@maudecode/talaria-web 1.17.2 → 1.19.0

package/dist/sessions/turn.js

@@ -6,37 +6,43 @@
  * `api/streaming.py::_run_agent_streaming`, `cancel_stream`).
  */
 import { randomUUID } from 'node:crypto';
-import { closeSync, fstatSync, readFileSync, rmSync } from 'node:fs';
+import { closeSync, fstatSync, readFileSync } from 'node:fs';
 import { openAnchoredFd } from '../workspace/fs.js';
 import { isWithin, resolvePathLikePython } from '../workspace/paths.js';
-import { SidecarError } from '../sidecar/client.js';
+import { runtimeFailureDetails, SidecarError } from '../sidecar/client.js';
 import { HttpFailure, markSessionTitleGenerated } from './service.js';
 import { StreamRegistry, SessionChannels } from './streams.js';
-import { ChatUsageSchema } from '@maudecode/talaria-web-contracts';
-import { PendingPrompts, clarifyReply } from './pending.js';
-import { RunJournal } from './journal.js';
-import { CONTEXT_USAGE_FIELDS, Session, contextUsage, titleFrom } from './session.js';
+import { ChatUsageSchema, ConnectorOperationSnapshotSchema } from '@maudecode/talaria-web-contracts';
+import { catalogNames, connectorOperation } from '../tools/connectors.js';
+import { PendingPrompts, clarifyReply, isVaultPrompt, vaultReply } from './pending.js';
+import { COMPRESSION_RECOVERY_ACTION_START_FOCUSED, COMPRESSION_RECOVERY_TERMINAL_STATE, CONTEXT_USAGE_FIELDS, Session, contextUsage, forgetContextFigures, titleFrom } from './session.js';
 import { anchorMessageKey, compressionReference, markerSummary, visibleMessagesForAnchor } from './compress.js';
 import { buildActiveTurnToken, completedToolIndex, publicToolFrame, redactNestedMessageContainers, redactSessionData, redactString, withToolId } from '../redact.js';
-import { dict } from '../config/agent-config.js';
+import { dict, findPersonality } from '../config/agent-config.js';
 import { ReasoningTitleTracker, reasoningEventPayload } from './reasoning-titles.js';
 import { messageWindowForDisplay, messagesForLimitedPayload, toolCallsForMessageWindow } from './window.js';
 import { UNSETTLED_TODO_KEY, attachTodoState, parseTodoToolResult } from './todo.js';
-import { withSessionWireFlags } from './list.js';
+import { SessionNotices } from './notices.js';
+import { SessionGoals } from './goals.js';
 import { hydrateAnchorActivityScenes, turnTerminalState, withTurnIds } from './anchor.js';
 import { persistentStateChanges, persistentStateSnapshot } from './state-saved.js';
-import { maxIterationsFromConfig, maxTokensFromConfig, processWakeupMaxIterations, reasoningConfigFromConfig, webuiEphemeralSystemPrompt, workspaceSystemMessage } from './turn-context.js';
-import { agentSteerText, assistantReplyAddedAfterCurrentTurn, attachedFilesPrompt, buildPartialMessage, dedupeContext, checkpointTurnStart, extractToolCallsFromMessages, hasNativeImages, injectMaxIterationSummaryFallback, isDict, mergeDisplayMessagesAfterAgentResult, messageIdentity, messageText, pendingUserRow, sanitizeMessagesForApi, sessionLacksFinalAssistantAnswer, splitDisplayText, stateDbSeenId, joinReasoning, reasoningFieldsText, stoppedTurnContext, toolOutcome, withAttachmentObjects, withBodyExcerpts, withDisplayMedia, withMarkerKinds, withToolCallOutcomes, workspaceContextPrefix, withoutMaxIterationSummaryRequest, withNativeImagesRestored, withoutNativeImages, withoutToolImages } from './merge.js';
-import { mayBecomeSilentReply, turnOrigin, withBackgroundUpdates } from './background-updates.js';
-import { fallbackTitleFromExchange, firstExchangeSnippets, isGenericFallbackTitle, latestExchangeSnippets, looksInvalidGeneratedTitle, looksLikeDefaultCliTitle, sanitizeGeneratedTitle, titleLanguageMismatch, titlePrompts } from './titles.js';
+import { agentBuildParams, webuiEphemeralSystemPrompt, workspaceSystemMessage } from './turn-context.js';
+import { agentSteerText, assistantReplyAddedAfterCurrentTurn, attachedFilesPrompt, buildPartialMessage, dedupeContext, checkpointTurnStart, extractToolCallsFromMessages, hasNativeImages, injectMaxIterationSummaryFallback, isDict, mergeDisplayMessagesAfterAgentResult, messageIdentity, messageText, pendingUserRow, sanitizeMessagesForApi, sessionLacksFinalAssistantAnswer, splitDisplayText, stateDbSeenId, joinReasoning, reasoningFieldsText, toolOutcome, withAttachmentObjects, newestSettledReply, withBodyExcerpts, withDisplayMedia, withMarkerKinds, withToolCallOutcomes, withoutMaxIterationSummaryRequest, withNativeImagesRestored, withoutNativeImages, withoutToolImages } from './merge.js';
