Skip to content

Complete Gemini bridge: prompt, streaming, MCP relay #17

@EightRice

Description

@EightRice

Goal

Finish the Gemini CLI ACP bridge so agents can run on the Gemini AI Pro subscription path end-to-end, matching the feature parity of the Claude Max and Codex Max bridges.

Current state

bridge/gemini-bridge.ts skeleton (committed in bbc3f45) handles:

  • Subprocess spawn with Windows CRLF TransformStream
  • ACP initialize handshake (with required fs capability stub)
  • session/new creation
  • Skip-authenticate-if-cached-creds-exist flow
  • NDJSON Python-facing protocol (ping / initialize / authenticate / new_session / shutdown)

What's missing:

  • prompt request handling — actually sending a user turn to Gemini and awaiting completion.
  • session/update streaming — propagating partial outputs back to the orchestrator.
  • MCP tool relay via mcpServers in newSession so agent tools work.
  • Python-side provider atn/providers/gemini_bridge.py mirroring claude_bridge.py / codex_bridge.py.
  • gemini_max provider entry in _KNOWN_PROVIDERS and friends.

Acceptance

  • An agent configured with provider=gemini_max, model=gemini-3-pro can complete a multi-turn task with at least one tool call.
  • Token usage is reported in Usage with the same disjoint-bucket convention as other bridges.

Notes

  • ACP loadSession: false for Gemini: cross-restart thread resume is not supported. Either inject history on every wake (see related issue) or document the restart-resets-context behavior.

Metadata

Metadata

Assignees

No one assigned

    Labels

    track:agentATN runtime, providers, orchestrator, bridgestype:featureNew capability

    Type

    No type

    Projects

    Status

    Backlog

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions