에이전트 코딩 실패 막는 '에반플로우(TDD)'

EvanFlow A TDD-driven iterative feedback loop for software development with Claude Code. 16 cohesive skills + 2 custom subagents walk an idea from brainstorm through implementation, with checkpoints throughout where you stay in control. One entry point: say "let's evanflow this" and the orchestrator runs the loop. brainstorm → plan → execute (sequential or parallel) → tdd → iterate → STOP The loop is conductor, not autopilot : real checkpoints at design approval, plan approval, and after iteration. The agent stops short of every git operation and waits for your direction. No auto-commits. No forced ceremony. No "must invoke a skill" tax. Quick Install The recommended path — Claude Code's plugin marketplace: /plugin marketplace add evanklem/evanflow /plugin install evanflow@evanflow Restart, then try: "Let's evanflow this — I want to add a small feature that does X." evanflow-go fires and walks the loop. The git-guardrails hook auto-activates with the plugin (no settings.json edit needed). Skills appear under the evanflow: namespace (e.g., /evanflow:evanflow-go ). See Installation below for two alternative paths. What Makes It a Feedback Loop The loop is built around discipline that compounds across iterations , not single-shot generation. Every step has a checkpoint that gates the next: Brainstorm clarifies intent, proposes 2–3 approaches with embedded grill (stress-test) → you approve the design Plan maps file structure first (deep modules, deletion test) → you approve the plan Execute runs task-by-task with inline verification → blockers stop the loop and surface to you TDD is vertical-slice only: one failing test → minimal impl → repeat. Tests verify behavior through public interfaces, so they survive refactors Iterate re-reads the diff with fresh eyes, runs quality checks, screenshots UI changes, and runs against a Five Failure Modes checklist (hallucinated actions, scope creep, cascading errors, context loss, tool misuse). Hard cap of 5 iterations STOP. Report. Await your direction. The agent never auto-commits, never auto-stages, never proposes a PR For plans with 3+ truly independent units, the loop forks into a parallel coder/overseer orchestration : one coder per unit (using vertical-slice TDD with a RED checkpoint), one overseer per coder (read-only review subagent that can't modify code), plus an integration overseer that runs named integration tests at every touchpoint. The integration tests are the executable contract — interfaces can't drift if both sides have to satisfy the same passing test. Hard Rules Baked Into the Loop Several rules come from 2025-2026 industry research on agentic coding failure modes and are baked into every skill: Never invent values — file paths, env vars, IDs, function names, library APIs. If unsure, the agent stops and asks. (Action-hallucination is the most dangerous agent failure.) Assertion-correctness warning — research shows 62% of LLM-generated test assertions are wrong. Both evanflow-tdd and the overseer review explicitly check whether a one-character bug in the implementation would still let the assertion pass. Watch for context drift — evanflow-compact triggers when symptoms appear (re-asking established questions, contradicting earlier decisions). Industry data: ~65% of enterprise AI coding failures trace to context drift, not raw token exhaustion. Five Failure Modes pass in iterate + overseer review — explicit check against hallucinated actions, scope creep, cascading errors, context loss, tool misuse. No skill tax — ad-hoc questions don't require a skill invocation. Skills are tools, not a tollbooth. The Skill Set Default Loop (5 skills) Skill Purpose evanflow-brainstorming Clarify intent, propose 2–3 approaches with embedded grill (stress-test). Mockup quick-mode for visual-only requests. evanflow-writing-plans File structure first, bite-sized tasks, embedded grill. Step 2.5 offers evanflow-coder-overseer if the plan is parallelizable. evanflow-executing-plans Task-by-task with inline verification. Step 0 re-offers parallel path. Hands off to iterate, then STOPS. evanflow-tdd Vertical-slice TDD. One test → one impl → repeat. Behavior through public interface. Assertion-correctness warning. evanflow-iterate Self-review loop after implementation. Re-read diff, fix issues, run quality checks, screenshot UI (via headless Chromium). Five Failure Modes checklist. Hard cap of 5 iterations. Special-Purpose (8 skills) Skill Purpose evanflow-go Single entry point. Say "let's evanflow this" and it walks the whole loop. evanflow-glossary Extract canonical domain terms into CONTEXT.md . Flag ambiguities and synonyms. evanflow-improve-architecture Surface refactor opportunities via the deletion test + deep-modules vocabulary. evanflow-design-interface "Design it twice" — spawn 3+ parallel sub-agents with radically different constraints, compare on depth/simplicity/efficiency. evanflow-debug Root-cause discipline. Hypothesis stated explicitly, embedded grill before fixing, failing test first. evanflow-review Both halves of code review (giving + receiving). Don't capitulate to feedback you can't justify. evanflow-prd Synthesize a PRD from existing context. For substantial new features. evanflow-qa Conversational bug discovery → issue draft. Asks before filing. Cross-Cutting (1 skill) Skill Purpose evanflow-compact Long-session context management. Strategies for proactive summarization at clean boundaries. Drift symptoms checklist. Meta (1 skill) Skill Purpose evanflow The index. Shared vocabulary + when to invoke each evanflow-* skill. Custom Subagents (2) In agents/ — invoked via Agent tool with subagent_type: parameter: Subagent Tool restrictions Purpose evanflow-coder Read, Edit, Write, Glob, Grep, Bash, TodoWrite Implementation subagent for evanflow-coder-overseer . Tools + system prompt prevent git ops, out-of-scope edits, value hallucination. evanflow-overseer Read, Grep, Glob (no Edit/Write/Bash) Read-only review subagent. Tools physically enforce "report findings, never fix." Bundled Hook hooks/block-dangerous-git.sh — PreToolUse hook that blocks destructive git ops ( git push , git reset --hard , git clean -f , git branch -D , git checkout . , git restore . ). Auto-activates with the plugin install path. Hard Rules (apply to every skill) Never auto-commit, never auto-stage, never auto-finish. Every git write op requires you to explicitly ask in the current turn. Never invent values. File paths, env vars, IDs, function names, library APIs — if unsure, the agent stops and asks. No skill tax. Ad-hoc questions don't require a skill invocation. Skills are tools, not a tollbooth. No forced spec/plan paths. Files live where you want them. Verify before claiming done. Quality checks (typecheck, lint, test) run before any "done" report. Requirements Claude Code (any recent version) Bash — for the bundled hook script (Linux, macOS, or Windows + WSL) jq — used by the hook script to parse Claude's JSON tool input. Install via apt install jq , brew install jq , or your platform's package manager. If jq is missing, the guardrail hook fails silently and dangerous git ops are NOT blocked. Optional but recommended: chromium or google-chrome — for evanflow-iterate 's visual verification of UI changes ( chromium --headless --screenshot=... ). Falls back gracefully if missing — the skill flags it and asks you to verify visually. Installation Three paths, in priority order. All three end with the same skill set in your .claude/skills/ . The plugin path additionally auto-wires the guardrail hook. Path 1 — Claude Code Plugin Marketplace (recommended) This is the cleanest install. Skills, agents, AND the guardrail hook all activate automatically. /plugin marketplace add evanklem/evanflow /plugin install evanflow@evanflow Restart Claude Code (or /reload-plugins ). Skills appear namespaced as /evanflow:evanflow-go , /evanflow:evanflow-tdd , etc. Auto-invocation via "let's evanflow this" still works regardless of namespace. To u