Skip to content

feat(server): cross-provider subagent dispatch and model-selection skill #21

Description

@nohat

Goal

Allow any T3-hosted agent to delegate work to any configured provider instance and model, independent of the parent agent's provider, model, or project. Pair the shared delegation capability with model-selection guidance grounded in recent T3 usage and available pricing.

Acceptance criteria

  • A provider-neutral T3 MCP capability exposes configured provider/model choices and launches delegated work in an explicitly selected project/workspace.
  • Delegation records a durable relationship to its parent and exposes status/result through T3's existing agent/thread surfaces.
  • Pricing comes from T3's existing cached rate service; unavailable/proprietary rates stay explicitly unpriced. Do not imply API-equivalent pricing is actual subscription spend.
  • A portable model-selection skill recommends among models already observed in recent T3 usage, shows sourced pricing, and is available to agents across providers/projects.
  • Focused tests cover provider/model/project selection, invalid/unconfigured targets, pricing availability, and completion/status.

Notes

The caller asked for this scope directly. Preserve T3's existing per-project runtime/access decisions; no provider/model/project allowlist is being added.

Activity

  1. nohat commented on Oct 5, 2026

    @nohat
    OwnerAuthor

    Implementation is starting in an isolated worktree after fleet admission cleared. Scope: universal MCP tools for model catalog and cross-provider delegated thread launch/status, plus portable model-selection guidance with T3 usage/rate provenance.

  2. nohat commented on Oct 5, 2026

    @nohat
    OwnerAuthor

    Implemented in the isolated worktree at /Users/nohat/code/t3code-provider-agnostic-subagents on feat/provider-agnostic-subagents. Changes are uncommitted and no PR was opened. The shared MCP server exposes list_agent_models, dispatch_subagent, and get_subagent_status to every provider session through the common agents capability. The model catalog includes registered projects, configured provider models, availability, recent 14-day T3 thread counts, and live pricing metadata. The model-selection skill includes the 2026-10-04 usage/rate survey and distinguishes API-list estimates from subscription spend. Same-project child threads reuse the parent worktree; different-project children use the selected project root. Verification: 118 focused tests passed; t3 server typecheck passed; scoped lint, format check, and diff check passed. Leaving this issue open for review and landing.

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

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions