From dcf7eeae3053004d0d27fae558ddbbd6b19c05d3 Mon Sep 17 00:00:00 2001 From: vesc Date: Mon, 17 Aug 2026 01:10:55 +0900 Subject: [PATCH] adr: ask for DeepSeek Harness as an inner engine Pitch only. No adapter code. --- adrs/deepseek-harness.md | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) create mode 100644 adrs/deepseek-harness.md diff --git a/adrs/deepseek-harness.md b/adrs/deepseek-harness.md new file mode 100644 index 00000000..428c1e4a --- /dev/null +++ b/adrs/deepseek-harness.md @@ -0,0 +1,17 @@ +# DeepSeek Harness as a harness + +Ask: add DeepSeek Harness (`dsh`) as another inner engine, same seat as pi / opencode / codex / claude. + +Not the DeepSeek model provider (that's PR 265). The harness itself: https://github.com/deepseek-ai/deepseek-harness + +Why we'd want it: it's the loop DeepSeek measures their agent benches in, and it already speaks JSON-RPC (and ACP). Transport-wise that looks like the Codex adapter, not a new protocol. + +The work, as we read it from the existing four: + +- a `dsh` adapter through `defineHarness` / `runTurn` +- rebind its tools onto QM's sandbox (same brain/hands split as the others) +- add `dsh` to `HARNESS_IDS`, or do issue 542 first so this doesn't touch the closed set + +We're not sending an implementation. If this is in scope we'd use it. If you'd rather keep the set at four, that's fine. + +Status: request, no code.