@animalabs/agent-framework 0.11.0 → 0.13.0

package/dist/src/framework.js

@@ -1,8 +1,12 @@
 import { join } from 'node:path';
+import { INLINE_WITHHELD_TEXT, classifyBlock, isInlineContradiction, referenceRegistry, referenceStubOrNull } from './mcpl/references.js';
+import { ReferenceFetcher, DEFAULT_FETCH_MAX_BYTES, EAGER_FETCH_TIMEOUT_MS } from './mcpl/reference-fetcher.js';
 import { appendFileSync, mkdirSync } from 'node:fs';
 import { JsStore } from '@animalabs/chronicle';
 import { MembraneError } from '@animalabs/membrane';
-import { ContextManager, PassthroughStrategy } from '@animalabs/context-manager';
+import { ContextManager, PassthroughStrategy, WindowedPassthroughStrategy } from '@animalabs/context-manager';
+import { SUBCONSCIOUS_TOOLS, SUBCONSCIOUS_TOOL_NAMES } from './tune-out/tools.js';
+import { TuneOutCoordinator, TUNE_OUT_DEFAULTS } from './tune-out/coordinator.js';
 import { ProcessQueueImpl } from './queue.js';
 import { REFUSAL_REACTIONS, REFUSAL_REACTION_FALLBACK } from './refusal-reactions.js';
 import { Agent } from './agent.js';
@@ -15,6 +19,7 @@ import { HookOrchestrator } from './mcpl/hook-orchestrator.js';
 import { PushHandler } from './mcpl/push-handler.js';
 import { parseProsePrefix, parseHybridProsePrefix } from './mcpl/prose-grammar.js';
 import { ProseStreamRouter } from './mcpl/prose-stream-router.js';
+import { detectKnownToolWrapperProse } from './tool-wrapper-prose-guard.js';
 import { InferenceRouter } from './mcpl/inference-router.js';
 import { ChannelRegistry } from './mcpl/channel-registry.js';
 import { ConversationRouter } from './mcpl/conversation-router.js';
@@ -24,6 +29,7 @@ import { randomUUID } from 'node:crypto';
 import { PyRunner, buildInjectedTools } from './code-execution/py-runner.js';
 import { buildCodeExecutionToolDefinition, CODE_EXECUTION_TOOL_NAME, } from './code-execution/tool-definition.js';
 import { splitProseSegments } from './prose-segments.js';
+import { cumulativeDelta } from './usage-accounting.js';
 /** Detect a supported image media type from magic bytes (the model API
  *  rejects mislabeled media types, so trust bytes over extensions).
  *  Returns undefined for non-image content. */
@@ -64,8 +70,23 @@ function sniffImageMediaType(data) {
 const SILENCING_TOOLS = new Set([
     'skip_reply', 'channel_publish', 'send_message', 'reply_message', 'send_dm',
 ]);
-/** World-surface publication names that outrank hybrid prose envelopes. */
-const HYBRID_PUBLICATION_TOOLS = new Set(['say', 'whisper']);
+/**
+ * World-surface publication tools (Eidoverse `say` / `whisper`). An explicit
+ * world utterance is the resident's chosen public speech for that round, so
+ * it silences adjacent auto-routed prose exactly like a channel send — in
+ * EVERY prose-routing mode. Until 2026-09-01 this set was consulted only in
+ * hybrid mode; in locus mode a round of ordinary text + `say` published twice
+ * (Cairn world seq 15146 = the say text, 15147 = the adjacent prose,
+ * byte-for-byte). Hybrid additionally lets these outrank a same-round
+ * `>>>destination` envelope.
+ */
+const WORLD_PUBLICATION_TOOLS = new Set(['say', 'whisper']);
+/** True when a tool call (possibly MCPL-prefixed) is an explicit delivery
+ *  that silences the round's auto-routed prose, regardless of mode. */
+const isSilencingTool = (name) => {
+    const bare = bareToolName(name);
+    return SILENCING_TOOLS.has(bare) || WORLD_PUBLICATION_TOOLS.has(bare);
+};
 /**
  * True when an injected message is real conversational input — something the
  * agent might actually be replying to — rather than ambient machinery.
@@ -504,6 +525,16 @@ export class AgentFramework {
     frameworkCancelledStreams = new Map();
     /** Active runEphemeralToCompletion runs, keyed by agent name. */
     ephemeralRuns = new Map();
+    /** Ephemeral namespaces/names are single-generation for this framework
+     * lifetime. Reusing one while an old stream drains would let name-keyed
+     * liveness/settlement state cross generations. */
+    usedEphemeralAgentNames = new Set();
+    /** One-shot generation tickets minted by createEphemeralAgent. */
+    ephemeralCandidates = new WeakMap();
+    /** Physical ephemeral frames deliberately disposed by their run watchdog/caller.
+     * The name is single-generation, so these frames still own their terminal
+     * typing/outgoing close even after runEphemeralToCompletion deregisters them. */
+    disposedEphemeralFrames = new WeakSet();
     /** Per-agent count of consecutive exhausted inferences (reset on any success).
      *  Drives hard-down escalation — see noteInferenceExhausted. */
     consecutiveInferenceFailures = new Map();
@@ -551,6 +582,14 @@ export class AgentFramework {
     primaryAgentName = null;
     // Messages deferred while an agent is waiting_for_tools (to preserve
     // tool_use → tool_result adjacency required by the Anthropic API).
+    /** The subconscious resident's registry name (issue #77), or null. */
+    subconsciousAgentName = null;
+    /** Its windowed strategy — the coordinator moves the anchor on tune-out entry. */
+    subconsciousStrategy = null;
+    /** Its config (speak_in_channel gate, voice block provenance). */
+    subconsciousConfig = null;
+    /** Tune-out coordinator (issue #77); non-null iff subconscious + channels. */
+    tuneOutCoordinator = null;
     deferredMessages = [];
     // Turn-alive markers (2026-07-31 Mythos phantom-skip incident): agentName →
     // token of the turn currently in progress. Set at startAgentStream ENTRY —
@@ -569,6 +608,21 @@ export class AgentFramework {
     // (KV bust). Invariant: nothing enters the window between a turn's dequeue
     // and its settle except that turn's own blocks.
     activeTurnTokens = new Map();
+    /**
+     * The InferenceRequest that started each agent's turn in progress — set and
+     * cleared exactly where the turn token is. Read-only outside: the host
+     * stamps gateway telemetry (why this call exists, in which channel, woken
+     * by whom) from it, so a household ledger can attribute wakes without
+     * guessing from the newest message in the window.
+     */
+    activeTurnTriggers = new Map();
+    /** The trigger of `agentName`'s turn in progress, if a turn is running. */
+    getActiveTurnTrigger(agentName) {
+        return this.activeTurnTriggers.get(agentName);
+    }
+    /** Structured calls executed in the current logical turn. Unlike the
+     * per-stream local this survives framework retries and budget restarts. */
+    logicalTurnToolCalls = new WeakMap();
     nextTurnToken = 1;
     // Undo/redo state
     turnCounters = new Map(); // agentName → next turnIndex
@@ -780,6 +834,12 @@ export class AgentFramework {
         for (const agentConfig of config.agents) {
             await framework.createAgent(agentConfig);
         }
+        // The subconscious resident (issue #77) registers after the residents so
+        // it can never become primary; it is excluded from every broadcast
+        // fan-out and woken only by the tune-out coordinator's explicit pushes.
+        if (config.subconscious?.enabled) {
+            await framework.createSubconsciousAgent(config.subconscious);
+        }
         // Finish any branch-local suppression interrupted after Chronicle switched
         // branches. This runs before modules, MCPL connections, or inbound traffic.
         await framework.resumePreparedDiscordSuppressions();
@@ -824,11 +884,11 @@ export class AgentFramework {
                 initialConfig: config.gate.config,
                 privilegedUsersPath: config.gate.privilegedUsersPath,
                 emitTrace: (e) => framework.emitTrace(e),
-                addMessage: (p, c, m) => framework.addMessage(p, c, m),
+                addMessage: (p, c, m, forAgent) => framework.addMessage(p, c, m, forAgent ? { forAgent } : undefined),
                 requestInference: (agentName, reason, source) => {
                     framework.pendingRequests.push({ agentName, reason, source, timestamp: Date.now() });
                 },
-                getAgentNames: () => [...framework.agents.keys()],
+                getAgentNames: () => [...framework.agents.keys()].filter((n) => n !== framework.subconsciousAgentName),
             });
         }
         // Liveness watchdog: fail hard if the main thread wedges (opt-in).
@@ -879,6 +939,30 @@ export class AgentFramework {
             }
             await framework.initializeMcpl(config.mcplServers, config.inferenceRouting);
         }
+        // Tune-out coordinator (issue #77): needs both the subconscious resident
+        // and the MCPL channel subsystem; created after both exist.
