Skillet

Spec-driven agent skills with mechanical evals.

Skillet is a small CLI that manages three artifacts per skill and proves the skill works:

• spec.md — the source of truth: intent, triggers, behaviors with WHEN/THEN scenarios, and constraints, in a strict-but-tiny markdown grammar humans review in PRs.
• SKILL.md — the instruction text agents load, rendered from the spec by your coding agent.
• evals/cases/*.yaml — declarative eval cases that run the skill through a real coding agent (codex or claude CLI, or any CLI you configure) in a fresh workspace and check what it actually did.

Skillet itself makes zero LLM calls. It scaffolds, validates, serves writing instructions to your agent, and runs evals mechanically. All generation happens in the coding agent you already use, driven by the skillet-authoring skill — so upgrading skillet upgrades every agent's behavior, and nothing here needs an API key.

Install

npm install -g @sentry/skillet   # or: npx @sentry/skillet
skillet init                     # installs the skillet-authoring skill for all your
                                 # agents via @sentry/dotagents (user scope, ~/.agents);
                                 # asks first — or manage it yourself, see below

Quickstart

skillet new commit-conventions   # scaffold spec.md + evals/ layout
cd commit-conventions
# fill in spec.md by hand, or ask your agent to author it (skillet-authoring skill)
# then have the agent render SKILL.md + eval cases
skillet validate                 # grammar, frontmatter, case schema, coverage — no LLM
skillet eval --trials 3 --baseline

The last command is the point of the tool: every case runs through a real agent with the skill installed and (with --baseline) without it, and skillet reports per-behavior pass rates and lift — the difference the skill actually makes. Illustrative output:

Behaviors:
  conventional-subject: 100% (3/3) | baseline 33% | lift +67%
  branch-safety:        100% (3/3) | baseline 0%  | lift +100%

Zero lift is a finding too: it means your agent already does this without the skill (note that harness CLIs still load your global agent config, so baseline measures your configured agent, not a bare model).

Under the hood, evals run through an embedded vitest-evals engine — no config or extra dependencies on your side. Add --report results.json to get a Vitest JSON artifact you can inspect locally (npx vitest-evals serve results.json) or publish in CI with the getsentry/vitest-evals GitHub Action.

The spec

# Commit Conventions

## Intent

Make the agent produce disciplined git commits...

## Triggers

- **SHOULD** trigger when the user asks to commit changes
- **SHOULD NOT** trigger when the user asks to review a diff

## Behaviors

### Behavior: Conventional subject

The agent SHALL write commit subjects as `<type>(<scope>): <description>`...

#### Scenario: Committing a staged bug fix

- **WHEN** the workspace has a staged bug fix and the user asks to commit
- **THEN** the commit subject starts with `fix` and stays under 70 characters

## Constraints

### Constraint: No history rewriting

The agent MUST NOT amend, rebase, or force-push unless explicitly asked.

Rules the validator enforces: every behavior has at least one scenario (####, exactly four hashes), behavior names slugify to unique ids (conventional-subject), and every scenario has WHEN/THEN bullets. Errors come with line numbers and fix hints.

Evals

One YAML file per case in evals/cases/, linked to a spec behavior by id:

behavior: conventional-subject
prompt: |
  I fixed the null check in app.js — please commit my staged change.
setup: |
  git init -q -b main && git add -A ...
checks:
  - shell: git log -1 --format=%s | grep -Eq '^(feat|fix|chore)...'
  - file_exists: some/artifact
  - judge: The commit message accurately describes the null-check fix.
trials: 1
timeout: 300

• Each trial gets a fresh temp workspace: the optional fixture: directory (evals/fixtures/<slug>/) is copied in, then setup: runs (the script itself never enters the workspace).
• file_exists and shell checks run in the workspace after the agent finishes — check artifacts, not phrasing. Transcript regexing is deliberately unsupported.
• judge: checks are graded by the same harness CLI with a strict pass/fail verdict protocol, and only run after all deterministic checks pass. No API keys, no thresholds; repeatability comes from --trials.
• A behavior with no case is a validation warning; a case referencing an unknown behavior or missing fixture is an error.

