スキル一覧に戻る

decode-minified-js-gates

pjt222
更新日 2 days ago
8 閲覧
17
2
17
GitHubで表示
開発general

について

このスキルは、難読化されたJavaScriptを分析し、様々な機能フラグゲートの実装を特定・分類します。リーダーバリアント、デフォルト値、論理構造などの詳細を抽出します。フラグの挙動が不明確な場合、複数のライブラリが使用されている場合、設定スキーマが複雑な場合など、リバースエンジニアリングのシナリオを想定して設計されています。出力は、ゲートの仕組みを構造化した記録であり、さらなる状態分析に利用されます。

クイックインストール

Claude Code

推奨
メイン
npx skills add pjt222/agent-almanac -a claude-code
プラグインコマンド代替
/plugin add https://github.com/pjt222/agent-almanac
Git クローン代替
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/decode-minified-js-gates

このコマンドをClaude Codeにコピー&ペーストしてスキルをインストールします

ドキュメント

Decode Minified JS Gates

Read call-site context around flag string in minified JavaScript bundle and produce gate-mechanics record: which reader variant, what default, what conjunction, what role. Where probe-feature-flag-state answers "is this gate on or off?", this skill answers prerequisite question — "what does this gate actually do?"

When Use

  • Flag surfaced by sweep-flag-namespace cannot be classified from name alone.
  • Binary uses more than one gate-reader function, need to know which one flag invokes.
  • Gate's "default" appears non-boolean ({}, null, numeric literal), need to decode actual reader variant.
  • Suspect kill-switch (inverted gate) but cannot confirm from flag name.
  • Predicate combines multiple gates with &&, need to enumerate co-gates before probing any.

Inputs

  • Required: minified JavaScript bundle file (.js, .mjs, .bun).
  • Required: target flag string to decode, in literal form.
  • Optional: list of known reader function names from prior decode pass — speeds Step 2.
  • Optional: context-window size override; default is 300 chars before, 200 chars after flag occurrence.

Steps

Step 1: Extract Context Window

Locate flag string and capture asymmetric window around each occurrence. Pre-context (before flag) is where reader function name lives; post-context (after) is where default value and conjunction live.

BUNDLE=/path/to/cli/bundle.js
FLAG=acme_widget_v3                   # synthetic placeholder
PRE=300
POST=200

# All byte offsets where the flag string occurs
grep -boE "\"${FLAG}\"" "$BUNDLE" | cut -d: -f1 > /tmp/decode-offsets.txt
wc -l /tmp/decode-offsets.txt

# Capture an asymmetric window per occurrence
while read -r offset; do
  start=$((offset - PRE))
  [ "$start" -lt 0 ] && start=0
  length=$((PRE + POST))
  echo "=== offset $offset ==="
  dd if="$BUNDLE" bs=1 skip="$start" count="$length" 2>/dev/null
  echo
done < /tmp/decode-offsets.txt > /tmp/decode-windows.txt

less /tmp/decode-windows.txt

For fast first pass, grep -oE with negative lookbehind via Perl-compatible regex catches same windows in one pipe.

Got: one or more context windows per flag occurrence, each ~500 chars. Multiple occurrences typically share same reader function but may differ in default or conjunction — inspect each independently.

If fail: if bundle too big for dd-per-occurrence (binary > 100MB or many occurrences), use rg -B 5 -A 3 "$FLAG" "$BUNDLE" for structured-output approximation. If windows look corrupted, bundle may be UTF-16 or have non-ASCII delimiters; use iconv or treat as binary.

Step 2: Identify Reader Variant

Minified gate libraries commonly expose 4–6 reader variants with different semantics. Reader function name is first cue; call signature is verifier.

Variant taxonomy (synthetic names — substitute actual minified identifiers from your bundle):

VariantSynthetic shapeReturnsCommon usage
Sync booleangate("flag", false) or gate("flag", true)booleanStandard on/off feature switches
Sync config-objectfvReader("flag", {key: value})JSON objectStructured config (delays, allowlists, model names)
Bootstrap-aware TTLttlReader("flag", default, ttlMs)boolean (cached)Startup-path gates before remote config arrives
Truthy-onlytruthyReader("flag")truthy/falsyQuick checks; no explicit default
Async bootstrapasyncReader("flag")Promise<boolean>Gates resolved post-bootstrap
Async bridgebridgeReader("flag")Promise<boolean>Bridge/relay-channel gates with separate evaluation path

Match each context window against variant patterns:

# Test for variant patterns. Replace the synthetic reader names with the
# actual minified identifiers found in the bundle.
grep -oE '\b(gate|fvReader|ttlReader|truthyReader|asyncReader|bridgeReader)\("acme_widget_v3"' /tmp/decode-windows.txt | sort | uniq -c

If multiple variants appear for same flag (rare but real — flag read both sync at startup and async post-bootstrap), record each occurrence's variant separately. Probe results may differ.

Got: every gate-call occurrence tagged with one variant. Variant counts across whole sweep make binary-level distribution (e.g., "60% sync boolean, 30% config-object, 10% TTL").