+        if (framework.subconsciousAgentName && framework.channelRegistry) {
+            framework.tuneOutCoordinator = new TuneOutCoordinator(framework.channelRegistry, framework.mcplServerRegistry, {
+                addMessage: (participant, content, metadata, forAgent) => framework.addMessage(participant, content, metadata, forAgent ? { forAgent } : undefined),
+                requestInference: (agentName, reason, source) => {
+                    framework.pendingRequests.push({ agentName, reason, source, timestamp: Date.now() });
+                },
+                subconsciousName: () => framework.subconsciousAgentName,
+                primaryName: () => framework.primaryAgentName,
+                getStoredMessages: () => {
+                    const primary = framework.primaryAgentName
+                        ? framework.agents.get(framework.primaryAgentName) : undefined;
+                    return primary ? primary.getContextManager().getAllMessages() : [];
+                },
+                currentSequence: () => framework.store.currentSequence(),
+                setSubconsciousAnchor: (sequence) => framework.subconsciousStrategy?.setAnchor(sequence),
+                isForkBound: (channelId) => [...framework.conversationAgentHomes.values()].includes(channelId),
+                isPrivilegedAuthor: (authorId) => framework.eventGate?.isPrivilegedUser(authorId) ?? false,
+                allowChannelSpeech: () => !!framework.subconsciousConfig?.allowChannelSpeech,
+                emitTrace: (e) => framework.emitTrace(e),
+            });
+            framework.tuneOutCoordinator.resumeActiveEpochs();
+        }
         // Diagnostics: `kill -USR2 <pid>` dumps live wake/inference state to stderr
         // (journal) without a restart — for catching the wake-wedge on the running
         // process. Shows the gate's `inferring` set + buffered-event count (the
@@ -943,6 +1027,7 @@ export class AgentFramework {
         this.running = false;
         this.providerAdmissionClosed = true;
         this.queue.close();
+        this.tuneOutCoordinator?.stop();
         // Kill running code_execution scripts before cancelling streams: a
         // zombie script must not keep firing side-effectful tool calls into a
         // framework that is shutting down.
@@ -1374,7 +1459,12 @@ export class AgentFramework {
      */
     getAllTools() {
         const moduleTools = this.moduleRegistry.getAllTools();
-        const channelTools = this.channelRegistry?.getChannelTools() ?? [];
+        // Copy: getChannelTools() returns the registry's shared definitions
+        // array; pushing onto it would append another tune_out on every call.
+        const channelTools = [...(this.channelRegistry?.getChannelTools() ?? [])];
+        if (this.tuneOutCoordinator) {
+            channelTools.push(AgentFramework.TUNE_OUT_TOOL);
+        }
         const gateTools = this.eventGate
             ? [
                 this.eventGate.getToolDefinition(),
@@ -1392,7 +1482,8 @@ export class AgentFramework {
             ...gateTools,
             this.buildAgentSettingsTool(),
             ...(this.getWorkspaceModule()
-                ? [AgentFramework.SAVE_IMAGE_TOOL, AgentFramework.READ_IMAGE_TOOL]
+                ? [AgentFramework.SAVE_IMAGE_TOOL, AgentFramework.READ_IMAGE_TOOL,
+                    AgentFramework.FETCH_REFERENCE_TOOL]
                 : []),
             ...(this.codeExecutionConfig
                 ? [buildCodeExecutionToolDefinition(this.codeExecutionConfig)]
@@ -1406,6 +1497,57 @@ export class AgentFramework {
             ? module
             : undefined;
     }
+    /** RFC-005 host-mediated reference fetcher (lazy; needs server configs +
+     *  a writable workspace mount for storage). */
+    referenceFetcherInstance;
+    getReferenceFetcher() {
+        if (!this.referenceFetcherInstance) {
+            this.referenceFetcherInstance = new ReferenceFetcher((serverId) => {
+                const cfg = this.mcplServerConfigs.get(serverId);
+                if (!cfg?.url)
+                    return undefined;
+                return { url: cfg.url, token: cfg.token, autofetch: cfg.autofetch };
+            }, async (fileName, data, mimeType) => {
+                const workspace = this.getWorkspaceModule();
+                if (!workspace)
+                    return null;
+                const mount = this.firstWritableMountName(workspace);
+                if (!mount)
+                    return null;
+                const path = `${mount}/refs/${fileName}`;
+                const res = await workspace.writeBinary(path, data, mimeType);
+                return res.success ? path : null;
+            });
+        }
+        return this.referenceFetcherInstance;
+    }
+    /** Register every RFC-005 reference in a tool-result content array against
+     *  its server, then eagerly fetch the ones inside the autofetch envelope
+     *  (below the eager ceiling, connection origin — never worse than the
+     *  server having inlined the bytes). Failures are non-fatal: the stub then
+     *  says fetch_reference instead of a path. */
+    async autoFetchReferences(content, serverId, eager) {
+        if (!content)
+            return;
+        const fetcher = this.getReferenceFetcher();
+        const eligible = [];
+        for (const block of content) {
+            const c = classifyBlock(block);
+            if (c.kind !== 'reference')
+                continue;
+            const record = referenceRegistry.register(c.testimony, serverId);
+            if (eager && fetcher.isEagerEligible(record))
+                eligible.push(record);
+        }
+        // Parallel with a count cap: N small fetches cost one round-trip of
+        // wall-clock, and anything past the cap stays lazy (stub says so).
+        const MAX_EAGER_PER_RESULT = 8;
+        await Promise.all(eligible.slice(0, MAX_EAGER_PER_RESULT).map(async (record) => {
+            const outcome = await fetcher.fetch(record, { timeoutMs: EAGER_FETCH_TIMEOUT_MS });
+            if (!outcome.ok)
+                console.error(`[rfc005] eager fetch of ${record.refId} failed: ${outcome.error}`);
+        }));
+    }
     /** Core agent_settings keys — extension keys must not collide with these. */
     static AGENT_SETTINGS_CORE_KEYS = [
         'context_budget_tokens',
@@ -1552,6 +1694,18 @@ export class AgentFramework {
         };
     }
     getToolsForAgent(agentName, snapshot) {
+        // The subconscious gets its own small surface (plus think/skip_reply/
+        // end_turn), not the residents' tool board: no channel lifecycle, no
+        // gate rules, no workspace — it observes, judges, and reports.
+        if (agentName === this.subconsciousAgentName) {
+            const basics = this.getAllTools()
+                .filter((t) => t.name === 'think' || t.name === 'skip_reply' || t.name === 'end_turn')
+                .map((t) => t.name === 'think'
+                ? this.buildThinkTool(snapshot?.sameRoundThinkTextPolicy
+                    ?? this.getAgentRuntimeSettings(agentName).sameRoundThinkTextPolicy)
+                : t);
+            return [...SUBCONSCIOUS_TOOLS, ...basics];
+        }
         return this.getAllTools().map((tool) => {
             if (tool.name === 'think') {
                 return this.buildThinkTool(snapshot?.sameRoundThinkTextPolicy
@@ -1907,21 +2061,35 @@ export class AgentFramework {
      * Call cleanup() when done to release the ContextManager.
      */
     async createEphemeralAgent(config) {
-        const namespace = `subagent/${config.name}`;
-        const contextManager = await ContextManager.open({
-            store: this.store,
-            namespace,
-            isolate: true,
-            strategy: config.strategy ?? new PassthroughStrategy(),
-            membrane: this.membrane,
-            debugLogContext: !!process.env.DEBUG_CONTEXT,
-        });
-        const agent = new Agent(config, contextManager, this.membrane);
-        const cleanup = () => {
-            // Don't close the store — it's shared. Just release the CM.
-            // Data persists in the store under the namespace for investigation.
-        };
-        return { agent, contextManager, cleanup };
+        // Names are Chronicle namespaces and generation identities. Reserve before
+        // opening the namespace so a second creation cannot append to an earlier
+        // generation before runEphemeralToCompletion has a chance to reject it.
+        if (this.usedEphemeralAgentNames.has(config.name) || this.agents.has(config.name)) {
+            throw new Error(`Ephemeral agent name \"${config.name}\" is already registered or has been used in this framework`);
+        }
+        this.usedEphemeralAgentNames.add(config.name);
+        try {
+            const namespace = `subagent/${config.name}`;
+            const contextManager = await ContextManager.open({
+                store: this.store,
+                namespace,
+                isolate: true,
+                strategy: config.strategy ?? new PassthroughStrategy(),
+                membrane: this.membrane,
+                debugLogContext: !!process.env.DEBUG_CONTEXT,
+            });
+            const agent = new Agent(config, contextManager, this.membrane);
+            this.ephemeralCandidates.set(agent, contextManager);
+            const cleanup = () => {
+                // Don't close the store — it's shared. Just release the CM.
+                // Data persists in the store under the namespace for investigation.
+            };
+            return { agent, contextManager, cleanup };
+        }
+        catch (error) {
+            this.usedEphemeralAgentNames.delete(config.name);
+            throw error;
+        }
     }
     createDeferred() {
         let resolve;
@@ -1942,6 +2110,18 @@ export class AgentFramework {
         }
         run.lastActivity = Date.now();
     }
+    /** Record structured calls for the current logical turn. Ignore late stream
+     * events after an ephemeral/conversation agent has been disposed. */
+    recordLogicalTurnToolCalls(agent, turnToken, count) {
+        if (count <= 0 || this.agents.get(agent.name) !== agent)
+            return;
+        const state = this.logicalTurnToolCalls.get(agent);
+        // The Agent object is the generation: a late stream from a disposed
+        // ephemeral/conversation fork cannot contaminate a same-name successor.
+        if (!state || state.turnToken !== turnToken)
+            return;
+        this.logicalTurnToolCalls.set(agent, { turnToken, count: state.count + count });
+    }
     recordEphemeralToolCalls(agentName, count) {
         const run = this.ephemeralRuns.get(agentName);
         if (!run)
@@ -1978,7 +2158,18 @@ export class AgentFramework {
      * The task message should already be in the context manager.
      */
     async runEphemeralToCompletion(agent, contextManager, watchdogs) {
-        // Register temporarily so the event loop can drive it
+        // Only a fresh object returned by createEphemeralAgent may enter this path.
+        // Never overwrite a resident/conversation owner or a concurrent run: their
+        // cleanup is name-keyed and could cancel/deregister the legitimate owner.
+        if (this.ephemeralCandidates.get(agent) !== contextManager) {
+            throw new Error(`Ephemeral agent "${agent.name}" has no fresh generation ticket from this framework`);
+        }
+        // Consume before any await/registration: the exact (Agent, ContextManager)
+        // generation is one-shot even if startup later fails.
