Run any Skill in Manus with one click

repo-offensive

Stars0

Forks0

UpdatedJune 21, 2026 at 06:16

Run a repo-wide quality offensive for <SCOPE>. Use when the user says "ultracode", "repo offensive", "quality offensive", "codebase cleanup", or explicitly requests a full-coverage quality offensive, pre-release hardening, or inherited-codebase cleanup.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

Simplereally

Simplereally/old-town

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

SKILL.md

readonly

More from this repository

same repository

implementation-tournament

Simplereally/old-town

Implement a feature using an implementation tournament with competing designs, judge scoring, and champion synthesis. Use when the user says "ultracode", "implementation tournament", "tournament", "design competition", or explicitly requests /implementation-tournament or a feature-building workflow with multiple competing strategies.

2026-06-210

quality-gate-council

Simplereally/old-town

Run a quality gate council before shipping. Use when the user says "ultracode", "quality gate", "gate council", "ship check", "pre-ship review", "before this ships", or explicitly requests /quality-gate-council or a pre-push/pre-merge validation council.

2026-06-210

adversarial-pr-review

Simplereally/old-town

Review a pull request or diff by fanning out one reviewer per changed file and adversarially verifying every issue before reporting. Use this WHENEVER the user asks to review a PR, review a diff, "look over these changes", check a branch before merge, or do a code review of more than a couple of files — even if they don't say the word "workflow". Prefer this over a single-pass review whenever the diff spans several files or the cost of missing a bug is high.

2026-06-210

bug-hunt

Simplereally/old-town

Sweep a codebase for real, reproducible bugs using diverse finder agents in a loop-until-dry, with every candidate confirmed by a multi-lens majority vote. Use this WHENEVER the user wants to "find bugs", do a bug hunt, audit a service/module for defects, hunt for a class of bug (races, null derefs, leaks, off-by-one), or shake out problems before a release — even without the word "workflow". Prefer this over reading files one-by-one whenever coverage across many files matters.

2026-06-210

dependency-audit

Simplereally/old-town

Audit every dependency in parallel — CVEs, breaking changes, maintenance health, and dead/unused packages — then rank by risk and produce an upgrade/remediation plan. Use this WHENEVER the user asks to audit dependencies, check for vulnerable/outdated packages, find unused deps, plan upgrades, or review package.json / requirements.txt / Cargo.toml / go.mod — even without the word "workflow". Prefer this over eyeballing a manifest whenever there are more than a handful of dependencies.

2026-06-210

incident-rca

Simplereally/old-town

Root-cause a hard incident by generating competing hypotheses, gathering evidence for each in parallel, adversarially refuting them, and synthesizing the surviving root cause with a minimal fix and a verification plan. Use this WHENEVER the user is debugging a production incident, a flaky/intermittent failure, a regression, a mysterious stack trace, or asks "why is this happening / what's the root cause" — even without the word "workflow". Prefer this over a single linear debugging pass whenever the cause is non-obvious and the stakes are high.

2026-06-210

name	repo-offensive
description	Run a repo-wide quality offensive for <SCOPE>. Use when the user says "ultracode", "repo offensive", "quality offensive", "codebase cleanup", or explicitly requests a full-coverage quality offensive, pre-release hardening, or inherited-codebase cleanup.
disable-model-invocation	true

Repo Offensive — Repo-Wide Quality Offensive

Default posture: Maximum throughput. Disregard cost. Exhaustiveness and correctness over efficiency. This skill loads into the main conversation. The deterministic workflow lives in .claude/workflows/repo-offensive.js and is invoked manually via the Workflow tool.

How to Invoke

Run the skill (loads instructions):

/repo-offensive

Run the workflow (deterministic execution):

/repo-offensive-workflow

Or invoke the script directly via the Workflow tool:

Workflow({ scriptPath: ".claude/workflows/repo-offensive.js" })

Ad-Hoc Orchestration Guide

When the user asks for a repo offensive, follow this pattern. Treat this as a deterministic workflow, not a normal chat.

1. Mission Framing

First frame the mission:

Quality bar
Non-goals
Stop conditions
Minimum severity
Unsafe paths

2. Recon

Recon the repo:

Stack and package manager
Canonical install/dev/build/test/lint/typecheck/format commands
Architecture map
Generated files
Likely hotspots

3. Discovery Waves

Run repeated discovery waves until dry. Each wave fans out independent specialist agents across:

Type soundness
Runtime crashes
Async/concurrency
Security/auth/input validation/secrets
Tests/flakiness
Performance
UI/accessibility
API/data contracts
Dead code/duplication
Dependency/config/CI drift
Error handling/observability
Architecture boundaries
Docs/spec drift

Every finding must include:

Files
Evidence
Why it matters
Proof/repro
Severity
Fix strategy
Risk

Dedupe aggressively. Send each surviving P0/P1/P2 through an evidence court.

4. Evidence Court

For every P0/P1/P2 finding, spawn 3 skeptics in parallel:

correctness skeptic: Is the claim actually true?
reproduction/tooling skeptic: Attack command assumptions, generated files, stale docs, false positives, missing context.
regression skeptic: Attack the proposed fix. Would it break behavior, tests, public API, performance, or UX?

Accept only findings with 2/3 non-low-confidence real votes.

5. Battle Plan

Convert verified findings into conflict-aware implementation slices:

Group by independent implementation slice, not by category.
Defer anything speculative, high-risk without user intent, or not locally verifiable.
Mark conflicts with dependsOn and parallelSafe.
Prefer small, coherent patches with tests/checks.

6. Surgical Fixes

Implement only low/medium-risk slices unless explicitly told otherwise.

Parallelize independent slices with worktree isolation.
Run relevant checks.
Repair introduced failures.

7. Validation Gauntlet

Finish with a validation gauntlet:

Commands run
Changed files
Fixes made
Remaining failures
Residual risk ledger
Next recommended workflow

Safety Rules

Never auto-invoke. This is manual-only (/repo-offensive).
Never fix high-risk (P0 refactor) issues automatically. Always defer.
Always use worktree isolation for parallel implementation slices.
Always run real checks after edits. If checks fail, attempt repair; if repair fails, revert and mark as pre-existing.
Never modify CI/CD configs, auth secrets, or infrastructure definitions during this offensive.
Always respect maxWaves: 4 — no infinite loops.

Workflow Script

The deterministic workflow is at .claude/workflows/repo-offensive.js. It is a resume-safe, loop-until-dry multi-agent offensive. Reference it from this skill when orchestrating.