/debug-bug — Bug Fix Lifecycle

Run /debug-bug to investigate and fix a bug with proper quality gates. The process adapts based on bug severity — minor bugs get a lightweight path, critical bugs get full board review.

Phase 0: Preflight

Read Architecture Context First

CRITICAL: Before touching any code, read these files to understand the system:

1. •CLAUDE.md — routing, principles, review process
2. •docs/REVIEW-SOP.md — 4-round SOP, severity levels
3. •.board/board/BOARD.md — agent invocations, auth strategy, isolation mode
4. •.claude/skills/setup/SKILL.md — Hard-Won Knowledge section (operational landmines)

Key paths (commonly confused):

• •.board/board/BOARD.md — operational source of truth (invocations, auth, proxy)
• •.board/board/{agent}/ — agent dirs (inbox, outbox, contexts, settings)
• •.board/board/DEFERRED_WORK.md — active deferred items (legacy, also check tasks.json status=deferred)
• •tasks.json — master task index (bugs, features, deferred work)
• •docs/tasks/ — per-task YAML files with phase gates and checklists
• •.claude/skills/ — skill definitions (what the orchestrator follows)

Run Diagnostic First

# Run /debug diagnostic script (Section 10)
# This catches 90% of common issues before investigation starts

Read .claude/skills/debug/SKILL.md Section 10 and run the quick diagnostic script.

Clean Working Tree

git status --porcelain

If dirty, commit or stash first.

Check for In-Progress Tasks

# Check tasks.json for in-progress bugs
cat tasks.json | grep '"status": "in-progress"'
# Check for existing task YAML files
ls docs/tasks/FB-*.yaml 2>/dev/null

If in-progress bugs exist, show them and ask: continue existing, or start new?

Create Task Entry

Ask user for bug name. Assign the next available FB-XXX ID from tasks.json.

1. •Add an entry to tasks.json with status: open, type: bug
2. •Create docs/tasks/{task-id}-{name}.yaml from docs/tasks/TEMPLATE.yaml:

bug: { name }
type: bug
severity: null # set in Phase 2
reported: { today }
status: pending

gates:
  investigation:
    status: pending
    root_cause: null
    files_affected: []
    touches_security: false
    date: null

  classification:
    status: pending
    severity: null
    escalated: false
    board_required: false
    date: null

  board_review:
    status: pending
    rounds: 0
    findings: []
    date: null

  fix:
    status: pending
    files_modified: []
    regression_test: false
    date: null

  smoke_test:
    status: pending
    method: null
    result: null
    all_agents_pass: false
    skipped_reason: null
    date: null

  code_review:
    status: pending
    rounds: 0
    findings: []
    date: null

  cleanup:
    status: pending
    artifacts_removed: []
    date: null

  final_signoff:
    status: pending
    date: null

Phase 1: Investigate

Run /debug Diagnostic

Start with the diagnostic script from /debug skill. This checks: BOARD.md exists, auth tokens, agent directories, lockfile, deferred items.

Reproduce the Bug

1. •Read the bug report / user description
2. •Run a single-agent smoke test from /debug skill
3. •Trace the execution path: BOARD.md invocation → sudo -u $BOARD_USER → CLI → agent reads CLAUDE.md → reads inbox → writes outbox

Read Relevant Source Files

Don't guess — read the actual code. Follow the flow:

• •Bare mode: BOARD.md command → sudo -u $BOARD_USER bash -c 'unset CLAUDECODE && ...' → CLI reads CLAUDE.md → reads inbox → writes outbox
• •Auth flow: CLI reads credentials from board user's home directory → authenticates with upstream API

Document Root Cause

Update task YAML:

investigation:
  status: done
  root_cause: "description of what's wrong and why"
  files_affected: [list of files]
  touches_security: true/false # auth, permissions, blind review enforcement

Phase 2: Classify Severity

Based on investigation, classify:

Escalation Check

If the fix touches ANY of these, escalate severity by one level:

• •Agent invocation pattern (BOARD.md commands, parallelism, timeout)
• •Review SOP (4-round process, lockfile, deferred items, blind review)
• •Setup skill templates (invocation commands that get written to BOARD.md)

Update task YAML: classification.status: done, severity: minor|major|critical

Phase 3: Board Review (Major/Critical only)

For minor bugs: set board_review.status: skipped and proceed to Phase 4.