+import { mayBecomeSilentReply, turnOrigin, recordBackgroundReview, recordBackgroundUpdate } from './background-updates.js';
+import { countExchanges, fallbackTitleFromExchange, firstExchangeSnippets, isGenericFallbackTitle, latestExchangeSnippets, looksInvalidGeneratedTitle, looksLikeDefaultCliTitle, sanitizeGeneratedTitle, titleLanguageMismatch, titlePrompts } from './titles.js';
 import { WORKSPACE_BUSY_MESSAGE } from '../workspace/git.js';
+import { CHECKPOINT_STORE_BUSY_MESSAGE } from '../workspace/rollback.js';
 import { str } from '../util.js';
 import { toolEditDiff } from './tool-display.js';
+import { recordWakeupPause } from './wakeup-pause.js';
+import { displayName } from '../providers/catalog.js';
 export const CHAT_LOCK_WAIT_SECONDS = 2;
 const IMAGE_MODE_TIMEOUT_MS = 15_000;
 const TERMINAL_SSE_VISIBLE_MESSAGE_LIMIT = 80;
 /** TAL-397: streamed text reports the live meter at most this often (seconds); a counter change reports at once. */
 const METERING_INTERVAL_S = 1;
+/** The Agent's content frames, which end a rate-limit wait (TAL-821) and a compression span (TAL-1006). A terminal frame ends the run, its wait, and its span. */
+const CONTENT_EVENTS = new Set(['token', 'reasoning', 'interim_assistant', 'tool', 'tool_complete']);
 /** How long a user's message waits for a stopped background turn to unwind before its own turn is admitted. */
 const BACKGROUND_UNWIND_WAIT_MS = 30_000;
 const COMPRESSION_EXHAUSTED = { label: 'Context compression exhausted', type: 'compression_exhausted', hint: 'The conversation context is too large to compress safely. Start a new conversation or retry with a narrower task.' };
@@ -63,6 +69,9 @@ export function classifyProviderError(errStr, opts = {}) {
     const lower = errStr.toLowerCase();
     if (opts.compressionExhausted)
         return COMPRESSION_EXHAUSTED;
+    // TAL-798: the Agent's local llama.cpp provider has no key to add; its server is off or still starting.
+    if (opts.condition === 'local_runtime_unavailable')
+        return { label: 'Local model server unavailable', type: 'local_runtime_unavailable', hint: 'The local model server is off or still starting. Turn it on in Hermes Agent, or pick another model.' };
     if (opts.condition === 'credential_missing')
         return { label: 'Authentication failed', type: 'auth_mismatch', hint: 'The selected model may not be supported by your configured provider or your API key is invalid. Run `hermes model` in your terminal to update credentials, then restart the WebUI.' };
     const cancelled = ['cancelled by user', 'canceled by user', 'user cancelled', 'user canceled', 'task cancelled', 'task canceled', 'cancellederror'].some((k) => lower.includes(k));
@@ -71,7 +80,7 @@ export function classifyProviderError(errStr, opts = {}) {
     if (['interrupted by user', 'response interrupted', 'operation interrupted', 'operation was interrupted', 'operation aborted', 'request was aborted', 'aborterror'].some((k) => lower.includes(k))) {
         return { label: 'Response interrupted', type: 'interrupted', hint: 'The run stopped before a provider response completed. If you did not cancel it, try again.' };
     }
-    const quota = ['insufficient_quota', 'insufficient credits', 'out of credits', 'credits exhausted', 'quota exceeded', 'usage limit', 'plan limit', 'billing'].some((k) => lower.includes(k));
+    const quota = ['insufficient_quota', 'insufficient credits', 'out of credits', 'credits exhausted', 'quota exceeded', 'usage limit', 'plan limit', 'billing', 'more credits', 'can only afford', 'fewer max_tokens', 'credit balance', 'exceeded your current quota', 'usage_limit_exceeded', 'reached the limit of messages', 'used up your usage'].some((k) => lower.includes(k));
     const poolEmpty = lower.includes('credential(s) exhausted') || lower.includes('credentials exhausted') || (lower.includes('credential') && lower.includes('exhausted'));
     const auth = !quota && !poolEmpty && (errStr.includes('401') || lower.includes('authentication') || lower.includes('unauthorized') || lower.includes('invalid api key') || lower.includes('invalid_api_key') || lower.includes('no cookie auth credentials'));
     const notFound = errStr.includes('404') || lower.includes('not found') || lower.includes('does not exist') || lower.includes('model_not_found') || lower.includes('invalid model') || lower.includes('does not match any known model') || lower.includes('unknown model');
@@ -125,8 +134,6 @@ export function sniffImageMime(bytes, declared = '') {
         return 'image/bmp';
     return null;
 }
-const COMPRESSION_RECOVERY_TERMINAL_STATE = 'compression_exhausted';
-const COMPRESSION_RECOVERY_ACTION_START_FOCUSED = 'start_focused_continuation';
 /** Python `build_compression_recovery_payload` + `stamp_compression_exhausted_recovery`. */
 export function stampCompressionExhaustedRecovery(s, message, details) {
     const positive = (v) => { const n = Math.trunc(Number(v) || 0); return n > 0 ? n : 0; };
@@ -209,8 +216,14 @@ export class TurnRunner {
     liveUsage = new Map();
     /** TAL-397: each running turn's latest todo list, which an error or cancel keeps on the session. */
     liveTodos = new Map();
+    /** The current presentation, scoped to its owning stream and released on every terminal path. */
+    runStatuses = new Map();
+    /** TAL-1006: each live stream's open compression span, by its start (epoch seconds). */
+    compressionSpans = new Map();
     /** Streams whose run completed (`done` emitted) and only await title work; a late cancel is a no-op for these. */
     settledStreams = new Set();
+    /** The last adaptive title refresh per session; a successor turn's refresh runs after it (TAL-589). */
+    titleRefreshes = new Map();
     /**
      * TAL-364: what a Stop needs to write the turn's model context: its starting context, the prompt the Agent got, and
      * whether a settlement already used a canonical checkpoint (which marks the stop boundary; later work never replaces it).
@@ -220,25 +233,134 @@ export class TurnRunner {
     stopContexts = new Map();
     /** Each turn's worker, so a stopped background turn can be awaited before the user's turn takes the session. */
     workers = new Map();
+    /** TAL-899: sessions whose user pressed Stop; no automatic turn starts in them until the user's next turn is admitted. */
+    stopHeld = new Set();
+    notices = new SessionNotices(() => this.deps.now());
+    goals = new SessionGoals(() => this.deps.now());
     constructor(deps) {
         this.deps = deps;
     }
     get registry() { return this.deps.registry; }
     // ── admission ────────────────────────────────────────────────────────────
+    /** TAL-899: whether a Stop holds the session's automatic turns (session `automatic_turns_held`). */
+    async automaticTurnsHeld(sessionId) {
+        if (this.stopHeld.has(sessionId))
+            return true;
+        try {
+            return (await this.deps.store.get(sessionId, { metadataOnly: true })).automatic_turns_stop_held;
+        }
+        catch {
+            return false;
+        }
+    }
+    /** TAL-803: the Agent's active notices for the session (session `notices`). */
+    sessionNotices(sessionId) {
+        const redact = this.deps.redactEnabled();
+        return this.notices.list(sessionId).map((n) => ({ ...n, text: redactString(n.text, redact) }));
+    }
+    /** TAL-808: the session's standing goal (session `goal`). */
+    sessionGoal(sessionId) {
+        const goal = this.goals.view(sessionId);
+        if (!goal)
+            return null;
+        const redact = this.deps.redactEnabled();
+        return { ...goal, text: redactString(goal.text, redact), reason: goal.reason === null ? null : redactString(goal.reason, redact) };
+    }
+    /**
+     * TAL-808: reads the session's goal from the Agent and, when its view changed, sends `session_goal` on `put` (else the
+     * session's live stream) and `sessions_changed`, which reloads the session for a client between turns.
+     */
+    async refreshGoal(s, put = async (event, data) => { (await this.emitToSession(s.session_id, event, data)); }) {
+        if (!(await this.readGoal(s)))
+            return;
+        await put('session_goal', { session_id: s.session_id, goal: this.sessionGoal(s.session_id) });
+        this.deps.events.publish('goal', { profile: s.profile, sessionId: s.session_id });
+    }
+    /** TAL-808: a session's first detail after a restart reads its goal before answering, so it tells no one else. */
+    async loadGoal(sessionId) {
+        if (this.goals.has(sessionId))
+            return;
+        let s;
+        try {
+            s = (await this.deps.store.get(sessionId, { metadataOnly: true }));
+        }
+        catch {
+            return;
+        }
+        await this.readGoal(s);
+    }
+    /** The Agent's `goals.get` into the cache, and a re-read when a timed park ends; whether the view changed. A failed read keeps the last one. */
+    async readGoal(s) {
+        const sidecar = this.deps.sidecar();
+        if (!sidecar)
+            return false;
+        let status;
+        try {
+            // Short: a session's first load waits on it.
+            status = await sidecar.call('goals.get', { session_id: s.session_id, profile_home: this.deps.profileHome(s.profile) }, { timeoutMs: 3_000 });
+        }
+        catch (error) {
+            this.deps.log(`[webui] WARNING: goal read failed for ${s.session_id}: ${error.message}`);
+            return false;
+        }
+        const changed = this.goals.set(s.session_id, status);
+        this.goals.rereadAfterPark(s.session_id, () => { void this.refreshGoal({ session_id: s.session_id, profile: s.profile }); });
+        return changed;
+    }
+    /** Hygiene sweep: a deleted session's hold, notices and goal go with it (TAL-899, TAL-803, TAL-808). */
+    async sweep() {
+        const exists = async (sid) => (await this.deps.store.exists(sid));
+        for (const sid of this.stopHeld)
+            if (!(await exists(sid)))
+                this.stopHeld.delete(sid);
+        await this.notices.sweep(exists);
+        await this.goals.sweep(exists);
+    }
+    /**
+     * Admit automatic turns through the same workspace recovery and model resolution as user turns. After the user's Stop
+     * none starts (TAL-899): the answer is a `held` 409, and the caller keeps its work for the user's next turn.
+     */
+    async startServerTurn(session, opts) {
+        if ((await this.automaticTurnsHeld(session.session_id)))
+            return { _status: 409, error: 'automatic turns are held after Stop', held: true };
+        try {
+            const service = this.deps.service();
+            const workspace = (await service.turnWorkspace(session));
+            const [model, modelProvider, normalizedModel] = service.turnModel(session);
+            return (await this.admit(session, { ...opts, workspace, model, modelProvider, normalizedModel }));
+        }
+        catch (error) {
+            if (error instanceof HttpFailure)
+                return { _status: error.status, error: error.message };
+            throw error;
+        }
+    }
+    /** A user's turn; admitting it releases the hold a Stop put on automatic turns (TAL-899). */
+    async start(session, opts) {
+        const response = (await this.admit(session, opts, true));
+        if (response._status === undefined)
+            this.stopHeld.delete(session.session_id);
+        return response;
+    }
     /** Python `_start_chat_stream_for_session`: persist pending state, register the channel, launch the worker. */
-    start(session, opts) {
+    async admit(session, opts, releaseStopHold = false) {
         const s = session;
+        const readOnly = (await this.deps.service().readOnlyReason(s));
+        if (readOnly)
+            return { error: readOnly, _status: 403 };
         if (this.deps.updateInProgress?.())
             return { error: 'Web is updating. Retry after it restarts.', _status: 503 };
         if (this.deps.profileDeleting?.(s.profile ?? null))
             return { error: `Profile '${str(s.profile)}' is being deleted.`, _status: 409 };
         if (this.deps.workspaceBusy?.(opts.workspace))
             return { error: WORKSPACE_BUSY_MESSAGE, _status: 409 };
+        if (this.deps.checkpointStoreBusy?.(s.profile ?? null))
+            return { error: CHECKPOINT_STORE_BUSY_MESSAGE, _status: 409 };
         const locked = s.active_stream_id;
         if (locked) {
             if (this.registry.liveIds.has(locked) || this.registry.activeRuns.has(locked))
                 return { error: 'session already has an active stream', active_stream_id: locked, _status: 409 };
-            this.deps.service().clearStaleStreamState(s);
+            await this.deps.service().clearStaleStreamState(s);
             if (s.active_stream_id)
                 return { error: 'session already has an active stream', active_stream_id: s.active_stream_id, _status: 409 };
         }
@@ -247,11 +369,21 @@ export class TurnRunner {
             if (blocking)
                 return { error: 'session already has an active stream', active_stream_id: blocking, _status: 409 };
         }
+        const accepted = isDict(s.extra.completion_acceptances) ? s.extra.completion_acceptances : {};
+        if (opts.completionKeys?.length && opts.completionKeys.every((key) => typeof accepted[key] === 'string' && accepted[key])) {
+            return { stream_id: str(accepted[opts.completionKeys[0]]), session_id: s.session_id };
+        }
         const streamId = randomUUID().replace(/-/g, '');
         const wasHiddenEmpty = !s.messages.length && !s.context_messages.length && !s.hasPendingPrompt;
         const attachments = opts.attachments ?? [];
         // A background wakeup stays one in a forked session too (TAL-460).
         const source = opts.source !== 'process_wakeup' && str(s.source_tag).toLowerCase() === 'fork' ? 'fork' : (opts.source ?? 'webui');
+        // TAL-1005: the context figures the last turn reported hold for a turn on the same model only, and are judged before
+        // admission clears the post-compression estimate. A `/btw` clone is thrown away after one answer, so it is never due.
+        const sameRuntime = str(s.model) === str(opts.model) && str(s.model_provider) === str(opts.modelProvider);
+        const compactionDue = !opts.ephemeral && sameRuntime && this.deps.service().compactionDue(s);
+        if (!sameRuntime)
+            forgetContextFigures(s);
         s.workspace = opts.workspace;
         s.model = opts.model;
         s.model_provider = opts.modelProvider;
@@ -261,6 +393,12 @@ export class TurnRunner {
         s.pending_attachments = attachments;
         s.pending_started_at = this.deps.now();
         s.pending_user_source = source;
+        // TAL-927: reasoning one provider issued can fail on another (hermes-agent#122831), so a provider switch moves the
+        // boundary its carriers replay from.
+        if (s.reasoning_replay_provider !== str(opts.modelProvider)) {
+            s.reasoning_replay_provider = str(opts.modelProvider);
+            s.reasoning_replay_from = s.pending_started_at;
+        }
         if (['', 'Untitled', 'New Chat'].includes(str(s.title).trim())) {
             const provisional = titleFrom([{ role: 'user', content: opts.msg }], s.title || 'Untitled');
             if (provisional && !['', 'Untitled', 'New Chat'].includes(provisional.trim()))
@@ -268,19 +406,39 @@ export class TurnRunner {
         }
         if (this.deps.saveMode() === 'eager')
             this.checkpointUserMessage(s, opts.msg, attachments, s.pending_started_at, source, streamId);
-        this.deps.store.save(s);
+        const previousStopHold = s.automatic_turns_stop_held;
+        if (releaseStopHold)
+            s.automatic_turns_stop_held = false;
+        const previousAcceptances = s.extra.completion_acceptances;
+        if (opts.completionKeys?.length)
+            s.extra.completion_acceptances = { ...accepted, ...Object.fromEntries(opts.completionKeys.map((key) => [key, streamId])) };
+        const previousUpdates = s.extra.background_updates;
+        if (opts.backgroundUpdate)
+            recordBackgroundUpdate(s, streamId, opts.backgroundUpdate);
+        const admission = { stream_id: streamId, session_id: s.session_id, started_at: this.deps.now(), phase: 'starting', workspace: opts.workspace, model: opts.model, provider: opts.modelProvider, ephemeral: Boolean(opts.ephemeral), origin: turnOrigin(source), admitting: true };
+        this.registry.registerActiveRun(admission);
+        try {
+            (await this.deps.store.save(s));
+        }
+        catch (error) {
+            this.registry.activeRuns.delete(streamId);
+            s.extra.background_updates = previousUpdates;
+            s.extra.completion_acceptances = previousAcceptances;
+            s.automatic_turns_stop_held = previousStopHold;
+            throw error;
+        }
         this.registry.writebackOwners.set(s.session_id, streamId);
         // A new chat appears in the list; every turn tells open clients a run started (TAL-434).
