#    /execute reads plan.json next to requirements.md. Do NOT touch plan.json
#    directly from bugfix — /execute handles task status and re-derivation.
#    If you need to reset failed tasks to pending so /execute will re-run them,
#    do it via the plan CLI:
#      hoyeon-cli plan list ${SPEC_DIR}
#      hoyeon-cli plan task ${SPEC_DIR} --status <task-id>=pending
#    In most cases /execute will resume on its own from the existing plan.json.

# 5. Re-invoke execute
→ Phase 3

Execute2 handles resume naturally:

• Reads existing plan.json and skips done tasks
• Context files (learnings.json, issues.json — in the context dir) retain previous failure info for the new worker
• Known Issues section in requirements.md carries failure context and retry_hint for the worker

Step 4.5: Circuit Breaker

Max attempts exceeded. Present escalation options to user.

First, save attempt records:

Bash: mkdir -p .hoyeon/debug

Write to .hoyeon/debug/{slug}.md:
  # Bugfix Report: {description}
  Date: {timestamp}
  Status: ESCALATED
  Attempts: {attempt count}
  Spec Dir: {SPEC_DIR}

  ## Debugger Analysis
  {debugger's full Bug Analysis Report}

  ## Attempt History
  {full ## Attempts section from debug-state.md}

  ## Assessment
  "{attempt} attempts failed. Architecture-level issue likely."

Update DEBUG_STATE:
  status: escalated

AskUserQuestion:
  header: "Circuit Breaker"
  question: "Fix attempts have failed. This may be an architecture-level issue."
  options:
  - "Switch to /specify (full planning)"
    → "requirements.md and debug report are available:
       Spec Dir: {SPEC_DIR}
       Report: .hoyeon/debug/{slug}.md
       /specify can reference this context for deeper analysis."
  - "Try once more"
    → attempt += 1, go to Phase 3 (no circuit breaker reset)
  - "Stop"

---

Phase 5: CLEANUP & REPORT

After execute completes successfully.

Step 5.1: Save Debug Report

Bash: mkdir -p .hoyeon/debug

Write to .hoyeon/debug/{slug}.md:
  # Bugfix Report: {description}
  Date: {timestamp}
  Status: RESOLVED
  Attempts: {attempt count + 1}
  Spec Dir: {SPEC_DIR}

  ## Root Cause
  {debugger's Root Cause analysis}

  ## Fix
  {proposed fix from debugger + result summary from /execute's final report}

  ## Verification
  {verification results from execute's final report}

Update DEBUG_STATE:
  status: resolved

Step 5.2: Final Summary

print("""
## Bugfix Complete

**Bug**: {description}
**Root Cause**: {file:line — 1-line description}
**Attempts**: {count}
**Spec Dir**: {SPEC_DIR}
**Report**: .hoyeon/debug/{slug}.md
""")

Step 5.3: QA Suggestion

After successful fix, offer QA verification via browser/app:

AskUserQuestion:
  header: "QA"
  question: "Fix complete. Run QA to verify in browser/app?"
  options:
  - "Yes — run /qa"
    → Skill("qa", args="Verify bugfix: {description}. Root cause was at {file:line}. Check that the fix works and no regressions.")
  - "No — done"
    → End

---

Escalation Path

/bugfix (diagnose + requirements.md + execute)
   ↓ circuit breaker (3 failures)
   ↓ requirements.md + .hoyeon/debug/{slug}.md saved
/specify (requirements.md enrichment, leveraging existing diagnosis context)
   ↓
/execute (enriched requirements execution)

Since requirements.md is the standard format, /specify can read and enrich the existing requirements on escalation. All diagnosis context (constraints, known issues, GWT scenarios) is preserved.

---

Agent Summary

Phase	Agent	Status	Condition	Role
1	debugger	existing	always	Root cause analysis, Bug Type classification
1	verification-planner	existing	always	Generate Auto items list (verification commands)
1	gap-analyzer	existing	always	Check for missed factors, risk assessment
1.5	Bash (git grep + wc -l)	new	always	Blast-radius scan + triage hint (no agent, direct grep)
3	/execute (Skill)	existing	always	requirements.md-based execution (worker, verify, commit, review)
5	/qa (Skill)	existing	user opts in	Browser/app QA verification of the fix

Phase 2 (REQUIREMENTS GENERATION) and Phase 4 (RESULT HANDLING) are handled directly by bugfix without agents (hoyeon-cli calls + Write tool + judgment logic).

---

Design Principles

This skill combines core patterns from 3 proven open-source projects:

Principle	Source	Application
Root cause before fix	superpowers (systematic-debugging)	Entire Phase 1
Backward call stack tracing	superpowers (root-cause-tracing)	debugger's Step 3
Defense-in-depth after fix	superpowers (defense-in-depth)	Optional worker application
Anti-pattern rationalizations	superpowers (common rationalizations)	debugger's checklist
Bug Type → Tool routing	oh-my-opencode (Metis intent classification)	debugger's tool table
Full investigation	oh-my-opencode (Momus "80% is good enough")	Always gap-analyzer + standard verify
Minimal diff (<5%)	oh-my-claudecode (executor/build-fixer)	requirements constraint
Circuit breaker (3 attempts)	oh-my-claudecode (debugger) + superpowers	Phase 4
requirements.md as universal format	internal (specify/execute unification)	Phase 2
Execute reuse	internal (single execution engine)	Phase 3
Advisory triage (grep blast-radius + user choice)	internal (bugfix lightweight escalation)	Phase 1.5 / Step 1.4