Behavioral Evals · google-gemini/gemini-cli · Skills Docs

Gemini

Skill 3 of 23

Behavioral Evals

Guidance for creating, running, fixing, and promoting behavioral evaluations.

1 minute · 196 words · 7 sections

Install

$ npx skills add google-gemini/gemini-cli --skill behavioral-evals

$ npx skills add google-gemini/gemini-cli

The first command installs just this skill, by the name in its SKILL.md; the second installs the whole repository.

Overview

Behavioral evaluations (evals) are tests that validate the agent’s decision-making (e.g., tool choice) rather than pure functionality. They are critical for verifying prompt changes, debugging steerability, and preventing regressions.

[!NOTE] Single Source of Truth: For core concepts, policies, running tests, and general best practices, always refer to evals/README.md.

🔄 Workflow Decision Tree

Does a prompt/tool change need validation?
- No -> Normal integration tests.
- Yes -> Continue below.
Is it UI/Interaction heavy?
- Yes -> Use appEvalTest (AppRig). See creating.md (opens in a new tab).
- No -> Use evalTest (TestRig). See creating.md (opens in a new tab).
Is it a new test?
- Yes -> Set policy to USUALLY_PASSES.
- No -> ALWAYS_PASSES (locks in regression).
Are you fixing a failure or promoting a test?
- Fixing -> See fixing.md (opens in a new tab).

📋 Quick Checklist

1. Setup Workspace

Seed the workspace with necessary files using the files object to simulate a realistic scenario (e.g., NodeJS project with package.json).

Details in creating.md (opens in a new tab)

2. Write Assertions

Audit agent decisions using rig.setBreakpoint() (AppRig only) or index verification on rig.readToolLogs().

Details in creating.md (opens in a new tab)

3. Verify

Run single tests locally with Vitest. Confirm stability locally before relying on CI workflows.

See evals/README.md for running commands.

📦 Bundled Resources

Detailed procedural guides:

creating.md (opens in a new tab): Assertion strategies, Rig selection, Mock MCPs.
fixing.md (opens in a new tab): Step-by-step automated investigation, architecture diagnosis guidelines.
promoting.md (opens in a new tab): Candidate identification criteria and threshold guidelines.

Appendix 3.1–3.6

6 files · 15 KB

Everything this skill ships beside its prose. All of it is set here, as subchapters of skill 3.

ReferencesMarkdown · 4 files

Documentation the agent loads on demand, rather than up front.

AssetsText · 2 files

Templates, schemas and fixtures the skill draws on.

About this skill

Trigger

Guidance for creating, running, fixing, and promoting behavioral evaluations. Use when verifying agent decision logic, debugging failures, debugging prompt steering, or adding workspace regression tests.

The verbatim description from this skill’s front matter — the string an agent matches on to decide whether to load it.

Colophon for this issue

Colophon

How this issue was made

Source

github.com/google-gemini/gemini-cli, branch main, last pushed 24 September 2026.

Licence

Apache-2.0 — the text of every skill is reproduced unmodified, frontmatter included, under the upstream licence.

Discovery

23 skills found by walking the repository tree for SKILL.md, not by matching a directory convention. 4 distinct layouts observed: .gemini/skills/*/SKILL.md, packages/core/src/skills/builtin/*/SKILL.md, tools/caretaker-agent/cloudrun/triage-worker/.gemini/skills/*/SKILL.md, tools/gemini-cli-bot/.gemini/skills/*/SKILL.md.

Authorship

4 of 23 skills are installed into the repository rather than published from it. They are shelved in the closing “Credited skills” part and remain the work of their own authors. A skill counts as credited when its only copy lives under one agent’s dot directory; visible directories, per-agent mirror sets, and skills installable from this repository on skills.sh all count as published.

Issue colours

Resolved from a curated brand profile — hue 288°, chroma 0.190. Two accent tones are generated per issue and each is proven against its own ground before it ships: a single accent that passes AA on both light and dark paper is arithmetically impossible.

Heading repairs

1 repair applied to this skill so the document has one h1 and no skipped levels:

Removed “Behavioral Evals”, a leading h1 that duplicated the skill title.

Spec compliance

2 editorial notes across 2 of 23 skills. They are printed in the margin of each skill rather than as errors here.

Code Explorer — 1 note