         this.deps.events.publish(wasHiddenEmpty ? 'session_new' : 'turn_started', { profile: s.profile, sessionId: s.session_id });
         try {
-            this.deps.workspaces.setLastWorkspace(opts.workspace, s.profile);
+            (await this.deps.workspaces.setLastWorkspace(opts.workspace, s.profile));
         }
         catch { /* best effort */ }
         const channel = this.registry.create(streamId, s.session_id);
+        admission.admitting = false;
         if (opts.goalRelated)
             this.registry.goalRelated.add(streamId);
-        this.registry.registerActiveRun({ stream_id: streamId, session_id: s.session_id, started_at: this.deps.now(), phase: 'starting', workspace: opts.workspace, model: opts.model, provider: opts.modelProvider, ephemeral: Boolean(opts.ephemeral), origin: turnOrigin(source) });
-        const worker = this.run(s.session_id, streamId, channel, opts).catch((error) => {
+        const worker = this.run(s.session_id, streamId, channel, opts, compactionDue).catch((error) => {
             this.deps.log(`[webui] ERROR turn ${streamId} crashed\n${error instanceof Error ? (error.stack ?? error.message) : String(error)}`);
         }).finally(() => { this.workers.delete(streamId); });
         this.workers.set(streamId, worker);
@@ -291,6 +449,10 @@ export class TurnRunner {
             response.effective_model_provider = opts.modelProvider;
         return response;
     }
+    /** Resolves once the stream's worker has ended (settled, failed, or torn down); it never rejects. */
+    finished(streamId) {
+        return this.workers.get(streamId) ?? Promise.resolve();
+    }
     /** Python `_checkpoint_user_message_for_eager_session_save`. */
     checkpointUserMessage(s, msg, attachments, startedAt, source, turnId) {
         const latest = s.messages[s.messages.length - 1];
@@ -299,20 +461,69 @@ export class TurnRunner {
         s.messages.push(pendingUserRow(msg, attachments, startedAt, source, turnId));
     }
     // ── worker ───────────────────────────────────────────────────────────────
-    async run(sessionId, streamId, channel, opts) {
+    async run(sessionId, streamId, channel, opts, compactionDue) {
         const { deps } = this;
         let writer = null;
         try {
-            writer = deps.journal.writer(sessionId, streamId);
+            writer = (await deps.journal.writer(sessionId, streamId));
             this.writers.set(streamId, writer);
         }
         catch (error) {
             deps.log(`[webui] WARNING: run journal degraded for stream ${streamId}: ${error.message}`);
         }
-        const put = (event, data, meta = {}) => {
+        const failure = { apperror: null, cancelled: false };
+        let waitDetail = null;
+        let providerPhase = null;
+        let statusEnded = false;
+        // TAL-1006: the open compression span closes on the Agent's end of it, else on the next content frame.
+        const closeCompression = async (createdAt) => {
+            if (!this.compressionSpans.has(streamId))
+                return;
+            await put('compressed', { session_id: sessionId, message: 'Compression finished' }, createdAt === undefined ? {} : { createdAt });
+        };
+        const put = async (event, data, meta = {}) => {
+            if (this.registry.cancelled.has(streamId) && !['cancel', 'apperror', 'steer_consumed', 'steer_withdrawn', 'run_status'].includes(event))
+                return;
+            const terminal = ['done', 'cancel', 'apperror', 'error', 'stream_end'].includes(event);
+            if (terminal)
+                statusEnded = true;
+            if (CONTENT_EVENTS.has(event) || terminal)
+                providerPhase = null;
+            if (event === 'rate_limit_wait')
+                waitDetail = data.wait;
+            // Both pre-turn compaction and Agent callbacks publish through this boundary.
