Scope

Channel-scope is the per-channel opt-in compatibility layer between ClawCode and messaging plugins that publish their own access governance (today: claude-whatsapp's historyScope). When a channel is armed (mode != off, governance resolvable), ClawCode's memory surfaces honor the upstream plugin's access rules; when a channel is mode: off (the default), behavior is identical to having no scope layer.

This is an OPTIONAL feature — see docs/channel-scope-compat.md. Enforcement covers memory_search, memory_get, memory_context, the QMD path, voice_transcribe, dream promote routing, and chat_inbox_read. Per-chat semantics flow through (a) a synthetic per-chat indexer over the upstream messages.db which produces per-chat chunks, and (b) a cross-plugin request envelope contract — claude-whatsapp embeds a requestEnvelopeToken in each inbound notification, and ClawCode resolves the token to a chat/sender binding for the current MCP call. Owner unlock (declarative identity = "owner" + out-of-band trust file) remains the always-available escape hatch; without an envelope (or with an invalid/expired one) under mode = enforce, calls fall through to guest [].

Talk to the user in the language they've been using on this turn — never default to a hard-coded language.

When to use

• After installing claude-whatsapp + pairing it: /agent:scope status confirms the adapter sees access.json.
• To turn enforcement on: /agent:scope wizard walks through the choices, or use the one-liner /agent:scope enable whatsapp shadow (or enforce).
• To turn it off: /agent:scope disable whatsapp.
• To inspect existing channel-derived content in shared memory: /agent:scope audit.

Steps

Step 1 — Parse the subcommand

The first word in $ARGS selects the path:

• status → show every configured channel's runtime state
• enable <channel> [shadow|enforce] → set config.scope.<channel>.mode
• disable <channel> → set mode to off
• wizard → interactive REPL flow via AskUserQuestion
• test <chatId> → dry-run probe against the adapter (see Step 6)
• audit → re-run mcp__clawcode__agent_doctor and surface only the scope-* rows

If $ARGS is empty or the subcommand is unknown, default to status.

Step 2 — status

1. Call mcp__clawcode__agent_config(action='get') and parse the scope block.
2. For each channel under scope, display:
   • mode (off / shadow / enforce)
   • identity (auto / owner / guest)
   • background.identity (deny / system-owner)
   • execGate.mode (off / shadow / enforce), execGate.policy (denylist / allowlist), and execGate.tools count (or "defaults" when omitted)
   • WhatsApp only: accessJsonPath, cwdExactMatchOnly
3. Surface trust file presence (read scope + exec):
   • <channel>-owner exists? (yes/no) — gates read scope owner unlock
   • <channel>-exec exists? (yes/no) — gates execGate "trust this machine" path
4. Then call mcp__clawcode__agent_doctor(action='check') and surface the scope rows (scope-pre-enforce-audit, scope-bypasses, scope-quarantine-pending, scope-execgate-status, scope-execgate-shadow-events).

If scope is absent in config, say so explicitly and recommend /agent:scope wizard.

Step 3 — enable <channel> [mode]

All scope-tree writes are refused by mcp__clawcode__agent_config(action='set') (any key starting with scope). The agent cannot silently elevate or relax the policy; every scope key goes through Bash, which surfaces a permission prompt to the user.

For enable <channel> <mode> the single Bash call covers ALL scope.<channel> keys. Substitute BOTH <channel> (validated against the shipped enum {whatsapp, telegram, discord, imessage, webchat}) AND <mode> ('shadow' or 'enforce'; default 'shadow' when omitted) from $ARGS. Refuse any other literal for either parameter — only those values are valid for enable. cwdExactMatchOnly preserves the prior value if the user had it true:

Bash('node -e "const fs=require(\"fs\"),p=\"agent-config.json\";const c=fs.existsSync(p)?JSON.parse(fs.readFileSync(p,\"utf-8\")):{};c.scope=c.scope||{};const ch=\"<channel>\";const cur=c.scope[ch]||{};c.scope[ch]=Object.assign({},cur,{mode:\"<mode>\",identity:\"auto\",accessJsonPath:cur.accessJsonPath||\"auto\",cwdExactMatchOnly:cur.cwdExactMatchOnly===true,background:Object.assign({},cur.background,{identity:\"deny\"})});fs.writeFileSync(p,JSON.stringify(c,null,2));console.log(\"wrote\",p);"')

The Bash call surfaces a permission prompt to the user — that's intentional. Default mode when omitted: shadow. Confirm by re-running status.

