cultivar

A CLI tool to help you write tests for skills, test them across agents, and iterate until they work, from the Pinecone DevRel team.

Test how well skills work against tasks, across agents, locally and remotely. Customize sandboxes for how agents should start, and graders for how agents should work.

Use traces to iteratively refine skills and optimize them against tasks.

Benchmark against skills, docs, and baselines. And, even run in parallel simulatenously for faster execution.

Two ways to run it: the cultivar CLI directly, or install the bundled skill and let your coding agent drive it:

npx skills add https://github.com/pinecone-io/cultivar --skill cultivar

Same engine either way — but when an agent drives it you won't watch the live run / --remote sandbox dashboard as directly as in your own terminal. Keep the skill project-scoped (not -g); it's never auto-tested.

Prerequisites

• Python 3.11+ and uv
• An Anthropic API key (the grader runs locally; agents in the sandbox use Modal-injected keys)
• A Modal account if you want --remote runs (recommended for parallelism + isolation).This is the recommended experience!

Install

uv tool install cultivar

# Or install from source:
uv tool install --from "git+https://github.com/pinecone-io/cultivar" cultivar

Modal setup (one-time)

--remote runs each eval in an isolated Modal sandbox — recommended for parallelism and clean auth state. Skip this section if you only need local runs.

# 1. Install Modal and authenticate
pip install modal
modal token new

# 2. Create the secret the sandbox reads at runtime
modal secret create eval-sandbox-secrets \
  ANTHROPIC_API_KEY=sk-ant-...
  # Add any keys your tasks need: GEMINI_API_KEY, COPILOT_GITHUB_TOKEN, etc.

# 3. Verify
modal secret list   # eval-sandbox-secrets should appear

The first --remote run builds the sandbox image (~3–5 min). Subsequent runs use the cached image (~5–10 s cold start).

Defaults you can override via env var:

For workspace sharing, custom images, and debugging sandbox failures, see docs/sandbox.md.

Quickstart: testing your own skill

1. Set up your working directory

mkdir ~/my-evals && cd ~/my-evals
cat > .env <<'EOF'
ANTHROPIC_API_KEY=sk-ant-...
# Optional — only for `--backend typesafe` grading (see docs/grader.md)
# TYPESAFE_API_KEY=...
EOF

2. Scaffold a task file

cultivar init my-skill

This writes ./tasks/my-skill.yaml and ./.claude/skills/my-skill/SKILL.md.

Tip: to test skills without your interactive coding agent auto-loading them, keep them outside .claude/ — e.g. set CULTIVAR_SKILLS_DIR=skills (or pass --skills-dir skills) and init scaffolds into ./skills/my-skill/. See where skills live.

3. Edit the skill (.claude/skills/my-skill/SKILL.md)

The skill file is what the agent sees when you invoke /my-skill. Write it like a concise brief: what the skill does, when to use it, and the key commands or patterns it should follow. Keep it tight — a few focused sections outperform a wall of text. If you're not sure where to start, drop your existing docs or a rough draft into Claude and ask it to write a SKILL.md for you.

4. Edit the tasks (tasks/my-skill.yaml)

Each task has an intent (what you'd say to the agent) and a criteria block (what PASS looks like, in plain English). A good criteria block names 2–3 concrete things that must be true and at least one common failure mode. Agents are good at this too: share a few examples of passing and failing behavior and ask Claude to draft the criteria.

For the full YAML schema and field reference, see docs/task-yaml.md.

5. Run + grade

cultivar run --skill my-skill --runner claude --remote --grade

Smoke test (post-install, no clone)

After uv tool install, verify the install works end-to-end with the packaged smoke:

cultivar hello                      # local: agent + grader (needs ANTHROPIC_API_KEY)
cultivar hello --remote             # also exercises Modal + eval-sandbox-secrets
cultivar hello --no-grade           # just exercise the runner (no API key needed)

hello runs a tiny "write hello.py" task that ships inside the wheel — no repo clone, no tasks/ setup. It exits 0 on PASS and prints diagnostics on FAIL. Use this to learn how to use cultivar.

Running remotely + inspecting results

# Single task, single variant
cultivar run --skill my-skill --runner claude --task my-task -v with-skill --remote

# All tasks + every applicable variant (with-skill, without-skill, and with-docs
# for tasks that declare context_refs)
cultivar run --skill my-skill --runner claude --remote

# 3 runs per (task, variant) for reliability, 5 sandboxes at once
cultivar run --skill my-skill --runner claude --remote --repeat 3 --parallel 5

# Raise the per-call wall-clock budget (default 90s; sandbox gets +60s buffer)
cultivar run --skill my-skill --runner claude --remote --timeout 180

# All three runners in parallel
cultivar run --skill my-skill --runner claude --remote &
cultivar run --skill my-skill --runner copilot --remote &
cultivar run --skill my-skill --runner gemini --remote &

# Run + grade in one shot
cultivar run --skill my-skill --runner claude --remote --grade

# Name a run so you can tell it apart later
cultivar run --skill my-skill --runner claude --remote --title baseline
cultivar run --skill my-skill --runner claude --remote --title after-tweak

What you get per run (results/<timestamp>[__title]/):

results/2026-04-22T11-31-47__baseline/
├── tasks.json                                 # task definitions used (for reproducibility)
├── notes.md                                   # --notes text, if any
├── grades.json                                # written by grader after `cultivar grade`
└── claude/                                    # one subdir per runner
    ├── my-task__with-skill.json               # structured result + stats (tokens, cost, timing, session_id)
    ├── my-task__with-skill.md                 # readable conversation trace
    ├── my-task__with-skill.jsonl              # raw event stream from the agent CLI
    ├── my-task__with-skill.stderr.log         # captured stderr (if any)
    ├── my-task__with-skill.setup.log          # setup/verify/teardown outputs (if those hooks ran)
    ├── my-task__with-skill.verify.log
    ├── my-task__with-skill.teardown.log
    └── my-task__with-skill.workdir/           # any files the agent wrote (code-gen tasks)
        └── hello.py

With --repeat N, files get a __1 / __2 / __N suffix. Without --title, the dir is just <timestamp>/.

