Problem
SWE-2 can return action announcements instead of usable client tool calls when a large coding-client prompt and tool catalog are sent through a cloud chat adapter. Client-specific fixes then spread into shared prompts, schemas, and defaults.
Proposed contribution
An opt-in DEVIN_SWE2_TRANSPORT=acp Chat Completions transport using the official signed-in Devin CLI. It exposes a small MCP catalog, retrieves full client tool schemas on demand, and returns ordinary tool calls for the client to execute. Native CLI tools remain denied; original instructions, model access policy, client permission rejections, and provider policy errors are preserved.
Medium/High/Max map to native CLI variants. Tool IDs survive OMO/OpenClaw normalization. Updated continuation instructions restart the prompt with the complete transcript before more work; initializing sessions cannot be evicted; the relay has no implicit five-minute header timeout. Other model routes are unchanged.
Prepared as an experimental draft with no npm runtime dependencies, focused protocol/route tests, and a real CLI/client-tool smoke test. The CLI account, in-memory session limits, text-only support, and rollback switch are documented. I will link the draft PR here.
AI-assisted implementation; no personal configuration, credentials, or conversation traces are included.
Problem
SWE-2 can return action announcements instead of usable client tool calls when a large coding-client prompt and tool catalog are sent through a cloud chat adapter. Client-specific fixes then spread into shared prompts, schemas, and defaults.
Proposed contribution
An opt-in
DEVIN_SWE2_TRANSPORT=acpChat Completions transport using the official signed-in Devin CLI. It exposes a small MCP catalog, retrieves full client tool schemas on demand, and returns ordinary tool calls for the client to execute. Native CLI tools remain denied; original instructions, model access policy, client permission rejections, and provider policy errors are preserved.Medium/High/Max map to native CLI variants. Tool IDs survive OMO/OpenClaw normalization. Updated continuation instructions restart the prompt with the complete transcript before more work; initializing sessions cannot be evicted; the relay has no implicit five-minute header timeout. Other model routes are unchanged.
Prepared as an experimental draft with no npm runtime dependencies, focused protocol/route tests, and a real CLI/client-tool smoke test. The CLI account, in-memory session limits, text-only support, and rollback switch are documented. I will link the draft PR here.
AI-assisted implementation; no personal configuration, credentials, or conversation traces are included.