[Feature]: Support Prime Agent as a provider (long-running RLM agent via its RPC mode)

[iPurpl3x]

Summary

Add a built-in driver + adapter for Prime Agent (https://github.com/PrimeIntellect-ai/prime-agent) so it can be controlled from T3 Code like the existing five providers. Prime Agent is an open-source, self-improving RLM (Recursive Language Model) agent for long-running coding and research tasks — daemon-backed background sessions that survive terminal detach, re-attachable, with persistent goals, heartbeats, schedules, and bounded autonomous mode. That is a capability gap none of the current providers fill, and T3 Code is the natural control surface for it.

Why it fits T3 Code

• T3 Code is provider-agnostic by design. Per docs/internals/providers.md, adding an agent is one Driver + Adapter entry in BUILT_IN_DRIVERS — no orchestration, contract, or client change required.
• Prime Agent ships a headless RPC mode (packages/coding-agent/docs/rpc.md) explicitly for "headless automation and integrations" — the ready-made transport for an adapter.
• Complementary overlap: T3 Code gives the human surface (threads, diffs, approval, phone remote); Prime Agent gives long-running autonomous research (persistent IPython runtime, programmatic subagents via rlm(...), /refine self-improvement, /autonomous budgets).

What it would take

• New driverKind primeAgent + adapter mapping Prime Agent's RPC/JSON streams onto ProviderAdapter.
• Config schema mirroring existing drivers. Auth is flexible: subscriptions (Claude / Codex / GitHub Copilot OAuth) or API keys (OpenAI, Anthropic, DeepSeek, OpenRouter, …) via ~/.prime/agent/auth.json.
• Notes: Prime Agent is daemon/TUI-first rather than a plain stdio CLI, so the adapter would talk to its RPC surface instead of an app-server like Codex. macOS/Linux only today; MIT licensed.

Considerations

• CONTRIBUTING notes large feature work is least likely to be accepted — this is a request for maintainers to consider, not a demand. Happy to test an RPC-mode integration or provide a minimal proof-of-concept adapter if that helps de-risk it.
• Possible long-term seam worth weighing: Grok already speaks ACP; a generic ACP-based driver could cover Prime Agent and future agents in one move, instead of per-agent custom adapters.

Environment

• T3 Code nightly 0.0.33, macOS

[rynfar]

I have a personal fork with this running. Maybe I can submit a pull request for it. @iPurpl3x

[t3dotgg]

Note

🤖 GPT-5.6 Sol responding on behalf of Theo

Related proposal issue #7326 adds constraints and reporter-local evidence for the Prime Agent integration.

The source proposes RPC as the only runtime boundary, with no automatic ACP fallback. Its stated reason is that Prime's ACP surface does not expose authentication status, model discovery, or session loading. A fallback could therefore lose login guidance, the model catalog, and resume or fork behavior while still appearing available.

The proposed compatibility rule starts at Prime Agent 0.7.2. It blocks known incompatible versions. An unknown newer version proceeds only after RPC starts, get_state succeeds, and the authenticated model catalog is not empty, with an advisory for the unknown version.

The reporter also describes two changes in a local branch. First, the probe accepted only a prefixed version on stdout, while the tested binary printed a bare version on stderr. Second, the reporter measured a 70-byte TMPDIR Unix socket path hanging on macOS while a 26-byte path answered in 1.4 seconds, reducing the observed probe from a 20.5-second timeout to about 1.9 seconds. These are the reporter's local observations. They are not a verified upstream fix or benchmark.

I am preserving this scope before issue #7326 closes during the policy pass. Prime Agent support has not landed.