diff --git a/claude/README.md b/claude/README.md
index 286c1aa..e299d1f 100644
--- a/claude/README.md
+++ b/claude/README.md
@@ -135,6 +135,41 @@ or the CLI call fails, the hook exits silently without affecting the session.
 | **PostCompact** | Records the compacted summary as a structured event and buffers a short post-compaction note for later recall. |
 | **SessionEnd** | Reuses the session-final memory capture path so Claude can flush the final checkpoint even when `Stop` is not the last lifecycle event observed. |
 
+### Permission and approval policy
+
+The PreToolUse hook applies a **tokenized** check against the actual `akm` argv (using a shared assessor that the OpenCode plugin also calls). It only fires on real shell invocations of `akm <subcommand>` — it does not match the phrase appearing in commit messages, release notes, or heredoc bodies. The full list of subcommands it blocks pending approval is in `shared/risky-command.ts`.
+
+For most users, the plugin's hook is enough. If you prefer Claude Code's built-in permission dialog (a one-time approve/deny prompt rather than a hook stderr block), add the matching rules to your **own** settings — Claude Code plugins cannot ship default permission rules, so this step is opt-in and manual.
+
+Add this block to `~/.claude/settings.json` (user-wide) or `.claude/settings.json` (current project):
+
+```json
+{
+  "permissions": {
+    "ask": [
+      "Bash(akm accept:*)",
+      "Bash(akm reject:*)",
+      "Bash(akm proposal accept:*)",
+      "Bash(akm proposal reject:*)",
+      "Bash(akm remove:*)",
+      "Bash(akm upgrade:*)",
+      "Bash(akm vault create:*)",
+      "Bash(akm vault set:*)",
+      "Bash(akm vault unset:*)",
+      "Bash(akm vault load:*)",
+      "Bash(akm vault show:*)"
+    ]
+  }
+}
+```
+
+Notes:
+
+- The `Bash(pattern:*)` matcher is **prefix-anchored** and only matches the actual bash invocation. It does not trip on the same phrase appearing inside argv (commit messages, heredoc bodies, etc.).
+- Patterns that cannot be expressed as a simple prefix — for example "`akm save` with `--push` anywhere in the flags" — are not coverable by this list. The plugin hook keeps handling those.
+- `akm remember` payload secret-scanning is content-aware (it inspects the literal arguments for things that look like secrets). The plugin hook keeps doing this; native permission rules cannot.
+- If you want a command to skip both the plugin hook and the permission dialog entirely, put it under `permissions.allow` instead of `permissions.ask` — but only do this for commands you genuinely want to auto-approve.
+
 ### Environment overrides
 
 | Variable | Default | Purpose |
diff --git a/claude/hooks/akm-hook.ts b/claude/hooks/akm-hook.ts
index ac16e66..d41868f 100644
--- a/claude/hooks/akm-hook.ts
+++ b/claude/hooks/akm-hook.ts
@@ -10,6 +10,7 @@ import { appendMemoryEvent, getEventLogPath, readJsonl, type AkmMemoryEvent } fr
 import { shouldRecall } from "../shared/recall-policy"
 import { redactObject, redactSecrets } from "../shared/redaction"
 import { extractAkmRefsFromString } from "../shared/ref-extraction"
+import { assessRiskyAkmCommand } from "../shared/risky-command"
 
 const COMMAND = process.argv[2] ?? ""
 const MODE = process.argv[3] ?? ""
@@ -455,20 +456,24 @@ function extractPostToolFields(raw: string, mode: string): { toolName: string; c
 }
 
 function assessRiskyClaudeCommand(command: string): string | undefined {
+  // Tokenized assessment of canonical risky `akm` subcommands (shared with the
+  // OpenCode plugin). Token-aware matching avoids false positives on prose
+  // mentions of these phrases inside commit messages, heredoc bodies, or
+  // release notes.
+  const assessment = assessRiskyAkmCommand(command)
+  if (assessment) {
+    return `${assessment.reason} ${assessment.approval} This requires explicit user approval.`
+  }
+
+  // Content-scan check: `akm remember` with a payload that looks like it
+  // contains secrets. This is intentionally regex-based — we want to fire on
+  // *any* occurrence of a secret in the argv, even if `akm remember` itself
+  // is being invoked safely. The redactor uses its own heuristics; we just
+  // gate on its verdict.
   const normalized = command.trim()
-  if (!normalized) return undefined
-  if (/\bakm\s+(?:proposal\s+)?(accept|reject)\b/.test(normalized)) return "Proposal acceptance/rejection requires explicit user approval."
-  if (/\bakm\s+save\b[\s\S]*--push\b/.test(normalized)) return "`akm save --push` requires explicit user approval."
-  if (/\bakm\s+remove\b/.test(normalized)) return "`akm remove` is destructive and requires explicit user approval."
-  if (/\bakm\s+update\b[\s\S]*--all\b/.test(normalized)) return "`akm update --all` requires explicit user approval."
-  if (/\bakm\s+upgrade\b(?:[\s\S]*--force\b|\b)/.test(normalized)) return "`akm upgrade` requires explicit user approval."
-  if (/\bakm\s+vault\s+(create|set|unset|load)\b/.test(normalized)) {
-    if (/^\s*eval\s+["']?\$\(akm\s+vault\s+load\s+/.test(normalized)) return undefined
-    return "Vault create/set/unset/load requires explicit approval, and raw `akm vault load` output must not be exposed in logs or chat."
+  if (/\bakm\s+remember\b/.test(normalized) && redactSecrets(normalized).redacted) {
+    return "Raw `akm remember` payload appears to include secrets; redact it before writing memory."
   }
-  if (/\bakm\s+config\s+set\s+(?:llm\.features\.|registries|searchPaths|stashDir)/.test(normalized)) return "AKM config mutations require explicit user approval."
-  if (/\bakm\s+(?:add|wiki\s+register)\b[\s\S]*--trust\b/.test(normalized)) return "Trusted source registration requires explicit user approval."
-  if (/\bakm\s+remember\b/.test(normalized) && redactSecrets(normalized).redacted) return "Raw `akm remember` payload appears to include secrets; redact it before writing memory."
   return undefined
 }
 
diff --git a/claude/shared/risky-command.ts b/claude/shared/risky-command.ts
new file mode 100644
index 0000000..d302b4a
--- /dev/null
+++ b/claude/shared/risky-command.ts
@@ -0,0 +1 @@
+export * from "../../shared/risky-command"
diff --git a/opencode/index.ts b/opencode/index.ts
index 2a0ee8d..c394c7d 100644
--- a/opencode/index.ts
+++ b/opencode/index.ts
@@ -10,6 +10,7 @@ import { appendMemoryEvent, getEventLogPath, readJsonl, type AkmMemoryEvent } fr
 import { shouldRecall } from "../shared/recall-policy"
 import { redactObject, redactSecrets } from "../shared/redaction"
 import { extractAkmRefsFromString } from "../shared/ref-extraction"
+import { assessRiskyAkmCommand, blockedCommandMessage, splitArguments, type RiskyCommandAssessment } from "../shared/risky-command"
 
 let resolvedAkmCommand = "akm"
 const moduleDir = path.dirname(fileURLToPath(import.meta.url))
@@ -979,52 +980,9 @@ async function recordRetrospectiveFeedback(client: LogCapableClient, sessionID:
   retrospectiveState.set(sessionID, { recentRefs })
 }
 
-type RiskyCommandAssessment = {
-  category: string
-  reason: string
-  approval: string
-}
-
-function assessRiskyAkmCommand(command: string): RiskyCommandAssessment | undefined {
-  const args = splitArguments(command)
-  const akmIndex = args.findIndex((arg) => arg === "akm" || arg.endsWith("/akm") || arg.endsWith("\\akm.exe"))
-  if (akmIndex === -1) return undefined
-  const tokens = args.slice(akmIndex + 1)
-  if ((tokens[0] === "proposal" && tokens[1] === "accept") || tokens[0] === "accept") {
-    return { category: "proposal-accept", reason: "Proposal acceptance changes curated AKM content.", approval: "Ask the user to approve `akm accept <id>`." }
-  }
-  if ((tokens[0] === "proposal" && tokens[1] === "reject") || tokens[0] === "reject") {
-    return { category: "proposal-reject", reason: "Proposal rejection is a durable curation decision.", approval: "Ask the user to approve `akm reject <id> --reason \"...\"`." }
-  }
-  if (tokens[0] === "save" && tokens.includes("--push")) {
-    return { category: "save-push", reason: "Pushing stash changes must be explicitly approved.", approval: "Ask the user to approve `akm save --push`." }
-  }
-  if (tokens[0] === "remove") {
-    return { category: "remove", reason: "Removing AKM sources is destructive.", approval: "Ask the user to approve the exact `akm remove ...` command." }
-  }
-  if (tokens[0] === "vault" && ["show", "load", "set", "unset"].includes(tokens[1] ?? "")) {
-    return { category: `vault-${tokens[1]}`, reason: "Vault access or mutation is sensitive.", approval: `Ask the user to approve the exact \`akm vault ${tokens[1]} ...\` command.` }
-  }
-  if (tokens[0] === "config" && tokens[1] === "set" && (tokens[2]?.startsWith("llm.features.") ?? false)) {
-    return { category: "config-llm-features", reason: "Changing AKM LLM feature flags alters autonomous behavior.", approval: "Ask the user to approve the exact `akm config set llm.features.* ...` command." }
-  }
-  if (tokens[0] === "update" && tokens.includes("--all")) {
-    return { category: "update-all", reason: "Updating all AKM kits changes many assets at once.", approval: "Ask the user to approve `akm update --all`." }
-  }
-  if (tokens[0] === "upgrade") {
-    return { category: "upgrade", reason: "Upgrading the AKM CLI changes the toolchain.", approval: "Ask the user to approve the exact `akm upgrade` command." }
-  }
-  return undefined
-}
-
-function blockedCommandMessage(command: string, assessment: RiskyCommandAssessment): string {
-  return [
-    `Blocked risky AKM command: ${command}`,
-    assessment.reason,
-    assessment.approval,
-    "Retry only after explicit user approval in this conversation.",
-  ].join("\n")
-}
+// Risk assessment for `akm` CLI invocations moved to shared/risky-command.ts.
+// Imported above so both the Claude Code plugin and the OpenCode plugin
+// apply the same rules and stay in sync.
 
 function queueFeedback(
   client: LogCapableClient,
@@ -2327,16 +2285,7 @@ async function getParentSessionID(
   }
 }
 
-function splitArguments(raw: string): string[] {
-  if (!raw.trim()) return []
-  const args: string[] = []
-  const re = /"([^"]*)"|'([^']*)'|`([^`]*)`|(\S+)/g
-  let match: RegExpExecArray | null
-  while ((match = re.exec(raw)) !== null) {
-    args.push(match[1] ?? match[2] ?? match[3] ?? match[4] ?? "")
-  }
-  return args
-}
+// splitArguments moved to shared/risky-command.ts (imported at top of file).
 
 function renderCommandTemplate(template: string, rawArguments: string): string {
   const args = splitArguments(rawArguments)
diff --git a/shared/risky-command.ts b/shared/risky-command.ts
new file mode 100644
index 0000000..8674d1f
--- /dev/null
+++ b/shared/risky-command.ts
@@ -0,0 +1,142 @@
+// Tokenized assessment of risky `akm` CLI invocations.
+//
+// Shared between the Claude Code plugin (claude/hooks/akm-hook.ts) and the
+// OpenCode plugin (opencode/index.ts) so both harnesses block the same set of
+// commands using the same logic.
+//
+// Why tokenized matching: a regex against the raw command string (e.g.
+// /\bakm\s+vault\s+set\b/) fires on any prose mention — commit messages,
+// release notes, heredoc bodies, README quotes — because the input is the
+// entire bash argv concatenated. Tokenized matching uses splitArguments() to
+// recover the shell argv, locates the actual `akm` executable token, then
+// inspects subcommand positions. Only real invocations trip it; prose does
+// not.
+
+export type RiskyCommandAssessment = {
+  category: string
+  reason: string
+  approval: string
+}
+
+// Quote-aware shell tokenizer. Splits on whitespace but respects single,
+// double, and backtick quotes. Intentionally simple: does not handle escaped
+// quotes inside the same quote style, $(...) expansion, or process
+// substitution. Sufficient for "did the user invoke `akm <sub>`" checks; not
+// a full shell parser.
+export function splitArguments(raw: string): string[] {
+  if (!raw.trim()) return []
+  const args: string[] = []
+  const re = /"([^"]*)"|'([^']*)'|`([^`]*)`|(\S+)/g
+  let match: RegExpExecArray | null
+  while ((match = re.exec(raw)) !== null) {
+    args.push(match[1] ?? match[2] ?? match[3] ?? match[4] ?? "")
+  }
+  return args
+}
+
+// Documented loader pattern: `eval "$(akm vault load vault:<name>)"`. The
+// outer `eval` is the real invocation; the inner `akm vault load` is the
+// vault read that the user has explicitly opted into by piping to eval.
+// Treat as safe — the user wrote this exact form.
+const EVAL_VAULT_LOAD_PATTERN = /^\s*eval\s+["']?\$\(akm\s+vault\s+load\s+/
+
+export function assessRiskyAkmCommand(command: string): RiskyCommandAssessment | undefined {
+  const normalized = command.trim()
+  if (!normalized) return undefined
+
+  if (EVAL_VAULT_LOAD_PATTERN.test(normalized)) return undefined
+
+  const args = splitArguments(normalized)
+  const akmIndex = args.findIndex((arg) => arg === "akm" || arg.endsWith("/akm") || arg.endsWith("\\akm.exe"))
+  if (akmIndex === -1) return undefined
+
+  const tokens = args.slice(akmIndex + 1)
+  const sub = tokens[0]
+  const subSub = tokens[1]
+
+  if ((sub === "proposal" && subSub === "accept") || sub === "accept") {
+    return {
+      category: "proposal-accept",
+      reason: "Proposal acceptance changes curated AKM content.",
+      approval: "Ask the user to approve `akm accept <id>`.",
+    }
+  }
+
+  if ((sub === "proposal" && subSub === "reject") || sub === "reject") {
+    return {
+      category: "proposal-reject",
+      reason: "Proposal rejection is a durable curation decision.",
+      approval: 'Ask the user to approve `akm reject <id> --reason "..."`.',
+    }
+  }
+
+  if (sub === "save" && tokens.includes("--push")) {
+    return {
+      category: "save-push",
+      reason: "Pushing stash changes must be explicitly approved.",
+      approval: "Ask the user to approve `akm save --push`.",
+    }
+  }
+
+  if (sub === "remove") {
+    return {
+      category: "remove",
+      reason: "Removing AKM sources is destructive.",
+      approval: "Ask the user to approve the exact `akm remove ...` command.",
+    }
+  }
+
+  if (sub === "vault" && (subSub === "create" || subSub === "set" || subSub === "unset" || subSub === "load" || subSub === "show")) {
+    return {
+      category: `vault-${subSub}`,
+      reason: "Vault access or mutation is sensitive.",
+      approval: `Ask the user to approve the exact \`akm vault ${subSub} ...\` command.`,
+    }
+  }
+
+  if (sub === "config" && subSub === "set" && tokens[2]) {
+    const key = tokens[2]
+    if (key.startsWith("llm.features.") || key === "registries" || key === "searchPaths" || key === "stashDir") {
+      return {
+        category: "config-mutation",
+        reason: "AKM config mutations alter autonomous behavior or trust boundaries.",
+        approval: `Ask the user to approve the exact \`akm config set ${key} ...\` command.`,
+      }
+    }
+  }
+
+  if (sub === "update" && tokens.includes("--all")) {
+    return {
+      category: "update-all",
+      reason: "Updating all AKM kits changes many assets at once.",
+      approval: "Ask the user to approve `akm update --all`.",
+    }
+  }
+
+  if (sub === "upgrade") {
+    return {
+      category: "upgrade",
+      reason: "Upgrading the AKM CLI changes the toolchain.",
+      approval: "Ask the user to approve the exact `akm upgrade` command.",
+    }
+  }
+
+  if ((sub === "add" || (sub === "wiki" && subSub === "register")) && tokens.includes("--trust")) {
+    return {
+      category: "trust-registration",
+      reason: "Trusted source registration bypasses safety checks.",
+      approval: "Ask the user to approve the exact command before adding a trusted source.",
+    }
+  }
+
+  return undefined
+}
+
+export function blockedCommandMessage(command: string, assessment: RiskyCommandAssessment): string {
+  return [
+    `Blocked risky AKM command: ${command}`,
+    assessment.reason,
+    assessment.approval,
+    "Retry only after explicit user approval in this conversation.",
+  ].join("\n")
+}