Stats
Actions
Tags
From perf-kp-sql
Kunpeng ARM64 + multi-database (MongoDB / MySQL / Redis) joint performance diagnosis. Runs SSH-based remote collection (50 OS metrics + per-engine runtime), evaluates 59-39 enabled rules from a sqlite knowledge base (FTS5 trigram + sqlite-vec 384-dim semantic search), and emits an impact-ranked HTML report with authoritative citations. Use when users report database slowness, CPU spikes, latency jitter, query timeouts, or are doing Kunpeng migration / config audit. Triggers include '数据库慢' / 'CPU 高' / '抖动' / 'mongo perf' / 'mysql 慢查询' / 'redis 延迟' / 'Kunpeng 性能' / similar phrases. First-time use:run `/perf-kp-sql-setup` to install native deps.
How this skill is triggered — by the user, by Claude, or both
Slash command
/perf-kp-sql:perf-kp-sqlThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
> **首次安装后**:跑 `/perf-kp-sql-setup` 完成 native 依赖检查与安装(better-sqlite3 / sqlite-vec / ssh2 / @xenova/transformers + knowledge.sqlite 完整性)。setup skill 会在缺依赖时给出 `npm install` 命令并自动执行。
首次安装后:跑
/perf-kp-sql-setup完成 native 依赖检查与安装(better-sqlite3 / sqlite-vec / ssh2 / @xenova/transformers + knowledge.sqlite 完整性)。setup skill 会在缺依赖时给出npm install命令并自动执行。
每次本 skill 触发,直接进入 Step 1 — /perf-kp-sql-setup 已经把 build 产物和 native 依赖都校验过了。运行时若仍出现 Cannot find module 'better-sqlite3' 或 NODE_MODULE_VERSION X != Y,提示用户重跑 /perf-kp-sql-setup。
ssh CLI (key auth recommended; password auth via sshpass on supported agents) runs per-engine batch commands on the remote host~/.ohsql/tmp/perf-kp-sql-<engine>-{os,db}-<ts>.txt (NOT /tmp — sandboxes vary)node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/diagnose.mjs --os-file ... --db-file ... --engine <name>${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/data/<engine>/ + data/common/node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_DEP_CPU_FLAMEGRAPH_ROOT}/scripts/capture.mjs ... (cpu-flamegraph plugin)node .../scripts/render-html-report.mjs + render-screen-footer.mjs⚠️ Some agent shell sandboxes reject heredoc /
>/</|/ 3+ consecutive quotes. Strictly use write-file → shell (no redirection) → write-file for the collection chain.
This skill SSHs to the target host multiple times. Use the agent's shell tool to invoke local ssh / scp CLI (no SshExec / FlameGraph / Sql* — those bind to one specific kernel). Both auth modes are supported:
Mode A · SSH key auth (recommended, all agents) — when privateKeyPath=<path> was provided:
ssh -o BatchMode=yes -o StrictHostKeyChecking=accept-new \
-i <privateKeyPath> -p <port> <user>@<host> '<command>'
Mode B · SSH password auth (Claude Code + ohsql only) — when password=<pw> was provided:
sshpass -p '<password>' ssh -o StrictHostKeyChecking=accept-new \
-p <port> <user>@<host> '<command>'
On OpenAI Codex CLI (sandbox blocks sshpass) — when only password=<pw> is given, stop the workflow and ask the user:
Codex CLI's sandbox blocks
sshpassso SSH password auth is unavailable here. Please run once on your local machine:ssh-copy-id -i ~/.ssh/id_ed25519.pub <user>@<host>Then re-invoke this skill withprivateKeyPath=~/.ssh/id_ed25519instead ofpassword=....
stdout-non-empty = success rule: if shell returns non-empty stdout, treat as success regardless of exitCode / red-rendering / stderr WARN/deprecated noise. Only when both stdout and stderr are empty does SSH actually fail (no auth / no route).
This skill runs 5 phases. Track them with the agent's task-list facility:
TodoWrite (5-element array, status field per item)update_plan (5 steps, status field per step)◻ ◼ ✔ markersEach phase transition (in_progress → completed; pending → in_progress) re-sends the full updated task list. Detailed phase titles + counts are defined in Step 1.4.
banner 必须在任何远端 SSH 命令(or remote scp / node ...flamegraph capture invocation)之前渲染。 本地参数收集(history load · prompts to user)不受限。
触发:slash args 缺 host 时。
Bash(command="node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/history.mjs --op load --max 5")
hosts 非空 → ask the user to pick one (even if there's only 1 entry — explicit confirmation):
请选择最近使用过的连接 · 或新建:
1. 124.70.180.36 · root · mongo · 上次 2 小时前 (port=22 · 累计 8 次)
2. 10.20.30.40 · ec2-user · mysql · 上次 3 天前 (port=22 · 累计 3 次)
3. 新连接 · 手动输入参数
Stop here and wait for the user's selection in the next turn. Once selected, fully decode the JSON (host / user / port / engine / password / privateKeyPath / mongo_user / mongo_password / auth_db) into the parameter set.
No stopping after selection: same turn → render banner → declare 5-phase task list → mark phase 1 in_progress → run SSH OS collection.
从用户任意措辞抽取:
host(IP/FQDN)、user、password(或 privateKeyPath)port(默认 22)、engine(mongo/mysql/redis)mongo_user、mongo_password、auth_db(默认 admin)抽取策略:严格 kv → 半结构化 → 自然语言 → 混合。抽取失败只问缺的字段,不重来整表。
两类 check,任一命中阻塞 banner:
Class 1 缺字段 — host / user / (password OR privateKeyPath) 任一缺:
Ask the user (with the missing field name as the topic):
━ kunpeng · 参数待补全 ━ 还缺:<缺字段名>(例:SSH 密码 / SSH 私钥路径) 请补充。Stop and wait for the next turn.
Class 2 格式非法 — host 非合法 IP/FQDN、port 非 1-65535、engine 不在支持集合:
Ask the user (with the bad field name as the topic):
━ kunpeng · 参数格式异常 ━ <字段名> 格式不对:<原值> → <期望格式 / 合法集合> 请重新提供。Stop and wait for the next turn.
<字段名> 只写具体名(SSH 密码 / 主机格式 / 端口格式),不用模糊词。
参数齐备后打:
[连接信息]
· host=124.70.180.36 · user=root · port=22 · engine=mongo
· password=YAN***216
· mongo_password=YAN***216 · auth_db=admin
password 前 3 + *** + 后 3 脱敏。后续 SSH 命令的 host/user/port/password/privateKeyPath 参数必须与 banner 字段一一对应。
Gate 2 — SSH 命令的 host/user/port/password/privateKeyPath 必须与 banner 字面一致。
Gate 3 — history 里 password 非空 → 必传;privateKeyPath 非空 → 必传;两者都空 → 先问凭据。
Gate 4 — SSH 命令返 stdout=stderr="" 时:
Declare a 5-phase task list using the agent's task-tracking facility (see "Task tracking pattern" in the Architecture section above). On Claude Code use TodoWrite; on OpenAI Codex CLI use update_plan; otherwise render as plain text.
If a previous task list exists from an earlier run in the same session, replace it (idempotent — both TodoWrite and update_plan overwrite the whole list). Do NOT append.
The 5 phases (English subject for portability, with Chinese display titles):
| # | Display title (subject) | Spinner verb (activeForm) | Count placeholder |
|---|---|---|---|
| 1 | OS/硬件采集 (50 项) | 采集 OS/硬件 | 50 (fixed) |
| 2 | 运行时采集 ( 项) | 采集运行时 | :mongo=18; mysql/redis omit (N 项) |
| 3 | 规则诊断 ( 条) | 诊断规则 | :mongo=59; mysql=38; redis=39 |
| 4 | 知识库检索 (54 篇) | 检索知识库 | 54 (fixed) |
| 5 | 报告渲染 | 渲染报告 | (no count) |
Exactly 5 phases. Do NOT create extras (no "cleanup" or "init" phases). Do NOT assume sequential IDs (1-5) — use whatever IDs the agent's task tool returns.
