lemonmmice/dsh-agent-toolchain--plugins-dsh-verify ↗★ 0
dsh-verify
将任务收尾清单交给机器裁决并记录失败样本 适合需要对Agent任务交付物进行自动化硬校验和质量把控的场景。
同名套件的其他儲存庫
安裝
此插件尚未提供可驗證的 bundle，或相容性檢查未通過。請先閱讀倉庫說明。 閱讀完整 README ↗
說明文件
閱讀完整 README ↗dsh-verify
The DSH closing-adjudication surface — a thin shell over lib/verify/report.mjs
(the same engine as the MCP verify_report tool). Zero logic lives here.
What it changes
Your task-closing summary stops being free text. You hand over a claims list (each claim = statement + evidence reference), and the machine adjudicates:
kind=build— reads the per-run build record (run-.json)kind=api— queries the shared capture store (filter+expect.min/all2xx)kind=file— checks an evidence artifact existskind=manual— explicit opt-out for what the system can't check
Verdict: pass / incomplete / fail. Claims contradicted by evidence
auto-record into the failure corpus as agent-misjudge — the only data
source for "caught the agent claiming success while the evidence disagrees".
The workflow (mandatory)
- Finish the work.
- Write down what you claim is done — as structured claims, not prose.
- Call
verify_report(runId, task, claims), get the verdict. - Report to the human: summary + verdict.
Skipping step 3 is a claim without verification. The trigger surface is the closing summary itself — an action you already perform every time — not a separate habit you have to remember.