AGENT SKILL

bc-drive-plus

Use when driving a LIVE Battle Cats (jp.co.ponos.battlecatsen, EN 15.4.1) on rooted BlueStacks beyond bc-drive. Adds bounded one-shot touch-dispatch capture to resolve the stripped NewButton ctor RVA, OmniParser V2 set-of-mark grounding for the DOM-less GLES canvas, minitouch/evdev low-latency input, and a save-snapshot ring with auto-rollback. Needs adb + frida-server; builds on bc-drive and bcsfe-save-roundtrip.

License 许可
MIT
最近更新
2026-11-02
Stars
★ 0
Forks
0

01 前言 Front Matter

name
bc-drive-plus
license
MIT
description
Use when driving a LIVE Battle Cats (jp.co.ponos.battlecatsen, EN 15.4.1) on rooted BlueStacks beyond bc-drive. Adds bounded one-shot touch-dispatch capture to resolve the stripped NewButton ctor RVA, OmniParser V2 set-of-mark grounding for the DOM-less GLES canvas, minitouch/evdev low-latency input, and a save-snapshot ring with auto-rollback. Needs adb + frida-server; builds on bc-drive and bcsfe-save-roundtrip.

02 资源文件 2

03 SKILL.md 正文

bc-drive-plus — advanced Battle Cats driving (dispatch-capture · grounding · input · snapshot-ring)

Additive layer over bc-drive (read the base skill first). Same target: a live Battle Cats EN 15.4.1, x86_64 libnative-lib.so, on BlueStacks (adb 127.0.0.1:5555, frida 17.11.0). Every enhancement rides the existing bus — the one-shot-NativeFunction / OS-tap / screenshot discipline — and never reintroduces a per-frame render hook (which crashes BC on scene transitions).

The four enhancements

# Capability File / status Risk
1 Bounded one-shot dispatch-capture (recover stripped NewButton ctor RVA) dispatch_capture.py ✅ built + unit-tested; live run needs device live-frida input-path hook — thread-selection safety honored
2 OmniParser V2 set-of-mark grounding for the GLES canvas recipe below (ADOPT) — rides adb screencap bus read-only (screenshot)
3 minitouch/evdev low-latency input recipe below (ADOPT) — falls back to adb tap actuation only on safe menu nav
4 Save-snapshot ring + auto-rollback (HGT-safe-restore) snapshot_ring.py ✅ built + unit-tested; live run needs device stateful save write — auto-rollback shadow-by-default

Shadow-mode is the universal discipline (council C5): any automated / stateful action defaults to OBSERVE-only (log "would do X", mutate nothing); only an explicit operator command actuates. Nothing here blocks or nags.


1 · Bounded one-shot dispatch-capture (dispatch_capture.py)

Closes bc-drive open gap #1. bc-drive documented that the NewButton class is known from RTTI in .rodata (callback void(NewButton*, NewButtonChangeState), vtable RVA 0xBED598) but its constructor RVA is unresolved because the game's own C++ symbols are stripped. This module recovers the runtime evidence to resolve it, without a per-frame hook.

Mechanism

The touch-dispatch function is already resolved read-only by bc-drive: 0x668C70 == getGlobal()->vtable[+0x40]. It fires only on touch events (not per frame). We attach a single-shot Interceptor to it; on the first tap we capture — minimally — (phase, x, y, pid), the return address (caller RVA), and a short accurate backtrace (module-relative RVAs), emit them asynchronously, then detach. With the known NewButton vtable RVA and (optionally) the touched node's own vtable, an offline xref pass localizes the ctor.

Honest boundary: the module yields the evidence (caller RVAs ranked by frequency + any hit-node vtable RVA + a match flag against 0xBED598); the final ctor RVA is confirmed by the documented offline step below — it is not hallucinated as a printed number.

Offline follow-up to confirm the ctor RVA

  1. In Ghidra/IDA on the pinned 15.4.1 libnative-lib.so, go to the NewButton vtable RVA 0xBED598 (from RTTI).
  2. Find the function that stores that vtable into a fresh object (mov [obj], <vtable_addr>) — that is the NewButton constructor.
  3. Cross-check it against the caller_rvas_ranked list from the capture (the live dispatch → widget callback path). If hit_node_matches_newbutton_vtable is true, node_ptr is a live NewButton and its ctor is the storer of its vtable.
  4. Feed the confirmed ctor RVA back into bc-drive's capture <ctorRVA> <ms> to dump live button objects + candidate rect floats (its infra was built and ready).

THREAD-SELECTION SAFETY (the flagged live-frida risk — how it is honored)

The touch dispatcher runs on the game's UI/input thread. An Interceptor callback runs on that same thread. The safety rules baked in:

  • No synchronous host RPC inside onEnter. A blocking recv()/RPC would stall the input thread → ANR/crash. The agent uses only async send(). This is the thread-selection flag.
  • Single-shot. The agent self-limits to max_hits (default 1); the driver detaches the instant the cap is hit.
  • Scene-guarded. arm_and_capture refuses to arm unless scene is in the out-of-battle menu cluster (97..101). It will never hook the input path mid-battle (hot, timing-sensitive). Scene is read via bc-drive's on-demand scene — if it can't be read, the guard refuses (fail-safe).
  • Hard timeout. The driver auto-detaches after timeout_ms even if no touch arrives, so a hook is never left live.
  • Not a per-frame hook. 0x668C70 ≠ the render path appUpdateDraw @0x33F7E0.

Usage

python dispatch_capture.py show               # print the generated single-shot agent JS
python dispatch_capture.py capture 6000        # SHADOW: arm 1-shot, YOU tap once (6 s window)
python dispatch_capture.py capture 6000 955 665 # ARMED: arm + drive ONE safe menu tap as trigger

Shadow (capture <ms>) hooks + records and waits for your manual tap — it never actuates. Passing <x> <y> is an explicit operator-armed run that drives one safe menu tap. Output includes ctor_evidence (ranked caller RVAs + node-vtable match + the exact offline follow-up). Env: BC_ADB, BC_SERIAL, BC_DC_SINK (bounded rotating shadow log). Optional node_field_off (Python API) reads a candidate hit-node pointer off arg0 and reports its vtable RVA directly.


2 · OmniParser V2 set-of-mark grounding (ADOPT recipe)

BC's canvas is a DOM-less custom GLES surface — there is no accessibility tree to query, so bc-drive identifies screens by screenshot + vision and carries a hand-curated per-screen widget-rect map. Set-of-Mark (SoM) grounding upgrades this: overlay numbered marks on detected interactive regions so the model refers to "tap 7" instead of guessing pixel coordinates (the technique that lifts GUI grounding accuracy on canvas UIs).

Rides the existing bus — the input is bc-drive's adb exec-out screencap; the output maps mark-number → device-px rect (1920×1080), reusing bc-drive's metrics device↔drawable scale (×1.6875) for drawable back-projection.

Adopt (local, GPU-optional):

# OmniParser V2 (MIT) = YOLO icon/box detector + a caption model. Run on cloud-hub
# GPU or locally; feed it the bc-drive screenshot, get boxes back, number them.
pip install ultralytics  # icon detector backbone
# weights: microsoft/OmniParser-v2.0 (HF)  ->  hf_hub_download

Pipeline: bc_drive.py screenshot → OmniParser detect → number boxes top-left→ bottom-right → build {mark: {device_px:[cx,cy], rect:[...], drawable:[...]}} → the operator/model picks a mark → bc_drive.py tap <cx> <cy>. Status: documented ADOPT, not yet wired to a live model here (honest — grounding accuracy must be validated against real BC screens before it supersedes the curated SCREENS map). Until then bc-drive's screenshot-derived map remains the source of truth; SoM is the roadmap to generalize beyond the three hand-mapped screens.


3 · minitouch / evdev low-latency input (ADOPT recipe)

bc-drive actuates with adb shell input tap, which spawns a process per tap (~tens–hundreds of ms). For faster/smoother sequences (swipes, multi-tap), a persistent low-latency channel helps:

  • minitouch — a tiny daemon writing directly to the touch evdev node; open one socket and stream d/m/u commands (down/move/up) with sub-ms dispatch.
  • raw evdev — su -c write EV_ABS/EV_KEY events to /dev/input/eventX.

Adopt:

# push a prebuilt minitouch matching the emulator ABI (x86_64), then:
adb -s 127.0.0.1:5555 push minitouch /data/local/tmp/ && adb shell chmod 755 /data/local/tmp/minitouch
adb shell "/data/local/tmp/minitouch -i"     # prints max-contacts + coordinate bounds
# stream: printf 'd 0 <x> <y> 50\nc\nu 0\nc\n' | nc <minitouch socket>

Coordinate space: minitouch/evdev use the touch device's ABS range, which may differ from the 1920×1080 display px — read the bounds from minitouch -i and scale bc-drive's device-px targets into that range. Status: documented ADOPT with an adb tap fallback — the safety envelope is unchanged (taps only on safe menu navigation; never gacha "Draw", stage "Start" confirm, or anything spending currency). Latency is an optimization, not a correctness requirement, so adb tap stays the default until minitouch is bound for the specific BlueStacks build.


4 · Save-snapshot ring + auto-rollback (snapshot_ring.py)

Encodes the HGT-safe-restore discipline from bcsfe-save-roundtrip as automation: keep a bounded rotating ring of known-good pre-edit SAVE_DATA snapshots so any broken/suspect state can be rolled back to the last verified save. ("HGT" = the anti-cheat ban trap; the guard is: always be able to restore a clean, pre-edit save.)

Rides the bcsfe bus exactly (do not reinvent):

  • Pull after a KEYCODE_HOME flush (backgrounding flushes RAM→disk; force-stop does not flush).
  • Restore in place via cat > so the inode + owner (uid 10065) + SELinux app_data_file context survive; am force-stop FIRST so autosave can't race; md5sum verify device == snapshot after.
  • Empty-wipe trap: a 0-byte/short read is never stored as good and never pushed back (the MSYS d41d8cd9… wipe class). MIN_SAVE_BYTES gates both ends; a backup whose on-disk md5 ≠ recorded md5 is refused before any push.

