One open contract for bounded scope, executable proof, sealed authority, portable handoff, and independent acceptance.
概览
Agents can write code. TASK-SPEC makes them earn done . One open contract for bounded scope, executable proof, sealed authority, portable handoff, and independent acceptance. A prompt tells an agent what you want. A Task-Spec also records what the agent may change, what observable behavior counts as success, what evidence must exist, who authorized that exact contract, and what an independent gate must verify afterward. The core stops at that boundary. It does not host models, store credentials, create a sandbox, or turn a weak eval into a wise oracle. Version 3.9 adds as an optional runtime: it can route and execute already-authorized leaves, but it cannot change the contract or accept work on its own terms.
README
以下为原文 README 节选,完整内容请查看原文。
The missing contract between intent and execution
A prompt tells an agent what you want. A Task-Spec also records what the agent may change, what observable behavior counts as success, what evidence must exist, who authorized that exact contract, and what an independent gate must verify afterward.
| Without Task-Spec | With Task-Spec |
|---|---|
| “Implement search and test it.” | One atomic leaf with explicit paths, behavior, evals, budget, and owner |
| The agent decides what “done” means while working | Humans review the contract; runnable evals decide the technical result |
| Scope changes disappear into the conversation | HMAC v3 seals a canonical task revision, including future fields by default |
| Every harness receives a different interpretation | Every harness receives the same attempt, revision, base commit, closure, scope, and budget |
| “Tests pass” is the final claim | Acceptance reruns proof, checks Git history and the worktree, binds receipts, and writes an auditable record |
The core stops at that boundary. It does not host models, store credentials, create a sandbox, or turn a weak eval into a wise oracle. Version 3.9 adds TaskMesh as an optional runtime: it can route and execute already-authorized leaves, but it cannot change the contract or accept work on its own terms.
Prove it in one command
After installation, run a complete lifecycle in a disposable repository:
$ taskspec demo
推荐工具
换一个关键词,或者移除筛选条件。
安装
npx skillfish add haraujo21/task-spec