OUM353

Graftree — Testing skill for Claude Code

Testing community

Tree-structured, test-first, multi-model solving for hard coding problems.

How to install Graftree

This entry records only its repository, not the path inside it, so there is no exact command to give. Open OUM353/graftree and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Graftree does

Tree-structured, test-first, multi-model solving for hard coding problems. Splits a problem into sub-tasks, writes acceptance tests first, solves each part several times with different models, and merges the verified results. Claude Code skill + CLI.

Alternatives in Testing

Deslop Reborn — One iteration of the IronClaw Reborn de-slop loop — take ONE Reborn crate, fan out parallel review sub-agents 12.6k ★
Debug — You are tasked with helping debug issues during manual testing or implementation 10k ★
Adversarial Spec — A Claude Code plugin that iteratively refines product specifications by debating between multiple LLMs until a 514 ★

README

graftree

**Tree-structured, test-first, multi-model problem solving for hard coding tasks.**

graftree splits a hard problem into a tree of independent sub-tasks. It writes acceptance tests *before* any code, then solves each leaf several times in isolated git worktrees, across different models. The verified results are merged back up into a single best solution. It trades speed and tokens for accuracy.

**⚠ It uses a lot of tokens.** A run makes many model calls: several attempts per sub-task, repairs, reviews and integrations, plus the planning and review done by the closer (e.g. Claude Code) itself. Expect roughly **5–30× the tokens of one agent solving the task directly**. `graftree plan` prints an estimate (`⚠ Cost: expect N–M worker calls`) before you approve, warnings fire during a run when usage gets high, and `report.md` lists what was spent. The closer's own usage is not included in those numbers; check it in your agent (for Claude Code, `/cost`).

**Use it where a single agent tends to get things subtly wrong:** many interacting requirements, tricky edge cases, a vague spec, an unexplained bug. On a small, clearly specified task a single agent is usually just as accurate. In our [kvstore example](examples/kvstore/) both scored 25/25 on a hidden test suite, and graftree cost 6× the worker calls. [examples/minisheet](examples/minisheet/) is a harder benchmark built to separate the two.

The agent that invokes it (Claude Code by default) is always the **closer**: it makes every final decision. Other models, such as DeepSeek via [CommandCode](https://www.npmjs.com/package/command-code), anything on OpenRouter, or local models, can do the planning, solving and review work.

            [problem]
                │  triage → plan + acceptance tests
                ▼
        ⏸ human approves the plan         (nothing is spent before this)
                │  tests locked by hash
       ┌────────┼────────┐
      [A]      [B]