Spider King skill

---
name: spider-king
description: "Pure-web protocol reverse skill: turn hostile browser clients into browser-free Python collectors. Auto-judge intake and tool path: artifact-only when samples suffice; fresh web live-target needs sequential fingerprint-baseline then debugger-trace (default chrome-devtools then js-reverse; if DP is already required js-reverse may own both roles on one host and must not launch a second Chrome; replay identity locks to installed curl impersonate; unknown risk-control does not bulk-add unproven extra routes; Camoufox/managed baseline only on fingerprint pressure; silent-value-capture is a first-class debugger-trace means when attach is missing or page-world capture is insufficient/observer-toxic). Continuation reuses the current gate. Route only mounted web MCP families. Out of scope as primary path: APK, native app, mini-program. Use for hostile web sign, token, cookie, WebSocket, GraphQL, protobuf, response-decode, fingerprint, WASM, challenge-bootstrap, dynamic-font, or protocol-collector flows."
---

# Spider King

## Mission

Turn hostile web clients into stable protocol collectors.

This is a pure-web protocol-recovery skill, not a browser-automation skill and not an APK/mini-program reverse skill. Use browser tooling only to gather web evidence. Deliver raw HTTP plus narrow local sign, bootstrap, decode, or transport helpers. If the primary target is APK, native app, or mini-program, state out-of-scope instead of inventing a web dual-browser first-pass.

## Non-Negotiables

- Before any tool choice on a fresh task, run the Auto Judge Card in this file and `references/mcp-routing-playbook.md`: decide intake, whether a browser is needed, which baseline host to use, and which debugger-trace means to use (`js-reverse` and/or `silent-value-capture`). Do not open tools first and rationalize later.
- Start every fresh web target classified as `live-target` with sequential evidence for both roles before claiming that the page/runtime is understood: `fingerprint-baseline` then `debugger-trace`. The only narrower claim is the `transport-only direct-endpoint exception` in `references/startup-triage-playbook.md`: it may accept a collector for one exact endpoint without claiming page/runtime understanding, and only after its strict negative-control and business-oracle gates pass.
- Default means are baseline `chrome-devtools` then debugger-trace `js-reverse`. If DP/initiator work is already required, `js-reverse` may own both roles on one host; do not launch a second Chrome. Record `launch` vs `attach` and never pretend launch is attach. Upgrade the baseline host to Camoufox or another managed profile only on proved fingerprint pressure or clean-baseline failure; never default Camoufox for ordinary low-risk work or because CDP `navigator.webdriver` is true.
- Satisfy `debugger-trace` with correlated initiator/mutation proof from `js-reverse`, correlated `silent-value-capture` cards under `references/silent-value-capture-playbook.md`, or both. Prefer `js-reverse` when attach exists.
- If attach is missing, record `debugger_attach_gap` and continue debugger-trace via `silent-value-capture` when a backend is available instead of abandoning the role. If a needed silent backend is also missing, record `silent_value_capture_gap` and continue offline or lower the delivery shape.
- Silent-value-capture never replaces baseline and never owns `compact-replay` or `collector` delivery. APK, native app, and mini-program primary tasks remain out of scope and must not invent paired-browser proof as ceremony. If a required role cannot be satisfied by any available means, report the blocker before claiming the live target is understood.
- Keep browser and engine-capture evidence collection target-serial: at most one browser or engine-capture family may be `TARGET_ACTIVE`, and never place two target owners in the same parallel tool batch. Apply the handoff gate in `references/tool-playbook.md`; park or retain baseline before granting silent-value engine capture; preserve unique unreplayable state with `RETAINED_EXCEPTION` instead of destroying it for cleanup.
- For `compact-replay` and `collector`, deliver a browser-free Python run path. Never use browser automation, Playwright, Selenium, CDP page-driving, page-context fetch, browser profiles, manual browser state, silent-value-capture backends, or engine instrumentation as the final replay path or fallback.
- Prefer pure Python for HTTP, orchestration, parsing, retries, persistence, and output. Keep JS/WASM only as a tiny local helper when a verified Python port would be riskier; a local bootstrap may emulate host semantics, but Python owns live HTTP and the report states browser-free/runtime-free truth. Replay identity locks to the installed curl impersonate chrome major; do not copy a newer live Chrome major onto a lower impersonate or restart risk-control reverse for that drift.
- Recover one stable business request before pagination, concurrency, submission, or runtime shrinking. Unknown risk-control reverses the rejected request first; unproven extra routes are at most one per round and only after omission changes the oracle.
- When a local JS/WASM/SDK helper is called repeatedly in one browser-shaped workflow, treat process/VM/context identity and hidden mutable state as moving parts; prove persistent-world versus fresh-world behavior before shrinking or widening the helper, and do not reuse a self-invalidating instance as a pager.
- Back every conclusion with the evidence its declared shape requires. Require repeated live replay only when claiming current live acceptance. Keep raw sensitive values local; redact credentials, tokens, personal data, and cookie values from chat, reports, and version control.
- Never hardcode rotating state before proving its writer, slot, scope, expiry, and refresh path; preserve one session chain for bootstrap-heavy flows until cross-session reuse is proven.
- Stop only when the declared shape passes its capability-specific gate or a real external blocker is proved. Do not label `evidence` or `local-proof` as a collector, and do not disguise incomplete automation as a temporary collector.
- `compact-replay` and `collector` delivery include a PyCharm right-click runnable Python entrypoint, normally project-root `main.py` or `collector/main.py`, with no required terminal arguments. `evidence` and `local-proof` do not require an entrypoint unless the requested artifact is executable code.
- Route only MCP families confirmed in the current tool registry/schema; disk source trees are not availability. Passive wire stores may coexist with `TARGET_ACTIVE` but do not replace live first-pass evidence, and `compact-replay`/`collector` remain free of browser runtime, page driving, profiles, capture backends, and engine instrumentation.

- Before hardening a gate-family label, widening runtime, or handing work to an installed specialist, pass `references/evidence-corroboration-gate.md`. One status code, cookie name, token name, old note, or plausible helper output is only a hypothesis.
- When a trusted positive sample passes and the local path fails, use `references/positive-sample-ablation-playbook.md` after sample hygiene. Name unfinished live failures with `references/failure-surface-taxonomy.md` instead of implying collector completion. When stuck, pick one Stuck next action from that file; do not bulk-add unproven extra routes.

For failure-shaped counterexamples, read `references/anti-patterns-playbook.md`. Before packaging a result, apply only the matching capability gate from `references/delivery-gate-playbook.md`. Graded degrade: import/`NameError` is our bug (alarm on stderr); external timeout is fallback. Print key responses to stderr; keep stdout machine JSON.

## Lightweight Dispatch

Use these labels internally to keep the first move small. Do not force a rigid machine header in normal user replies.