evolve

evolve is a Go CLI for evaluating coding-agent plugins and plugin repositories. It validates plugin structure, checks whether skills trigger for the right prompts, runs behavioral eval suites in throwaway workspaces, and writes committed Markdown/JSON rollups for review and CI.

The pipeline is split into three tiers:

• Tier 0 checks: static validation of manifests, schemas, skill metadata, and repository shape.
• Tier 1 triggers: prompt-level checks that verify the expected skill activates.
• Tier 2 evals: behavioral cases that run real agent CLIs and grade the result.

Tip

New to evolve? Read the full docs at oss.bitwisemedia.uk/evolve — getting started, authoring evaluations (triggers, behavioral evals, fixtures, and how they run), the configuration reference, and the TUI guide.

Supported repositories

evolve auto-detects these layouts, or you can force one with --layout:

Layout	Marker	Skill paths	Eval paths
single	.claude-plugin/plugin.json	skills/<skill>/	evals/<skill>/
multi	plugins/*/.claude-plugin/plugin.json	plugins/<plugin>/skills/<skill>/	plugins/<plugin>/evals/<skill>/
marketplace	.claude-plugin/marketplace.json at root	plugins/<plugin>/skills/<skill>/	plugins/<plugin>/evals/<skill>/

Each eval directory may contain:

• triggers.<ext> for trigger-accuracy prompts.
• evals.<ext> for behavioral eval cases.
• results.<ext> for stored model results.

Supported data formats are json, jsonc, yaml, and yml; for a given basename, only one matching file may exist.

Harnesses and providers

evolve distinguishes the harness — the agent CLI it drives — from the provider that owns and prices the model the harness runs. Models are provider-qualified, and each is bound to the harness that can run it; evals execute once per model, through that harness.

Built-in harnesses, each needing its runner CLI on PATH and whatever credentials that CLI requires:

Harness	Runner CLI
Claude Code	claude
OpenAI Codex	codex
Gemini	gemini
Cursor	cursor-agent
GitHub Copilot	copilot
Antigravity	agy

Built-in providers (the model vendors) are Anthropic, OpenAI, Google, and Cursor — the last is both a provider (it owns Composer) and a harness. Run evolve doctor from a plugin repository to check the environment, credentials, and runner CLIs, and evolve models to see the effective provider / model / harness matrix.

Install

Install with Homebrew on macOS and Linux:

brew install --cask bitwise-media-group/tap/evolve

Build from source with Go:

go install github.com/bitwise-media-group/evolve/cmd/evolve@latest

Or build this checkout:

make build
./evolve version

Quick start

From the root of a plugin repository:

evolve doctor
evolve run checks
evolve run triggers
evolve run evals
evolve report

To run the full pipeline:

evolve run all

To make evaluation failures fail CI:

evolve run all --strict
evolve report --check

By default, run commands warn about failed checks or evals but exit 0 when the run itself completes. --strict changes those failures to exit 1; usage, configuration, and runtime errors exit 2.

Running evals

evolve run checks performs static validation only. It does not start agent CLIs.

evolve run checks

evolve run triggers runs each authored trigger prompt several times and records whether the expected skill activated.

evolve run triggers --model anthropic,openai --runs 5

evolve run evals runs behavioral cases in temporary workspaces, then grades the outputs with deterministic assertions and any configured LLM judge.

evolve run evals --model anthropic,openai --jobs 4 --max-turns 12 --timeout 900

Useful run filters and debug flags:

• --plugin a,b (alias --plugins): restrict the run to one or more plugins. Repeatable, or comma-separated.
• --skill x,y (alias --skills): restrict the run to one or more skills. Repeatable, or comma-separated.
• --model anthropic,openai (alias --models): pick providers / model ids, or all. Repeatable, or comma-separated.
• --eval case-id: restrict run evals to one behavioral case.
• --new: run only work with missing or stale stored results.
• --modified: rerun only cases whose authored content changed since their stored results (trigger frontmatter or definition; eval skill files or definition), fingerprinted alongside the results.
• --keep-workspaces: leave temporary workspaces behind for debugging.
• --count-only: compute token usage without running agents.
• --stale-results keep|drop: decide what to do with stored results outside the models restriction.

Interactive TUI

On an interactive terminal, evolve run triggers, run evals, and run all open a full-screen TUI: first a selection form to scope the run, then a live dashboard that streams results as agents finish. Pass --no-tui (or set EVOLVE_NO_TUI=1) for the plain line-based output used in CI and non-TTY pipes — both paths drive the same engine, so the run is identical either way.