+        this.ephemeralCandidates.delete(agent);
+        if (this.agents.has(agent.name) || this.ephemeralRuns.has(agent.name)) {
+            throw new Error(`Ephemeral agent "${agent.name}" is already registered or running`);
+        }
         this.agents.set(agent.name, agent);
         const run = {
             settle: this.createDeferred(),
@@ -2037,8 +2228,26 @@ export class AgentFramework {
                 clearTimeout(startupWatchdog);
             if (completionWatchdog)
                 clearInterval(completionWatchdog);
-            this.ephemeralRuns.delete(agent.name);
-            this.agents.delete(agent.name);
+            // Mark before cancellation: an adapter may synchronously settle its
+            // iterator, and this physical frame still owns terminal typing/outgoing
+            // closure even though the ephemeral name is about to be deregistered.
+            this.disposedEphemeralFrames.add(agent);
+            // Cancel before deregistration. Adapters may still yield a queued event;
+            // driveStream's Agent-identity check above discards it before dispatch.
+            agent.cancelStream();
+            if (this.ephemeralRuns.get(agent.name) === run)
+                this.ephemeralRuns.delete(agent.name);
+            if (this.agents.get(agent.name) === agent) {
+                // Exact-generation disposal owns these name-keyed entries. Remove them
+                // before deregistration; the late driveStream finally will see the
+                // generation mismatch and leave any future owner untouched.
+                this.activeStreams.delete(agent.name);
+                this.pendingAssistantBlocks.delete(agent.name);
+                // The late stream finalizer intentionally refuses name-keyed cleanup once
+                // deregistered, so disposal must release EventGate liveness itself.
+                this.eventGate?.onInferenceEnded(agent.name);
+                this.agents.delete(agent.name);
+            }
             // Spawn-and-dispose bookkeeping (main, d453165/fee96a7): without this,
             // ephemeral agents leave checkpoint-tree keys and diagnostics map
             // entries behind for the life of the store/session.
@@ -2049,6 +2258,8 @@ export class AgentFramework {
             // unbounded map growth. Blind delete is safe: the agent is already out
             // of the map, and a late driveStream finally token-match no-ops.
             this.activeTurnTokens.delete(agent.name);
+            this.activeTurnTriggers.delete(agent.name);
+            this.logicalTurnToolCalls.delete(agent);
         }
     }
     /**
@@ -2061,6 +2272,44 @@ export class AgentFramework {
      * Query inference logs.
      * Returns entries with summary info (doesn't resolve blobs).
      */
+    /** Tune-out (issue #77): divert a channel to your subconscious. Present
+     *  only when the subconscious resident is configured. */
+    static TUNE_OUT_TOOL = {
+        name: 'tune_out',
+        description: 'Divert a channel to your subconscious for a while: its traffic stops ' +
+            'entering your context and stops waking you. Your subconscious watches ' +
+            'it instead — summarizing on a cadence in its own voice, judging ' +
+            'whether mentions merit interrupting, and able to cancel. People who ' +
+            'mention you get an automatic reaction so they know you saw nothing. ' +
+            'Cancelling (mode "cancel") delivers the diverted backlog. The ' +
+            'tune-out auto-cancels after maxWakes mention-bursts.',
+        inputSchema: {
+            type: 'object',
+            properties: {
+                channelId: { type: 'string' },
+                mode: { type: 'string', enum: ['enter', 'cancel'], description: 'Default: enter.' },
+                cadenceSeconds: {
+                    type: 'number',
+                    description: `Summary cadence in seconds (default ${TUNE_OUT_DEFAULTS.cadenceSeconds}; floors at 30).`,
+                },
+                backlogCap: {
+                    type: 'number',
+                    description: `Max raw messages delivered at cancel (default ${TUNE_OUT_DEFAULTS.backlogCap}). ` +
+                        '0 = deliver no raw backlog at all: your subconscious\'s report is the whole delivery.',
+                },
+                maxWakes: {
+                    type: 'number',
+                    description: `Wake budget before auto-cancel (default ${TUNE_OUT_DEFAULTS.maxWakes}; floors at 1).`,
+                },
+                durationSeconds: {
+                    type: 'number',
+                    description: 'Optional bounded duration: the tune-out auto-cancels (backlog ' +
+                        'delivered) after this many seconds; floors at 60. Unset = until cancelled.',
+                },
+            },
+            required: ['channelId'],
+        },
+    };
     /** Synthesized sleep/wake tool definitions (present when a gate is wired). */
     static SLEEP_TOOLS = [
         {
@@ -2265,6 +2514,26 @@ export class AgentFramework {
             required: ['path'],
         },
     };
+    /** Synthesized fetch_reference tool — RFC-005 on-demand dereference.
+     *  Present when a workspace module is registered (storage target). */
+    static FETCH_REFERENCE_TOOL = {
+        name: 'fetch_reference',
+        description: 'Fetch a referenced payload (a [ref_…] id from a tool result or event) ' +
+            'to a workspace file via the host-mediated fetcher. The fetch is ' +
+            'origin-bound to the server that sent the reference, size-capped, and ' +
+            'digest-verified when the reference carries one. Returns the workspace ' +
+            'path plus verified size/type; then use ordinary workspace tools on it. ' +
+            'Small payloads may already be fetched automatically — their stubs show ' +
+            'a saved path instead.',
+        inputSchema: {
+            type: 'object',
+            properties: {
+                ref_id: { type: 'string', description: 'The reference id, e.g. "ref_1a_x9z2".' },
+                max_bytes: { type: 'number', description: 'Optional per-call byte ceiling (default host policy).' },
+            },
+            required: ['ref_id'],
+        },
+    };
     /** Synthesized read_image tool — present when a workspace module is
      *  registered. Returns the image as a native image block in the tool
      *  result, so the agent SEES it in the live turn (history keeps a compact
@@ -3292,6 +3561,15 @@ export class AgentFramework {
             // streaming agent keeps the original Membrane.
             membrane: this.auxiliaryMembraneFor(config.name),
             debugLogContext: !!process.env.DEBUG_CONTEXT,
+            // Tune-out (issue #77): messages stamped at ingestion as diverted
+            // (metadata.tuneOut = { epochId }) never enter a resident's compiled
+            // view — not at emission, not via chunking into the memory pyramid.
+            // The stamp is permanent: after cancel, the backlog arrives as one
+            // appended <tuned-out-backlog> dump instead of retro-inserting into
+            // the timeline (KV-prefix-stable delivery). The stored originals
+            // remain in the shared slot for the subconscious's merged view,
+            // fetch_history, and audit.
+            viewFilter: (message) => !message.metadata?.tuneOut,
         });
         const agent = new Agent(config, contextManager, this.membrane);
         const restoredSettings = this.readAgentRuntimeSettings(config.name);
@@ -3306,6 +3584,68 @@ export class AgentFramework {
         }
         return agent;
     }
+    /**
+     * Create and register the subconscious resident (issue #77): a persistent
+     * side-agent with an isolated slot (`subconscious/<primary>`) whose
+     * strategy view merges the residents' shared timeline with its own —
+     * including the tuned-out-stamped messages the residents' viewFilter
+     * excludes; the backlog is precisely what it exists to see. Windowed
+     * passthrough, no memory pyramid: its durable output is what it delivers
+     * into the resident's window, which accumulates there.
+     */
+    async createSubconsciousAgent(cfg) {
+        const primaryName = this.primaryAgentName;
+        const primaryConfig = primaryName ? this.agentConfigs.get(primaryName) : undefined;
+        if (!primaryName || !primaryConfig) {
+            throw new Error('subconscious requires a primary agent to attend to');
+        }
+        const name = cfg.name ?? 'Subconscious';
+        if (this.agents.has(name)) {
+            throw new Error(`subconscious name "${name}" collides with a registered agent`);
+        }
+        const strategy = new WindowedPassthroughStrategy({
+            reAnchorFraction: cfg.reAnchorFraction,
+            // Mirror the principal's per-message truncation ceiling so both see
+            // oversized tool results / attachments under the same policy.
+            maxMessageTokens: primaryConfig.strategy?.maxMessageTokens,
+        });
+        const contextManager = await ContextManager.open({
+            store: this.store,
+            namespace: `subconscious/${primaryName}`,
+            isolate: true,
+            // The residents' shared un-namespaced slot, merged read-only and
+            // UNfiltered (no viewFilter here — stamped messages included).
+            auxiliaryMessageViews: [{}],
+            strategy,
+            membrane: this.membrane,
+            debugLogContext: !!process.env.DEBUG_CONTEXT,
+        });
+        // Inherit the principal's inference settings wholesale (fork-template
+        // idiom): reasoning effort (thinking), provider params, token ceilings,
+        // caching, refusal handling, timezone — a same-model side-process must
+        // not silently run under different inference conditions than the
+        // resident it serves. Only identity- and surface-defining fields are
+        // overridden below.
+        const agentConfig = {
+            ...primaryConfig,
+            name,
+            model: cfg.model ?? primaryConfig.model,
+            systemPrompt: cfg.systemPrompt,
+            strategy: undefined, // its CM owns the windowed strategy instance
+            // Prose is never auto-routed: a cadence-triggered turn carries no
+            // locus, and bare prose must not fall through to the default channel.
+            // Speech happens only via speak_in_channel (its own marked voice).
