Visitar URL original
agent-plugins/README.md at main · bioinfornatics/agent-plugins · GitHub
Skip to content

Latest commit

 

History

History
98 lines (76 loc) · 5.01 KB

README.md

File metadata and controls

98 lines (76 loc) · 5.01 KB

Servier Powered

Agent Plugins

Reusable authoring instructions and comparative Skill/Plugin evaluation.

0.6.0 candidate — not published. This document describes the new candidate, not a claim that the current remote main or an existing installation contains it. Goose is the verified reference host for plugin installation; other hosts and native model/profile compatibility require separate evidence.

Distribution contract

  • Five Skills remain: skill-creator, agent-creator, hook-creator, plugin-creator, and plugin-script-packaging.
  • Only skill-creator contains an executable runtime. Its existing scripts/skill-creator.mjs launches agent-plugin with evaluation, history and inspect commands, using Node.js >=22.22.2.
  • The other Skills retain authoring instructions, but their old CLI references are migration-pending. They do not have bundled creator executables. Do not treat those references as working commands.
  • Old validation, packaging, grading, score, HTML-report and historical-import commands are retired. No standalone native creator executable is provided.

Installation channel

For the published Git-backed distribution, Goose supports:

goose plugin install https://github.com/bioinfornatics/agent-plugins.git
# Optional auto-update installation instead:
goose plugin install --auto-update https://github.com/bioinfornatics/agent-plugins.git
# Update an already installed Git-backed plugin:
goose plugin update agent-plugins

These commands consume the remote distribution; they do not publish or select this unpublished candidate. Install trusted sources only. Goose imports namespaced Skills such as agent-plugins:skill-creator. Manually copied plugins are discovered but are not managed by the Git-backed update command.

A complete Git distribution needs no build, dependency installation, Bun or runtime download. Retain the whole Skill directory when copying a standalone Skill, and follow the host's documented discovery rules. A missing runtime is a distribution defect, not an invitation to repair it with npm install.

Author a Skill

Use agent-plugins:skill-creator to write .agents/skills/api-review for reviewing OpenAPI changes. Keep metadata and instructions portable, add realistic scenarios, review the files, and report what remains unverified. Do not launch models.

Authoring and static review are instruction-driven; no bundled validation or packaging CLI is promised. Custom agent, hook, plugin and script-packaging authoring are separate Skill roles, subject to the migration warning above.

Compare and inspect

Resolve <skill-root> to the installed skill-creator directory:

node <skill-root>/scripts/skill-creator.mjs evaluation --skills /path/to/skill --dry-run
node <skill-root>/scripts/skill-creator.mjs evaluation --skills /path/to/skill --plugin /path/to/plugin --dry-run
# Executes model trials; obtain explicit authority first:
node <skill-root>/scripts/skill-creator.mjs evaluation --skills /path/to/skill --model <model>
node <skill-root>/scripts/skill-creator.mjs history --json
node <skill-root>/scripts/skill-creator.mjs inspect <campaign-id> --json
node <skill-root>/scripts/skill-creator.mjs evaluation --resume <campaign-id>

Discover suites exactly at Skill evals/evals.json. Plugins use immediate skills/*/SKILL.md children and their suites: skip missing suites, fail if all are absent or any present suite is malformed. Repeated/combinable selectors form one deduplicated profile; every scenario compares that whole profile with its absence, not merely the scenario's source child.

Dry-run prepares a validated matrix without host calls. Normal execution stores SQLite records and campaign-isolated evidence from bounded parallel Goose trials in private HOME/XDG, which is not an OS sandbox. Defaults: 3 repetitions, 3 workers, 50 maximum turns, timeout 0 (disabled). There is no monetary budget cap. Model compatibility and native activation remain unverified beyond explicitly recorded evidence. History and inspect are read-only; they do not grade or verify evidence hashes. Resume uses the frozen plan, with no selectors or parameter overrides; --retry-uncertain explicitly permits new attempts without bypassing integrity.

See Skill Creator command contract for supported parameters. Repository regression suites are development-only, not installed user evaluation inputs. Missing evidence is unavailable, never a pass.

Learn more

main is generated installation material, not an authoring branch. Publication is separate from source changes, local projection and verification.