kimi-plugin-cc 2.0.7

package/dist/hooks/install-paths.js

@@ -0,0 +1,504 @@
+// Shared canonical-path helpers for the kimi-plugin-cc PreToolUse hook.
+//
+// Why a separate module:
+//
+//   `runtime/commands/setup.ts` is the installer; `runtime/hooks/install.ts`
+//   is the per-call verifier. Audit (reports 27 + 28) found two
+//   convergent issues:
+//
+//     1. The verifier's drift gate was opt-in (callers had to pass
+//        `expectedHookPath`). rescue.ts called it WITHOUT the path, so
+//        a managed block referencing a stale hook script silently passed.
+//        kimi-code's spawn of the stale path exited 127 (or
+//        MODULE_NOT_FOUND), which the hook runner treats as ALLOW —
+//        rescue's workspace-bound allowlist bypassed in production.
+//
+//     2. Even when callers DID pass `expectedHookPath`, the verifier
+//        used `commandPath.includes(expectedHookPath)` (substring), so a
+//        crafted command like `true # /path/to/approval-hook.js` would
+//        pass: `/bin/sh -c "true # ..."` runs only `true` (exit 0,
+//        no-op allow), then kimi-code treats exit 0 as ALLOW.
+//
+//   Fix: every verifier path now reconstructs the canonical expected
+//   shell command from the resolved Node binary + hook script path, and
+//   does EXACT equality. This module owns the single source of truth for
+//   how that command is built.
+//
+//   Both setup.ts (write side) and install.ts (verify side) import from
+//   here. The probe in setup.ts also uses `buildHookShellCommand` so the
+//   shell probe runs the exact byte string the managed block writes.
+//   These three call sites cannot drift without a compile error.
+import { createHash } from "node:crypto";
+import { realpathSync } from "node:fs";
+import path from "node:path";
+import { fileURLToPath } from "node:url";
+import { RuntimeError } from "../errors.js";
+/**
+ * Prefer a non-version-stamped path to the SAME Node binary.
+ *
+ * `process.execPath` is fully symlink-resolved, so under Homebrew it is
+ * `/opt/homebrew/Cellar/node/<version>/bin/node` — a path that changes on
+ * every `brew upgrade node`. The managed block is verified by byte-exact
+ * equality, so that churn silently invalidates EVERY host's hook at once and
+ * forces a re-run of `/kimi:setup` with no user-visible cause (diagnosed
+ * 2026-07-25; a June config backup held `node/26.0.0` and `node/26.3.0`
+ * side by side, i.e. this had already fired before).
+ *
+ * `process.argv0` preserves the path the interpreter was actually invoked
+ * with, and `scripts/companion.sh` execs `$(command -v node)` — the stable
+ * `/opt/homebrew/bin/node` symlink. Prefer it, but ONLY when it provably
+ * denotes the same executable: absolute, and `realpath()`-identical to
+ * `process.execPath`. That guard is the whole safety argument — it can never
+ * select a DIFFERENT binary, only a more stable NAME for the one already
+ * running. Any doubt falls back to `execPath`, so the result is always an
+ * absolute path to a real interpreter.
+ *
+ * Side benefit: this keeps the hook and the companion on the same interpreter
+ * by construction, since companion.sh resolves `node` through that same
+ * stable name. The old behavior could diverge — companion running the new
+ * node while the hook stayed pinned to a Cellar path brew had already
+ * deleted, which spawns nothing and (exit 127) reads as ALLOW.
+ *
+ * Scope: helps symlink-based installs (Homebrew, and any manager exposing a
+ * stable `node`). nvm/asdf/mise expose version-stamped paths directly, so
+ * argv0 == execPath there and behavior is unchanged, not worse.
+ */
+export function preferStableNodePath(execPath, argv0 = process.argv0, realpath = realpathSync) {
+    if (argv0 === undefined || argv0.length === 0 || !path.isAbsolute(argv0)) {
+        return execPath;
+    }
+    try {
+        return realpath(argv0) === realpath(execPath) ? argv0 : execPath;
+    }
+    catch {
+        return execPath;
+    }
+}
+/**
+ * Resolve the absolute path to the Node binary used in the PreToolUse
+ * hook command. kimi-code spawns hooks via `/bin/sh -c "<command>"`; a
+ * bare `node` would rely on the shell's PATH at execution time, which
+ * fails under GUI/LaunchAgent launches with sanitized PATH. Require an
+ * absolute path — either a stable alias of the in-process interpreter
+ * (see `preferStableNodePath`) or an explicit `KIMI_PLUGIN_CC_NODE_BIN`
+ * override.
+ */
+export function resolveNodeBinary(env) {
+    const override = env.KIMI_PLUGIN_CC_NODE_BIN;
+    if (override === undefined || override.length === 0) {
+        return preferStableNodePath(process.execPath);
+    }
+    if (!path.isAbsolute(override)) {
+        throw new RuntimeError("SETUP_NODE_BIN_NOT_ABSOLUTE", [
+            `KIMI_PLUGIN_CC_NODE_BIN must be an absolute path; got ${JSON.stringify(override)}.`,
+            `kimi-code spawns hooks via /bin/sh -c, where a bare command relies on the shell's PATH at hook execution time.`,
+            `Use an absolute path so the hook keeps firing under sanitized-PATH launches.`,
+        ].join(" "), "setup.node-bin", { details: { override } });
+    }
+    return override;
+}
+/**
+ * Build the exact shell command string that `/bin/sh -c "<command>"`
+ * needs to spawn the hook. Single source of truth for:
+ *
+ *   - what `/kimi:setup` writes into kimi-code's config
+ *     (`command = "..."` inside [[hooks]])
+ *   - what the shell probe runs via `spawn("/bin/sh", ["-c", ...])`
+ *   - what the verifier (`evaluateInstalled`) equality-checks the
+ *     installed `command = "..."` against on every command spawn
+ *
+ * Single-quoting both tokens means a path containing spaces or
+ * apostrophes round-trips cleanly. The probe and managed block ARE the
+ * same byte string — drift between them would break safety.
+ */
+export function buildHookShellCommand(hookScriptPath, env) {
+    const nodeBin = resolveNodeBinary(env);
+    return `${shellSingleQuote(nodeBin)} ${shellSingleQuote(hookScriptPath)}`;
+}
+/**
+ * POSIX shell single-quote a string. Inner `'` are escaped as `'\''`
+ * (close-quote, escaped quote, re-open-quote). Always safe — no shell
+ * metacharacters survive the encoding.
+ */
+export function shellSingleQuote(value) {
+    return `'${value.replaceAll("'", "'\\''")}'`;
+}
+/**
+ * Resolve the absolute path to the compiled hook script.
+ *
+ * Resolution order:
+ *
+ *   1. `KIMI_PLUGIN_CC_HOOK_SCRIPT` override — tests / advanced users.
+ *   2. Sibling resolution from this file's URL. This module lives at
+ *      `<root>/{runtime,dist}/hooks/install-paths.{ts,js}`. The hook
+ *      artifact lives at `<root>/dist/hooks/approval-hook.js`. Walk up
+ *      to `<root>` and append the canonical hook artifact path.
+ */
+export function resolveHookScriptPath(env) {
+    const override = env.KIMI_PLUGIN_CC_HOOK_SCRIPT;
+    if (override !== undefined && override.length > 0) {
+        if (!path.isAbsolute(override)) {
+            // kimi-code spawns hooks via `/bin/sh -c "<command>"` with a
+            // cwd that may not match the companion's. A relative path here
+            // would resolve against the kimi-code shell's working dir at
+            // hook execution time — different from the path resolved at
+            // install time. The mismatch would let the verifier bless a
+            // path that doesn't actually run. Match the NODE_BIN_NOT_ABSOLUTE
+            // contract by requiring an absolute override. Audit re-review
+            // (report 34 Codex MEDIUM) flagged this.
+            throw new RuntimeError("SETUP_HOOK_SCRIPT_NOT_ABSOLUTE", [
+                `KIMI_PLUGIN_CC_HOOK_SCRIPT must be an absolute path; got ${JSON.stringify(override)}.`,
+                `kimi-code spawns hooks via /bin/sh -c with a cwd that may differ from the companion's.`,
+                `Use an absolute path so the verifier and the runtime spawn refer to the same file.`,
+            ].join(" "), "setup.hook-script-path", { details: { override } });
+        }
+        return override;
+    }
+    const here = fileURLToPath(import.meta.url);
+    const parts = here.split(path.sep);
+    // Pin to the canonical suffix `{runtime|dist}/hooks/install-paths.{ts,js}`
+    // — anchoring to a specific tail keeps ancestor directories named
+    // "runtime" or "dist" from confusing the lookup.
+    if (parts.length < 3) {
+        throw resolveHookFailure(here);
+    }
+    const tailParent = parts[parts.length - 2];
+    const tailGrandparent = parts[parts.length - 3];
+    if (tailParent !== "hooks" || (tailGrandparent !== "runtime" && tailGrandparent !== "dist")) {
+        throw resolveHookFailure(here);
+    }
+    const pluginRoot = parts.slice(0, parts.length - 3).join(path.sep) || path.sep;
+    return path.join(pluginRoot, "dist", "hooks", "approval-hook.js");
+}
+function resolveHookFailure(here) {
+    return new RuntimeError("SETUP_RESOLVE_HOOK_FAILED", `Could not infer plugin root from install-paths module path ${here}. Set KIMI_PLUGIN_CC_HOOK_SCRIPT to the absolute path of dist/hooks/approval-hook.js.`, "setup.resolve-hook", { details: { here } });
+}
+/**
+ * Parse a hook shell command of the canonical `'<nodeBin>' '<hookScript>'`
+ * shape (two POSIX single-quoted tokens, space-separated) back into its two
+ * tokens. The exact inverse of `buildHookShellCommand` →
+ * `shellSingleQuote(nodeBin) + " " + shellSingleQuote(hookScript)`.
+ *
+ * Returns `null` for any command that isn't exactly two single-quoted tokens
+ * (e.g. a legacy bare-`node` form, a crafted/garbage command, or anything with
+ * unquoted bare words) — callers treat `null` as "can't classify; use the
+ * generic mismatch message". Strict by design: it only recognizes the canonical
+ * single-quoted form this module emits, so it never mis-attributes a hand-rolled
+ * command's tokens. Tokens with embedded spaces or apostrophes round-trip
+ * (spaces live inside the quotes; `'` is encoded as `'\''`).
+ */
+export function parseHookShellCommand(command) {
+    const tokens = parseSingleQuotedTokens(command);
+    if (tokens === null || tokens.length !== 2) {
+        return null;
+    }
+    return { nodeBin: tokens[0], hookScript: tokens[1] };
+}
+/**
+ * Tokenize a string of POSIX single-quoted tokens as produced by
+ * `shellSingleQuote` (`'...'` with inner `'` encoded as `'\''`). Returns the
+ * decoded token list, or `null` if the input contains any unquoted bare
+ * character or an unterminated quote — i.e. anything outside the grammar this
+ * module emits.
+ */
+function parseSingleQuotedTokens(input) {
+    const tokens = [];
+    let i = 0;
+    const n = input.length;
+    while (i < n) {
+        if (input[i] === " ") {
+            i += 1;
+            continue;
+        }
+        let token = "";
+        let consumedAny = false;
+        while (i < n && input[i] !== " ") {
+            const ch = input[i];
+            if (ch === "'") {
+                // A '...'  quoted segment. Single-quoted content can never contain a
+                // literal "'", so the next "'" is always the real close.
+                const close = input.indexOf("'", i + 1);
+                if (close === -1)
+                    return null; // unterminated quote
+                token += input.slice(i + 1, close);
+                i = close + 1;
+                consumedAny = true;
+            }
+            else if (ch === "\\") {
+                // The `\'` half of a `'\''` escape (an apostrophe in the value).
+                if (i + 1 >= n)
+                    return null;
+                token += input[i + 1];
+                i += 2;
+                consumedAny = true;
+            }
+            else {
+                // A bare unquoted character — not part of the canonical grammar.
+                return null;
+            }
+        }
+        if (!consumedAny)
+            return null;
+        tokens.push(token);
+    }
+    return tokens;
+}
+/**
+ * H4 — classify a hook-command MISMATCH into an actionable diagnosis. When the
+ * managed block is structurally valid but its `command` differs from what this
+ * companion would write, the raw "expected X; got Y" dump is hard to act on. The
+ * common real cause is environment drift: a Node upgrade or version-manager
+ * (nvm/asdf/mise/fnm/Homebrew) switch moved the pinned interpreter, or a plugin
+ * update changed the version-stamped hook-script path. This names which token
+ * drifted and, for Node, whether the old binary still exists on disk (a gone
+ * binary is the unambiguous "your Node moved" signal).
+ *
+ * Returns `undefined` when it can't classify (either command isn't the canonical
+ * two-single-quoted-token shape, or the tokens are somehow equal) — the caller
+ * then falls back to the generic mismatch message. Pure except for the injected
+ * `nodeExists` predicate (so the fs probe stays at the call site). Does NOT alter
+ * the verifier's exact-equality decision — only the human/LLM-facing reason.
+ */
+/**
+ * Machine-readable sibling of `describeHookCommandDrift`. Same inputs, same
+ * parse, but returns WHICH token moved instead of prose — so an LLM caller can
+ * branch on `axis` rather than pattern-matching a human sentence (LLM-caller
+ * discipline: load-bearing context goes in structured state, never in prose).
+ *
+ * `"hook-script"` is the recoverable case: install paths are version-stamped,
+ * so a plugin update moves it and re-running setup on THAT host re-pins it.
+ * `"node-bin"`/`"both"` are reported but deliberately NOT auto-recoverable by
+ * a caller retry — a moved interpreter can mean the pinned one is gone, and a
+ * hook that fails to spawn exits 127, which kimi-code reads as ALLOW.
+ *
+ * Returns `undefined` when either command isn't the canonical two-token shape
+ * or the commands are equal. Never influences the installed decision.
+ */
+/**
+ * Do two paths name the same file? Used ONLY to sharpen a drift *message*
+ * (distinguishing the v1.9.0 stable-path re-pin from a real interpreter move).
+ * Never load-bearing: any throw means "can't tell", and the caller falls back
+ * to the more conservative wording.
+ */
+function defaultSameFile(a, b) {
+    try {
+        return realpathSync(a) === realpathSync(b);
+    }
+    catch {
+        return false;
+    }
+}
+export function classifyHookCommandDrift(installedCommand, expectedCommand, sameFile = defaultSameFile) {
+    const installed = parseHookShellCommand(installedCommand);
+    const expected = parseHookShellCommand(expectedCommand);
+    if (installed === null || expected === null) {
+        return undefined;
+    }
+    const nodeDrift = installed.nodeBin !== expected.nodeBin;
+    const hookDrift = installed.hookScript !== expected.hookScript;
+    if (!nodeDrift && !hookDrift) {
+        return undefined;
+    }
+    const axis = nodeDrift && hookDrift ? "both" : nodeDrift ? "node-bin" : "hook-script";
+    // A node token that changed SPELLING but still names the same file is not an
+    // interpreter move. Distinguishing this matters because it is precisely the
+    // v1.9.0 upgrade shape (the pin moves off the version-stamped path, and the
+    // plugin's own version-stamped script path moves at the same time → `both`),
+    // which every existing install hits exactly once. Without this the caller
+    // would be told "do not retry, the interpreter may be gone" while the human
+    // message says "your Node did not move" — a contradiction where the
+    // structured field, being the load-bearing one, was the wrong half.
+    return {
+        axis,
+        installedCommand,
+        expectedCommand,
+        ...(nodeDrift ? { nodeInterpreterUnchanged: sameFile(installed.nodeBin, expected.nodeBin) } : {}),
+    };
+}
+export function describeHookCommandDrift(installedCommand, expectedCommand, nodeExists, sameFile = defaultSameFile) {
+    const installed = parseHookShellCommand(installedCommand);
+    const expected = parseHookShellCommand(expectedCommand);
+    if (installed === null || expected === null) {
+        return undefined;
+    }
+    const nodeDrift = installed.nodeBin !== expected.nodeBin;
+    const hookDrift = installed.hookScript !== expected.hookScript;
+    if (!nodeDrift && !hookDrift) {
+        return undefined;
+    }
+    const parts = [];
+    if (nodeDrift) {
+        if (!nodeExists(installed.nodeBin)) {
+            parts.push(`Node binary drift: the installed hook pins ${installed.nodeBin}, which no longer exists on disk ` +
+                `(this companion runs ${expected.nodeBin}). This is the classic Node-upgrade / version-manager ` +
+                `(nvm, asdf, mise, fnm, Homebrew) drift — the pinned interpreter moved, so kimi-code can no longer ` +
+                `spawn the hook and read-only enforcement silently degrades.`);
+        }
+        else if (sameFile(installed.nodeBin, expected.nodeBin)) {
+            // Both paths name the SAME interpreter — a spelling change, not a moved
+            // binary. This is the v1.9.0 migration: the pin moved off the
+            // symlink-resolved (version-stamped) path onto the stable one, so every
+            // pre-1.9.0 install drifts exactly once. Blaming a "version-manager
+            // switch between runs" here would be actively misleading — the user
+            // changed nothing about their Node.
+            parts.push(`Node path spelling changed: the installed hook pins ${installed.nodeBin} and this companion runs ` +
+                `${expected.nodeBin} — these are the SAME interpreter reached by different paths. Since v1.9.0 the ` +
+                `hook pins the stable path instead of the version-stamped one, so a Node upgrade no longer breaks it. ` +
+                `This is the expected one-time re-pin after upgrading the plugin; your Node did not move.`);
+        }
+        else {
+            parts.push(`Node binary changed: the installed hook pins ${installed.nodeBin}, but this companion runs ` +
+                `${expected.nodeBin} (both exist on disk — likely a Node version-manager switch between runs).`);
+        }
+    }
+    if (hookDrift) {
+        parts.push(`Hook script path drift: the installed hook points at ${installed.hookScript}, but this companion's ` +
+            `hook is ${expected.hookScript} (likely a plugin update or move — the install path is version-stamped).`);
+    }
+    parts.push("Run /kimi:setup to re-pin the managed block to this companion's current paths.");
+    return parts.join(" ");
+}
+/**
+ * Best-effort: compute the canonical expected shell command for the
+ * current env. Returns `undefined` if either path can't be resolved
+ * (caller treats this as "managed block is unverifiable; do not assume
+ * installed"). Never throws.
+ *
+ * This is the helper the verifier uses on every plugin command spawn.
+ */
+export function tryBuildExpectedHookCommand(env) {
+    try {
+        const hookScriptPath = resolveHookScriptPath(env);
+        const command = buildHookShellCommand(hookScriptPath, env);
+        // Surfaced so the verifier can probe the pinned interpreter for
+        // executability. DERIVED FROM `command` BY PARSING IT BACK, never by calling
+        // `resolveNodeBinary` a second time.
+        //
+        // Re-resolving would be a latent bypass: `resolveNodeBinary`'s no-override
+        // path runs `preferStableNodePath`, which calls `realpathSync` and falls
+        // back to `execPath` on ANY throw. So it is NOT a pure function of `env` —
+        // it reads live filesystem state. If the pinned symlink were unlinked
+        // between the two calls (exactly what `brew upgrade node` does), call #1
+        // would embed the symlink in `command` while call #2 returned the old
+        // realpath. The verifier would then compare the config against the symlink
+        // (match) but probe the OLD Cellar path for executability (may still exist)
+        // — reporting installed=true for a hook whose actual pinned interpreter is
+        // gone. `/bin/sh -c` would exit 127, which kimi-code reads as ALLOW.
+        //
+        // Parsing the command back is the exact inverse of how it was built, so
+        // `nodeBin` is byte-identical to the token in `command` BY CONSTRUCTION —
+        // the guarantee is structural, not probabilistic. The `??` is unreachable
+        // (we just built this string with the very quoting `parseHookShellCommand`
+        // decodes) and exists only so a future change to either side degrades to
+        // the old behavior instead of throwing.
+        const nodeBin = parseHookShellCommand(command)?.nodeBin ?? resolveNodeBinary(env);
+        return { command, hookScriptPath, nodeBin };
+    }
+    catch (err) {
+        if (err instanceof RuntimeError) {
+            return { error: err };
+        }
+        return {
+            error: new RuntimeError("SETUP_RESOLVE_HOOK_FAILED", `Unexpected error resolving hook path: ${err.message}`, "setup.resolve-hook", err instanceof Error
+                ? { cause: err, details: {} }
+                : { details: {} }),
+        };
+    }
+}
+// ----- Host identity -----------------------------------------------------
+//
+// Claude Code and Codex install this plugin to DIFFERENT, version-stamped,
+// host-specific paths but SHARE one `~/.kimi-code/config.toml`:
+//
+//   Claude: ~/.claude/plugins/cache/kimi-marketplace/kimi/<ver>/dist/hooks/approval-hook.js
+//   Codex:  ~/.codex/plugins/cache/kimi-marketplace/kimi/<ver>/dist/hooks/approval-hook.js
+//
+// The managed block is host-scoped (marker suffix `:<host-id>`) so each host
+// owns and verifies its OWN PreToolUse block without clobbering the other's.
+// The host id must be VERSION-INDEPENDENT so a plugin upgrade REFRESHES the
+// same host's block instead of accumulating one block per version.
+/**
+ * Resolve a stable, version-independent host id for the managed-block marker.
+ *
+ * Order: an explicit `KIMI_PLUGIN_CC_HOST_ID` override (slugified) wins — used
+ * by tests and the live-repair path. Otherwise derive from the resolved hook
+ * script path. Pass the already-resolved `hookScriptPath` when you have it
+ * (the verifier + setup do) so we don't re-resolve and risk a second throw.
+ */
+export function resolveHostId(env, hookScriptPath) {
+    const override = env.KIMI_PLUGIN_CC_HOST_ID;
+    if (override !== undefined && override.trim().length > 0) {
+        return slugifyHostId(override);
+    }
+    const resolved = hookScriptPath ?? resolveHookScriptPath(env);
+    return hostIdFromHookScript(resolved);
+}
+/**
+ * Derive a host id from a hook-script path. `~/.claude/...` → `claude-code`,
+ * `~/.codex/...` → `codex` (both literal + version-independent), else a stable
+ * `host-<sha1(hookDir)[:8]>` for dev checkouts / unrecognized layouts.
+ */
+export function hostIdFromHookScript(hookScriptPath) {
+    const norm = hookScriptPath.split(path.sep).join("/");
+    if (norm.includes("/.claude/"))
+        return "claude-code";
+    if (norm.includes("/.codex/"))
+        return "codex";
+    // Dev checkouts are not version-stamped, so hashing the hook's directory is
+    // stable across runs (and across plugin upgrades, which don't move it). Use
+    // 16 hex chars (64 bits): an 8-char prefix is only 32 bits, where two
+    // distinct dev roots can collide (Codex review found a concrete pair) and the
+    // second setup would treat the first host's block as its own and clobber it.
+    const digest = createHash("sha1").update(path.dirname(hookScriptPath)).digest("hex");
+    return `host-${digest.slice(0, 16)}`;
+}
+/**
+ * Normalize an arbitrary host-id override into the `[a-z0-9-]+` slug the
+ * marker regex accepts. Empty results fall back to `host` so the marker is
+ * always well-formed.
+ */
+export function slugifyHostId(value) {
+    const slug = value
+        .trim()
+        .toLowerCase()
+        .replace(/[^a-z0-9]+/g, "-")
+        .replace(/^-+|-+$/g, "")
+        .slice(0, 40);
+    return slug.length > 0 ? slug : "host";
+}
+/**
+ * True when a (decoded) hook `command` string is unambiguously THIS plugin's
+ * approval hook — a canonical two-single-quoted-token command whose script is
+ * `approval-hook.js` living under a `kimi-plugin-cc` / `kimi-marketplace`
+ * install tree. Used to prune orphaned, marker-less `[[hooks]]` entries left by
+ * older installs. Deliberately strict: a hand-rolled or non-canonical command
+ * returns false, so we never remove a user's own hook.
+ */
+/**
+ * Derive the host id that OWNS a hook command, from its script path — the
+ * host-scoped counterpart of `hostIdFromHookScript`. Returns `null` when the
+ * command isn't the canonical two-single-quoted-token shape (a stale/bare
+ * legacy command whose owner can't be determined). Callers treat `null` as
+ * "claimable by the current host." Lets a legacy (un-suffixed) block be
+ * attributed to whichever host actually wrote it, so one host's `/kimi:setup`
+ * never adopts or removes another host's block.
+ */
+export function hostIdFromHookCommand(command) {
+    const parsed = parseHookShellCommand(command);
+    if (parsed === null)
+        return null;
+    return hostIdFromHookScript(parsed.hookScript);
+}
+export function isOurApprovalHookCommand(decodedCommand) {
+    const parsed = parseHookShellCommand(decodedCommand);
+    if (parsed === null)
+        return false;
+    const script = parsed.hookScript.split(/[\\/]/);
+    if (script[script.length - 1] !== "approval-hook.js")
+        return false;
+    const normalized = parsed.hookScript.replace(/\\/g, "/");
+    // Require a real path SEGMENT, not an arbitrary substring — otherwise a
+    // user hook at `/opt/acme/kimi-plugin-cc-wrapper/approval-hook.js` would be
+    // misclassified as ours and pruned. (Codex review.)
+    return normalized.includes("/kimi-plugin-cc/") || normalized.includes("/kimi-marketplace/");
+}