+            proseRouting: 'explicit',
+            allowedTools: [...SUBCONSCIOUS_TOOL_NAMES, 'think', 'skip_reply', 'end_turn'],
+        };
+        const agent = new Agent(agentConfig, contextManager, this.membrane);
+        this.agents.set(name, agent);
+        this.agentConfigs.set(name, agentConfig);
+        this.subconsciousAgentName = name;
+        this.subconsciousStrategy = strategy;
+        this.subconsciousConfig = cfg;
+        return agent;
+    }
     async runLoop() {
         while (this.running) {
             try {
@@ -3430,12 +3770,11 @@ export class AgentFramework {
                     // queued, they flush at the target's own next boundary — with
                     // injection — or in driveStream's finally when its turn ends.
                     const midTurnInjections = [];
-                    if (this.deferredMessages.length > 0) {
-                        const target = (this.primaryAgentName ? this.agents.get(this.primaryAgentName) : undefined) ?? agent;
-                        if (target === agent) {
-                            const deferred = this.deferredMessages.splice(0);
+                    {
+                        const deferred = this.drainDeferredFor(agent.name);
+                        if (deferred.length > 0) {
                             for (const msg of deferred) {
-                                target.getContextManager().addMessage(msg.participant, msg.content, msg.metadata);
+                                agent.getContextManager().addMessage(msg.participant, msg.content, msg.metadata);
                                 // Injection guards: tool blocks would corrupt the tool-cycle
                                 // structure the membrane enforces, and a message named as the
                                 // agent itself would render as an ASSISTANT turn on the wire
@@ -3732,7 +4071,7 @@ export class AgentFramework {
             // Broadcast requests exclude conversation forks — they are driven by
             // their own channel's messages, not by framework-wide events.
             const targetAgents = response.requestInference === true
-                ? Array.from(this.agents.keys()).filter((n) => !this.conversationAgentHomes.has(n))
+                ? Array.from(this.agents.keys()).filter((n) => !this.conversationAgentHomes.has(n) && n !== this.subconsciousAgentName)
                 : response.requestInference;
             for (const agentName of targetAgents) {
                 const agent = this.agents.get(agentName);
@@ -3799,11 +4138,31 @@ export class AgentFramework {
                 Object.assign(metadata, invitation.metadataPatch);
             }
         }
+        // Tune-out divert (issue #77): stamped, stored, no resident wake. The
+        // coordinator has already handled wake bookkeeping (coalesced
+        // subconscious invocation, durable count, suppression ack) before we
+        // stamp; the subconscious reads the message through its merged view.
+        const divert = this.tuneOutCoordinator?.onIncoming(event.serverId, event.channelId, event.messageId, event.tags, event.author?.id, 
+        // The resident's gate verdict rides the event; it preconditions
+        // subconscious wakes (gate composes, it is not replaced).
+        event.triggerInference !== false) ?? null;
+        if (divert) {
+            metadata.tuneOut = { epochId: divert.epochId };
+        }
         const id = this.addMessage('user', incomingContent, metadata);
         this.emitTrace({ type: 'message:added', messageId: id, source: 'mcpl:channel-incoming' });
-        if (event.triggerInference) {
+        if (event.triggerInference && !divert) {
             const addressed = isAddressedMessage(event.tags, event.metadata);
-            for (const agentName of this.agents.keys()) {
+            // Honor explicit targeting, mirroring the push-event path: an event
+            // that names its agents wakes exactly those; an untargeted event
+            // keeps the historical broadcast. (targetAgents was declared on
+            // McplChannelIncomingEvent from the start but never honored here —
+            // tune-out's wake routing is the first setter.)
+            const targetAgents = event.targetAgents
+                ?? [...this.agents.keys()].filter((n) => n !== this.subconsciousAgentName);
+            for (const agentName of targetAgents) {
+                if (!this.agents.has(agentName))
+                    continue;
                 this.pendingRequests.push({
                     agentName,
                     reason: 'mcpl:channel-incoming',
@@ -3812,6 +4171,7 @@ export class AgentFramework {
                     // Route this turn's auto-published speech back to THIS channel, not
                     // the global most-recent-inbound locus (item-3 redux, trunk agents).
                     channelId: event.channelId,
+                    counterparty: event.author?.id ? `${event.serverId}:user:${event.author.id}` : undefined,
                     // Addressed messages outrank ambient chatter when a batched wake
                     // picks the turn's frozen speech locus.
                     addressed,
@@ -3896,6 +4256,7 @@ export class AgentFramework {
                 // A fork's home channel wins in routeSpeech regardless, but carry the
                 // triggering channel too so the trunk/active path stays consistent.
                 channelId: event.channelId,
+                counterparty: event.author?.id ? `${event.serverId}:user:${event.author.id}` : undefined,
                 addressed: isAddressedMessage(event.tags, event.metadata),
             });
         }
@@ -3914,38 +4275,48 @@ export class AgentFramework {
         if (!template || !templateConfig) {
             throw new Error(`conversation template agent "${router.templateAgent}" not found`);
         }
-        const contextManager = await ContextManager.open({
-            store: this.store,
-            namespace: `conversations/${name}`,
-            isolate: true,
-            // Strategy instances are stateful — never share the template's.
-            strategy: router.strategyFactory?.() ?? new PassthroughStrategy(),
-            // Dynamic conversation forks retain the established provider policy in
-            // this bounded first slice. Provider cooldown ownership belongs to the
-            // persistent resident only; generation-unique forks must not leak gates.
-            membrane: this.membrane,
-            debugLogContext: !!process.env.DEBUG_CONTEXT,
-        });
-        // Seed with the template's compiled context, renaming the template
-        // participant so the fork reads its inheritance as its own history.
-        // Guard: only seed a genuinely fresh namespace. Generation counters are
-        // persisted precisely so names aren't reused, but if this namespace has
-        // history anyway (counter state lost, crash between spawn and persist),
-        // seeding again would stack another template copy on top of it.
-        const { messages: existing } = await contextManager.compile();
-        if (existing.length === 0) {
-            const { messages: compiled } = await template.getContextManager().compile();
-            for (const msg of compiled) {
-                const participant = msg.participant === template.name ? name : msg.participant;
-                contextManager.addMessage(participant, msg.content);
-            }
-        }
-        const config = { ...templateConfig, name, strategy: undefined };
-        const agent = new Agent(config, contextManager, this.membrane);
-        this.agents.set(name, agent);
-        this.agentConfigs.set(name, config);
-        this.conversationAgentHomes.set(name, channelId);
-        return agent;
+        if (this.usedEphemeralAgentNames.has(name) || this.agents.has(name)) {
+            throw new Error(`Conversation agent name "${name}" is already registered or reserved`);
+        }
+        this.usedEphemeralAgentNames.add(name);
+        try {
+            const contextManager = await ContextManager.open({
+                store: this.store,
+                namespace: `conversations/${name}`,
+                isolate: true,
+                // Strategy instances are stateful — never share the template's.
+                strategy: router.strategyFactory?.() ?? new PassthroughStrategy(),
+                // Dynamic conversation forks retain the established provider policy in
+                // this bounded first slice. Provider cooldown ownership belongs to the
+                // persistent resident only; generation-unique forks must not leak gates.
+                membrane: this.membrane,
+                debugLogContext: !!process.env.DEBUG_CONTEXT,
+            });
+            // Seed with the template's compiled context, renaming the template
+            // participant so the fork reads its inheritance as its own history.
+            // Guard: only seed a genuinely fresh namespace. Generation counters are
+            // persisted precisely so names aren't reused, but if this namespace has
+            // history anyway (counter state lost, crash between spawn and persist),
+            // seeding again would stack another template copy on top of it.
+            const { messages: existing } = await contextManager.compile();
+            if (existing.length === 0) {
+                const { messages: compiled } = await template.getContextManager().compile();
+                for (const msg of compiled) {
+                    const participant = msg.participant === template.name ? name : msg.participant;
+                    contextManager.addMessage(participant, msg.content);
+                }
+            }
+            const config = { ...templateConfig, name, strategy: undefined };
+            const agent = new Agent(config, contextManager, this.membrane);
+            this.agents.set(name, agent);
+            this.agentConfigs.set(name, config);
+            this.conversationAgentHomes.set(name, channelId);
+            return agent;
+        }
+        catch (error) {
+            this.usedEphemeralAgentNames.delete(name);
+            throw error;
+        }
     }
     /** Persist the router's generation counters (see hydration in create()). */
     persistConversationRouterState() {
@@ -3969,10 +4340,13 @@ export class AgentFramework {
     disposeConversationAgent(agentName) {
         this.closingConversationAgents.delete(agentName);
         const channelId = this.conversationAgentHomes.get(agentName);
+        const agent = this.agents.get(agentName);
         this.agents.delete(agentName);
         this.agentConfigs.delete(agentName);
         this.conversationAgentHomes.delete(agentName);
         this.evictTurnCheckpoints(agentName);
+        if (agent)
+            this.logicalTurnToolCalls.delete(agent);
         this.emitTrace({
             type: 'mcpl:conversation-disposed',
             agentName,
@@ -4172,7 +4546,7 @@ export class AgentFramework {
         if (event.triggerInference) {
             // Default broadcast excludes conversation forks (channel-driven).
             const targetAgents = event.targetAgents
-                ?? [...this.agents.keys()].filter((n) => !this.conversationAgentHomes.has(n));
+                ?? [...this.agents.keys()].filter((n) => !this.conversationAgentHomes.has(n) && n !== this.subconsciousAgentName);
             for (const agentName of targetAgents) {
                 this.pendingRequests.push({
                     agentName,
@@ -4354,6 +4728,10 @@ export class AgentFramework {
                 // exactly this class of silent drop. One stderr line per drop.
                 console.error(`[inference-dropped] agent=${agentName} reason=policy-skip ` +
                     `requests=${requests.length} triggers=${requests.map((r) => r.reason).join(',')}`);
+                // A context-budget restart inherits the predecessor's EventGate liveness.
+                // If policy drops the queued successor, no driveStream remains to release it.
+                if (budgetRestart)
+                    this.eventGate?.onInferenceEnded(agentName);
                 const gate = this.providerGates.get(agentName);
                 if (gate?.primaryPending && this.providerAccelerationRecoveries.has(agentName)) {
                     gate.primaryPending = false;
@@ -4380,21 +4758,29 @@ export class AgentFramework {
             // newest (2026-07-21 Cairn lounge misroute, turn-start variant).
             // Non-channel wakes (heartbeats, module events, reactions — which never
             // carry channelId) leave both undefined → global fallback.