Shadow vs enforce (council C5)

  • snapshot (read-only pull) and an explicit rollback are always real.
  • Automatic rollback (auto_rollback(is_broken)) defaults to SHADOW: if the health probe says broken, it logs which snapshot it would restore and mutates nothing. Only enforce=True actually restores. A flaky probe that raises is treated as not broken (never triggers a destructive restore on its own error).

Usage

python snapshot_ring.py snap [label]      # flush+pull a byte-exact snapshot into the ring
python snapshot_ring.py list              # ring contents, newest first
python snapshot_ring.py rollback [index]  # EXPLICIT restore (index from `list`; default newest good)
python snapshot_ring.py verify            # md5 device vs newest snapshot
python snapshot_ring.py prune             # enforce capacity + drop orphan files

Env: BC_ADB, BC_SERIAL, BC_RING_DIR (default <skill>/.save-ring), BC_RING_CAP (default 10). Python API: SnapshotRing(io, ring_dir).auto_rollback( probe, enforce=False) for a shadow health-guard loop; inject a custom SaveIO for testing or a non-adb transport.

Live-game care (inherited): read-only discovery is safe; prefer reversible save-edits over forcing unverified in-memory fields. On a broken state: kill all hooks + restart BC first, then rollback to the newest verified snapshot.


What works vs what's inferred (honesty ledger)

Capability Status
dispatch_capture agent-JS generation (valid JS, single-shot, no sync RPC) ✅ unit-tested (node --check clean)
Scene guard refuses to arm outside menu cluster ✅ unit-tested
Single-shot detach + hard-timeout detach (hook never left live) ✅ unit-tested (fake device)
Shadow = no auto-tap; explicit arm fires exactly one tap ✅ unit-tested
resolve_ctor_candidates ranks caller RVAs + flags vtable match ✅ unit-tested
Live dispatch-capture on a running BC + actual ctor RVA ⚠️ needs a device run + the offline xref step (documented)
Snapshot ring rotation, corruption guard, integrity-gated rollback ✅ unit-tested (fake SaveIO)
Shadow auto-rollback logs-not-mutates; flaky probe is safe ✅ unit-tested
Live pull/restore round-trip (inode/uid/SELinux preserved) ⚠️ needs a device run (rides proven bcsfe recipe)
OmniParser V2 set-of-mark grounding ? documented ADOPT recipe; not yet wired to a live model
minitouch/evdev low-latency input ? documented ADOPT recipe; adb tap remains default

Files

  • dispatch_capture.py — bounded single-shot touch-dispatch capture + ctor-evidence resolver + CLI.
  • snapshot_ring.py — save-snapshot ring + HGT-safe auto-rollback + CLI.
  • (grounding + input are ADOPT recipes above; they ride bc-drive's screenshot/adb bus.)

Safety summary

Read-mostly. Zero per-frame hooks. The one input-path hook is single-shot, scene-guarded, hard-timeout-bounded, and async-only (thread-selection safe). Save writes are inode/context-preserving, force-stop-guarded, md5-verified, and empty-wipe-guarded; automated rollback is shadow-by-default. Taps only on safe menu navigation — never gacha "Draw", stage "Start" confirm, or anything that spends currency or alters the save.

04 同仓库更多 Skills

Prekzursil
Prekzursil/agent-skills-toolchain

clarify-requirements

Resolve genuine pre-execution ambiguity by scanning repo context, generating 1-6 grounded options, and asking 1-3 focused questions via AskUserQuestion. NOT for mid-task use (injection risk); NOT when the goal is already unambiguous; use to disambiguate vague prompts before any file edits.

★ 0
Prekzursil
Prekzursil/agent-skills-toolchain

openmm-md

Run FREE, open-source, Windows-native GPU molecular dynamics via the openmm-md MCP (OpenMM + CUDA on the local NVIDIA GPU). Use for protein / protein-ligand MD when you want a free stack or are on Windows without a Schrodinger license. Covers the full workflow (prepare, solvate, minimize, equilibrate, production, analyze, mmgbsa) AND an honest decision guide vs Schrodinger Desmond, including how to run Desmond legitimately on Linux or a cloud GPU with a real license. This is a deliberately LESSER free alternative and does NOT replace the schrodinger MCP tools. Do NOT use it to pick a free tool when a validated FEP+, WaterMap, IFD-MD, or Prime MM-GBSA result is actually required.

Skill与智能体开发
★ 0
Prekzursil
Prekzursil/agent-skills-toolchain

visual-review

Take screenshots of web pages and UI using Playwright for visual review and iteration

演示文稿测试与质量
★ 0
Prekzursil
Prekzursil/agent-skills-toolchain

status

Diagnostic status report — shows metaswarm installation state, project setup, and potential issues

★ 0
Prekzursil
Prekzursil/agent-skills-toolchain

start

Use when starting work on any task, when the user mentions metaswarm, or when the user wants to begin tracked development work

Skill与智能体开发
★ 0
Prekzursil
Prekzursil/agent-skills-toolchain

setup

Interactive project setup — detects your project, configures metaswarm, writes project-local files

★ 0