Ikalus1988/MisakaNet
Failure-recovery memory: search and record failure-recovery lessons from real engineering sessions, with BM25 + semantic RAG retrieval and a lessons knowledge base.
MisakaNet searches 310+ failure lessons so your agent skips known bugs. It is a recovery-memory layer for AI agents: a local MCP server with 6 tools (search, get_lesson, submit_feedback, status, doctor, smoke) over a SQLite-backed index of past failure cases, each lesson carrying the error signature, the fix, and the evidence trail. For DeepSeek Harness the thin mcp_deepseek_adapter.py naming layer exposes the same tools under deepseek.recovery.* names, so the agent can look up 'pip install timeout SSL error' before re-attempting a task. The lessons corpus and index are maintained in the repo (build via python3 scripts/build_sag_index.py).
Install
git clone https://github.com/Ikalus1988/MisakaNet.git && python3 MisakaNet/scripts/mcp_deepseek_adapter.pyMisakaNet is a failure-recovery memory layer exposed as an MCP server. For DeepSeek Harness, clone the repo and register the adapter as an MCP server in your DSH MCP config: { "mcpServers": { "misakanet-recovery": { "command": "python3", "args": ["/path/to/MisakaNet/scripts/mcp_deepseek_adapter.py"] } } } (README 2026-08-31). Requires Python 3.10+ and the repo's dependencies (pip install -r requirements.txt from the repo root).
Compatibility
DeepSeek Harness via MCP config; Python 3.10+; also works with Codex/Cursor/Gemini CLI/OpenCode as a generic MCP server; MIT.
Details
- Repo: Ikalus1988/MisakaNet
- Category: Skills
- Stars: 426
- Version: MCP adapter (repo Ikalus1988/MisakaNet, 426★)
- Last push: 2026-08-25
- First seen: 2026-04-29
Recent updates
310+ failure lessons with evidence; 6 MCP tools; DSH adapter (deepseek.recovery.*); SQLite FTS index; Smithery deploy; JSON-LD structured data; CI badges.
FAQ
- Does MisakaNet work with DeepSeek Harness?
- Yes — the repo ships docs/integration/deepseek-harness.md with a ready MCP config that registers misakanet-recovery as a DSH MCP server; the adapter delegates everything to the core MCP server.
- What happens when a search returns nothing?
- Rebuild the search index with python3 scripts/build_sag_index.py, then check deepseek.recovery.doctor for health. Some SQLite FTS keywords (e.g. 'off', 'and') conflict and need rephrasing.
- Is my data sent anywhere?
- No — the server runs locally and only reads the bundled lessons corpus; feedback you submit is stored locally.
Alternatives
Relistencode/dsh-recall · Relistencode/dsh-extension-hub · perrylink/dsh-library