+            if (event === 'compressing' && !statusEnded) {
+                const startedAt = this.compressionSpans.get(streamId) ?? meta.createdAt ?? deps.now();
+                this.compressionSpans.set(streamId, startedAt);
+                data = { ...data, started_at: startedAt };
+            }
+            else if (event === 'compressed')
+                this.compressionSpans.delete(streamId);
+            if (CONTENT_EVENTS.has(event)) {
+                if (waitDetail) {
+                    await put('rate_limit_wait', { session_id: sessionId, wait: null }, meta.createdAt === undefined ? {} : { createdAt: meta.createdAt });
+                }
+                await closeCompression(meta.createdAt);
+            }
+            const status = event === 'run_status' ? data.status : statusEnded ? null : this.compressionSpans.has(streamId) ? { phase: 'compressing' }
+                : waitDetail ? { phase: 'rate_limit_wait', wait: waitDetail } : providerPhase ?? { phase: 'working' };
+            if (JSON.stringify(this.runStatuses.get(streamId) ?? null) !== JSON.stringify(status)) {
+                if (status)
+                    this.runStatuses.set(streamId, status);
+                else
+                    this.runStatuses.delete(streamId);
+                await put('run_status', { session_id: sessionId, status }, meta);
+            }
+            // The sidecar phase is private; only the selected status crosses the client boundary.
+            if (event === 'provider_phase')
+                return;
+            if (event === 'apperror')
+                failure.apperror = data;
+            else if (event === 'cancel')
+                failure.cancelled = true;
             for (const steerEvent of this.takeSteerEventsBefore(streamId, event))
-                put(steerEvent[0], steerEvent[1]);
-            if (this.registry.cancelled.has(streamId) && !['cancel', 'apperror', 'steer_consumed', 'steer_withdrawn'].includes(event))
+                (await put(steerEvent[0], steerEvent[1]));
+            if (this.registry.cancelled.has(streamId) && !['cancel', 'apperror', 'steer_consumed', 'steer_withdrawn', 'run_status'].includes(event))
                 return;
             // `cancel()` already wrote the terminal row and closed the stream: the worker's unwind adds no second one.
             if (event === 'cancel' && !this.registry.streams.has(streamId)) {
@@ -325,7 +536,7 @@ export class TurnRunner {
             let eventId = null;
             if (writer) {
                 try {
-                    eventId = writer.appendSseEvent(event, data, meta).event_id;
+                    eventId = (await writer.appendSseEvent(event, data, meta)).event_id;
                 }
                 catch (error) {
                     this.registry.degradedJournals.add(streamId);
@@ -347,25 +558,43 @@ export class TurnRunner {
         const activeRun = this.registry.activeRuns.get(streamId);
         let s;
         try {
-            s = deps.store.get(sessionId);
+            s = (await deps.store.get(sessionId));
         }
         catch {
-            put('apperror', { type: 'error', terminal_state: 'error', message: 'Session not found', session_id: sessionId });
-            this.teardown(sessionId, streamId);
+            await put('apperror', { type: 'error', terminal_state: 'error', message: 'Session not found', session_id: sessionId });
+            await this.teardown(sessionId, streamId);
             return;
         }
         const msgText = opts.msg;
-        const previousMessages = structuredClone(s.messages);
-        // TAL-493: one state.db read builds the history and marks where this turn's own rows begin.
-        const startRead = deps.service().stateDbRead(s);
-        const previousContext = structuredClone(deps.service().modelContext(s, startRead.rows));
-        // Python `_sanitize_messages_for_agent`: the model never sees display-only rows or a replayed cancelled prompt, and
-        // the Agent gets back the exact bytes it sent (TAL-541).
-        const apiHistory = sanitizeMessagesForApi(previousContext, { preserveApiContent: true });
-        const workspaceCtx = workspaceContextPrefix(opts.workspace);
+        const history = () => {
+            // TAL-493: one state.db read builds the history and marks where this turn's own rows begin.
+            const startRead = deps.service().beginStateDbTurn(s);
+            const previousMessages = structuredClone(s.messages);
+            // TAL-578: an eager save already put this turn's prompt in the transcript; `user_message` carries it, so the history
+            // leaves it out.
+            const previousContext = structuredClone(deps.service().modelContext(s, startRead.rows)).filter((m) => m._turn_id !== streamId);
+            // Python `_sanitize_messages_for_agent`: the model never sees display-only rows or a replayed cancelled prompt, and
+            // the Agent gets back the exact bytes it sent (TAL-541).
+            const apiHistory = sanitizeMessagesForApi(previousContext, { preserveApiContent: true, issuedReasoningSince: s.reasoning_replay_from });
+            this.stopContexts.set(streamId, { previousContext, historyLength: apiHistory.length, prompt: msgText, msgText, checkpointed: false, stateDbStartId: startRead.ok ? stateDbSeenId(startRead.rows) ?? 0 : null, stateDbStartGeneration: startRead.generation });
+            return { previousMessages, previousContext, apiHistory };
+        };
         // Before the first await: a Stop can land at any point after admission.
