skill-up

Evaluate Agent Skills, agents, and workspaces. Evolve Skills with evidence.

English | 中文

## Overview **skill-up** is a CLI for evaluating Agent Skills, agents, and their workspaces. It runs declarative cases against an Agent Engine, grades the response and workspace changes, and produces reports locally or in CI. - **Skill evaluation** measures a Skill's behavior across cases and can compare runs with and without the Skill. - **Agent evaluation** runs the same cases without installing a Skill, so you can assess an engine or custom agent on its own. - **Workspace evaluation** checks how an agent works with files and repositories, using case fixtures or an existing local workspace. - **Skill evolution** uses **skill-upper** to inspect failures, repair or expand the eval suite, and rerun skill-up through conversation. ![The Skill evaluation and evolution workflow with skill-upper](docs/public/skill-up-overview.png) ## Features - **Eval-to-Evolution Loop with skill-upper**: Create evals through natural conversation, diagnose failures, automatically repair or expand cases, and rerun skill-up until the eval suite evolves. - **Declarative Eval Config**: Define evaluation environment, engine, model, and cases through YAML (`eval.yaml` + `cases/*.yaml`). - **Environment Provisioning**: Prepare a local, Docker, or OpenSandbox runtime for each case, install configured real or mocked MCP servers into supported Agent Engines, and upload `context.repo_fixture` and `context.files` into the workspace. - **Skill-Optional Runs and Workspace Cases**: Evaluate an agent without a Skill, prepare repository fixtures per case, or run against an existing local workspace with `--workspace`. - **Multi-Engine Support**: Works with Qoder CLI, Claude Code, and Codex as built-in Agent Engines, plus user-defined agents via `engine.custom` (local commands or HTTP services — see [docs/design/custom-engine.md](docs/design/custom-engine.md)). - **DeepSeek Harness Plugin**: Use the bundled `skill-upper` Skill, opt-in durable observation capture, approval-gated regression cases, isolated runs, and status comparison to verify evidence-backed Skill improvements. See [`plugins/dsh-skill-up`](plugins/dsh-skill-up/README.md). - **Multi-Turn Conversations and User Simulation**: Run fixed turns, mix scripted and simulated replies, or let an independent user model drive follow-ups and decide when to stop. - **Role-Specific Engines and Models**: Configure the evaluated agent, Agent Judge, and user simulator separately, including independent judging of custom agents. - **Flexible Judging**: Supports `rule_based`, `script`, and `agent_judge` evaluation strategies. - **Structured Reports**: Outputs Anthropic-compatible `grading.json`, `benchmark.json`, `benchmark.md`, plus `result.json`, JUnit XML, and HTML reports. - **Anthropic Compatible**: Import `evals.json` via `skill-up import`, or auto-detect with `--auto`. - **CI-Ready**: Designed for local development and continuous integration pipelines. ## Why skill-up [Harness engineering](https://openai.com/index/harness-engineering/) treats an agent's environment, tools, constraints, and feedback loops as parts of the system to design and improve. [Behavioral evals](https://developers.googleblog.com/the-anatomy-of-harness-engineering-how-to-evaluate-iterate-and-guard-ai-coding-agents/) make expected agent actions observable and help catch regressions. `skill-up` contributes evaluation and feedback to this work: repeatable prompts and workspace setups, checks on responses, tool calls, and file changes, and structured reports. For Skills, the official [Agent Skills evaluation guide](https://agentskills.io/skill-creation/evaluating-skills) describes a further loop: run with and without the Skill, grade outputs, aggregate results, and iterate. `skill-up` supports these workflows: - Replaces ad hoc run folders with a declarative `eval.yaml` + `cases/*.yaml` format. - Closes the improvement loop: skill-upper can interpret failed reports, repair or add eval cases, and drive the next skill-up run through conversation. - Automates workspace setup, optional Skill installation, Agent Engine invocation, judging, and report generation. - Supports multiple engines (`claude_code`, `codex`, `qodercli`, `qwen_code`) instead of tying the workflow to one client. - Keeps compatibility with Anthropic-style `evals.json` while adding richer judges, CI-friendly commands, and structured reports. ## Quick Start: Evolve a Skill with skill-upper For Skill evolution, use **skill-upper**, the Agent Skill shipped in this repository. It lets your AI agent create evals, run skill-up, understand failures, fix the Skill or its evals, add regression coverage, and repeat the loop through conversation. For repository development, `skills/skill-upper/` is the only editable source. Plugin bundles are generated into ignored build directories rather than tracked copies or symlinks; run `make bundle-plugins` when preparing them locally. `make package-plugins VERSION=` creates self-contained Codex and DSH archives. Tagged GitHub releases attach both archives and their checksums; this repository does not publish either plugin to an external marketplace or npm. ### 1. Install skill-upper ```bash # Codex, global install npx skills add https://github.com/alibaba/skill-up/tree/main/skills/skill-upper -g -a codex -y # Claude Code, global install npx skills add https://github.com/alibaba/skill-up/tree/main/skills/skill-upper -g -a claude-code -y ``` You normally do not need to install skill-up first. skill-upper checks for the CLI when it runs and guides the agent through installation if needed. ### 2. Create and run the first evals Open a project that contains your Skill's `SKILL.md` in Codex, Claude Code, or another compatible Agent, then ask: ```markdown Use skill-upper to evaluate this Skill. Read SKILL.md, identify its most important behaviors, create realistic eval cases with appropriate judges, validate the configuration, and run skill-up. Summarize the results and the highest-impact failures. ``` skill-upper creates the declarative eval suite and drives the CLI for you: ```text my-skill/ SKILL.md evals/ eval.yaml cases/ .yaml my-skill-workspace/ iteration-1/ result.json ``` ### 3. Fix, regress, and iterate Continue in the same conversation: ```markdown Use skill-upper to improve this Skill from the latest evaluation results. Follow its evolution workflow, compare the original and candidate, add regression coverage, and summarize which changes the evidence supports. ``` The comparison and acceptance rules live in [skill-upper's evolution workflow](skills/skill-upper/SKILL.md#step-8-evolve-the-skill-when-requested). This prompt invokes that workflow; keeping the original is a valid outcome when a candidate does not show a supported improvement. ### Prefer manual setup? You can still install the CLI directly and hand-write `eval.yaml` and case files: ```bash curl -fsSL https://raw.githubusercontent.com/alibaba/skill-up/main/install.sh | bash ``` See the official documentation for [Getting Started](https://alibaba.github.io/skill-up/guide/getting-started), [Writing Evals](https://alibaba.github.io/skill-up/guide/writing-evals), [CLI Reference](https://alibaba.github.io/skill-up/guide/cli-reference), and [User Configuration](https://alibaba.github.io/skill-up/guide/user-config). Windows-specific setup and limitations are covered in the [Windows guide](https://alibaba.github.io/skill-up/guide/windows). ## Evaluate an agent or workspace without a Skill An agent-only eval does not need `SKILL.md`. Set `skills: []` explicitly: when `skills` is omitted and skill-up finds a `SKILL.md` above the config, it installs that Skill automatically. Put `eval.yaml` and case files under `evals/`, then pass the config path explicitly. Case and fixture paths resolve from the discovered Skill root, or from the directory containing `evals/` if none is found. ```yaml # evals/eval.yaml schema_version: v1alpha1 environment: type: none skills: [] engine: name: codex cases: files: - evals/cases/agent-response.yaml ``` ```yaml # evals/cases/agent-response.yaml input: prompt: Reply with READY. expect: must_contain: [READY] ``` ```bash skill-up validate ./evals/eval.yaml skill-up run ./evals/eval.yaml ``` For repository tasks, a case can use `context.repo_fixture` to prepare a fresh workspace and `expect` or a judge to check the result. To evaluate an agent in an already prepared local directory, use a case written for that directory and run `skill-up run ./evals/eval.yaml --workspace /path/to/project --parallelism 1`. The existing directory is preserved, but the agent and case setup can change its contents. This option requires `environment.type: none`, one case at a time, and benchmark mode off. See [Writing Evals](docs/guide/writing-evals.md) and the [CLI Reference](docs/guide/cli-reference.md#skill-up-run). ## Multi-Turn Conversations and Role