-            let ambientChannel;
-            let addressedChannel;
+            // Track the winning REQUEST, not just its channel: channel, addressed
+            // and counterparty must come from the same message, or a batch of
+            // "ambient from A, then addressed from B" would report B's channel
+            // with A's author (review finding on the provenance change).
+            let ambientReq;
+            let addressedReq;
             for (const r of requests) {
                 if (!r.channelId)
                     continue;
-                ambientChannel = r.channelId;
+                ambientReq = r;
                 if (r.addressed)
-                    addressedChannel = r.channelId;
+                    addressedReq = r;
             }
-            const triggerChannel = addressedChannel ?? ambientChannel;
-            const triggerAddressed = addressedChannel !== undefined;
+            const channelReq = addressedReq ?? ambientReq;
             await this.startAgentStream(agent, {
                 ...trigger,
-                channelId: triggerChannel,
-                addressed: triggerAddressed,
+                channelId: channelReq?.channelId,
+                addressed: addressedReq !== undefined,
+                // A context-budget restart continues the same logical turn: it keeps
+                // the channel for routing but names no author — the restart is its
+                // own cause, and borrowing another request's author would be false
+                // provenance.
+                counterparty: budgetRestart ? undefined : channelReq?.counterparty,
             });
         }
     }
@@ -4862,6 +5248,7 @@ export class AgentFramework {
         // no-ops instead of clobbering a successor's marker.
         const turnToken = this.nextTurnToken++;
         this.activeTurnTokens.set(agent.name, turnToken);
+        this.activeTurnTriggers.set(agent.name, trigger);
         let tokenHandedOff = false;
         try {
             tokenHandedOff = await this.beginAgentTurn(agent, trigger, attempt, turnToken, ownsProviderGate);
@@ -4869,6 +5256,7 @@ export class AgentFramework {
         finally {
             if (!tokenHandedOff && this.activeTurnTokens.get(agent.name) === turnToken) {
                 this.activeTurnTokens.delete(agent.name);
+                this.activeTurnTriggers.delete(agent.name);
             }
             if (!tokenHandedOff && ownsProviderGate)
                 this.releasePrimaryProviderGate(agent.name);
@@ -4897,9 +5285,8 @@ export class AgentFramework {
         if (attempt === 0 &&
             trigger?.reason !== 'context_budget_restart' &&
             this.deferredMessages.length > 0) {
-            const target = (this.primaryAgentName ? this.agents.get(this.primaryAgentName) : undefined) ?? agent;
-            if (target === agent) {
-                const deferred = this.deferredMessages.splice(0);
+            {
+                const deferred = this.drainDeferredFor(agent.name);
                 for (const msg of deferred) {
                     // Per-message try/catch (mirrors announceLocusIfChanged): one
                     // poison message — or a transient store-write failure — must not
@@ -4953,9 +5340,22 @@ export class AgentFramework {
         // BEFORE this turn compiles, so the agent always knows where its voice
         // goes (announce-on-change only — no per-turn chatter, append-only for
         // KV stability).
+        if (trigger?.reason === 'context_budget_restart') {
+            const previousLogicalToolState = this.logicalTurnToolCalls.get(agent);
+            this.logicalTurnToolCalls.set(agent, { turnToken, count: previousLogicalToolState?.count ?? 0 });
+        }
         if (trigger?.reason !== 'context_budget_restart') {
             if (attempt === 0)
                 this.maybePrimeProseMode(agent);
+            const previousLogicalToolState = this.logicalTurnToolCalls.get(agent);
+            if (attempt === 0) {
+                // A true new turn gets a fresh generation and count.
+                this.logicalTurnToolCalls.set(agent, { turnToken, count: 0 });
+            }
+            else {
+                // A framework retry is a new physical stream in the same logical turn.
+                this.logicalTurnToolCalls.set(agent, { turnToken, count: previousLogicalToolState?.count ?? 0 });
+            }
             // Fresh turn: forget the previous turn's explicit-send engagements and
             // prose deliveries — both are strictly turn-scoped (a restart continues
             // the same logical turn and keeps them).
@@ -4993,7 +5393,11 @@ export class AgentFramework {
             // context-budget restart, which skips the re-pin but re-emits this).
             channelId: this.turnLocusPins.get(agent.name),
         });
-        this.eventGate?.onInferenceStarted(agent.name);
+        // A budget restart continues the same logical inference window. The
+        // predecessor keeps EventGate liveness until its successor terminates.
+        if (trigger?.reason !== 'context_budget_restart') {
+            this.eventGate?.onInferenceStarted(agent.name);
+        }
         this.lastInferenceAt.set(agent.name, { ...this.lastInferenceAt.get(agent.name), startedAt: Date.now() });
         // Typing indicator, started at TURN START. It says "attending", which is
         // true from the moment the turn is accepted — but it used to start inside
@@ -5047,8 +5451,32 @@ export class AgentFramework {
                     console.error('beforeInference hook error:', error);
                 }
             }
-            const { stream, request: compiledRequest } = await agent.startStreamWithInjections(tools, injections);
-            const handle = this.driveStream(agent, stream, requestSnapshot, trigger, attempt, compiledRequest, ownsProviderGate);
+            // An ephemeral watchdog may dispose this Agent while hooks/compile await.
+            // Do not create a provider stream after its generation lost ownership.
+            if (this.agents.get(agent.name) !== agent) {
+                this.channelRegistry?.stopTyping();
+                this.eventGate?.onInferenceEnded(agent.name);
+                return false;
+            }
+            // The subconscious's standing dispositions ride a fixed system-position
+            // injection (antra, #tuneout-talk): never repeated in its timeline, and
+            // — since it has no long-term memory — structurally unable to wash out
+            // of its window. Bytes change only when a disposition changes.
+            if (agent.name === this.subconsciousAgentName && this.tuneOutCoordinator) {
+                const dispositionInjection = this.tuneOutCoordinator.getDispositionsInjection();
+                if (dispositionInjection) {
+                    injections = injections ? [...injections, dispositionInjection] : [dispositionInjection];
+                }
+            }
+            const { stream, request: compiledRequest, takeKvSubmission, drainKvSubmissionIds, } = await agent.startStreamWithInjections(tools, injections);
+            if (this.agents.get(agent.name) !== agent) {
+                stream.cancel();
+                agent.cancelStream();
+                this.channelRegistry?.stopTyping();
+                this.eventGate?.onInferenceEnded(agent.name);
+                return false;
+            }
+            const handle = this.driveStream(agent, stream, requestSnapshot, trigger, attempt, compiledRequest, ownsProviderGate, takeKvSubmission, drainKvSubmissionIds, new Set(tools.map((tool) => tool.name)));
             this.activeStreams.set(agent.name, handle);
             // Handoff: driveStream captured the token in its synchronous prefix;
             // its finally now owns clearing it. The caller's finally must not.
@@ -5088,6 +5516,7 @@ export class AgentFramework {
                 // the eventGate `inferring` leak) — every exit path must clear it.
                 if (this.activeTurnTokens.get(agent.name) === turnToken) {
                     this.activeTurnTokens.delete(agent.name);
+                    this.activeTurnTriggers.delete(agent.name);
                 }
                 this.settleAgent(agent.name, {
                     stopReason: 'exhausted',
@@ -5120,15 +5549,26 @@ export class AgentFramework {
             return false;
         }
     }
-    async driveStream(agent, stream, requestSnapshot, trigger, attempt = 0, compiledRequest, ownsProviderGate = false) {
+    async driveStream(agent, stream, requestSnapshot, trigger, attempt = 0, compiledRequest, ownsProviderGate = false, takeKvSubmission, drainKvSubmissionIds, knownToolNames = new Set()) {
         const startTime = Date.now();
         const requestId = `${agent.name}-${startTime}-${Math.random().toString(36).slice(2, 8)}`;
         const myStreamId = agent.streamId;
+        // Membrane usage events are cumulative across the native/XML tool loop.
+        // Keep the previous cumulative sample so consumers that operate at the
+        // physical provider-call boundary (estimator calibration and kv receipt
+        // acceptance) see this call only, not an ever-growing turn total.
+        let previousUsage = {
+            inputTokens: 0,
+            outputTokens: 0,
+            cacheCreationTokens: 0,
+            cacheReadTokens: 0,
+        };
         // This turn's alive-marker, set at startAgentStream entry. Safe to read
         // from the map here: the turn-alive busy check in processInferenceRequests
         // means no successor turn can have replaced it while we compiled.
         const myTurnToken = this.activeTurnTokens.get(agent.name);
         let hadToolCalls = false;
+        let generationLost = false;
         // ---- Present-while-acting turn state ---------------------------------
         // Output locus for the WHOLE logical turn: frozen in startAgentStream
         // (home → addressed trigger → global default) before this stream began,
@@ -5203,13 +5643,20 @@ export class AgentFramework {
         // failed, a framework-cancelled stream sets aborted. Best-effort
         // notifications — consumers dedupe by inferenceId and keep a timeout.
         let lifecyclePhase = 'completed';
+        let preserveEventGateForSuccessor = false;
         this.hookOrchestrator?.emitLifecycle({
             inferenceId: outgoingInferenceId,
             conversationId: agent.name,
             turnIndex: 0,
             phase: 'started',
         });
-        const proseStream = this.channelRegistry && agent.proseRouting !== 'disabled'
+        // The wrapper guard needs the complete response before it can distinguish
+        // an exact invocation-shaped body from ordinary XML/prose. Buffer guarded
+        // turns until completion: otherwise speculative outgoing chunks could
+        // expose the wrapper before the fail-closed classifier runs.
+        const proseStream = this.channelRegistry &&
+            agent.proseRouting !== 'disabled' &&
+            !agent.toolWrapperProseGuard
             ? new ProseStreamRouter({
                 mode: agent.proseRouting === 'explicit' ? 'explicit' : agent.proseRouting === 'hybrid' ? 'hybrid' : 'locus',
                 initialTarget: typingChannel,
@@ -5235,6 +5682,14 @@ export class AgentFramework {
         };
         try {
             for await (const event of stream) {
+                // Ignore every late event from an Agent generation that no longer owns
+                // this name (ephemeral watchdog/disposal, conversation replacement).
+                if (this.agents.get(agent.name) !== agent || agent.streamId !== myStreamId) {
+                    generationLost = true;
+                    lifecyclePhase = 'aborted';
+                    stream.cancel();
+                    break;
+                }
                 this.touchEphemeralRun(agent.name, true);
                 switch (event.type) {
                     case 'tokens':
@@ -5297,6 +5752,7 @@ export class AgentFramework {
                     case 'tool-calls': {
                         adoptInjectedRound();
                         hadToolCalls = true;
+                        this.recordLogicalTurnToolCalls(agent, myTurnToken ?? -1, event.calls.length);
                         this.recordEphemeralToolCalls(agent.name, event.calls.length);
                         this.emitTrace({
                             type: 'inference:tool_calls_yielded',
@@ -5374,8 +5830,7 @@ export class AgentFramework {
                             const roundToolNames = event.calls.map((c) => c.name);
                             const hasSameRoundPrivateThink = roundToolNames.includes('think') &&
                                 requestSnapshot.sameRoundThinkTextPolicy === 'private';
-                            if (roundToolNames.some((n) => SILENCING_TOOLS.has(bareToolName(n)) ||
-                                (agent.proseRouting === 'hybrid' && HYBRID_PUBLICATION_TOOLS.has(bareToolName(n))))) {
+                            if (roundToolNames.some(isSilencingTool)) {
                                 turnSilenced = true;
                             }
                             if (roundContent && roundContent.length > 0) {
@@ -5479,7 +5934,29 @@ export class AgentFramework {
                         const terminalContent = lastToolIdx >= 0
                             ? response.content.slice(lastToolIdx + 1)
                             : response.content;
-                        if (lastToolIdx >= 0) {
+                        // This is a whole-response boundary, not a trailing-prose
+                        // classifier. If the turn executed any genuine structured tool
+                        // call, preserve its later prose exactly as ordinary history.
+                        const logicalToolState = this.logicalTurnToolCalls.get(agent);
+                        const logicalTurnHadToolCalls = logicalToolState?.turnToken === myTurnToken && (logicalToolState?.count ?? 0) > 0;
+                        const guardedWrapperTool = agent.toolWrapperProseGuard && !logicalTurnHadToolCalls
+                            ? detectKnownToolWrapperProse(response.content, knownToolNames)
+                            : null;
+                        if (guardedWrapperTool) {
+                            // AF #133: containment only. Never execute wrapper text, never
+                            // publish it, and never retain it as the freshest assistant
+                            // self-seed. Exact raw output remains in the provider ledger.
+                            agent.getContextManager().addMessage('user', [{
+                                    type: 'text',
+                                    text: '[tool-boundary] No tool was called.',
+                                }], {
+                                system: true,
+                                kind: 'tool-wrapper-prose-contained',
+                                toolName: guardedWrapperTool,
+                            });
+                            console.error(`[tool-boundary] ${agent.name}: contained whole-response prose wrapper for registered tool ${guardedWrapperTool}; no tool called`);
+                        }
+                        else if (lastToolIdx >= 0) {
                             if (terminalContent.length > 0) {
                                 agent.addAssistantResponse(terminalContent);
                             }
@@ -5502,7 +5979,9 @@ export class AgentFramework {
                         // convention remains "text is thoughts"; on a text-only turn all
                         // text is speech.
                         const isTextBlock = (block) => block.type === 'text';
-                        const allText = response.content.filter(isTextBlock);
+                        const allText = guardedWrapperTool
+                            ? []
+                            : response.content.filter(isTextBlock);
                         const speechContent = hadToolCalls ? [] : allText;
                         const thoughts = hadToolCalls ? allText : [];
                         const du = response.details?.usage;
@@ -5514,19 +5993,30 @@ export class AgentFramework {
                                 cacheRead: du.cacheReadTokens,
                             }
                             : undefined;
+                        // Async completion work may overlap a replacement stream. Recheck
+                        // physical ownership before any terminal state/settlement side effect.
+                        if (this.agents.get(agent.name) !== agent || agent.streamId !== myStreamId) {
+                            generationLost = true;
+                            lifecyclePhase = 'aborted';
+                            stream.cancel();
+                            return;
+                        }
                         // Reset agent state before emitting inference:completed. Traces are
                         // observability-only, but external synchronous listeners should
                         // still see the terminal state at the terminal trace boundary.
                         // Speech dispatch happens after but doesn't depend on the status
                         // field.
                         agent.reset();
-                        this.eventGate?.onInferenceEnded(agent.name);
+                        if (this.agents.get(agent.name) === agent && agent.streamId === myStreamId)
+                            this.eventGate?.onInferenceEnded(agent.name);
                         this.settleAgent(agent.name, {
                             stopReason: 'completed',
-                            speech: terminalContent
-                                .filter((block) => block.type === 'text')
-                                .map((block) => block.text)
-                                .join('\n'),
+                            speech: guardedWrapperTool
+                                ? ''
+                                : terminalContent
+                                    .filter((block) => block.type === 'text')
+                                    .map((block) => block.text)
+                                    .join('\n'),
                         });
                         this.emitTrace({
                             type: 'inference:completed',
@@ -5745,8 +6235,7 @@ export class AgentFramework {
                                 .filter((n) => typeof n === 'string');
                             const silenced = liveProseRouting
                                 ? turnSilenced
-                                : turnSilenced || toolNames.some((n) => SILENCING_TOOLS.has(bareToolName(n)) ||
-                                    (agent.proseRouting === 'hybrid' && HYBRID_PUBLICATION_TOOLS.has(bareToolName(n))));
+                                : turnSilenced || toolNames.some(isSilencingTool);
                             const segments = splitProseSegments(liveProseRouting ? terminalContent : response.content);
                             // Preserve in-channel ordering: everything enqueued live must
                             // land before the trailing prose. Awaited even when silenced —
@@ -5868,7 +6357,8 @@ export class AgentFramework {
                         agent.reset();
                         if (ownsProviderGate && this.holdProviderAcceleration(agent, err, trigger)) {
                             lifecyclePhase = 'failed';
-                            this.eventGate?.onInferenceEnded(agent.name);
+                            if (this.agents.get(agent.name) === agent && agent.streamId === myStreamId)
+                                this.eventGate?.onInferenceEnded(agent.name);
                             break;
                         }
                         const action = this.errorPolicy.onInferenceError(err, agent.name, attempt);
@@ -5892,7 +6382,8 @@ export class AgentFramework {
                                 // itself) means retrying the same context can never succeed.
                                 ...this.classifyInferenceError(err),
                             });
-                            this.eventGate?.onInferenceEnded(agent.name);
+                            if (this.agents.get(agent.name) === agent && agent.streamId === myStreamId)
+                                this.eventGate?.onInferenceEnded(agent.name);
                             if (action.emit) {
                                 this.pushEvent(action.emit);
                             }
@@ -5913,7 +6404,14 @@ export class AgentFramework {
                             if (cancelKind !== undefined) {
                                 this.frameworkCancelledStreams.delete(cancelKey);
                                 lifecyclePhase = 'aborted'; // §10.5 — terminal emitted in finally
-                                this.eventGate?.onInferenceEnded(agent.name);
+                                // A budget restart continues the same EventGate window. Do not
+                                // end it here or in finally; the replacement owns final release.
+                                if (cancelKind === 'budget_restart') {
+                                    preserveEventGateForSuccessor = true;
+                                }
+                                else if (this.agents.get(agent.name) === agent && agent.streamId === myStreamId) {
+                                    this.eventGate?.onInferenceEnded(agent.name);
+                                }
                                 // endTurn IS a logical turn end — earlier rounds may have
                                 // live-routed prose (narrate → skip_reply is a real shape),
                                 // so settle the delivery chain and drop the receipt. A
@@ -5960,7 +6458,8 @@ export class AgentFramework {
                                 request: compiledRequest ?? { note: 'streaming request aborted' },
                                 durationMs,
                             });
-                            this.eventGate?.onInferenceEnded(agent.name);
+                            if (this.agents.get(agent.name) === agent && agent.streamId === myStreamId)
+                                this.eventGate?.onInferenceEnded(agent.name);
                         }
                         break;
                     }
@@ -5980,14 +6479,46 @@ export class AgentFramework {
                         // loop (5 calls x ~160k reported as 884k), which is not a
                         // window-shaped number and drove the multiplier to 2.37 before
                         // the guards caught it.
+                        const cumulativeUsage = {
+                            inputTokens: event.usage.inputTokens ?? 0,
+                            outputTokens: event.usage.outputTokens ?? 0,
+                            cacheCreationTokens: event.usage.cacheCreationTokens ?? 0,
+                            cacheReadTokens: event.usage.cacheReadTokens ?? 0,
+                        };
+                        const perCallUsage = {
+                            inputTokens: cumulativeDelta(cumulativeUsage.inputTokens, previousUsage.inputTokens),
+                            outputTokens: cumulativeDelta(cumulativeUsage.outputTokens, previousUsage.outputTokens),
+                            cacheCreationTokens: cumulativeDelta(cumulativeUsage.cacheCreationTokens, previousUsage.cacheCreationTokens),
+                            cacheReadTokens: cumulativeDelta(cumulativeUsage.cacheReadTokens, previousUsage.cacheReadTokens),
+                        };
+                        previousUsage = cumulativeUsage;
+                        const strat = agent.getContextManager?.()?.getStrategy?.();
                         try {
-                            const realTotal = (event.usage.inputTokens ?? 0) +
-                                (event.usage.cacheCreationTokens ?? 0) +
-                                (event.usage.cacheReadTokens ?? 0);
-                            const strat = agent.getContextManager?.()?.getStrategy?.();
+                            const realTotal = perCallUsage.inputTokens +
+                                perCallUsage.cacheCreationTokens +
+                                perCallUsage.cacheReadTokens;
                             strat?.reportRealInputTokens?.(realTotal);
                         }
                         catch { /* calibration is best-effort */ }
+                        const kvSubmission = takeKvSubmission?.();
+                        if (kvSubmission) {
+                            try {
+                                strat?.reportKvUnifiedAccepted?.({
+                                    submissionId: kvSubmission.submissionId,
+                                    acceptedAt: Date.now(),
+                                    wireReceipt: kvSubmission.wireReceipt,
+                                });
+                            }
+                            catch {
+                                // The queue item has already been consumed. Explicitly close
+                                // the receipt flight so one persistence error cannot wedge
+                                // every later provider call on this branch.
+                                try {
+                                    strat?.reportKvUnifiedFailed?.(kvSubmission.submissionId);
+                                }
+                                catch { /* teardown only */ }
+                            }
+                        }
                         this.emitTrace({
                             type: 'inference:usage',
                             agentName: agent.name,
@@ -6004,6 +6535,14 @@ export class AgentFramework {
             }
         }
         catch (error) {
+            // A superseded physical stream may throw after its replacement starts.
+            // It has no authority to settle, reset, gate-release, or publish failure.
+            if (this.agents.get(agent.name) !== agent || agent.streamId !== myStreamId) {
+                generationLost = true;
+                lifecyclePhase = 'aborted';
+                stream.cancel();
+                return;
+            }
             // Stream itself threw. Organization acceleration is deferred from a
             // fresh compile; every other throw keeps the ordinary exhausted path.
             const err = error instanceof Error ? error : new Error(String(error));
@@ -6016,7 +6555,8 @@ export class AgentFramework {
                 this.abortAgentScript(agent.name, 'provider acceleration cooldown');
                 lifecyclePhase = 'failed';
                 agent.reset();
-                this.eventGate?.onInferenceEnded(agent.name);
+                if (this.agents.get(agent.name) === agent && agent.streamId === myStreamId)
+                    this.eventGate?.onInferenceEnded(agent.name);
                 return;
             }
             this.emitTrace({
@@ -6051,9 +6591,33 @@ export class AgentFramework {
             this.abortAgentScript(agent.name, 'stream threw');
             lifecyclePhase = 'failed'; // §10.5 — terminal emitted in finally
             agent.reset();
-            this.eventGate?.onInferenceEnded(agent.name);
+            if (this.agents.get(agent.name) === agent && agent.streamId === myStreamId)
+                this.eventGate?.onInferenceEnded(agent.name);
         }
         finally {
+            const unsettledKvSubmissions = drainKvSubmissionIds?.() ?? [];
+            if (unsettledKvSubmissions.length > 0) {
+                try {
+                    const strategy = agent.getContextManager().getStrategy();
+                    for (const submissionId of unsettledKvSubmissions) {
+                        strategy.reportKvUnifiedFailed?.(submissionId);
+                    }
+                }
+                catch { /* receipt cleanup must not mask inference teardown */ }
+            }
+            // A framework retry/budget restart can overlap physical streams on the
+            // same Agent object. Name + Agent identity is not enough: only the
+            // current stream generation may mutate name-keyed teardown state.
+            const ownsPhysicalStream = this.agents.get(agent.name) === agent && agent.streamId === myStreamId;
+            // Ephemeral completion resolves its caller before this finally runs, so
+            // runEphemeralToCompletion may already have deregistered the name. The
+            // terminal frame still owns its typing/outgoing completion unless it was
+            // genuinely superseded or handed EventGate liveness to a budget restart.
+            const disposedEphemeralWithoutSuccessor = generationLost &&
+                this.disposedEphemeralFrames.has(agent) &&
+                !this.agents.has(agent.name);
+            const frameReachedTerminal = (!generationLost || disposedEphemeralWithoutSuccessor) &&
+                !preserveEventGateForSuccessor;
             // §10.5: exactly one terminal per `started`, on every exit path the
             // host controls. Which one was decided by the path taken (default
             // completed; catch → failed; framework-cancel → aborted). A host
@@ -6072,28 +6636,35 @@ export class AgentFramework {
             // all incoming events → the agent silently stops waking on messages
             // (typing still stops, compression still runs — matching the observed
             // wedge). onInferenceEnded is idempotent, so a redundant call is safe.
-            this.eventGate?.onInferenceEnded(agent.name);
-            this.lastInferenceAt.set(agent.name, { ...this.lastInferenceAt.get(agent.name), endedAt: Date.now() });
+            if (ownsPhysicalStream && !preserveEventGateForSuccessor) {
+                this.eventGate?.onInferenceEnded(agent.name);
+            }
+            if (!generationLost && ownsPhysicalStream) {
+                this.lastInferenceAt.set(agent.name, { ...this.lastInferenceAt.get(agent.name), endedAt: Date.now() });
+            }
             if (ownsProviderGate)
                 this.releasePrimaryProviderGate(agent.name);
             // Stop the typing indicator on every exit path (complete, error,
             // exhausted, abort) so it never sticks after the turn ends.
-            this.channelRegistry?.stopTyping();
+            if (frameReachedTerminal)
+                this.channelRegistry?.stopTyping();
             // Spec 14.3: flush any held line-start text, then close each streamed
             // channel with its final moderated content — the consumer's signal to
             // finalize (end the TTS utterance, settle the rendered message).
-            if (proseStream) {
+            if (frameReachedTerminal && proseStream) {
                 emitOutgoing(proseStream.finish());
                 for (const [channelId, text] of proseStream.byChannel()) {
                     this.channelRegistry.sendOutgoingComplete(channelId, agent.name, outgoingInferenceId, text);
                 }
             }
             this.frameworkCancelledStreams.delete(`${agent.name}:${myStreamId}`);
-            this.activeStreams.delete(agent.name);
-            this.pendingAssistantBlocks.delete(agent.name);
+            if (ownsPhysicalStream) {
+                this.activeStreams.delete(agent.name);
+                this.pendingAssistantBlocks.delete(agent.name);
+            }
             // A conversation fork whose TTL closure turn just finished is done for
             // good — dispose it so the agent map doesn't grow monotonically.
-            if (this.closingConversationAgents.has(agent.name)) {
+            if (ownsPhysicalStream && this.closingConversationAgents.has(agent.name)) {
                 this.disposeConversationAgent(agent.name);
             }
             // The turn is torn down — release the turn-alive marker BEFORE the
@@ -6105,12 +6676,16 @@ export class AgentFramework {
             // injection / teardown delivers the messages at a correct position.
             if (myTurnToken !== undefined && this.activeTurnTokens.get(agent.name) === myTurnToken) {
                 this.activeTurnTokens.delete(agent.name);
-            }
-            // Flush any deferred messages (e.g. if stream failed while tools were pending)
-            if (this.deferredMessages.length > 0 && this.pendingAssistantBlocks.size === 0) {
-                const deferred = this.deferredMessages.splice(0);
+                this.activeTurnTriggers.delete(agent.name);
+            }
+            // Flush any deferred messages (e.g. if stream failed while tools were
+            // pending). Only THIS agent's messages: other targets' entries wait
+            // for their own boundaries (re-adding via addMessage re-defers if the
+            // target has meanwhile started a turn).
+            if (frameReachedTerminal && this.deferredMessages.length > 0 && this.pendingAssistantBlocks.size === 0) {
+                const deferred = this.drainDeferredFor(agent.name);
                 for (const msg of deferred) {
-                    this.addMessage(msg.participant, msg.content, msg.metadata);
+                    this.addMessage(msg.participant, msg.content, msg.metadata, msg.forAgent ? { forAgent: msg.forAgent } : undefined);
                 }
             }
         }
@@ -6760,6 +7335,23 @@ export class AgentFramework {
         }
         if (result.data === undefined)
             return '';
+        // RFC-005 "eagerly lazy": a script receiving a result is the strongest
+        // available signal that the bytes are wanted (scripts process, they do
+        // not browse) — materialize every referenced payload now, so the stub
+        // the script receives carries a workspace path it can open directly.
+        if (Array.isArray(result.data)) {
+            const fetcher = this.getReferenceFetcher();
+            for (const block of result.data) {
+                const c = classifyBlock(block);
+                if (c.kind !== 'reference')
+                    continue;
+                const record = referenceRegistry.findByUri(c.testimony.uri)
+                    ?? referenceRegistry.register(c.testimony);
+                const outcome = await fetcher.fetch(record);
+                if (!outcome.ok)
+                    console.error(`[rfc005] script-demand fetch of ${record.refId} failed: ${outcome.error}`);
+            }
+        }
         // Scripts get (nearly) full data — the script environment IS the spill:
         // filtering big results in code is the whole point. 5MB protocol safety
         // cap only; no context-cap truncation, no file spill.
@@ -6864,6 +7456,11 @@ export class AgentFramework {
             if (!raw || typeof raw !== 'object')
                 return null;
             const b = raw;
+            if (isInlineContradiction(b)) {
+                // RFC-005 vector 2: withhold inline data claiming bulk disposition.
+                blocks.push({ type: 'text', text: INLINE_WITHHELD_TEXT });
+                continue;
+            }
             if (b.type === 'text' && typeof b.text === 'string') {
                 let text = b.text;
                 if (maxChars && text.length > maxChars) {
@@ -6881,6 +7478,15 @@ export class AgentFramework {
                     source: { type: 'base64', data: b.data, mediaType: b.mimeType },
                 });
             }
+            else if (b.type === 'resource' || ((b.type === 'image' || b.type === 'audio') && typeof b.uri === 'string')) {
+                // RFC-005 reference block → bounded stub. Previously this bailed the
+                // whole array to JSON, which both inlined the reference verbatim and
+                // demoted any adjacent image block to stringified base64.