If fail: if context window has no recognizable reader pattern, flag may not actually be gate-called — recheck call-site classification from sweep-flag-namespace Step 2. If window has reader name not in this taxonomy, document as new variant in research artifacts and decide whether warrants separate handling path.

Step 3: Extract Default Value

Default is second positional arg to reader (or absent for truthy-only / async variants). Capture exact literal — false, true, null, 0, string, or JSON config object.

# Boolean default extraction (sync boolean and TTL variants)
grep -oE '\b(gate|ttlReader)\("acme_widget_v3",\s*(true|false)' /tmp/decode-windows.txt

# Config-object default — match the opening brace and capture until the
# matching brace at the same nesting depth. For minified bundles this is
# usually safe with a non-greedy match because objects rarely span lines.
grep -oE 'fvReader\("acme_widget_v3",\s*\{[^}]*\}' /tmp/decode-windows.txt

# Numeric default (rare but real for TTL or threshold gates)
grep -oE '\b(gate|ttlReader)\("acme_widget_v3",\s*[0-9]+' /tmp/decode-windows.txt

For config-object defaults, inspect JSON structure — keys often hint at gate's purpose (e.g., {maxRetries: 3, timeoutMs: 5000} is retry-policy config, not feature toggle).

Got: exact literal default per occurrence. Booleans unambiguous; config-objects need manual read of structure.

If fail: if config-object's matching brace falls outside context window, increase post-context size in Step 1. If default appears to be variable reference (e.g., gate("flag", x)), default computed at runtime — note as DYNAMIC and probe actual returned value via probe-feature-flag-state.

Step 4: Detect Conjunctions and Kill Switches

Many gates participate in compound predicates. Conjunctions (&&) and inversions (!) change gate's effective role.

# Conjunction detection: gate-call followed by `&&` and another gate-call
# within the same predicate window
grep -oE '(gate|fvReader|ttlReader|truthyReader|asyncReader|bridgeReader)\("acme_widget_v3"[^)]*\)\s*&&\s*(gate|fvReader|ttlReader|truthyReader|asyncReader|bridgeReader)\("acme_[a-zA-Z0-9_]+"' /tmp/decode-windows.txt

# Kill-switch detection: leading `!` before the gate-call
grep -oE '!\s*(gate|fvReader|ttlReader|truthyReader|asyncReader|bridgeReader)\("acme_widget_v3"' /tmp/decode-windows.txt

For each detected conjunction, list co-gate flag names. They are now part of probe scope — if target flag's evaluation depends on co-gates, probing target alone makes incomplete state.

For each detected inversion, mark flag as kill switch in gate-mechanics record. Kill switches flip meaning of default: kill switch with default=false is "feature on by default" (because !false === true), while normal gate with default=false is "feature off by default."

Got: conjunction list (possibly empty) and inversion flag (boolean) per occurrence.

If fail: if conjunction has more than 2 co-gates, predicate complex enough that regex misses structure. Read context window manually and document predicate shape verbatim in gate-mechanics record.

Step 5: Classify Gate's Role

Synthesize Steps 2–4 into role classification. Roles drive different probe strategies and different integration risk.

RoleSignatureImplication
Feature switchsync boolean, no inversion, no conjunctionStandard on/off; probe directly
Config providersync config-object (fvReader)Read returned object; default-empty {} ≠ feature off
Lifecycle guardbootstrap-aware TTL or async bootstrapState depends on bootstrap timing; probe at multiple points
Kill switchinverted gate, default-falseFeature on for users by default; flag flips it OFF
Conjunction memberany variant with && co-gateCannot evaluate alone; co-gates are part of the probe scope
Bridge gateasync bridge variantProbe must occur over the bridge channel, not the main path

Got: every gate-call occurrence has exactly one primary role. Some flags appear in multiple roles across occurrences (e.g., feature switch in one call site, conjunction member in another) — record each role independently.

If fail: if role does not fit table, binary uses gate library not yet documented in this skill. Add row with synthetic identifiers and contribute variant back to skill (or project-specific extension) for future investigators.

Step 6: Produce Gate-Mechanics Record

Combine per-flag findings into structured record. JSONL convenient because each flag becomes one line, easy to merge with sweep-flag-namespace inventory.

{"flag":"acme_widget_v3","variant":"sync_boolean","default":false,"role":"feature_switch","conjunctions":[],"inverted":false,"occurrences":3}
{"flag":"acme_retry_policy","variant":"sync_config_object","default":{"maxRetries":3,"timeoutMs":5000},"role":"config_provider","conjunctions":[],"inverted":false,"occurrences":1}
{"flag":"acme_legacy_path","variant":"sync_boolean","default":false,"role":"kill_switch","conjunctions":[],"inverted":true,"occurrences":2}
{"flag":"acme_beta_feature","variant":"sync_boolean","default":false,"role":"conjunction_member","conjunctions":["acme_beta_program_active"],"inverted":false,"occurrences":1}

Gate-mechanics record feeds probe-feature-flag-state Step 2 (gate-vs-event split): variant + role + conjunction list determines what observations count as evidence of LIVE / DARK / INDETERMINATE state.

Got: one JSONL record per flag (or per flag-occurrence if single flag has multiple distinct mechanics). Record reproducible — running procedure again against same binary makes same record.

If fail: if records vary across runs, upstream step non-deterministic. Most often regex in Step 1 missing or over-matching occurrences. Lock regexes for duration of campaign.

Checks

  • Step 1 produces one context window per flag occurrence; windows ~500 chars
  • Step 2 tags each occurrence with exactly one reader variant from taxonomy
  • Step 3 captures exact default literal (boolean, config-object, or DYNAMIC)
  • Step 4 surfaces all conjunctions and kill-switch inversions present in windows
  • Step 5 assigns one role per occurrence, drawn from role table
  • Step 6 makes JSONL gate-mechanics record that diffs cleanly across re-runs
  • All worked examples use synthetic placeholders (acme_*, gate, fvReader, etc.) — no real flag names, real reader names, or real config-object schemas
  • Record consumable by probe-feature-flag-state (same flag identifiers, compatible field names)

Pitfalls

  • Reading "default" as "behavior": gate with default=true is on by default in this binary, but server-side overrides may flip. Default tells baseline; runtime probe (probe-feature-flag-state) tells state.
  • Confuse config-object empty default with feature off: fvReader("flag", {}) returns empty object as default — but flag is on (gate evaluates to truthy). Treating {} as "off" misclassifies config-providers as feature switches.
  • Miss kill switches: leading ! before gate-call inverts meaning. Skipping Step 4 makes record that says "default=false, feature off by default" when truth is "default=false, feature ON by default because of inversion."
  • Probe one half of conjunction: if acme_widget_v3 && acme_user_in_cohort is predicate, probing only acme_widget_v3 and finding it LIVE does not mean feature is live — conjunction may still gate it off via cohort flag.
  • Trust reader names across versions: minified identifiers can change between major versions. Taxonomy in Step 2 is by signature (call shape, return type, default position), not by name. When binary version changes, re-derive reader names from fresh decode pass.
  • Window too narrow: 200/100 split misses config-object defaults that span 300+ chars. Defaults of 300/200 or 400/300 safer; tighten only if bundle huge and window cost matters.
  • Leak real reader names: minified reader names sometimes look like nonsense (a, b, Yc1) and feel safe to paste verbatim. They are still findings — substitute synthetic placeholders before publishing methodology.

See Also

  • probe-feature-flag-state — uses gate-mechanics record to interpret runtime observations
  • sweep-flag-namespace — makes candidate flag set this skill decodes
  • monitor-binary-version-baselines — tracks reader-name changes across binary versions; re-derive Step 2 patterns when baselines flip
  • redact-for-public-disclosure — how to publish gate-decoding methodology without exposing real reader names or schemas
  • conduct-empirical-wire-capture — validates gate-mechanics record against runtime behavior

GitHub リポジトリ

pjt222/agent-almanac
パス: i18n/caveman/skills/decode-minified-js-gates
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

関連スキル

qmd

開発

qmdは、BM25、ベクトル埋め込み、およびリランキングを組み合わせたハイブリッド検索を用いて、ローカルファイルのインデックス作成と検索を可能にするローカル検索・インデックス作成CLIツールです。コマンドラインでの使用と、Claudeとの統合のためのMCP(Model Context Protocol)モードの両方をサポートしています。このツールは埋め込みにOllamaを使用し、インデックスをローカルに保存するため、ターミナルから直接ドキュメントやコードベースを検索するのに最適です。

スキルを見る

subagent-driven-development

開発

このスキルは、各独立したタスクに対して新規のサブエージェントを起動し、タスク間でコードレビューを実施しながら実装計画を実行します。レビュープロセスを通じて品質基準を維持しつつ、迅速な反復を可能にします。同一セッション内で主に独立したタスクに取り組む際に本スキルをご利用いただくことで、組み込まれた品質チェックを伴う継続的な進捗を確保できます。

スキルを見る

mcporter

開発

mcporterスキルは、開発者がClaudeから直接Model Context Protocol(MCP)サーバーを管理および呼び出せるようにします。このスキルは、利用可能なサーバーの一覧表示、引数を指定したツールの呼び出し、認証およびデーモンのライフサイクル管理を行うコマンドを提供します。開発ワークフローにおいてMCPサーバーの機能を統合およびテストする際に、このスキルをご利用ください。

スキルを見る

adk-deployment-specialist

開発

このスキルは、A2Aプロトコルを使用してVertex AI ADKエージェントをデプロイおよびオーケストレーションし、AgentCardの発見、タスク送信、およびコード実行サンドボックスやメモリバンクなどのサポートツールを管理します。Python、Java、またはGoで、順次、並列、またはループのオーケストレーションパターンを用いたマルチエージェントシステムの構築を可能にします。Google Cloud上でADKエージェントのデプロイやエージェントワークフローのオーケストレーションを求められた際にご利用ください。

スキルを見る