Configuration A simulated user can react to what the agent actually says, so follow-up messages need not be scripted in advance. For example, evaluate a Twenty Questions player: - The **evaluated agent** asks natural-language yes/no questions and guesses an object. - The **simulated user** knows the hidden object and answers from the conversation history. - An **independent Agent Judge** reviews the complete dialogue against explicit criteria. This **case YAML** excerpt lets the simulator drive all replies after the opening: ```yaml id: umbrella title: Guess a hidden object through simulated user replies user_simulator: scenario: | You host Twenty Questions. The hidden object is an ordinary folding umbrella. Answer questions using common knowledge, with only yes, no, or unsure. Do not reveal the object or volunteer hints. Stop immediately when the player guesses correctly, or after the final permitted guess. input: prompt: | I have a common everyday object in mind. Guess it by asking yes/no questions; I will answer yes, no, or unsure. Ask only one question per turn, and ask at least two questions before guessing. You have 20 questions or incorrect guesses in total, followed by one final guess. Use natural language. constraints: max_turns: 22 ``` The hidden answer belongs to the simulator's scenario, not the player's prompt. For example, a question about rain protection can lead to a different next question than one about electronics; the simulator responds to the actual question. The 22-turn runtime cap leaves room for 20 questions or incorrect guesses, a final guess, and a closing response. The judge separately checks the game's rules; the runtime cap alone does not enforce them. Other conversation patterns remain available: | Pattern | Case configuration | Behavior | | --- | --- | --- | | Autonomous simulation (above) | `input.prompt` plus `user_simulator.scenario`, without `input.turns` | Generate subsequent user messages and decide when to stop | | Mixed | `content` and `respond` in `input.turns`, plus `user_simulator.scenario` | Generate selected user replies from the actual conversation history and turn instruction | | Fixed | `input.turns[].content` | Send predefined user messages in order | Real turn-by-turn interaction requires session resumption. `claude_code`, `codex`, `qodercli`, and `opencode` support it; custom engines must set `engine.custom.conversation_mode: stateful` and resume using `session_id`. Ordinary fixed-turn cases fall back to a batch prompt on engines without resumption; simulated cases require resumption. The opening prompt counts as turn one in an autonomous conversation, and `max_turns` bounds the total. A simulator stop ends the dialogue; the configured judge still determines PASS or FAIL. ### Select the evaluated agent, judge, and simulator independently | Role | Configuration in `eval.yaml` | Independence | | --- | --- | --- | | Evaluated agent | `engine.name`, `engine.model`, optional `engine.custom` | Executes tasks through a built-in engine or your own local command / HTTP service | | Agent Judge | `judge.engine`, or each member's `engine` in `judges[]` | Can use a separate engine, model, provider, and endpoint; rule and script judges need no model | | Simulated user | `user_simulator.provider`, `protocol`, `model` | Uses its own model connection to generate replies from the scenario and history; does not select an Agent CLI engine | This **eval.yaml** excerpt shows role configuration. Model IDs are placeholders; replace them with IDs supported by each provider and configure case files and environment for your evaluation. These criteria grade the game above; `pass_threshold: 1.0` requires all criteria to pass: ```yaml engine: name: codex model: provider: openai name: your-agent-model-id judge: type: agent_judge engine: name: claude_code model: provider: anthropic name: your-judge-model-id criteria: - "The player asked at least two questions before guessing and correctly identified an umbrella" - "The player asked only one question per turn and respected the 20-attempt limit plus one final guess" - "The simulated user answered consistently with a folding umbrella and did not reveal the answer before a correct guess" pass_threshold: 1.0 context: transcript: file_ref user_simulator: provider: simulation protocol: openai model: your-simulator-model-id timeout_seconds: 60 ``` `judge.engine` does not inherit the evaluated agent's model, CLI overrides, or connection