dumbtv
dumbtv takes the ads, screen tracking and update nags out of your Fire TV over ADB. No root, none of your apps touched, every step can be undone. Free and open source.
翻譯準備中
將每個 AI 代理行為──模型呼叫、工具呼叫、批准、編輯──轉換為時間序列、雜湊雜湊的證據封包。可供稽核員驗證,且不儲存原始敏感資料。
AI 翻譯
Turn every AI agent action — model calls, tool calls, approvals, redactions — into a chronological, hash-anchored evidence packet. Provable to an auditor, with no raw sensitive data stored.
安全TypeScript · JavaScriptGitHub @ktlesr上線中營運者未確認
試用過了嗎? 點一下就好,不需要登入。
尚未偵測到更新。GitHub 發布版本後會顯示在這裡。
個人 GitHub 帳號下公開儲存庫的擁有者可以登入並驗證所有權,自行管理介紹,無需管理員核准。
登入並產生驗證檔案,在 30 分鐘內將其提交至此儲存庫的預設分支。其他分支或分叉儲存庫中的檔案無效。然後在下方檢查檔案,即可無需管理員核准直接管理專案。請保留檔案以維持管理權限。
如需登入,將前往登入頁面。登入後請回到此處,再次點擊所選的驗證按鈕。
在專案資料夾中執行 /nomorevibe verify,即可驗證所有權並自行管理介紹。
一分鐘完成驗證Agent Evidence Recorder
翻譯準備中
取自儲存庫 README
正在載入完整 README。 以下為已儲存的摘錄。
Audit-ready evidence layer for AI agent runs. Turns every critical agent action
(model call, tool call, human approval, redaction, error, output) into a redacted,
hash-anchored audit packet.
Scope of those words: redaction covers the patterns and secret-named fields AER recognises, not
all personal data; the hash chain proves internal consistency, not resistance to an operator who
rewrites the database (no external seal yet). Details: docs/hardening/.檢視完整 README ↗Audit-ready evidence layer for AI agent runs. Turns every critical agent action (model call, tool call, human approval, redaction, error, output) into a redacted, hash-anchored audit packet.
Scope of those words: redaction covers the patterns and secret-named fields AER recognises, not
all personal data; the hash chain proves internal consistency, not resistance to an operator who
rewrites the database (no external seal yet). Details: docs/hardening/.
For document-grounded assistants, a run can link to the previous turn and an event can record the context used, the retrieved candidates and the cited sources (with version and content hash). This records what the application reported; it does not show that a source is correct, current or legally applicable, and it is not a record of the model's internal reasoning.
Sealing a run returns a receipt (optionally Ed25519-signed). Exported JSON packets can be checked
offline with @ktlsr/aer-verify — no network, no AER account. Without a
receipt you kept or a key you trust, a passing check means "internally consistent", not "unaltered
since sealing": there is no independent external timestamp or seal service yet.
Agents can request a verified human approval for an action; only a signed-in user of the workspace can approve it in the dashboard, and the agent consumes it with the exact parameters right before acting. AER records and checks this, but it does not block an action whose code skips the check.
Not an observability tool. See
docs/PRODUCT.mdfor positioning.
apps/web Next.js (App Router) + TS strict + Tailwind + shadcn/ui + Prisma
packages/collector-js Collector SDK (startRun / run.event / run.complete)
packages/demo-agent Customer Data Deletion demo agent
docs/ PRODUCT.md (product positioning)
DATABASE_URL in apps/web/.envpnpm install # install all workspaces
pnpm dev # run apps/web (http://localhost:3000)
pnpm build # build apps/web
pnpm lint # lint apps/web
pnpm db:migrate # prisma migrate dev (apps/web)
pnpm db:seed # seed one demo run (apps/web)
pnpm demo # run the demo agent end-to-end
Positioning and product direction: docs/PRODUCT.md.
The remaining documents (spec, data model, API contract, security notes, agent roles, build runbook, design system) are internal and deliberately kept out of this repository, so they are not linked here.