Skip to content

BEC01/team-fix-skill

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 

Repository files navigation

Team Fix

Systematic codebase investigation and fix execution for Claude Code. Designed for production hardening, security audits, technical debt cleanup, and multi-issue fix campaigns where 5+ issues span multiple files.

Complements autopilot as the dedicated multi-issue fix methodology it relies on internally.

Install

# Project-level (recommended)
cp skill.md /path/to/your/project/.claude/skills/team-fix/skill.md

# Global (available in all projects)
cp skill.md ~/.claude/skills/team-fix/skill.md

Requires Claude Code CLI or desktop app.

Usage

# After a security audit, code review, or production incident
/team-fix harden this codebase before launch

# With a specific list of items
/team-fix implement these 12 items from the audit

# For cleanup campaigns
/team-fix sweep the codebase for reliability issues

The skill activates automatically on triggers like "harden", "production ready", "fix all issues", "audit and fix", or whenever the user provides 5+ items to implement.

When to use it

Use case Fit
Production hardening / pre-launch security audit Yes
Technical debt cleanup campaigns Yes
Implementing code review feedback with many items Yes
Any task requiring investigation of 5+ issues across multiple files Yes
Single-file fix or fewer than 5 items No — use direct edits instead

How it works

Phase 1: Investigate
  → Parallel Explore agents (one per domain: security, reliability,
    performance, observability, backend logic, frontend)
  → Each returns a compact findings table (max 15 items)

Phase 2: Consolidate
  → Deduplicate into a single prioritized table
  → Group into 4-5 item batches by theme and file ownership

Phase 3: Context Gate
  → Read actual source for every item (exact line numbers, current code,
    fix patterns from already-fixed siblings)
  → Blocks Phase 4 until ALL items have precise context

Phase 4: Skeptical Verification
  → 1 agent per item (no batching)
  → Independent CONFIRM / SKIP / MODIFY verdict
  → Typical rate: ~30-50% of items adjusted or dropped
  → Blocks Phase 5 until every item has a verdict

Phase 5: Sequential Team Execution
  → Teams execute one at a time (not parallel — isolation)
  → File ownership enforced (no two teammates touch the same file)
  → Max 2 fix rounds per team before escalation
  → Commit after each team, clean git before the next

Phase 6: Deep Post-Implementation Audit
  → Independent agents audit committed code line by line
  → Verify old pattern is gone, error paths handled, no collateral damage
  → Required for critical/high severity items; optional for trivial ones

Key ideas

The Context Gate (Phase 3)

Investigation agents return file names and approximate line numbers, but that's not enough. Teammates given imprecise descriptions waste turns searching instead of fixing. Phase 3 forces exact line numbers, current code, and fix patterns from already-fixed siblings BEFORE any team is created.

Skeptical Verification (Phase 4)

Every item gets an independent agent review before implementation. The agent's only job is to disprove the issue. Even items that "look obviously real" turn out to be false positives 30-50% of the time — often handled by upstream guards, contradicting intentional design decisions, or sitting at wrong line numbers from recent code changes.

No exceptions. Batching defeats the purpose — each item gets its own dedicated verification agent.

Sequential teams, not parallel

Teams execute one at a time. This isolates failures (a broken team doesn't derail the others), lets the next team build on the previous team's work, and keeps git history clean with one commit per team. File ownership within a team is exclusive to prevent merge conflicts.

Deep post-implementation audit (Phase 6)

Tests passing is necessary but not sufficient. Code can pass all tests and still be wrong. Phase 6 sends independent agents to read the committed code line by line, verifying the old pattern is fully eliminated, error paths are handled, and no collateral damage was introduced.

Critical/high severity items always get Phase 6. Trivial one-line changes with good test coverage can skip it.

What it orchestrates

Per round, team-fix typically coordinates:

  • 3-6 investigation agents (Phase 1)
  • Context gate agents (Phase 3, one per team group)
  • 1 verification agent per item (Phase 4)
  • 2-3 builder agents per team (Phase 5)
  • Deep audit agents (Phase 6)

All agents use Opus. The orchestrator delegates all code — it makes decisions, not edits.

Why sequential over parallel teams

Parallel teams sound faster but create problems at scale:

  • Merge conflicts when teams touch related files
  • Harder to diagnose which team introduced a regression
  • Git history with concurrent commits is hard to bisect
  • One broken team blocks the others in review

Sequential execution with per-team commits is slower wall-clock but dramatically faster to debug, review, and roll back. In practice, a well-ordered sequential run finishes faster than a tangled parallel one.

Limitations

  • Not for single-file or fewer than 5 items — the coordination overhead isn't worth it
  • Investigation agents work best on codebases with existing structure — greenfield code is harder to audit
  • Phase 6 is required for anything security-critical — skipping it defeats the point

Design principles

  • Verify before implementing — 30-50% of items are false positives or need adjustment; independent agents catch what self-review misses
  • Context before action — fix patterns from already-fixed siblings beat generic descriptions every time
  • Sequential over parallel — isolation and commit discipline over throughput
  • Audit after implementing — tests passing doesn't mean the fix is right

License

MIT

About

Systematic codebase investigation and fix execution for Claude Code. 6-phase methodology with skeptical verification, sequential agent teams, and deep post-implementation audit.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors