Repository navigation
Releases: kobubu/Rag-Coding
Release list
Code Memory 0.9.1 — local context for DeepSeek Harness
Code Memory 0.9.1
GitHub installation fix and the latest repository cleanup for DeepSeek Harness 0.2.0-rc.2. Unofficial, independently maintained, and not endorsed by DeepSeek.
GitHub installation
dsh plugin --profile web add github:kobubu/Rag-Coding#v0.9.1The tag contains the compiled plugin, CLI entry points, declarations, and static Atlas. GitHub installation no longer depends on a local TypeScript/Vite build or approval of a CM prepare script. Development uses npm run compile; CI compares the committed runtime to the build. Older tags still need their release tarball or a source build.
Apache Arrow is now an explicit runtime dependency for LanceDB. This matters in Harness profiles, which disable automatic peer installation. CI checks required peers of direct runtime dependencies so npm's implicit peer installation cannot hide this class of packaging error.
Alternatively, install dsh-plugin-code-memory-0.9.1.tgz from this release. Prepare the project's index and pinned local MiniLM model before indexed retrieval; Harness reference.
Windows installer
Download CM-0.9.1-Windows.zip and SHA256SUMS.txt, verify the checksum, and extract the ZIP. Run:
powershell -NoProfile -ExecutionPolicy Bypass -File .\install_cm.ps1 -ProjectRoot 'C:\path\to\your-project' -Clients harnessThe ZIP contains the bootstrap, ready npm package, guide, manifest, and checksums. It installs Harness and prepares project data; -Clients harness,codex,claude adds project MCP connections. Native-client automatic-context hooks remain a separate manual setup. Windows guide.
Cleanup included
- Removes nine unused functions, an obsolete benchmark type, and an unused import. Active evaluation runners and original A/B evidence remain available.
- Moves React/ReactDOM to development dependencies; static Atlas already includes them. The previous standalone package check measured 8,325,104 bytes excluded from runtime dependencies, including scheduler. This is disk saving, not a new token or speed benchmark.
- Enables unused-local/parameter checks and declares existing direct imports explicitly without upgrading dependency versions.
Retrieval, embeddings, graph routing, prompts, and model settings are unchanged. No new model A/B or savings claim is introduced. Existing benchmark and limitations.
Scope
Node.js 24 or 22.19+ within 22.x; pinned Harness; JavaScript/TypeScript; 500 source files and 8 MiB source. Native dependencies follow the package manager's policy. The Windows bootstrap is separate from direct GitHub installation. Models, project indexes, credentials, evaluation workspaces, and developer test artifacts are excluded. A cloud provider still receives selected source context.
Code Memory 0.9.0 — local context for DeepSeek Harness
Code Memory 0.9.0
First public, packaged community release for DeepSeek Harness 0.2.0-rc.2. Unofficial; independently maintained and not endorsed by DeepSeek.
Included
- Automatic source context before the first Harness model step, with direct file/symbol routes, local MiniLM/LanceDB retrieval, and dependency graph expansion.
- Freshness checks and suppression of repeated unchanged whole-file content already supplied to Harness.
- Local MCP server, CM Atlas, and the optional native-client context hook and warm daemon already present in the preceding source commits.
- English README and Windows installation guide; the previous Russian README is preserved.
- A Windows ZIP containing the bootstrap and prebuilt package. No source checkout or local TypeScript build is required to install it.
The automatic-context procedure is enabled by the installer for Harness. Codex and Claude Code receive MCP connections; their UserPromptSubmit hook remains a separate manual setup. This release does not change the retrieval/model procedure or add new model A/B results.
Download and installation
Download CM-0.9.0-Windows.zip and SHA256SUMS.txt from this release. Verify the ZIP against the checksum file, extract it, then run:
powershell -NoProfile -ExecutionPolicy Bypass -File .\install_cm.ps1 -ProjectRoot 'C:\path\to\your-project' -Clients harnessSetup prints launch paths and a diagnostic command. Initial npm dependencies and embedding weights require internet access. A compatible Node.js is reused or a private runtime is downloaded; administrator rights are not required. Indexed project code is not executed. Full installation guide.
Evidence and scope
The existing Harness benchmark reports 45–66% fewer total model tokens and 18–37% lower session time across three small evaluation series. This is one JavaScript/TypeScript project and six task types. Installation/indexing are separate costs; successes and failures are retained. These percentages are not a promise for other projects or clients. Native-client results and earlier unsuccessful approaches remain available.
The release packaging is checked separately by the required build/docs/unit/fixture suite, actual Harness ToolRuntime smoke, and installation from the extracted ZIP on a disposable project, followed by MCP diagnostics and retrieval. Publication checks verify the tag's commit and the downloaded asset hashes; they do not constitute a new model benchmark.
Limits
Windows bootstrap; pinned Harness version; JavaScript/TypeScript only; corpus limit of 500 files and 8 MiB source. macOS/Linux bootstrap, automatic updates/rollback, cloud subscriptions, and guaranteed savings are not included. Local retrieval does not prevent a cloud model provider from receiving selected source context.