Harnesses

The default harness is codex (codex exec); claude (claude -p) is built in. Pick per run with --harness, add a model with a suffix (--harness claude:sonnet, harness: codex:gpt-5 in config), or configure any CLI in .skillet.yaml:

harness:
  name: my-agent
  command: "my-agent run --dir {workspace} {prompt}"
  skill_dir: "{workspace}/.my-agent/skills"   # where the skill gets installed

Skill installation uses each agent's native mechanism: .claude/skills/ for claude, the workspace AGENTS.md for codex (which has no skill mechanism), skill_dir for custom harnesses. --baseline runs the same trials with no installation at all.

Sandboxed evals

By default, harness agents run directly on your machine with full access (codex --dangerously-bypass-approvals-and-sandbox, claude --dangerously-skip-permissions). Workspaces are disposable tempdirs, but the agent itself is not confined — the default trusts that you wrote the skill and evals you're running. Agent-native sandboxes aren't used because they distort the measurement (codex's own sandbox blocks git commit, which reads as a skill failure).

For skills you don't trust — or CI — wrap every harness invocation (trials and judges) in a container instead:

docker build -t skillet-eval sandbox/   # once; recipe ships in this repo
skillet eval --sandbox docker

The agent keeps full freedom inside the container while the host stays untouched; checks still run on the host against the mounted workspace, so results are identical in shape. Configure in .skillet.yaml:

sandbox:
  enabled: true            # or opt in per run with --sandbox docker
  image: skillet-eval
  mount_auth: ["~/.codex", "~/.claude", "~/.claude.json"]   # default: whichever exist
  network: true            # false -> --network none
  env: ["ANTHROPIC_API_KEY"]   # host env passed through by name

Caveat: on macOS, Claude Code keeps OAuth credentials in the Keychain, which can't be mounted — use the codex harness in the sandbox, or pass ANTHROPIC_API_KEY through env.

Commands

| Command | What it does | |---|---| | skillet init [--no-prompt] | Install the skillet-authoring skill for your agents via @sentry/dotagents (user scope; asks first) | | skillet new <name> | Scaffold a skill directory with a templated spec.md | | skillet status [path] | Artifact state and the single next step, derived from disk | | skillet instructions <spec\|skill\|evals> | Template + writing rules for one artifact (what the authoring skill consumes) | | skillet validate [path] | Spec grammar, SKILL.md frontmatter, case schema, coverage — exit 1 on errors | | skillet eval [path] [--case id] [--behavior id] [--trials n] [--baseline] [--harness x] [--sandbox docker] [--dry] [--out dir] [--verbose] [--keep-workspaces] | Run cases through the harness; per-behavior pass rates and lift. --dry finds cases a do-nothing agent would pass; --out persists per-case results incrementally and resumes interrupted runs | | skillet show [path] | Pretty-print the parsed spec with coverage |

Every command takes --json: one JSON object on stdout (failures emit {ok: false, error}), prose on stderr, exit 0/1.

Agent integration

Skillet's authoring workflow ships as a skill: skills/skillet-authoring/ — itself a spec-driven skillet skill, with evals. Once installed, asking your agent to create, improve, or migrate a skill lands on the skillet status / skillet instructions --json loop automatically.

skillet init installs it for every agent dotagents supports (claude, codex, cursor, opencode) in user scope, and always asks before writing. Prefer to own delivery? Skip init entirely:

npx @sentry/dotagents add getsentry/skillet skillet-authoring && npx @sentry/dotagents install

(project scope; add --user for global) — or install skills/skillet-authoring/ by any other means.

Migrating from skillet v0

The v0 formats (spec.yaml, generated evals/*.eval.ts, the create/improve/spec commands, and all SKILLET_* env vars) are gone. skillet status detects legacy skills and directs the migration: intent converts to spec.md and eval intent to YAML cases. See examples/commit-conventions/ for a complete current-format skill.