dsh-llm-kiro

If you already have kiro-cli installed and logged in locally, and want it to serve as the LLM backend for DeepSeek Harness out of the box, dsh-llm-kiro is built for exactly that. It routes the harness ctx.llm seam to a local kiro-cli ACP (Agent Client Protocol) stdio child process, fully reusing your existing IAM Identity Center or Builder ID session—no API key, settings section, or credential file required. Once installed, kiro simply appears in the model picker.

Key capabilities include zero runtime dependencies (the ACP client is pure Node child_process plus a hand-written JSON-RPC layer of roughly 200 lines, with no postinstall scripts and no agent SDK); long-lived warm sessions (each host session maps to one kiro-cli subprocess, with LRU eviction beyond the maxSessions cap); tool bridging via a per-session loopback HTTP MCP server with callId pairing and a 120-second timeout; native tool isolation (kiro starts in zero-trust mode, so only host MCP tools are permitted); and real context-usage reporting (contextUsagePercentage from kiro is converted to inputTokens, so the UI context ring and auto-compaction threshold reflect true occupancy rather than character estimates). It also supports live model catalogue retrieval with TTL caching and static fallback, five reasoning effort levels, automatic kiro-proxy discovery, and overflow error classification that hands recovery back to harness.

Ideal for developers who have kiro-cli deployed and authenticated on their machine and want to leverage those existing credentials and the model catalogue inside DeepSeek Harness without re-entering API keys or managing global proxy variables. Just confirm kiro-cli whoami works on your system and you are ready to go.