[MiloMMIN]

Adds omp (https://omp.sh) as a first-class coding agent. omp exposes an Agent Client Protocol server (omp acp), so it reuses the existing ACP harness the same way Gemini and Qwen do — no new protocol code.

What's included

• Executor (crates/executors/src/executors/omp.rs): spawn + follow-up via AcpAgentHarness with a dedicated omp_sessions namespace, supervised approvals through the existing ExecutorApprovalBridge, and --yolo for the AUTO permission policy.
• Model discovery: shells out to omp models --json and groups the catalog per provider, so the model picker shows exactly what the local binary can route to (reasoning levels come from each model's thinking list).
• Wiring: CodingAgent::Omp in the executor enum, MCP passthrough adapter, approval bridge, default_profiles.json, MCP task-server executor list, and type/schema generation.
• UI: agent name/icon mapping, onboarding priority, omp-dark.svg / omp-light.svg.
• Docs: docs/agents/omp.mdx registered in docs/docs.json, plus entries in the supported-agents page, both agent-configuration pages, and the MCP server executor list.

Verification

• cargo check --workspace, cargo clippy -p executors --all-targets, tsc for web-core/local-web, and cargo run --bin generate_types -- --check all pass.
• End-to-end in the built app: registered a repo, created a workspace from the composer with omp selected, and the agent (driven over ACP) wrote a file and committed it in its worktree; a follow-up prompt continued the same session and produced a second commit. Conversation entries, tool-call rendering, and the model picker (290 models across agnes/apifarm/magpie) all populated correctly.

Known gaps

• Token/context usage isn't surfaced for ACP executors, so CONTEXT_USAGE is intentionally not advertised for omp.
• Reset-to-message (session fork) works at the session level but reset_to_message_id is ignored, matching the Gemini/Qwen ACP behaviour.

[Copilot]

The job was not started because the account is locked due to a billing issue.