configuration; prepare the required engines and credentials separately. It is mutually exclusive with legacy `judge.model`, which reuses the evaluated agent's engine. The simulator above resolves `SIMULATION_API_KEY` and `SIMULATION_BASE_URL`, or the same provider in the credentials file; `protocol` supports `openai` (Chat Completions) and `anthropic` (Messages). An independent judge still uses the case runtime; there is no `judge.environment`. The host agent/model that drives skill-upper's diagnosis and edits is configured in your host application, outside these evaluation fields. See [Writing Evals](https://alibaba.github.io/skill-up/guide/writing-evals) for complete fields, per-turn assertions, and the session support matrix. See the [user simulator example](examples/user-simulator/README.md) for runnable configurations and validation scope, and the [Custom Engine design](docs/design/custom-engine.md) for custom agent integration. ## User config skill-up auto-loads an optional user-level config that supplies default OpenTelemetry env vars and per-environment runtime kwargs. The embedded defaults are empty; downstream consumers maintain their own config file. ### Discovery chain (lowest to highest precedence) ``` embed (empty) < user (~/.config/skill-up/config.yaml) < project ($PWD/.skill-up.yaml) < explicit (--config) ``` | Source | Path | | ---------- | ------------------------------------------------------------------------------------------------------- | | `embed` | empty `Config{}` — no vendor defaults baked in | | `user` | `$SKILL_UP_CONFIG`, else `$XDG_CONFIG_HOME/skill-up/config.yaml`, else `~/.config/skill-up/config.yaml` | | `project` | `$PWD/.skill-up.yaml` | | `explicit` | `--config ` (must exist) | Missing files at the `user` and `project` layers are silently skipped; a missing `--config` path is a hard error. A corrupt config at any layer also fails the run. ### Quickstart ```bash skill-up init # writes a template to ~/.config/skill-up/config.yaml (XDG-aware) skill-up init --local # writes a template to $PWD/.skill-up.yaml skill-up init --print # prints the template to stdout skill-up init --force # overwrite an existing file skill-up init --config foo.yaml # reads foo.yaml, writes it to ~/.config/skill-up/config.yaml skill-up init --config foo.yaml --local # reads foo.yaml, writes it to $PWD/.skill-up.yaml ``` With `--config `, `init` reads that file (validating it as a skill-up config) and writes its raw bytes to the target — comments and formatting are preserved. Without `--config`, `init` writes a commented YAML template. ### Schema ```yaml schema_version: v1alpha1 kind: SkillUpConfig telemetry: service_name: skill-up # OTEL_SERVICE_NAME traces_exporter: otlp # OTEL_TRACES_EXPORTER traces: endpoint: http://localhost:4317 # OTEL_EXPORTER_OTLP_TRACES_ENDPOINT (4317 for grpc, 4318/v1/traces for http/protobuf) protocol: grpc # OTEL_EXPORTER_OTLP_TRACES_PROTOCOL (grpc | http/protobuf); skill-up defaults to grpc resource_attributes: # serialized into OTEL_RESOURCE_ATTRIBUTES deployment.environment: local verbose: false # if true, also enables OTEL_LOG_* payload capture env: # arbitrary defaults, applied only-if-unset OTEL_EXPORTER_OTLP_HEADERS: authorization=${OTLP_TOKEN} runtime_kwargs: # keyed by environment.type opensandbox: base_url: http://localhost:8080 # extensions: '{}' ``` ### Precedence For environment variables: any value already set in the process environment wins; the config only fills in missing keys. For `runtime_kwargs`: explicit `--runtime-kwarg` on `run` > `eval.yaml` `environment.kwargs` > user-config `runtime_kwargs[environment.type]`. ### Secrets Prefer `${ENV_VAR}` references inside the config file rather than baking secret literals. The redaction mechanism (`userconfig.Redact`) masks fields tagged `secret:"true"` when printing; currently no Config field carries the tag, but the mechanism is in place for future fields. ## Importing `evals.json` Use `skill-up import` to migrate an Anthropic-compatible `evals.json` into the YAML layout used by this repo: ```bash skill-up import ./evals/evals.json --output ./evals ``` ## CLI Overview | Command | Description | | ------------------------------------ | ------------------------------------------- | | `skill-up run [path]` | Run evaluation cases and produce reports | | `skill-up validate [path]` | Validate `eval.yaml` and case files | | `skill-up list-cases [path]` | List all cases referenced by the config | | `skill-up