Inspecting what actually happened:

Quick debugging recipes:

# Read one run end-to-end (replaces jq/less incantations)
cultivar show latest -r claude -t my-task

# Just the grader's verdict + remediation suggestions on a failure
cultivar show latest -t my-task --grader

# Pipe-friendly conversation transcript (ASCII fallback when not a TTY)
cultivar show latest -t my-task --conversation-only > convo.txt

# Full summary table for the latest run (no regrading)
cultivar report

# Regrade after editing criteria or adding calibration examples
cultivar grade --report

# Grade with TypeSafe System One instead of Claude — ~30x cheaper, no written
# reasoning or quoted evidence. Good for CI gates; see docs/grader.md.
# Needs the optional extra: pip install 'cultivar[typesafe]'
cultivar grade latest --backend typesafe

# Preview what TypeSafe would be sent, per section, without spending anything
cultivar grade latest --backend typesafe --dry-run

# Drop down to raw artifacts when needed
jq . results/<run>/claude/my-task__with-skill.jsonl | less
ls results/<run>/claude/my-task__with-skill.workdir/

# Resume a Claude session interactively
claude --resume $(jq -r .session_id results/<run>/claude/my-task__with-skill.json)

Handing off to a coworker

Want to use cultivar with a team, but don't want to make everyone have different Modal workspaces?

Easiest path (assumes you have a Modal workspace set up):

1. Invite them to the Modal workspace (Modal dashboard → Settings → Members). They inherit the eval-sandbox-secrets secret group, so they don't need to set up their own Anthropic/Pinecone/etc. keys for remote runs.
2. On their machine:

   uv tool install cultivar
   modal token new                     # personal Modal token
   modal profile activate <workspace>  # if they belong to multiple workspaces
   echo "ANTHROPIC_API_KEY=sk-ant-..." > .env   # auto-loaded from cwd
   cultivar init my-skill            # scaffolds tasks/my-skill.yaml
   cultivar run --skill my-skill --runner claude --remote --grade
   

3. Billing accrues to your Modal account regardless of who runs what — set an expected budget if needed.

The only key your coworkers personally need is ANTHROPIC_API_KEY (the grader runs locally) — or TYPESAFE_API_KEY instead, if your tasks grade with --backend typesafe. For local (non-remote) agent runs they also need whatever the relevant agent CLI requires (Claude OAuths via claude on first run; Copilot needs COPILOT_GITHUB_TOKEN with the "Copilot Requests" fine-grained PAT scope; Gemini needs GEMINI_API_KEY).

Supported Agents

We're always interested in adding more agents. If you have one that's not here, please let us know by opening an Issue!

Each runner advertises three variants:

• with-skill — skill loaded, agent invoked via /<skill-name>
• without-skill — same agent, no skill loaded and no Use the /<skill> prefix in the prompt
• with-docs — same as without-skill, but the task's context_refs files are prepended to the prompt as raw reference material (local files or http(s):// URLs, fetched live and cached under ./.docs_cache). Only runs for tasks that declare context_refs.

Two deltas to read:

With --remote, each (task, variant, repeat) runs in its own Modal sandbox in parallel — three variants on one task means three sandboxes, run concurrently up to --parallel N (default 5). Apples-to-apples baseline; same image, only the prompt + skill mounting differ. See docs/concepts.md for the full discussion and docs/task-yaml.md for how to add context_refs to a task.

Testing docs, not skills (Claude runner only): three more real --variant choices, opt-in, not part of the default sweep above.

• without-docs — identical to without-skill, named for docs-testing clarity
• self-navigate — no injected reference material, but WebFetch is enabled and the task's self_navigate_refs gives it a starting page; tests whether the agent can find the right doc on its own (with-docs tests whether the content is good once handed over)
• with-docs:<label> — set ground_truth.doc_versions: {label: [refs]} instead of flat context_refs and --variant with-docs auto-expands into one real run per version, graded side by side, so comparing two doc versions (e.g. before/after a rewrite) is one command, not two runs you diff by hand

Also pin the agent's model with --model <id> (e.g. claude-sonnet-5, Claude runner only) to re-run the same task set under a different model; unset uses the CLI's own default.

Docs

• docs/concepts.md — start here if you're new: what cultivar measures, why, and how to read the results
• docs/task-yaml.md — task YAML schema, every field, worked examples
• docs/grader.md — how grading works, the Claude and TypeSafe backends and when to use each, calibration examples, the prompt anatomy
• docs/sandbox.md — Modal sandbox setup (DIY), lifecycle, what's controllable
• docs/runners/claude.md, gemini.md, copilot.md — per-runner specifics

For any subcommand: cultivar <cmd> --help.