Tell the user: "Run /mcp reconnect clawcode for changes to take effect."

Step 4 — disable <channel>

First validate <channel> is one of the shipped channel names (whatsapp, telegram, discord, imessage, webchat). Reject any other value with an error message — this defends against future channel IDs that might contain shell metacharacters. Currently shipped names are alphanumeric and safe.

scope.<channel>.mode AND scope.<channel>.execGate.mode are both on the security-sensitive blocklist, so this also goes through Bash. A disable resets BOTH read scope AND execGate to off (the user expected "turn the channel off" — they don't expect read scope to flip but exec scope to remain active):

Bash('node -e "const fs=require(\"fs\"),p=\"agent-config.json\";const c=fs.existsSync(p)?JSON.parse(fs.readFileSync(p,\"utf-8\")):{};c.scope=c.scope||{};const cur=c.scope.<channel>||{};c.scope.<channel>=Object.assign({},cur,{mode:\"off\",execGate:{mode:\"off\"}});fs.writeFileSync(p,JSON.stringify(c,null,2));console.log(\"disabled\");"')

Also remove BOTH trust files (read-scope owner trust AND exec trust) for THIS workspace. As of 1.7.0 the trust files live under a per-workspace fingerprint subdir; the fingerprint is computed by the bridge script scripts/print-workspace-fingerprint.mjs so Bash and TS agree byte-exact:

Bash('set -euo pipefail; [ -n "${CLAUDE_PLUGIN_ROOT:-}" ] || { echo "CLAUDE_PLUGIN_ROOT unset"; exit 1; }; TS_HASH=$(node "$CLAUDE_PLUGIN_ROOT/node_modules/tsx/dist/cli.mjs" "$CLAUDE_PLUGIN_ROOT/scripts/print-workspace-fingerprint.mjs" "$PWD"); [[ "$TS_HASH" =~ ^[0-9a-f]{32}$ ]] || { echo "bad TS_HASH=$TS_HASH"; exit 1; }; rm -f -- "$HOME/.claude/agent/scope-trust/$TS_HASH/<channel>-owner" "$HOME/.claude/agent/scope-trust/$TS_HASH/<channel>-exec"')

so both unlocks are dropped along with the mode flip. Both Bash calls surface a permission prompt — intentional: turning scope off and dropping trust are user-visible state changes. Cross-workspace isolation: this only removes THIS workspace's trust; other workspaces are untouched.

Step 5 — wizard (interactive)

Load AskUserQuestion via ToolSearch(query='select:AskUserQuestion') first.

Then walk the user through the choices, ONE question per AskUserQuestion call. Sample flow (substitute the user's language):

1. Which channel?
   • "WhatsApp (claude-whatsapp paired)" → whatsapp
   • "Cancel" → exit, no changes (Skip channels that are not installed/paired — only offer those whose agent_doctor reports installed+authenticated.)
2. Mode?
   • "Shadow — collect stats but don't filter (recommended for first week)"
   • "Enforce — filter denied chunks immediately"
   • "Cancel" → exit
3. ¿Sos el owner de este WhatsApp? (owner-unlock primitive — two-factor: identity = "owner" config + out-of-band trust file)
   • "Sí, soy el owner — quiero que ClawCode pueda buscar en mis chats sin restricción" → identity: owner AND step 6 will create a per-workspace trust file (as of 1.7.0) under ~/.claude/agent/scope-trust/<workspace-fingerprint>/whatsapp-owner via Bash. Both writes are required: config alone (which the agent can write) does NOT unlock — the trust file (a separate Bash call the user approves) is what makes the unlock real. The trust is scoped to THIS workspace; granting in workspace A does NOT unlock workspace B. This closes the prompt-injection escalation surface where an agent could otherwise write a config and simulate being the owner.
   • "No / no estoy seguro — mantener el techo conservador" → identity: auto (default; user still has to set WHATSAPP_OWNER_BYPASS=1 env if they want unlock, or pick "Sí" later)
   • "Soy un guest explícito — denegar siempre canales WA" → identity: guest (NO trust file needed — guest is a deny posture, no escalation surface)
   • "Cancel" → exit
4. Background lane (dreams / indexer)?
   • "Deny — never read scoped chunks in the dream lane (default, safest)" → background.identity: deny (NO trust file)
   • "System-owner — let the dream lane read as if owner (only if you control 100% of the channel and want full dream coverage)" → background.identity: system-owner AND step 6 will create the trust file just like the foreground owner case. Background system-owner without trust file silently degrades to deny in resolveAllowed, so the wizard MUST touch the file or the user gets surprising "no dream coverage" behavior.
   • "Cancel" → exit

4a. ¿Activar gate de ejecución? (execution gate — separate from read scope; blocks destructive tools when the current turn was triggered by a non-owner inbound)

• "Sí, recomendado — bloquear Bash y writes destructivos si un mensaje de un no-owner los dispara" → execGate.mode: shadow initially (will collect would-block events for review)
• "Sí, modo enforce desde el principio (avanzado)" → execGate.mode: enforce
• "No, dejar abierto (el agente puede ejecutar lo que sea aunque venga de un grupo)" → execGate.mode: off
• "Cancel" → exit

4b. ¿Qué política? (only if execGate mode is shadow/enforce)

• "Denylist (recomendado) — bloquear solo tools destructivos (Bash, Write, Edit, MultiEdit, NotebookEdit, agent_config, skill_install/remove, dream)" → execGate.policy: denylist, execGate.tools omitted (defaults applied)
• "Allowlist (estricto) — solo permitir tools de lectura/mensajería (Read, Grep, Glob, TodoWrite, WebFetch, WebSearch, reply, react, memory_search/get/context, voice_speak/transcribe)" → execGate.policy: allowlist, execGate.tools omitted (defaults applied)
• "Cancel" → exit

4c. ¿Confiar este equipo para ejecutar tools cuando un no-owner mensaje vino del canal armado? (only if execGate mode is shadow/enforce — this creates the <channel>-exec trust file, separate from the <channel>-owner read-scope trust file)

• "Sí, soy el owner — confío en que este equipo puede ejecutar tools incluso cuando un mensaje de otro entró en el window de 60s" → step 6 will create a per-workspace exec trust file (as of 1.7.0) under ~/.claude/agent/scope-trust/<workspace-fingerprint>/<channel>-exec via Bash. The exec-trust file is SEPARATE from the read-scope owner trust file (<channel>-owner) — having one does NOT imply the other. This is intentional: a user can read their own chats without granting the agent the right to run shell commands from non-owner-triggered turns. The trust is scoped to THIS workspace; if you have execGate enforce in multiple workspaces, you must opt-in to exec-trust in each one independently.
• "No, dejar el gate firmado — bloquear cualquier no-owner-inducido tool destructivo hasta que cree el archivo manualmente" → no exec trust file. Useful when the user wants to observe shadow events first before deciding.
• "Cancel" → exit

5. Confirmation preview: show a summary of the writes that will happen, AND a count of currently-visible channel chunks that would be filtered out under the chosen mode.

   To compute the count, call:

   Bash('node -e "const m=require(\"./lib/memory-db.ts\");const db=new m.MemoryDB(process.cwd(),[],{ headless:true,quietBoot:true});const n=db.countSyntheticChunksForChannel(\"whatsapp\");db.close();console.log(\"PREVIEW_COUNT=\"+n);"')

   Show the user: "Under the chosen mode, approximately N chunks from this channel would be filtered for non-owner queries (owner mode sees them all)."

   This is a preview — exact filtering depends on per-chat allowlists and is computed at query time.

   • The scope-policy writes go through Bash (a node -e one-liner that updates agent-config.json) because agent_config(action='set') refuses scope keys by design. The user gets one Bash permission prompt for the JSON edit.
   • If foreground owner OR background system-owner was chosen, an additional Bash('mkdir -p && touch && chmod 600') runs to create the trust file. Second permission prompt.
   • Explain in plain language: "El trust file es la prueba 'out-of-band' de que vos (no el agente vía MCP) declaraste ownership. Sin el archivo el unlock no aplica, aunque la config diga identity: owner o background.identity: system-owner. Las dos confirmaciones (config + trust file) son intencionales: cierran el agujero donde un agente prompt-inyectado podía escribir la config y simular ser owner."
   • Caveat to surface explicitly: "Si tenés Bash en auto-allow para esta sesión, el trust file se crea sin prompt — la protección depende de que aprobés cada Bash interactivamente. Si dudás, salí del wizard y corré los comandos a mano."
   • Confirm via AskUserQuestion:
     • "Apply" → run the Bash config edit + (conditionally) the trust-file touch + close with status
     • "Cancel" → exit, no changes
6. Apply: execute the calls in this order:

   • Single consolidated config write through Bash (because agent_config(action='set') refuses scope.* keys). Substitute <channel>, <mode>, <identity>, <bg>, <egMode>, <egPolicy> from the wizard answers:

     Bash('node -e "const fs=require(\"fs\"),p=\"agent-config.json\";const c=fs.existsSync(p)?JSON.parse(fs.readFileSync(p,\"utf-8\")):{};c.scope=c.scope||{};const ch=\"<channel>\";const cur=c.scope[ch]||{};const bg=Object.assign({},cur.background,{identity:\"<bg>\"});const next=Object.assign({},cur,{mode:\"<mode>\",identity:\"<identity>\",accessJsonPath:cur.accessJsonPath||\"auto\",cwdExactMatchOnly:cur.cwdExactMatchOnly===true,background:bg});const eg=\"<egMode>\";next.execGate=eg===\"off\"?{mode:\"off\"}:{mode:eg,policy:\"<egPolicy>\"};c.scope[ch]=next;fs.writeFileSync(p,JSON.stringify(c,null,2));console.log(\"wrote\",p);"')

     Notes:

     • cwdExactMatchOnly preserves the prior value if the user had it true — never silently flipped to false.
     • execGate.tools is intentionally OMITTED so the resolver applies defaults (DEFAULT_DENYLIST_TOOLS / DEFAULT_ALLOWLIST_TOOLS from lib/scope/exec-gate.ts). If the user picked custom tools in a future wizard branch, set next.execGate.tools = ["…"] accordingly.
     • execGate.mode = "off" is ALWAYS written so /agent:scope status surfaces the explicit "currently off" state and a later wizard run can detect intent.

   • If identity = "owner" OR background.identity = "system-owner", create the per-workspace owner trust file. As of 1.7.0 trust files live under a fingerprint subdir keyed by realpath(workspaceRoot) so trust in workspace A does NOT unlock workspace B. The bridge script computes the same hex the runtime uses — no Bash-side crypto reimplementation:

     Bash('set -euo pipefail; [ -n "${CLAUDE_PLUGIN_ROOT:-}" ] || { echo "CLAUDE_PLUGIN_ROOT unset"; exit 1; }; TS_HASH=$(node "$CLAUDE_PLUGIN_ROOT/node_modules/tsx/dist/cli.mjs" "$CLAUDE_PLUGIN_ROOT/scripts/print-workspace-fingerprint.mjs" "$PWD"); [[ "$TS_HASH" =~ ^[0-9a-f]{32}$ ]] || { echo "bad TS_HASH=$TS_HASH"; exit 1; }; umask 077; DIR="$HOME/.claude/agent/scope-trust/$TS_HASH"; mkdir -p -- "$DIR"; chmod 700 -- "$DIR"; touch -- "$DIR/<channel>-owner"; chmod 600 -- "$DIR/<channel>-owner"; echo "trust file created: $DIR/<channel>-owner"')

   • If user opted to trust the machine for execGate (step 4c "Sí"), create the per-workspace exec trust file. SEPARATE trust file from the read-scope owner trust; agent cannot create either via MCP — both go through Bash + user permission prompt. The doctor row scope-execgate-status reflects trust validity (mode/uid/symlink checks), not just file presence:

     Bash('set -euo pipefail; [ -n "${CLAUDE_PLUGIN_ROOT:-}" ] || { echo "CLAUDE_PLUGIN_ROOT unset"; exit 1; }; TS_HASH=$(node "$CLAUDE_PLUGIN_ROOT/node_modules/tsx/dist/cli.mjs" "$CLAUDE_PLUGIN_ROOT/scripts/print-workspace-fingerprint.mjs" "$PWD"); [[ "$TS_HASH" =~ ^[0-9a-f]{32}$ ]] || { echo "bad TS_HASH=$TS_HASH"; exit 1; }; umask 077; DIR="$HOME/.claude/agent/scope-trust/$TS_HASH"; mkdir -p -- "$DIR"; chmod 700 -- "$DIR"; touch -- "$DIR/<channel>-exec"; chmod 600 -- "$DIR/<channel>-exec"; echo "trust file created: $DIR/<channel>-exec"')

7. After applying, run status to display the new state. Then ALSO surface the first-run banner to the user once, in their language:

   Scope active. MCP-level filtering covers memory_search, memory_get, memory_context, voice_transcribe, dream promote, and chat_inbox_read. It does NOT cover native Read, Grep, or direct SQLite reads over channel log files — those bypass scope by design (MCP scope ≠ filesystem sandbox). If you want hard isolation, that lives at the OS / filesystem-permissions layer.

