Skip to content

Add Base Agent V2 capability discovery - #91

Open
affandar wants to merge 1 commit into
mainfrom
codex/base-v2-capability-discovery
Open

affandar wants to merge 1 commit into
mainfrom
codex/base-v2-capability-discovery

Conversation

@affandar

Copy link
Copy Markdown
Owner

Generic sessions currently have no guided way to discover visible capabilities, so models often reach for durable agents even when a native task or progressively loaded skill/tool would be enough. This adds an opt-in agents.base_v2 prompt profile that prefers native tasks and progressive capability discovery, while leaving the V1 prompt and all existing named-agent composition intact when the flag is off.

The additive capability layer searches visible static, published, and curated skills plus authored agent guidelines. Sessions can load skills or agent workflow guidance and explicitly bind selected published-package tools and MCP servers at a turn boundary. Packages may contain skills, tools, and MCP servers without any agent entry point. Base V2 is off by default, allows a user override, and is effective only when copilot.native_tasks is effective for the same owner.

Validation:

  • npm run build
  • npm run lint
  • published-package E2E: Base V2 on/off generic sessions load a real skill and invoke a real worker tool and stdio MCP server
  • Horizon provider matrix: 2,067 passed, 3 skipped
  • baseline provider matrix: 2,053 passed, 16 skipped; one unchanged handoff lease test timed out under suite load and passed 4/4 unchanged in isolation and in the Horizon phase
  • isolated published-package E2E: 2/2 passed
  • PilotSwarm AKS live test: all 8 workers adopted a zero-agent package; V1 used default and explicit discovery; V2 used default-v2, activated the package, and invoked its worker and MCP tools
  • rollout health: workers 8/8, portal 1/1, MCP 1/1, zero restarts; deployment configuration, secrets, and model catalog hashes preserved

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant