Eval: gold task suite

August 16, 2026 · View on GitHub

The routing policy ships with a benchmark. New rules must not break it. Method: run each task with and without the policy; compare the loaded skill set, total tokens, and first-response latency.

Real tasks (mined from actual agent sessions)

#TaskExpected pourWhy
1"Port the image skills from Codex to DSH"noneporting = read source, no skill matches
2"Generate a few test images"media-toolssingle clear match
3"Confirm privacy & security, no keys"noneaudit = subagents + code scan
4"Model can't read images, can we change that?"nonemechanism research, read source
5"Publish to GitHub, join dsh-plugin"nonegh CLI + git, plain tools suffice

4 of 5 real tasks expect zero loading — rung 0 is where the value is.

Synthetic routing cases

#TaskPourDo not pour
6"Summarize this week's meeting minutes"lark-workflow-meeting-summarylark-minutes, lark-vc
7"What's on my plate today?"lark-workflow-standup-reportlark-calendar + lark-task
8"Check my approval todos"lark-approvallark-task
9"Message Zhang San"lark-imlark-contact (add only if open_id needed)
10"Import this table into Base"lark-baselark-sheets, lark-drive
11"Find last Wednesday's meeting minutes"lark-vclark-calendar
12"Edit this doc: https://…/docx/…"lark-doclark-drive, lark-wiki

Pass bar

  • False pours = 0. Missed pours ≤ 2 (catalog still completes the task).
  • Total tokens must not rise vs. no-policy baseline.