sandbase-harness
sandbaseai
Local-first, self-hosted AI agent runtime and MCP bridge with sandboxed sessions, memory, credentials, audit/replay, and a local Console.
PROJECT TOPICS
INSTALL REFERENCE
dsh plugin --profile web add github:Oscar-Williams/dsh-deepcanary
该命令指向仓库当前默认分支;尚无绑定当前 commit 的完整验证结果。
PROJECT README
Let tasks run quietly. Know when they need you. Return to the right session.
English · 简体中文 · Documentation · Changelog
DeepCanary adds a local task inbox to DeepSeek Harness. It brings questions, approvals, failures and completion summaries together, with quiet notification policies and a direct return to the relevant session.
Version: 0.1.4-rc.1 · Host baseline: DSH 0.2.0-rc.2. This is a release candidate. The npm latest and next channels both select this RC; their channel names describe installation defaults. See compatibility and device coverage for the supported scope.

Open DSH Plugins, use its official package installation control, and enter:
dsh-deepcanary@0.1.4-rc.1
Apply the change using the manager's enable/restart instructions. Save active work before restarting. Desktop manages its own reserved profile; use the application's manager for installation, update and removal.
Check that your installed host is DSH 0.2.0-rc.2, then install into the web profile:
dsh --version
dsh plugin --profile web add dsh-deepcanary@0.1.4-rc.1
dsh --profile web
Open the authenticated URL printed by DSH. The fixed package version makes installation reproducible. You can also pass the downloaded .tgz from GitHub Releases to plugin add. Use the official npm registry if a mirror has not synchronized the release.
From a built checkout at dsh-v0.2.0-rc.2 (commit 639ed015397290b3745d163aafe02ffee4aa3f84), run the checkout's CLI script:
npx --yes pnpm@11.7.0 run dsh plugin --profile web add dsh-deepcanary@0.1.4-rc.1
npx --yes pnpm@11.7.0 run dsh --profile web
Follow DSH's source setup instructions to prepare that checkout. The plugin repository uses its own npm lockfile; its development commands are in development.
To update the Web profile, rerun the fixed-version installation command and follow DSH's reload instructions. To remove it:
dsh plugin --profile web remove dsh-deepcanary
In a source checkout, use npx --yes pnpm@11.7.0 run dsh plugin --profile web remove dsh-deepcanary. Removal preserves the host profile, sessions and saved reminder data.
| Level | Meaning | Behavior |
|---|---|---|
| C1 | Suggested check | Routine outcomes and completion summaries remain available in Inbox. |
| C2 | Needs attention | Failures and possible stalls are evaluated against your notification policy. |
| C3 | Immediate human attention | Requires authoritative Host/runtime evidence, including open human questions and approval boundaries. |
Related events are grouped and deduplicated. Healthy long tools keep bounded progress grace; parallel tool completion updates that grace. Timed questions remain answerable after timeout, and accepted replies close their reminder. Completion summaries survive recovery and reload.
The resizable Inbox has opaque light/dark surfaces, keyboard controls and expandable evidence. Navigation uses DSH's public session-opening capability. When a target is unavailable, the item explains its availability and preserves the summary.
Settings use DSH's revision-aware configuration service. A concurrent change produces a conflict while retaining your draft; discard and reread before editing against the current snapshot.
Reminder state stays in the local profile. Stored fields are bounded metadata, hashed references, explicit feedback and the local session handle used for navigation. Conversation text, prompts, model output, tool arguments and API credentials remain outside captured reminder state.
Plugin HTTP endpoints use DSH's authenticated shared API, with Host/Origin checks, bounded JSON input and sanitized errors. Feedback collection is optional. Open and acknowledge actions describe interaction; usefulness is supplied by explicit feedback. See security and feedback calibration.
See compatibility for coverage and troubleshooting, architecture for the implementation, and contributing to help improve it.
MIT. Built on the public DeepSeek Harness plugin APIs.
CLASSIFICATION EVIDENCE
系统优先读取 GitHub Topics,再与站内分类词典和词根规则比对。当前命中: agent-observability。