Add Codex API, ChatGPT, and remote SSH runtimes

[eladrave]

Summary

Adds Codex CLI app-server as a managed-agent compatibility runtime with three connection profiles:

• Codex API: an OpenAI Responses-compatible base URL, API key, and model
• Codex ChatGPT: Codex's native ChatGPT device authorization and persisted CODEX_HOME, without an OpenAI API key
• Codex Remote SSH: a fixed codex app-server --listen stdio:// command on another machine, authenticated by password or private key

The change includes the runtime bridge container, profile/controller APIs, runtime configuration UI, Compose registration, template metadata, deployment documentation, and event-protocol tests.

Dependency

Depends on #493 for scalable completed-history replay and mobile session access. This branch intentionally keeps that general reliability fix in its own PR and should be rebased onto main after #493 lands.

Architecture

templates/codex/ is a self-contained compatibility bridge implementing the Anthropic Managed Agents API already consumed by the control plane. It:

• pins Codex CLI because app-server is currently an experimental protocol
• runs Codex app-server over stdio and maps managed-agent sessions to durable Codex threads
• translates Codex JSONL notifications into normalized SSE runtime events
• persists agents, sessions, and ordered events in SQLite
• hosts multiple isolated profiles behind one controller registration
• preserves the legacy root API for existing deployments

The control plane registers the controller once, then creates individual API, ChatGPT, or SSH runtime aliases through server-side controller calls.

Security model

• the controller API key remains server-side and is not returned to the browser
• profile credentials are encrypted at rest with AES-256-GCM
• ChatGPT mode uses Codex account RPCs and never requires an OpenAI API key
• SSH starts a fixed app-server command rather than accepting an arbitrary command
• SSH host keys are pinned using either an operator-supplied fingerprint or trust on first use
• profile responses return masked metadata rather than passwords, API keys, or private keys
• this follows the control plane's current single-operator/master-key model; it does not add per-user ChatGPT or SSH credential isolation

Validation

• cargo fmt --check
• cargo test --release --locked: 212 passed; credential-dependent live tests remained ignored
• npm run lint: zero errors and one pre-existing React hook warning
• npm run build
• Codex bridge unit suite: 13/13 passed
• applicable browser suite: 8/8 passed, including API, ChatGPT, and Remote SSH configuration forms and native ChatGPT sign-in controls
• Docker build and authenticated live smoke through Cloudflare
• live model discovery, agent CRUD, session creation, SSE delivery, exact model response, and terminal idle completion
• live OpenAI-compatible endpoint with gpt-5.6-sol-high
• live Codex ChatGPT profile creation, authenticated account state, new turn, persisted session, and replay after reload

Generated file

src/model_prices_backup.json was refreshed by the repository's committed pre-commit hook during the feature commits. The feature does not change model-pricing behavior.

[deepankarmalhan-springhealth]

Is this adding support for ChatGPT managed agents? And if so, is there a general ETA around this?

[eladrave]

This allows you to use chatgpt as a harness/runtime. You can enable chatgpt either using an openai key, login to your chatgpt account (and use your subscription) or connect to a remote chatgpt harness using ssh. You can then create any agent you want to using the litellm ACP, and it will use chatgpt to run the agent. E

…

On Thu, Aug 6, 2026, 22:08 Deepankar Malhan - Spring Health < ***@***.***> wrote: *deepankarmalhan-springhealth* left a comment (LiteLLM-Labs/litellm-agent-control-plane#494) <#494 (comment)> Is this adding support for ChatGPT managed agents? And if so, is there a general ETA around this? — Reply to this email directly, view it on GitHub <#494?email_source=notifications&email_token=AUXLD5X5MKXHCLVTTF5MXBD5IU2Y5A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMRRGEZDKOBQGU2KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-5211258054>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AUXLD5U4MYKCZEFVQWO5QZT5IU2Y5AVCNFSNUABGKJSXA33TNF2G64TZHMYTEMZSGM3TANZRGQ5US43TOVSTWNJQG4ZTCNZTGM3DPILWAI> . You are receiving this because you authored the thread.Message ID: ***@***.***>