v0.5.1 命名规则:subject 名词在前 + 动词在后(X采集 / X诊断 / X检索 / X渲染)· ≤ 5 字 · activeForm 动词在前(spinner verb 自然中文)· 不挂 engine 前缀。
数字硬编表(数据来源 = 实测 · 不要编):
| engine | task 1 (OS/硬件采集) | task 2 (运行时采集) | task 3 (规则诊断) | task 4 (知识库检索) |
|---|---|---|---|---|
| mongo | 50 项 | 18 项 | 59 条 | 54 篇 |
| mysql | 50 项 | (不挂) | 38 条 | 54 篇 |
| redis | 50 项 | (不挂) | 39 条 | 54 篇 |
src/shared/os-collector.ts 中 out["..."] = ... 的去重 key 数(50)src/engines/<engine>/collector.ts 中 out["..."] = ... 的去重 key 数(mongo=18 · mysql/redis 用结构体返回不平摊 · 不挂)sqlite3 data/knowledge.sqlite "SELECT count(*) FROM rules WHERE enabled=1 AND (engine='<engine>' OR engine='any')"sqlite3 data/knowledge.sqlite "SELECT count(DISTINCT doc_id) FROM knowledge"mysql/redis 的 task 2 项数无法平摊 · 整段 (<N> 项) 删掉 · 不要瞎编。
Mark phase 1 (OS/硬件采集 (50 项)) as in_progress in the task list. Re-send the full 5-element list with status updated.
Do not print a literal "phase 1 / step 2" header — let the task-list UI render the spinner.
Read(file_path="${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/data/collect-cmds.json")
强制用 ~/.ohsql/tmp/(绝对路径),不用 /tmp。TS = YYYYMMDD-HHMMSS,全程一致。Write 自动创建父目录。
SSH 命令之前打 4 行分组活动行(v0.5.1 · 跟 task 1 detail 视觉对齐 · 替代旧单行 · 采集 OS / 硬件 35 项指标 (...)):
· 内存 · THP / 大页 / swappiness / ...
· 网络 · somaxconn / keepalive / ...
· 磁盘 · await / 调度器 / ...
· CPU · LSE / BIOS / NUMA / ...
Run via the SSH execution pattern (see Architecture section), substituting <command> with the literal osBatchCmd string read from data/collect-cmds.json (do NOT improvise the bash command):
# Mode A · key auth (recommended):
ssh -o BatchMode=yes -o StrictHostKeyChecking=accept-new \
-i <privateKeyPath> -p <port> <user>@<host> '<osBatchCmd>'
# Mode B · password auth (CC + ohsql):
sshpass -p '<password>' ssh -o StrictHostKeyChecking=accept-new \
-p <port> <user>@<host> '<osBatchCmd>'
# Codex CLI with password-only: stop and ask user to use ssh-copy-id (see Architecture).
Set timeout ≈ 60 seconds.
stdout-non-empty = success, proceed immediately. Ignore exitCode / red rendering / stderr WARN/deprecated noise.
Only both stdout and stderr empty → SSH not established; run Gate 4 self-check.
First failure · compact retry — ask the user:
━ SSH 未连通 · 已带全凭据仍空 ━ 可能原因:port 错 / 密码错 / 私钥未授权 / 防火墙阻断 请确认 host/user/port/(密码或私钥) 是否正确,或在目标主机跑
ss -lntp | grep ssh确认 sshd 端口。
Stop and wait for the next turn.
Second failure → expand the detailed troubleshooting checklist (port / firewall / sshd config).
拿到 osStdout 后 Write 落盘 (注意：必须提供 file_path 和 content 参数，绝对禁止空参数调用 Write()):
Write(
file_path="/Users/<yourlogin>/.ohsql/tmp/perf-kp-sql-os-<TS>.txt",
content="<此处必须填写你从上一步 SSH 命令刚拿到的全部 stdout 文本，别漏了>"
)
Write 顺序硬约束:SSH 命令 → Write 落盘 → discover.mjs,三步缺一不可。
Bash(command="node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/ssh.mjs --op discover --os-file <path> --hint-engine <engine|skipped>")
选择诊断实例 as topic): list each candidate as <engine> @ <bind>:<port> (pid=<pid>) and stop for selection.单实例活动行根据 _notes 决定:
port-inferred-from-default 且 pid 空/0 → ✓ 数据库实例 · <engine> @ <bind>:<port>(默认端口推断)✓ 数据库实例 · <engine> @ <bind>:<port> (pid=<pid>)SSH 命令之前打分组活动行(v0.5.1 · 跟 task 2 detail 对齐 · 替代旧单行 · 采集 MongoDB 19 项指标 (...)):
# engine=mongo · 3 行
· WiredTiger · cache / 并发 ticket / ...