report ` | Generate reports from a previous run | | `skill-up import ` | Import Anthropic `evals.json` to YAML cases | | `skill-up debug judge ` | Debug judge module with a JSON input | | `skill-up debug report ` | Debug report module with a JSON input | ## GitHub Action Run your Agent Skill evals in CI on every pull request — and check the same skill **across engines** (`claude_code` / `codex` / `qodercli` / `qwen_code`) in one step. This repo ships an action at its root ([`action.yml`](action.yml)): ```yaml # .github/workflows/skill-eval.yml name: Skill Eval on: pull_request: paths: ['skills/**', 'evals/**', '**/SKILL.md'] jobs: eval: runs-on: ubuntu-latest # Docker container action — Linux only steps: - uses: actions/checkout@v4 - uses: alibaba/skill-up@main # see "Versioning" below with: engine: claude_code # or codex / qodercli / qwen_code; empty = let eval.yaml decide api-key: ${{ secrets.ANTHROPIC_API_KEY }} base-url: https://api.anthropic.com # your model endpoint skill-target: evals/eval.yaml ``` Requirements for the caller: a **Linux** runner (it's a Docker container action), and your model credential stored as a repo secret. The runner image is public, so no extra registry auth is needed. Key inputs: `engine`, `model`, `provider`, `api-key`, `base-url`, `skill-target`, `parallelism`. The action prebuilds skill-up + the three engine CLIs into its runner image, so a run is just "pull image, eval". See [`action.yml`](action.yml) for the full input/output reference. ### Bundled skill-up version The container image includes a deliberately pinned skill-up CLI version. The version is not resolved from `latest` when a workflow starts, so a given image digest always runs the same CLI. The `skill-up-version` input is only a fallback for a custom image that does not already contain the `skill-up` binary. The official image contains the binary, so this input cannot override its bundled version. To find the effective version, check the `skill-up --version` line in the Action log. Publishing a new skill-up CLI release does not automatically update the GitHub Action image. Maintainers must synchronize the pinned version, publish and test a new runner image, and update the image digest in `action.yml`. The complete maintainer procedure is documented in the [CI maintenance runbook](docs/guide/ci-maintenance.md#manual-skill-up-version-synchronization). The production Action must use an immutable `sha256:` image digest. Do not replace it with `skill-up-runner:latest`; a mutable tag would allow existing Action references to change behavior without a repository commit. ### Versioning `uses:` points at any git ref that contains `action.yml`. Pin a **release tag** (the first release that includes the action onward) or a commit SHA for stability; `@main` always tracks the latest. Release tags published **before** the action was added do not contain `action.yml` and cannot be used as the ref. A CLI release tag captures the `action.yml` and runner-image digest that existed when that tag was created. Because the current runner image is refreshed manually after CLI release assets become available, do not assume that a CLI tag automatically contains an Action image with the same CLI version. Until a separate Action release tag process is introduced, use a post-refresh commit SHA for an immutable reference or `@main` when intentionally following Action updates. ## Project Structure ```text skill-up/ ├── cmd/skill-up/ # CLI entry point ├── internal/ # Private implementation packages │ ├── cli/ # Commands and flags │ ├── evaluator/ # Case execution and judging orchestration │ ├── evalevent/ # Evaluation event model, lifecycle, publisher, and JSONL sink │ ├── runner/ # End-to-end run orchestration │ └── report/ # JSON, JUnit, HTML, and benchmark reports ├── pkg/ # Publicly importable APIs ├── plugins/ # Agent host plugin bundles │ ├── codex-skill-up/ # Codex adapter; release bundles add shared assets │ └── dsh-skill-up/ # DSH bundle with a generated dist/ Skill copy ├── schemas/ # Evaluation-event and host-neutral observation contracts ├── skills/skill-upper/ # Canonical source for the distributable workflow Skill ├── docs/ # VitePress documentation ├── e2e/ # End-to-end tests and fixtures └── examples/ # Example evaluations and fixtures ``` See [CONTRIBUTING.md](CONTRIBUTING.md#directory-overview) for the complete directory map and architectural responsibilities. ## License Apache License 2.0 — see [LICENSE](LICENSE).