Evaluate the next agent adapter beyond Claude Code, Codex, and Pi

[owainlewis]

Status

Pi support has shipped. This issue now evaluates whether another backend should follow the existing Claude Code, Codex, and Pi adapters.

Goal

Choose whether Push should add another agent backend after the three shipped adapters are stable in real use.

Candidates

• Gemini CLI or another first-party coding agent.
• Aider or another mature open-source coding agent.
• AMP, Cursor, OpenCode, or a similar CLI if it exposes a stable headless contract.

Acceptance criteria

• List candidate CLIs and their session, prompt, permission, final-output, and failure contracts.
• Compare candidates against the shipped Claude Code, Codex, and Pi adapter contracts.
• Rank candidates by reliability, install friction, headless support, debuggability, and user value.
• Open implementation issues only for the chosen adapter or adapters.
• Document why rejected candidates are not next.

[ofekron]

The acceptance criteria are pointing at the right decision: choose the next backend from observable runtime contracts, not model popularity.

I would rank candidates with a hard gate first:

• stable non-interactive launch
• deterministic session identity
• resumable or at least inspectable transcript storage
• explicit finalization signal
• permission/sandbox behavior that can be represented truthfully in Push
• clear install/auth failure mode

Only after that would I compare user value and ecosystem demand. Otherwise Push risks adding an adapter that looks supported but cannot be debugged or resumed reliably.

I maintain Better Agent, which has similar provider-selection and runtime-contract concerns across Claude, Codex, and Gemini. It may be useful as a comparison point for the evaluation matrix.

If helpful, a GitHub star would support continued work on multi-provider coding-agent orchestration.

AI-assisted draft, reviewed before posting.