· 连接池 · 当前 / 可用上限 / ...
· 锁与断言 · global lock / asserts / ...
# engine=mysql · 2 行
· 引擎状态 · variables / status / processlist / ...
· 连接 · max_connections / threads_connected / ...
# engine=redis · 2 行
· INFO 指标 · memory / stats / clients / ...
· slowlog / 大 key / ...
然后 run via the SSH execution pattern with <command> = the substituted dbBatchTemplates[engine] string from collect-cmds.json:
# Same SSH pattern as Step 2.3 (Mode A key auth or Mode B password+sshpass).
# Substitute the engine's batch template after placeholder replacement.
Placeholder substitution:__BIND__、__DB_PORT__、__USER__、__PWD__.
MongoDB-specific:__MONGO_USER__、__MONGO_PWD__、__AUTH_DB__、__AUTH_ARGS__(if both mongo_user and mongo_password provided, fill auth args; else leave empty).
Set timeout ≈ 60 seconds.
MongoDB authentication failure auto-collect: if stderr contains requires authentication / Authentication failed, print a translated Chinese activity line and ask the user for credentials:
━ MongoDB 账号 · 密码 · 认证库 ━ 请提供:
- mongo_user(默认空 = 匿名)
- mongo_password
- auth_db(默认 admin)
提供后将自动重试,最多 3 轮。
Stop and wait for the next turn.
Engine discovery succeeded → MUST read the engine's dbBatchCmd template from collect-cmds.json, substitute placeholders, and run as the SSH command. Do NOT improvise mongosh / mysql client / redis-cli invocations.
最多 3 轮。拿到 dbStdout → Write 落盘 (同样必须提供 file_path 和 content 参数):
Write(file_path="/Users/<yourlogin>/.ohsql/tmp/perf-kp-sql-<engine>-db-<TS>.txt", content="<此处填入上一步 SSH 命令采集到的全部数据库状态 stdout 文本>")
In the task list, mark phase 1 as completed and phase 2 as in_progress (single update, both transitions in one re-send of the full list).
打硬件/OS 识别行:
· 硬件 · <cpu_model> · <arch> · <cpu_cores> core · <GB>
· 操作系统 · <os_id> <os_version> · Linux <kernel_version> · <virt>
填值:total_mem_mb 除 1024 得 GB(1位小数);numa_nodes ≥ 2 才显示;cpu_model 空写"CPU 未识别"。
After printing the data lines, mark phase 2 as completed.
SSH + DB 都成功后静默调:
Bash(command="node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/history.mjs --op save --host <host> --user <user> --port <ssh_port> --engine <engine>")
失败静默忽略,不打屏。
Mark phase 3 (规则诊断 (<R> 条)) as in_progress in the task list.
打 3 行严重度分组活动行(v0.5.1 · 跟 task 3 detail 对齐 · 替代旧单行 · 加载 N 条规则(蒸馏自 X 篇 ...)):
· critical <C> · <代表项 1> / <代表项 2> / <代表项 3> / ...
· warning <W> · <代表项 1> / <代表项 2> / <代表项 3> / ...
· info <I> · <代表项 1> / <代表项 2> / <代表项 3> / ...
数字硬编(实测 · 不要编):
| engine | 总规则 R | critical C | warning W | info I |
|---|---|---|---|---|
| mongo | 59 | 5 | 35 | 19 |
| mysql | 38 | 3 | 20 | 15 |
| redis | 39 | 4 | 21 | 14 |
来源 sqlite3 data/knowledge.sqlite "SELECT severity, count(*) FROM rules WHERE enabled=1 AND (engine='<engine>' OR engine='any') GROUP BY severity"。代表项从 diagnose.mjs 输出 top issues 抽 · 优先 critical · 不要瞎编。
Bash(command="node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/diagnose.mjs --engine <engine> --summary-only --os-file <os-path> --db-file <db-path> --out-json <diag-json-path>")
--summary-only 避免 LLM 上下文爆炸。返回 summary JSON < 2KB。完整 JSON 写到 --out-json 供后续脚本消费。
矩阵不在此处渲染,推迟到 Step 4.3。
Mark phase 3 as completed and phase 4 (知识库检索 (54 篇)) as in_progress.
首选 read 目录 + INDEX.md:
Read(file_path="${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/data/<engine>/INDEX.md")
Read(file_path="${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/data/<engine>/<file>.md")
可选 Grep(报 rg 未安装则回退 Read)。
执行位置:Step 4.0 kb-stats 之后、Step 4.2 render-html-report 之前。一次诊断只调一次。
Print activity line:
· 检查远端 perf / offcputime-bpfcc 依赖
Run via the SSH execution pattern (Architecture section) with <command>:
echo "perf=$(command -v perf 2>/dev/null || echo MISSING)"; \
echo "offcpu=$(command -v offcputime-bpfcc 2>/dev/null || echo MISSING)"
Bash(command="node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_DEP_CPU_FLAMEGRAPH_ROOT}/scripts/capture.mjs \
--host=<ip> --user=<user> --password='<真实pw>' [--port=<n>] \
--process=mongod --type=oncpu --duration=3 --engine=mongo")
凭据用真实未脱敏密码,单引号包裹。offcpu 齐则再调一次 --type=offcpu。
拿到 JSON 后暂存 flame-json 路径,不立即 echo terminalReport。等 Step 4.3 时传给 render-screen-footer.mjs 处理。同时记录 artifacts.serverSvgPath 用于拉 SVG。
必须立即使用 Write 工具将该完整的 JSON 落盘至 /Users/<yourlogin>/.ohsql/tmp/flame-json-<TS>.json。
Bash(command="node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/ssh.mjs --op exec \
--host <ip> --user <user> --password <pw> [--port <n>] \
--command 'cat <artifacts.serverSvgPath>' \
--output-file /Users/<yourlogin>/.ohsql/flame/<TS>-oncpu.svg")
LLM 不接触 SVG 内容。双采各拉一次。
任一环节失败 → 整段跳过,静默进 Step 4。
Bash(command="node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/kb.mjs --op stats --engine <engine>")
返回 subtitle 字段存入报告 metadata。
Mark phase 4 as completed and phase 5 (报告渲染) as in_progress.
node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/render-html-report.mjs \
/Users/<yourlogin>/.ohsql/reports/perf-kp-sql-<engine>-<TS>.html \
--from-diagnose <diag-json-path> \
--from-flame-json /Users/<yourlogin>/.ohsql/tmp/flame-json-<TS>.json
After the shell returns ok, mark phase 5 as completed (final state: all 5 ✔).
对每条 critical/warning rule:
Bash(node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/kb.mjs --op query --rule-id <rule_id> --engine <engine> --top-k 3)
query-kb 失败 → 跳过,不影响主体。
Bash(command="node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/render-screen-footer.mjs \
--from-diagnose <diag-json-path> \
--report-path <html-报告绝对路径> \
[--from-flame-json <flame-json-path>]")
LLM 必须用 Bash 工具执行上述命令。 然后用 echo -e 原样输出该脚本的 stdout(必须直接输出 markdown pipe 表格原样文本，绝对不许重绘为 box-drawing ┌─┬─┐)。
不再需要 LLM 单独 echo 火焰图的 terminalReport，因为脚本已经读取了 flame-json 并自动把它合并到了 footer 里面，而且参考编号也一起合并了。
[!NOTE] 报告要求
- LLM 不许擅自用 Markdown 重写 Footer。完全依赖 render-screen-footer.mjs 的原生输出。
- 不许单独包裹火焰图内容为代码块。
用户下一轮触发,非主流程。
关键词:采 / 火焰图 / profile / on-cpu / off-cpu 等。
type=both,duration=3,凭据从上轮拿。
与 Step 3.3.2 完全一致,调 capture.mjs。决策树同 3.3.1。
拿到后 echo terminalReport,拉 SVG 到本地(同 3.3.3)。
Read 原报告 → 追加火焰图节 → Write 回写。
对 self ≥ 5% 的热点:函数名+库 → 推断 → 关联本次 findings → 建议。
每行必带 [F<rank>](采样证据),每条解读必带 [参考N](KB 背书)。
查 KB 流程:
Bash(node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/kb.mjs --op query --flame-function <func> --module <mod> --engine mongo --top-k 3)
三层 fallback:flame_pattern_regex → context-ancestors → module → 模板 B(无 [参考N])。
建议只能落在:配置/参数、架构、硬件、OS/内核。
skill 加载后,任何非 / 命令的自然语言输入一律走 query-kb:
Bash(node ${CLAUDE_PLUGIN_ROOT:-$OHSQL_PLUGIN_ROOT}/scripts/kb.mjs --op query --q "<原话>" --engine <engine> --top-k 5)
results 非空 → 模板 A(从 content_zh 提炼,句末 [参考N],末尾列 URL)。
results 空 → 模板 B:
非常抱歉 · 您的 KB 中并未收录「<主题>」的相关资料。
如果基于 KB 之外的一般常识回答(请注意 · 非 KB 来源 · 需独立验证):
<一句话 · ≤50字 · 末尾 "请独立验证">
模板 B 不挂 [参考N]。涉及命令/代码/阈值的回答,要么 KB 背书带 [参考N],要么走模板 B 带 disclaimer。
ssh / scp CLI(see Architecture · SSH execution pattern); 不调任何 agent 的私有 SSH tool(SshExec / SshUpload / FlameGraph 等)python3 -c ... 行内 hack 替代写文件find 脚本路径;MODULE_NOT_FOUND 唯一根因是缺 build 产物/tmp/ 落盘| 文件 | 用途 |
|---|---|
data/collect-cmds.json | 3 engine SSH 采集命令 |
data/common/ | Kunpeng + OS 通用知识库 |
data/<engine>/INDEX.md | per-engine KB 目录 |
data/<engine>/rules.json | per-engine 规则 JSON |
data/<engine>/*.md | per-engine 文档 |
scripts/ssh.mjs --op exec | SSH 远端执行 |
scripts/ssh.mjs --op discover | 实例发现 |
scripts/history.mjs --op load | 读历史连接 |
scripts/history.mjs --op save | 写历史连接 |
scripts/diagnose.mjs | 本地分析脚本 |
scripts/kb.mjs --op query | KB 混合检索 |
scripts/kb.mjs --op stats | KB 统计 |
scripts/render-html-report.mjs | HTML 报告生成 |
scripts/render-screen-footer.mjs | 屏幕 footer 生成 |
perf-kp-sql-<engine>-<TS>.html,TS = YYYYMMDD-HHMMSSFor agents that don't render the argument-hint frontmatter as UI hints
(e.g. OpenAI Codex CLI), this section duplicates the parameters in body
prose so the LLM can quote them back to the user.
Required:
host=<ip> — target host (IP or FQDN; e.g. 10.0.0.1)user=<user> — SSH user (e.g. root, ec2-user)privateKeyPath=<path> (recommended) OR password=<pw>
privateKeyPath: SSH key file path (e.g. ~/.ssh/id_ed25519) — works on all agentspassword: SSH password — Claude Code + ohsql only (uses sshpass); OpenAI
Codex CLI sandbox blocks it, so Codex users must run ssh-copy-id once and
switch to privateKeyPathOptional:
engine=<mongo|mysql|redis> — database engine; auto-detected if omittedport=<ssh_port> — SSH port (default: 22)mongo_user=<user> — MongoDB auth user (auto-asked on auth failure)mongo_password=<pw> — MongoDB auth passwordauth_db=<db> — MongoDB auth database (default: admin)Examples:
/perf-kp-sql host=10.0.0.1 user=root privateKeyPath=~/.ssh/id_ed25519 engine=mongo
/perf-kp-sql host=10.0.0.1 user=root password=secret engine=mysql port=2222
/perf-kp-sql host=db.internal user=ec2-user privateKeyPath=~/.ssh/aws-prod
Searches MemPalace before answering questions about past work, people, projects, or prior decisions. Returns verbatim stored content instead of guessing from model memory.
Guides Payload CMS config (payload.config.ts), collections, fields, hooks, access control, APIs. Debugs validation errors, security, relationships, queries, transactions, hook behavior.
Implements vector databases with Pinecone, Weaviate, Qdrant, Milvus, pgvector for semantic search, RAG, recommendations, and similarity systems. Optimizes embeddings, indexing, and hybrid search.
npx claudepluginhub zlxtqbdgdgd/ohsql-plugin --plugin perf-kp-sql