   Then also create a once-only marker so future wizard runs in this workspace don't repeat the banner:

   Bash('mkdir -p ~/.claude/agent && touch ~/.claude/agent/.scope-banner-shown')

   Skip the banner emission if ~/.claude/agent/.scope-banner-shown already exists — the user has already seen it.

   Finally remind the user to /mcp reconnect clawcode for changes to take effect.

If invoked outside an interactive REPL session, abort with: "The wizard requires an interactive REPL — use /agent:scope enable <channel> [shadow|enforce] instead."

Step 6 — test <chatId> (read-scope dry-run) and test exec <senderJid> <toolName> (execGate dry-run)

Show what a given chat-id (or sender+tool combo) would resolve to under the current scope config — read-only, no writes.

Read-scope dry-run (test <chatId>):

1. Parse <chatId> from $ARGS. If missing, prompt the user for one.
2. Call mcp__clawcode__agent_config(action='get') and pull scope.<channel> (default whatsapp).

3. Use Bash to run a dry-run probe:

   Bash('node -e "const r=require(\"./lib/scope/runtime.ts\");const c=require(\"./lib/config.ts\");const ctx=require(\"./lib/scope/context.ts\");const cfg=c.loadConfig(process.cwd());const rt=r.detectScopeRuntime(cfg,process.cwd());const adapter=r.getScopeAdapter(\"whatsapp\");if(!adapter){console.log(JSON.stringify({error:\"adapter not armed\",mode:cfg.scope?.whatsapp?.mode}));process.exit(0);}const c1=ctx.makeForegroundContext(\"dry-run\");const allowed=adapter.allowedChatIds(c1);const chatIdToTest=process.argv[1]||\"unknown\";const visible=allowed===null||allowed.includes(chatIdToTest);console.log(JSON.stringify({chatId:chatIdToTest,mode:rt.channels.whatsapp?.mode,armed:rt.channels.whatsapp?.armed,allowedShape:allowed===null?\"unrestricted (owner)\":allowed.length===0?\"deny-all (guest)\":\"partial (\"+allowed.length+\" chats)\",visible}));" "<chatId>"')

4. Surface the JSON output to the user in their language. Translate visible: true to "this chat WOULD be visible to the agent under the current scope" and visible: false to "this chat would be hidden". Add a line: "This is a dry-run — no config or memory state changed."

ExecGate dry-run (test exec <senderJid> <toolName>):

Show what the execution gate would decide for a given sender + tool combo. Useful for verifying allowlist/denylist semantics + hard-deny (Bash/Task) + trust file unlock before flipping shadow → enforce.

1. Parse <senderJid> <toolName> from $ARGS. If missing, prompt the user for both.

2. Invoke the real pure-function resolver (lib/scope/exec-gate.ts:resolve) with an injected envelope reader that returns a synthetic non-owner envelope. The resolver honors the actual trust-file check, hard-deny set (Bash, Task), and protected-paths logic.

   IMPORTANT: this snippet calls into .ts files directly so it MUST run under tsx, but invoking via npx tsx from the user's workspace cwd is unreliable (npx searches the local node_modules/.bin first and fetches from the registry on miss). Use the plugin-local tsx CLI explicitly via node "$CLAUDE_PLUGIN_ROOT/node_modules/tsx/dist/cli.mjs":

   Bash('node "$CLAUDE_PLUGIN_ROOT/node_modules/tsx/dist/cli.mjs" -e "

const eg = require("./lib/scope/exec-gate.ts"); const c = require("./lib/config.ts"); const r = require("./lib/scope/runtime.ts"); const wa = require("./lib/scope/whatsapp.ts"); const path = require("path");

const cfg = c.loadConfig(process.cwd()); const senderJid = process.argv[1]; const toolName = process.argv[2]; const channel = "whatsapp";

const execGate = eg.execGateConfigForChannel(cfg.scope, channel); const channelDir = r.resolveWhatsappChannelDir(cfg, process.cwd()) || ""; const ownerJids = []; if (channelDir) { try { const access = wa.loadAccess(path.join(channelDir, "access.json"), new Map()); if (access.resolvable && access.access) ownerJids.push(...(access.access.ownerJids || [])); } catch {} }

