sentry01/agentcouncil
v2.0.0MIT
Multi-model council for Copilot CLI. Three subagents on different model families (Anthropic, OpenAI, Google) answer independently, then build on each other or attack the leading answer before an orchestrator writes the final result. Ships the agentcouncil skill and the AgentCouncil agent.
Changelog
All notable changes to this project are documented here. Versions follow Semantic Versioning.
[Unreleased]
Added
- Release workflow: merging a version bump to
maintags it and publishes a GitHub release, with notes built from this changelog byscripts/release-notes.mjs.
[2.0.0] - 2026-10-17
Breaking
- The plugin, skill and marketplace are now named
agentcouncil. The nameagent-councilis already used by a different plugin in the awesome-copilot marketplace. - When installed as a plugin, the agent ID is
agentcouncil:agentcouncil(copilot --agent agentcouncil:agentcouncil).--agent AgentCouncilonly worked for manually copied files. plugin.jsonuses the Agent Plugins 1.0 format. The agent moved tocom.github.copilot/agents/.- Explicit invocations (a council prefix such as
council:, or the agent) always convene. Simple tasks run the new light path instead of being answered directly.
Added
- Light effort path: one concise pass per seat plus a short synthesis (4 subagent calls).
- Context brief sent to every seat. Seats are told to work read-only.
- Anonymized, reordered inputs for every orchestrator call. Adversarial triage now runs on the orchestrator seat instead of the session model.
- Per-seat fallback lists, a family-diversity rule, user model overrides and a footer that discloses the models that actually ran.
- Routing evals (
evals/routing.json,scripts/eval-routing.mjs), a validation script (scripts/validate.mjs) and a CI workflow. .github/plugin/marketplace.json, so this repository is its own plugin marketplace.- LICENSE (MIT), CONTRIBUTING.md and this changelog.
Changed
- Default models: Alpha
claude-opus-5.5, Betagpt-6.1-sol, Gammagemini-4-argon, Orchestratorclaude-opus-5.5. The previous defaults (claude-opus-4.6,gpt-5.4,gemini-3.1-pro) are retired or superseded. - Mode routing checks the prefix first, then the user's intent, instead of matching keywords as substrings. "VS Code" and "the challenge is" no longer trigger adversarial mode, and choose-between and go/no-go questions now route to adversarial.
- Domain rules: data model and service API design is architecture, not code.
- Confidence labels must state what they rest on. The orchestrator weights evidence, not labels.
- The collaborative improve round always ends with a Tensions section, and unadopted first-draft ideas reach the synthesis.
SKILL.mdis a short router. Phase prompts live inreferences/and are loaded only for the chosen path.- The agent is a thin wrapper that loads the skill. Its tools are reduced to read, search, agent, web, todo and skill.
- Documentation rewritten in plain style.
Removed
- The duplicated protocol in the agent file.
- Manual install instructions that copied files into
~/.copilot.
[1.0.0] - 2026-09-16
- First plugin release:
plugin.json, collaborative and adversarial modes, complexity gate, domain detection, confidence labels and partial-consensus attacks. Earlier history (from 2026-02-09) is in the git log.