For major/critical bugs: submit investigation + proposed fix to the board.

Brief must include:

• •Bug description and reproduction steps
• •Root cause analysis
• •Proposed fix with code snippets
• •Files affected
• •Why the bug matters (auth impact, agent failure, data risk)

Round structure: Same as review SOP (blind → consolidation → deliberation → confirmation).

Retry enforcement: If any agent fails to produce a report in any round, retry that agent at least once before proceeding. If the retry also fails, do NOT silently continue — inform the user and get explicit approval to proceed with reduced coverage. With 3 agents, losing one means the tiebreaker is gone. Log all failures and retries in the task YAML.

Update task YAML: board_review.status: done, rounds: N

Phase 4: Fix

Implement the fix. For each change, document what file was modified and why.

Verification during fix (not after):

• •Run /debug diagnostic after each significant change
• •Single-agent smoke test after auth changes
• •Check BOARD.md invocation commands match actual bare-mode commands being tested

Update task YAML: fix.status: done, files_modified: [...]

Phase 5: Smoke Test

DO NOT skip this. ALL agents must pass.

Full Smoke Test (Bare Mode)

BOARD=/path/to/FrontierBoard/.board
BOARD_USER=llmuser  # or whatever board user was created during setup

# Write test brief to all agents
for agent in pragmatist systems-thinker skeptic; do
  echo "Smoke test: Confirm identity. Write one sentence to outbox/report.md." > $BOARD/board/$agent/inbox/brief.md
  echo "Smoke test context." > $BOARD/board/$agent/inbox/context.md
  rm -f $BOARD/board/$agent/outbox/report.md
done

# Run each agent (copy parallelism pattern from BOARD.md)
for agent in pragmatist systems-thinker skeptic; do
  AGENT_DIR="$BOARD/board/$agent"
  sudo -u $BOARD_USER bash -c "unset CLAUDECODE && cd $AGENT_DIR && claude --dangerously-skip-permissions -p 'Read inbox/brief.md. Follow its instructions. Write your response to outbox/report.md.'" &
done
wait

# Verify ALL reports
for agent in pragmatist systems-thinker skeptic; do
  [ -s "$BOARD/board/$agent/outbox/report.md" ] && echo "PASS: $agent" || echo "FAIL: $agent"
done

Every agent must produce a report. If any agent fails, the bug is not fixed.

Update task YAML:

smoke_test:
  status: done
  method: 'all 3 agents in parallel, bare mode'
  result: 'all passed / agent X failed with ...'
  all_agents_pass: true/false

Phase 6: Code Review (Major/Critical only)

For minor bugs: set code_review.status: skipped and proceed to Phase 7.

For major/critical bugs: send git diff to the board.

Update task YAML: code_review.status: done, rounds: N

Phase 7: Ship

1. •Commit with bug reference: git add -A && git commit -m "fix: {description}"
2. •Update task YAML: final_signoff.status: approved, status: approved

Phase 8: Cleanup

Remove temporary artifacts created during debugging. The bug checklist stays (it's the audit trail).

Artifacts to Remove

BOARD=/path/to/FrontierBoard/.board

# Smoke test briefs and reports (not real review artifacts)
for agent in $BOARD/board/*/; do
  name=$(basename "$agent")
  # Only clean if brief is a smoke test
  grep -q "Smoke test" "$agent/inbox/brief.md" 2>/dev/null && rm -f "$agent/inbox/brief.md" "$agent/inbox/context.md" "$agent/outbox/report.md"
done

# Stale review lockfile (if left from crashed test)
rm -f $BOARD/.review-lock

Update task YAML: cleanup.status: done, artifacts_removed: [list]

Severity Examples

Minor:

• •Agent report formatting inconsistent
• •Diagnostic script false positive on a directory
• •Skill documentation typo

Major:

• •Agent auth fails (OAuth token race condition, expired credentials)
• •Codex config format incompatible with new CLI version
• •Agent model not supported with ChatGPT account

Critical:

• •Credential leakage to unauthorized processes
• •Agent reads sibling agent's outbox (blind review violation)
• •OAuth token written to shared location accessible by all agents
• •Review lockfile bypass allows concurrent reviews corrupting reports