+                const stub = referenceStubOrNull(b, 'from tool result');
+                if (stub === null)
+                    return null;
+                blocks.push({ type: 'text', text: stub });
+            }
             else {
                 return null; // unknown shape — bail to JSON path
             }
@@ -7048,11 +7654,36 @@ export class AgentFramework {
             this.dispatchReadImageToolCall(agentName, enrichedCall);
             return;
         }
+        if (enrichedCall.name === 'fetch_reference') {
+            this.dispatchFetchReferenceToolCall(agentName, enrichedCall);
+            return;
+        }
         // Route gate_status tool
         if (enrichedCall.name === 'gate_status' && this.eventGate) {
             this.dispatchGateToolCall(agentName, enrichedCall);
             return;
         }
+        // Route the tune-out tool (residents) and the subconscious's surface
+        if (enrichedCall.name === 'tune_out' && this.tuneOutCoordinator) {
+            this.dispatchTuneOutToolCall(agentName, enrichedCall);
+            return;
+        }
+        if (this.tuneOutCoordinator &&
+            agentName === this.subconsciousAgentName &&
+            SUBCONSCIOUS_TOOL_NAMES.includes(enrichedCall.name)) {
+            void this.tuneOutCoordinator
+                .handleSubconsciousTool(enrichedCall.name, enrichedCall.input)
+                .then((result) => {
+                this.queue.push({
+                    type: 'tool-result',
+                    callId: enrichedCall.id,
+                    agentName,
+                    moduleName: 'tune-out',
+                    result,
+                });
+            });
+            return;
+        }
         // Route sleep / wake tools
         if ((enrichedCall.name === 'sleep' || enrichedCall.name === 'wake') && this.eventGate) {
             this.dispatchSleepToolCall(agentName, enrichedCall);
@@ -7181,12 +7812,19 @@ export class AgentFramework {
             });
         });
     }
-    addMessage(participant, content, metadata) {
-        // Route to the primary agent's context manager (not ephemeral subagents).
-        const agent = this.primaryAgentName
-            ? this.agents.get(this.primaryAgentName)
-            : this.agents.values().next().value;
+    addMessage(participant, content, metadata, opts) {
+        // Route to the named agent, else the primary (not ephemeral subagents).
+        const agent = opts?.forAgent
+            ? this.agents.get(opts.forAgent)
+            : this.primaryAgentName
+                ? this.agents.get(this.primaryAgentName)
+                : this.agents.values().next().value;
         if (!agent) {
+            if (opts?.forAgent) {
+                console.error(`[addMessage] delivery target "${opts.forAgent}" is not a registered agent — message dropped ` +
+                    `(participant=${participant})`);
+                return '';
+            }
             throw new Error('No agents configured');
         }
         // Defer non-tool_result messages while a tool cycle is mid-flight
@@ -7206,15 +7844,39 @@ export class AgentFramework {
         // ALSO injected into the live stream — hear-while-acting), or in
         // driveStream's finally when the turn ends.
         const hasToolResult = content.some(b => b.type === 'tool_result');
+        // Explicitly-targeted deliveries scope the tool-cycle check to the
+        // target (pendingAssistantBlocks is keyed by agent); the default path
+        // keeps the historical GLOBAL check byte-for-byte (conservative: any
+        // agent's pending cycle defers primary-bound messages).
+        const midToolCycle = opts?.forAgent
+            ? this.pendingAssistantBlocks.has(agent.name)
+            : this.pendingAssistantBlocks.size > 0;
         if (!hasToolResult &&
-            (this.pendingAssistantBlocks.size > 0 ||
+            (midToolCycle ||
                 this.activeTurnTokens.has(agent.name) ||
                 this.activeStreams.has(agent.name))) {
-            this.deferredMessages.push({ participant, content, metadata });
-            return ''; // Deferred — flushed at the next boundary
+            this.deferredMessages.push({ participant, content, metadata, forAgent: opts?.forAgent });
+            return ''; // Deferred — flushed at the target's next boundary
         }
         return agent.getContextManager().addMessage(participant, content, metadata);
     }
+    /**
+     * Splice out the deferred messages addressed to `agentName` (explicitly
+     * via forAgent, or implicitly when it is the primary), leaving other
+     * targets' messages queued for their own boundaries.
+     */
+    drainDeferredFor(agentName) {
+        if (this.deferredMessages.length === 0)
+            return [];
+        const mine = [];
+        const rest = [];
+        for (const msg of this.deferredMessages) {
+            const target = msg.forAgent ?? this.primaryAgentName;
+            (target === agentName ? mine : rest).push(msg);
+        }
+        this.deferredMessages = rest;
+        return mine;
+    }
     editMessage(id, content) {
         const agent = this.primaryAgentName
             ? this.agents.get(this.primaryAgentName)
@@ -8856,6 +9518,13 @@ export class AgentFramework {
             }
         }
         server.sendToolsCall(toolName, args, stateParams)
+            .then(async (result) => {
+            // RFC-005: register references and eagerly fetch the small ones
+            // before serialization, so their stubs carry a saved path.
+            await this.autoFetchReferences(result.content, serverId, true)
+                .catch((e) => console.error('[rfc005] autofetch error:', e.message));
+            return result;
+        })
             .then((result) => {
             const durationMs = Date.now() - startTime;
             this.emitTrace({ type: 'tool:completed', module: `mcpl:${serverId}`, tool: toolName, callId: call.id, durationMs });
@@ -9191,6 +9860,50 @@ export class AgentFramework {
      * suppression window, optionally announces in the sticky channel, and ends
      * the turn (the agent goes idle immediately). `wake` clears sleep.
      */
+    /** tune_out tool (residents): enter or cancel a channel tune-out. */
+    dispatchTuneOutToolCall(agentName, call) {
+        const coordinator = this.tuneOutCoordinator;
+        const input = (call.input ?? {});
+        const channelId = String(input.channelId ?? '');
+        const entry = this.channelRegistry?.listChannelsRaw()
+            .find((e) => e.descriptor.id === channelId);
+        let result;
+        if (!entry) {
+            result = { success: false, error: `unknown channel: ${channelId}`, isError: false };
+        }
+        else if (input.mode === 'cancel') {
+            const r = coordinator.cancel(entry.serverId, channelId, 'agent-tool', 'agent request');
+            result = r.ok
+                ? { success: true, data: { cancelled: channelId } }
+                : { success: false, error: r.error, isError: false };
+        }
+        else {
+            const r = coordinator.enter(entry.serverId, channelId, {
+                cadenceSeconds: input.cadenceSeconds,
+                backlogCap: input.backlogCap,
+                maxWakes: input.maxWakes,
+                durationSeconds: input.durationSeconds,
+            }, 'agent-tool');
+            result = r.ok
+                ? {
+                    success: true,
+                    data: {
+                        tunedOut: channelId,
+                        cadenceSeconds: r.params.cadenceSeconds,
+                        backlogCap: r.params.backlogCap,
+                        maxWakes: r.params.maxWakes,
+                    },
+                }
+                : { success: false, error: r.error, isError: false };
+        }
+        this.queue.push({
+            type: 'tool-result',
+            callId: call.id,
+            agentName,
+            moduleName: 'tune-out',
+            result,
+        });
+    }
     dispatchSleepToolCall(agentName, call) {
         this.emitTrace({ type: 'tool:started', module: 'gate', tool: call.name, callId: call.id, input: call.input });
         const gate = this.eventGate;
@@ -9264,6 +9977,43 @@ export class AgentFramework {
      * blocks) delivers the actual image to the model; history storage keeps
      * the standard compact `[image: …]` placeholder.
      */
+    /** RFC-005 on-demand dereference: [ref id] → workspace file, via the
+     *  origin-bound, ceiling-bounded, digest-verified host fetcher. */
+    dispatchFetchReferenceToolCall(agentName, call) {
+        this.emitTrace({ type: 'tool:started', module: 'workspace', tool: call.name, callId: call.id, input: call.input });
+        const startTime = Date.now();
+        void (async () => {
+            const args = (call.input ?? {});
+            const refId = String(args.ref_id ?? '');
+            const record = referenceRegistry.get(refId);
+            let result;
+            if (!record) {
+                // Vector 20: a stale or unknown id is a defined miss, never a
+                // different record.
+                result = { success: false, isError: true,
+                    error: `unknown reference: ${refId || '(missing ref_id)'} — it may have been evicted; ids are never reused` };
+            }
+            else {
+                const maxBytes = typeof args.max_bytes === 'number'
+                    ? Math.min(args.max_bytes, DEFAULT_FETCH_MAX_BYTES) : undefined;
+                const outcome = await this.getReferenceFetcher().fetch(record, { maxBytes });
+                result = outcome.ok
+                    ? { success: true, data: JSON.stringify({ path: outcome.path, bytes: outcome.bytes,
+                            mime_type: outcome.mimeType, digest_verified: outcome.digestVerified ?? 'no digest claimed' }) }
+                    : { success: false, isError: true, error: outcome.error };
+            }
+            this.emitTrace({
+                type: result.isError ? 'tool:failed' : 'tool:completed',
+                module: 'workspace', tool: call.name, callId: call.id,
+                durationMs: Date.now() - startTime,
+                ...(result.isError ? { error: result.error } : {}),
+            });
+            this.pushEvent({
+                type: 'tool-result', callId: call.id, agentName, moduleName: 'workspace',
+                result: { ...result, isError: result.isError ?? false },
+            });
+        })();
+    }
     dispatchReadImageToolCall(agentName, call) {
         this.emitTrace({ type: 'tool:started', module: 'workspace', tool: call.name, callId: call.id, input: call.input });
         const finish = (result) => {