// Inject a fake envelope reader so the resolver "sees" the supplied senderId // as an in-window non-owner inbound. recordShadow is a no-op for dry-run. const fakeReader = { load: (cd, token, now) => ({ version: 1, token, chatId: senderJid, senderId: senderJid, ts: now, expiresAt: now + 60000 }) }; const fakeFs = { readdirSync: () => ["dryrun.json"], statSync: () => ({ isFile: () => true, mtimeMs: Date.now() }) };

const decision = eg.resolve({ toolName, toolInput: toolName === "Bash" ? { command: "dryrun" } : {}, pluginRoot: process.cwd(), workspaceRoot: process.cwd(), memoryDir: path.join(process.cwd(), "memory"), armed: [{ channel, channelDir: channelDir || "/tmp/dryrun", ownerJids, execGate }], reader: fakeReader, fsImpl: fakeFs, effects: { recordShadow: () => {} } });

console.log(JSON.stringify({ senderJid, toolName, channel, execGateMode: execGate.mode, policy: execGate.policy, ownerJidsKnown: ownerJids.length, isOwner: ownerJids.includes(senderJid), decision: decision.decision, reason: "reason" in decision ? decision.reason : null }, null, 2)); " "<senderJid>" "<toolName>"')

3. Surface the JSON output. Translate `decision` to plain language (`allow` / `block` / `shadow`). Note: hard-deny tools (`Bash`, `Task`) block regardless of policy. Trust file `<channel>-exec` (if present + valid uid/mode) unlocks any non-owner-induced call. Add: "This is a dry-run — the resolver pure function was invoked but the hook subprocess was NOT spawned. To verify end-to-end including stderr surfacing, send a real test message from `<senderJid>` and observe the agent's response."

### Step 7 — `audit`

Pass-through to `mcp__clawcode__agent_doctor(action='check')`. Filter the doctor card to only the `scope-*` rows and surface them. Reference `docs/channel-scope-compat.md` for what each row means.

## Limitations

- **Owner unlock for WhatsApp**: set `scope.whatsapp.identity = "owner"` via wizard for declarative per-workspace unlock. The wizard also creates the out-of-band trust file `~/.claude/agent/scope-trust/<workspace-fingerprint>/whatsapp-owner` (Bash-prompted; per-workspace as of 1.7.0). Both are required: config alone does NOT unlock. `WHATSAPP_OWNER_BYPASS=1` env is an alternative escape-hatch. Without any of these, foreground calls hit the owner-only ceiling. Trust granted in one workspace does NOT unlock another workspace.
- **Per-chat binding for non-owner senders**: claude-whatsapp publishes a per-inbound `requestEnvelopeToken` in the MCP notification meta. The agent forwards the token to ClawCode memory tools, ClawCode validates the envelope file under `<channel-dir>/.request-envelopes/<token>.json` and emits a per-chat allowlist mirroring upstream `claude-whatsapp/scope.ts:scopedAllowedChats` byte-exact.
- **Residual risks (mirror of `docs/scope-envelope-contract.md` threat model, both kept in sync)**:
- **Same-uid filesystem forge** (architectural, out of scope): any code running as your user can plant or tamper with channel-state files including the request envelope. The uid match check in the reader rules out cross-user tampering but not same-user adversaries. Scope is a privacy/safety layer between MCP tool calls and the agent, NOT a defense against OS-level adversaries already running as you.
- **Token confusion across concurrent inbounds**: if two inbounds arrive within the 60s TTL, the agent holds two valid tokens. A prompt injection in inbound A could induce the agent to forward inbound B's token (leaking B's scope). Bounded by recent-inbound set + TTL; not closeable without per-tool-call authority binding.
- **Within-TTL replay**: same token re-used within 60s by the same agent (required for multi-tool flows like `memory_search` → `memory_get`).
- **Reply-egress taint**: once a snippet reaches the agent, voice/dream output is not taint-tracked.
- **MCP-level filtering ≠ filesystem sandbox**: native `Read`/`Grep`/SQLite over channel logs always bypass scope by design. If hard isolation is required, that lives at the OS layer.

## Reference

- `docs/channel-scope-compat.md` — full design doc + roadmap
- `docs/doctor.md` — `scope-pre-enforce-audit`, `scope-bypasses`, `scope-quarantine-pending`
- Plan: `~/.claude/plans/como-ves-poder-tener-agile-dove.md`