# Remvu for DeepSeek Harness This bundle connects DeepSeek Harness **0.2.1-rc.2** to the complete standalone Remvu app. Users do not install Python. It adds observable local recording, bounded memory context and five tools: `remvu_save`, `remvu_resume`, `remvu_check`, `remvu_events` and `remvu_show`. The official plugin install and an actual SDK-profile boot were checked on macOS without a model request. Contract tests exercise recording, failures, pending saves, plan/read-only behavior and workspace boundaries with the published Cordis/DSH packages and the real bundled engine. These checks do not certify model-driven handoff, the Web/Desktop UI or Windows DSH use. ## Install Have your agent perform these steps using reviewed downloads and a reviewed Remvu checkout. Keep the standalone application folder intact. `REMVU` below means its absolute executable, or the stable executable returned by installation; it is not a Python script. 1. Install the complete app for the project. Choose `--platform none` for DSH-only use; choose `both` when the same project also uses Claude Code/Codex. ```sh REMVU install --repo PROJECT --platform none ``` 2. In an existing DSH profile, add a reviewed pinned repository commit: ```sh dsh plugin --profile YOUR_PROFILE add github:caotaiyi/remvu#REVIEWED_COMMIT_SHA ``` Replace REVIEWED_COMMIT_SHA with the full tested source commit. The repository root declares the installable DSH bundle and ships JavaScript; no build script is run. For offline/local use, replace the GitHub spec with the absolute reviewed checkout path, or its adapters/dsh directory. For a fresh Web profile, create it **once before adding the plugin** with `dsh --profile remvu --from-default-profile web --dump-config`. Do not use `--from-default-profile` to reset an existing user profile. Then use `remvu` as `YOUR_PROFILE`. 3. Generate a private overlay using the **installed stable executable**, so moving the download does not break it: ```sh REMVU dsh-config --repo PROJECT dsh --profile YOUR_PROFILE --patch PROJECT/.remvu/dsh.patch.yml ``` The overlay holds the exact executable and project path. Review it before starting DSH. It refuses an existing destination; for an updated installation, retain the old overlay and generate a new filename using `--out`. There is no need to export environment variables. Advanced deployments may use `REMVU_EXECUTABLE` and `REMVU_PROJECT` instead. The package contains JavaScript and declares `dsh.bundle`; it needs no prepare/build script or build-script allowance. A packaged `.tgz` can also be installed through `dsh plugin ... add /path/dsh-remvu-0.3.0-beta.2.tgz`. Source users need the checkout kept in a stable location because local `add` links it. Published candidate plugin archives are generated by the [standalone workflow](https://github.com/caotaiyi/remvu/actions/workflows/standalone.yml); GitHub may require login to download CI artifacts. No npm registry publication is claimed. In PowerShell use `& 'C:\absolute\remvu.exe'` for `REMVU` and quote each argument literally. Follow DSH's own official setup requirements, including its Node/pnpm installation. ## Use Ask “Remvu save” before switching agents. The agent should inspect `remvu_resume`, `remvu_events` and relevant `remvu_show` results, draft a reviewed proposal, then call `remvu_save`. Its optional `proposal` is a JSON string containing `base_revision`, the actually reviewed `base_event_count` and typed `entries`. Without a proposal, it exports existing state and retains pending observations. See [the memory workflow](../../skills/remvu/references/engine-workflow.md) and [format](../../skills/remvu/references/format.md). Ask “Remvu resume” to load saved context and check code freshness. The receiving agent can also read the private handoff or an explicitly shared `REMVU.md` without installing this plugin. Public sharing remains an explicit standalone `share` operation with the existing privacy checks. No slash-command registration is claimed. Tools are registered through DSH's native `defineTool` API. Context uses `agent/pre-step`; events use `session/event` and final `tools/result`, not raw reasoning streams. Failure is retained as failure; a successful tool transport alone does not become a passing project check. ## Behavior and limits - Local projects only. Sessions outside the configured project, nested repositories and worktrees are excluded. Symlinked adapter paths are refused. Remote filesystem/cloud agents need a separate provider-aware adapter. - Plan/read-only mode defers writes; workspace-write requires the memory root inside the host's resolved writable workspace. The plugin is trusted host code and does not bypass DSH authorization. Keep `readOnly: true` in the plugin config when a deployment wants explicit read-only behavior. - Observations queue in order with a bounded backlog. Final durability checkpoints drain the queue. Failed or dropped records leave payload-free observation-gap markers; abrupt process death can still lose queued work. Complete semantic recording is never inferred from a clean process exit. - Unchanged memory is not reinjected on every step. Stop does not block or request another model call. Normal maintenance calls no model API; the current agent writes semantic summaries when requested. - `dsh plugin --profile YOUR_PROFILE remove dsh-remvu` removes the plugin layer. Retain `.remvu/`. Remove the extra `--patch` argument from your launch command. Upgrade the app through its lifecycle tools, then generate a new reviewed overlay. The adapter targets the pinned DSH prerelease because the host API is changing. Upgrading DSH outside `0.2.1-rc.2` requires revalidation; the bundle's engine range prevents claiming broad compatibility. Official references: [plugin bundles](https://github.com/deepseek-ai/deepseek-harness/blob/master/docs/user/develop/basic/publish.md), [tools](https://github.com/deepseek-ai/deepseek-harness/blob/master/docs/user/develop/basic/tool.md), [events](https://github.com/deepseek-ai/deepseek-harness/blob/master/docs/user/develop/framework/events.md).