dsh-immune-system@2.3.0
CAUTION· startup onlyPublisher claims: "Self-evolving pitfall immunity for coding agents (Claude Code, Codex, Cursor, Google Antigravity, Gemini CLI, Windsurf, ZCode, DSH, MCP): capture failed-then-fixed tool episodes as SQLite+FTS5 antibodies and inject them before the agent repeats the mistak" Observed in sandbox: no network, touches nothing outside the project, writes no memory.
Coverage
Startup only; authenticated tool behavior not tested.- Tested
- server startup over stdio (initialize sent) · every file read, file written and host contacted by what ran
- Not tested
- the tool list (the server exited before answering) · what the tools do when called with real credentials
What this means for the label: Caution here covers what ran. The server's tools were not exercised with a real credential, so nothing about their behavior is asserted; the credential summary below says what to do with that.
Before you give it your token
- It will ask for
IMMUNE_LLM_API_KEYIMMUNE_LLM_MAX_TOKENSIMMUNE_MAGPIE_API_KEYConfiguration, not credentials: DSH_HOME, IMMUNE_ALLOW_PRIVATE_ENDPOINT, IMMUNE_DB_PATH, IMMUNE_DISTILL_BUDGET_MS, IMMUNE_GUARD, IMMUNE_LLM_BACKOFF_MS, IMMUNE_LLM_BASE_URL, IMMUNE_LLM_MIN_INTERVAL_MS, IMMUNE_LLM_MODEL, IMMUNE_LLM_TEMPERATURE, IMMUNE_LLM_TIMEOUT_MS, IMMUNE_LOCALE, IMMUNE_MAGPIE_MODEL, IMMUNE_MAGPIE_URL, IMMUNE_OLLAMA_MODEL, IMMUNE_PREWARM, IMMUNE_SYNC_DOC, IMMUNE_SYNC_DOC_MIN_HITS, ZCODE_PROJECT_DIR- Where it talks to
- no vendor host named in the code
- What the honeypot saw
- no credential files read · no network
- What we could not test
- the tool list (the server exited before answering) · what the tools do when called with real credentials
Connect it with a scoped credential, under the guard. It asks for IMMUNE_LLM_API_KEY, IMMUNE_LLM_MAX_TOKENS, IMMUNE_MAGPIE_API_KEY and talks to its vendor. Give the credential to the guard, not the server: the value never enters the server's environment and only travels to the host you bind. What the tools do with a real credential was not observed.
Never paste a credential into a chat, a ticket or a report, including ours. How the guard holds it.
Is [email protected] safe to install?
KillerSkills ran it in the KillerSkills honeypot sandbox (2 scenarios; startup only; authenticated tool behavior not tested) on 2026-11-18 and scored it 72/100 (Caution). It is not known to be malicious, but it has behaviour or code a security policy should decide on before install. Notable findings: Access to sensitive file paths; Access to sensitive file paths; Collapsed <details> block carrying instructions. Summary: no network, touches nothing outside the project, writes no memory.
Behavior trace
Honeypot run · 20 s · 2 scenarios · 0 deviationsStatic findings
// npx cache sweep cannot break them. Tests pass scriptDir explicitly andFix: Pin versions and sources; vendor anything fetched at install time.symptom: '没有 .dockerignore,node_modules/.git/.env 全被塞进构建上下文和镜像',Fix: Stop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.* this, `npx dsh-immune-system@latest connect …` writes a _npx cache pathFix: Pin versions and sources; vendor anything fetched at install time.* family: a build failing under npm and succeeding under npx is the sameFix: Pin versions and sources; vendor anything fetched at install time.* this, `npx dsh-immune-system@latest connect …` writes a _npx cache pathFix: Pin versions and sources; vendor anything fetched at install time.* 3. `ignored_by_agent` — valid proven/generic rule ignored by model → escalate delivery, don't killFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.{ pattern: /环境变量|loadenv|import\.meta\.env|vite_app|process\.env/i, expansions: ['loadEnv', '环境变量', 'vite.config'] },Fix: Stop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.immune sync-doc # update <!-- immune:generated --> block in AGENTS.md or CLAUDE.mdFix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.<summary><b>Native DSH & ZCode install details + manual MCP JSON</b></summary> **DeepSeek Harness (DSH)**: ```bash dsh plugin --profile <profile> add dsh-imFix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.* an npx cache cleanup or an upgrade of the copy that ran `connect`. WithoutFix: Pin versions and sources; vendor anything fetched at install time.prescription: '加 .dockerignore(node_modules、.git、dist、.env),镜像更小也更安全',Fix: Stop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.* family: a build failing under npm and succeeding under npx is the sameFix: Pin versions and sources; vendor anything fetched at install time.what: 'git clean -f 会永久删除未跟踪文件(-x 连 .env/node_modules 一起删)',Fix: Stop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.* - Clean scoped command tactics (`npm run build -> npx tsc`) score 55..75.Fix: Pin versions and sources; vendor anything fetched at install time.* an npx cache cleanup or an upgrade of the copy that ran `connect`. WithoutFix: Pin versions and sources; vendor anything fetched at install time.<summary><b>展开查看全部 21 项 MCP Server 工具</b></summary> | MCP 工具名 | 功能说明 | | --- | --- | | `immune_match` | 针对当前任务检索匹配抗体并记录投递回执 | | `immune_list` | 只读浏览抗体库(绝不虚增命中计数Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.* - Clean scoped command tactics (`npm run build -> npx tsc`) score 55..75.Fix: Pin versions and sources; vendor anything fetched at install time.* another (`npm run build` fails, `npx vite build` works; `pnpm` replaced byFix: Pin versions and sources; vendor anything fetched at install time.Cost to the agent
Tokens estimated at four characters each. The tool list is re-read by the agent on every turn, so it is the number to compare between servers that do the same job: ranked across the index.
Known vulnerabilities
Not checked against OSV yet; the weekly check has not reached this version.
What would make this Verified
8 points to 80 · weights: behavior 45%, static 25%, prompt 15%, publisher 15%- Unclaimed publisher≈ +8 on the totalClaim meat-daemon-404: a verified publisher gets +50 on publisher trust. It is free and usually instant with GitHub sign-in.
- Hidden instructions≈ +6 on the totalRemove HTML comments, zero-width characters and white-on-white text that carry instructions. Everything the agent is meant to read should be visible to the person reading the file.
- Access to sensitive file paths≈ +4 on the totalStop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.
- New publisher≈ +3 on the totalTime clears this on its own after 14 and 30 days. Claiming the handle (+50) is the fastest way past it.
- npx invocation of non-well-known package≈ +2 on the totalPin versions and sources; vendor anything fetched at install time.
- Process kill command≈ +2 on the totalBuild commands from fixed arguments; do not interpolate untrusted strings into a shell.