Suggérer

Spec, TDD and review for coding agents

For developers and tech leads who want coding agents to work like careful engineers: guardrails first, then interview, spec, test-first build, debug, verify, review. Most run in Claude Code or Codex; the three Anthropic plugins are Claude Code only. Agent review assists; a person approves merges.

12 ressourcesMise à jour le 2026-10-12

  1. CC Safety NetPluginSet up first: blocks git reset --hard, rm -rf on risky paths, reads of .env or SSH keys. Node 18+; not a sandbox.kenryu42/cc-safety-net1.6k
  2. Anthropic Sandbox Runtime (srt)CLISet up first: Anthropic's OS-level file and network limits. Reads open by default; Linux needs bubblewrap; preview.anthropics/sandbox-runtime5.6k
  3. Security GuidancePluginAlways on: flags risky edits; Opus reviews diffs and commits (extra usage). Claude Code 2.1.144+, Python 3.8+.anthropics/claude-plugins-official40.1k
  4. Grill MeSkillStep 1: the agent interviews you about the plan until every decision is settled. You start it; install grilling too.mattpocock/skills294k
  5. Spec KitCLIStep 2: GitHub's toolkit turns your answers into a spec, plan and tasks, then builds against them. Python 3.11+ and uv.github/spec-kit142k
  6. Context7MCPStep 3: feed the agent current, version-specific library docs. Upstash service; a free API key raises limits.upstash/context763.1k
  7. Test-Driven DevelopmentSkillStep 4: strict red-green-refactor; code written before a failing test is deleted. Also in the Superpowers plugin.obra/superpowers301k
  8. Diagnosing BugsSkillStep 5, when it breaks: repro loop, hypotheses, fix, then a regression test. Redacts credentials in output.mattpocock/skills294k
  9. Verification Before CompletionSkillStep 6: the agent must run the full checks and read the output before it claims it is done. Also in Superpowers.obra/superpowers300k
  10. Code ReviewPluginStep 7: four parallel agents review a GitHub PR and keep issues scored 80+. Claude Code only; needs a logged-in gh.anthropics/claude-code151k
  11. PR Review ToolkitPluginStep 7: six reviewers for tests, silent failures, comments, type design, style and simplification. Claude Code only.anthropics/claude-code151k
  12. Supply Chain Risk AuditorSkillBefore merge: flags advisories, abandoned repos, install scripts in npm, PyPI, Go deps. Needs uv; skips pnpm/yarn locks.trailofbits/skills7.6k
Version Markdown (pour les agents)