Coding Discipline — Testing skill for Claude Code

coding-discipline

**Chinese** → [README.zh-CN.md](README.zh-CN.md)

**Guardrails, not a workflow.** Quiet on routine work; firm when a named failure mode is at risk.

AI coding agents tend to fail in two expensive ways:

1. They drift off-goal. A decision is confirmed in chat, but the next session reads a stale spec and builds the old target.
2. They skip steps. They code before requirements are aligned, patch symptoms before finding the root cause, or claim success without current evidence.

`coding-discipline` is a lightweight, workflow-agnostic plugin that addresses both:

• Spec sync writes confirmed goals, non-goals, constraints, and acceptance criteria back to the current source of truth when later sessions would otherwise act on stale intent.
• Risk-triggered discipline invokes firm engineering guardrails only when their conditions match: design alignment, TDD, systematic debugging, prioritized review, evidence-before-done, disciplined git flow, and context hygiene.

It does **not** replace Plan mode, launch subagents, orchestrate worktrees, choose models, or force every task through a spec workflow. Small changes stay small. When a risk trigger matches, the corresponding skill stays firm.

8 skills + 2 hooks. Install once, apply globally. Supports Claude Code and Codex.

Why this exists

[Superpowers](https://github.com/obra/superpowers) proved that engineering discipline can live in agent skills — this plugin even reuses its cross-platform hook pattern (see Credits). But Superpowers is a complete workflow: its bootstrap asks the agent to route every response, even a clarifying question, through the skill system first.

coding-discipline starts from the opposite default. Skills stay dormant until a named failure mode is at risk; routine work runs untouched. The fixed context injected per session is ~1.2k tokens and cannot silently grow — a CI test fails the build when it does. The cost that actually differs between pl