> ## Documentation Index
> Fetch the complete documentation index at: https://kensa.sh/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Skill

> The kensa-evals skill drives the eval lifecycle from your coding agent.

`kensa init` installs the Kensa skill set into your coding agent, led by the **`kensa-evals`** orchestrator. Ask the agent to set up or extend evals and `kensa-evals` walks the full lifecycle: wire the harness, confirm readiness, import evidence, inspect it into eval ideas, and generate and run evals.

## Installation

```bash theme={null}
kensa init
```

`kensa init` detects your coding agent from project markers and writes the skills to the right place. The `kensa-evals` entrypoint lands at:

| Agent       | Markers                 | Skill path                            |
| ----------- | ----------------------- | ------------------------------------- |
| Claude Code | `.claude/`, `CLAUDE.md` | `.claude/skills/kensa-evals/SKILL.md` |
| Codex       | `.agents/`, `AGENTS.md` | `.agents/skills/kensa-evals/SKILL.md` |
| Cursor      | `.cursor/`              | `.cursor/skills/kensa-evals/SKILL.md` |

The `kensa-setup`, `kensa-inspect`, and `kensa-generate` skills install alongside it in the same `skills/` directory. If no agent is detected, `kensa init` prints a copyable setup prompt instead. Re-run `kensa init` to refresh the skills after upgrading Kensa.

## Lifecycle

`kensa-evals` starts with state detection and routes to the first incomplete stage of the Kensa lifecycle:

```text theme={null}
setup  →  evidence  →  inspect  →  approval  →  generate  →  verify
```

Its checklist:

1. **Wire the harness** — connect `tests/evals/conftest.py::kensa_run(case)` to the real agent, then run `kensa doctor` and resolve any warnings.
2. **Import evidence** — `kensa import --from <provider>` to pull bounded trace evidence.
3. **Inspect** — mine imports into a YAML eval-idea queue under `.kensa/inspect/`, read with `kensa inspect list`.
4. **Approve** — mark the ideas worth keeping `status: approved`, then validate with `kensa inspect lint`.
5. **Generate** — materialize approved ideas as `tests/evals/test_<id>.py`.
6. **Verify** — `kensa eval` runs the suite and reports verdicts.

## The skill set

`kensa-evals` is the only entrypoint you invoke; it hands off to phase skills at each stage:

| Skill            | Role                                                                                     |
| ---------------- | ---------------------------------------------------------------------------------------- |
| `kensa-evals`    | Orchestrator — detects state and routes the lifecycle from setup through verification    |
| `kensa-setup`    | Connects the pytest harness to the real local agent boundary until `kensa doctor` passes |
| `kensa-inspect`  | Reads redacted TraceView evidence and writes a schema-validated YAML queue of eval ideas |
| `kensa-generate` | Writes and maintains `tests/evals/test_*.py` files from approved ideas                   |

## Guardrails

The skill is built to keep evals honest. `kensa doctor` flags harness patterns that fake a passing run — stub or mock agent classes, constructor bypasses, swallowed exceptions, hard-coded output fallbacks. The goal is evals wired to your real agent, so a green suite actually means the behavior held.