Selection form

The form is a set of focusable panes you tab between to choose what runs:

• Filters — the same new / modified / failed scoping that the run flags expose.
• Harnesses — the agent CLIs to drive; any whose CLI is off PATH is shown disabled.
• Models — individual models grouped under a per-provider header row, so you can toggle one model or a whole provider at once. Models unsupported by the enabled harnesses are shown disabled.
• Plugins / Skills / Cases — a tree of every trigger and behavioral case. Each row shows whether it is forced on, forced off, or auto-queued for all / some / none of the enabled models; a legend under the tree names every glyph.

Move between panes with tab / shift+tab (or 1–4 to jump), ↑↓ / jk to move within a pane, ←→ / hl to fold the tree, space to toggle, and g / G for the ends. Tab on to the RUN / CANCEL buttons, or just press r to run and esc to cancel. The form previews exactly what will execute — it and the engine resolve through the same plan, so they cannot drift.

Live dashboard

Once a run starts, the dashboard streams progress:

• A title bar with running pass / fail / error tallies, elapsed time, rolled-up cost, and an overall progress bar.
• An Execution tree (plugin → skill → model → case) carrying per-node rollup columns.
• A tabbed Rollup (Summary / Providers / Plugins / Skills), a Runs log of every execution in plan order, and a Details pane showing in-flight cases and the selected case's authored spec.

Selecting a run in any pane moves the selection everywhere; f follows the live execution, enter jumps to its detail, and g / G plus ^d / ^u scroll. See DESIGN.md → TUI for the full wiring.

Reports

evolve report rebuilds repository-level rollups from stored per-skill results:

evolve report
evolve report --check

The report command writes EVALUATION.md plus a machine-readable rollup using the configured results format. In marketplace and multi-plugin repositories, it also includes per-plugin detail pages.

Thresholds can be set in .evolve.<ext> or passed directly:

evolve report --check --min-triggers-pass-rate 0.95 --min-evals-pass-rate 0.90

Commands

Top-level commands:

• evolve doctor: check harness runner CLIs, credentials, and counting APIs.
• evolve models: show the effective provider / model / harness matrix and pricing metadata.
• evolve report: regenerate evaluation rollups from stored results.
• evolve run: run static checks, trigger checks, behavioral evals, or the full pipeline.
• evolve version: print build metadata.

Run-tier commands:

• evolve run checks
• evolve run triggers
• evolve run evals
• evolve run all

Common global flags:

• --root PATH: repository root to operate on.
• --layout auto|single|multi|marketplace: repository layout.
• --results-format json|jsonc|yaml: results and rollup format.
• --json: emit machine-readable JSONL progress.
• -v, --verbose: enable debug logging.

See docs/cli/evolve.md for the generated command reference.

Configuration

evolve reads at most one config file from the repository root:

• .evolve.yaml
• .evolve.yml
• .evolve.json
• .evolve.jsonc

Settings are layered in this order:

1. Built-in defaults.
2. The config file.
3. EVOLVE_* environment variables.
4. Explicit CLI flags.

Common settings:

• layout
• models
• harnesses
• cache_dir
• results_format
• max_turns
• stale_results
• checks.*
• report.thresholds.*
• providers.<name>.models

Read docs/config/index.md for the full generated configuration reference and annotated example configs.

Development

Common targets:

make fmt
make test
make lint
make docs
make smoke
make pr

Notes:

• make docs regenerates committed CLI, manpage, and config docs under docs/.
• make smoke runs the live end-to-end test in e2e/ and requires the relevant provider CLI and credentials.
• tools/ is a separate Go module for pinned developer CLIs.
• e2e/ is a separate Go module for live smoke coverage and fixture repositories.

Project layout

cmd/evolve/   cobra CLI entrypoint and subcommands
internal/     core packages by concern
docs/         generated CLI, manpage, and config reference
schemas/      JSON Schemas for eval and report data
e2e/          separate module for end-to-end smoke coverage
tools/        separate module for pinned developer tooling
security/     code-scanning and security notes

Further reading

• Documentation site — getting started, authoring evaluations, configuration, and the TUI guide.
• DESIGN.md for architecture, engine boundaries, and TUI wiring.
• docs/cli/evolve.md for generated command documentation.
• docs/config/index.md for the full config surface.