-        // An eager save already put this turn's prompt in the transcript; the Stop fallback appends it once itself.
-        this.stopContexts.set(streamId, { previousContext: previousContext.filter((m) => m._turn_id !== streamId), historyLength: apiHistory.length, prompt: workspaceCtx + msgText, msgText, checkpointed: false, stateDbStartId: startRead.ok ? stateDbSeenId(startRead.rows) ?? 0 : null });
+        let { previousMessages, previousContext, apiHistory } = history();
+        const controller = new AbortController();
+        this.abortControllers.set(streamId, controller);
+        // TAL-1005: an over-threshold session compacts in its joinable compression job before the turn reads its history, so
+        // a Stop ends only this wait and a resend joins the same job. The turn then reads the compacted session afresh.
+        const compacted = await deps.service().ensureCompacted(sessionId, { turn: streamId, signal: controller.signal, due: compactionDue, onWait: async () => { (await put('compressing', { session_id: sessionId, message: 'Compressing context' })); } });
+        if (compacted) {
+            try {
+                s = (await deps.store.get(sessionId));
+            }
+            catch { /* the cancel or error path below reports a deleted session */ }
+            ;
+            ({ previousMessages, previousContext, apiHistory } = history());
+            await put('compressed', this.compressedFrame(s));
+        }
         const activeTurnToken = buildActiveTurnToken(streamId, s.pending_started_at);
         const sidecar = deps.sidecar();
         const partialText = this.registry.partialText.get(streamId) ?? [];
@@ -382,21 +611,19 @@ export class TurnRunner {
         // TAL-397: the live meter. Counters start from the session's and follow the Agent's; the rate counts text deltas.
         const liveUsage = { input_tokens: s.input_tokens, output_tokens: s.output_tokens, estimated_cost: s.estimated_cost, cache_read_tokens: s.cache_read_tokens, cache_write_tokens: s.cache_write_tokens };
         const meter = { deltas: 0, first: 0, last: 0, sentAt: null };
-        const putMetering = (force) => {
-            const now = deps.now();
+        const putMetering = async (force, now = deps.now()) => {
             if (!force && meter.sentAt !== null && now - meter.sentAt < METERING_INTERVAL_S)
                 return;
             meter.sentAt = now;
             const tps = meter.deltas && meter.last > meter.first ? Math.round((meter.deltas / (meter.last - meter.first)) * 10) / 10 : null;
-            put('metering', { session_id: sessionId, usage: meteringUsage(liveUsage), tps, tps_available: tps !== null, estimated: false });
+            await put('metering', { session_id: sessionId, usage: meteringUsage(liveUsage), tps, tps_available: tps !== null, estimated: false });
         };
-        const meterDelta = () => {
-            const now = deps.now();
+        const meterDelta = async (now = deps.now()) => {
             if (!meter.deltas)
                 meter.first = now;
             meter.last = now;
             meter.deltas += 1;
-            putMetering(false);
+            await putMetering(false, now);
         };
         // TAL-460: a background turn's text is held back while it could still be a silence marker, so one never flashes by.
         // ponytail: checks the turn's whole text, so a marker after earlier prose streams until the settled row hides it;
@@ -405,13 +632,11 @@ export class TurnRunner {
         let heldText = '';
         const titles = new ReasoningTitleTracker();
         let capturedTerminalError = null;
-        const controller = new AbortController();
-        this.abortControllers.set(streamId, controller);
         const attachments = opts.attachments ?? [];
         // TAL-545: one image mode decides both this turn's upload and the replayed history, which a text-mode turn sends
         // without native images; only a turn with an image on either side asks for it.
         const imageMode = attachments.some(isImageAttachment) || hasNativeImages(apiHistory) ? await this.imageInputMode(sessionId, s, opts, controller.signal) : null;
-        const userMessage = this.buildUserMessage(workspaceCtx, msgText, attachments, opts.workspace, sessionId, imageMode);
+        const userMessage = this.buildUserMessage(opts.prompt ?? msgText, attachments, opts.workspace, sessionId, imageMode);
         const conversationHistory = imageMode === 'text' ? withoutNativeImages(apiHistory) : apiHistory;
         // The prompt the Agent actually gets, native image parts included.
         const stop = this.stopContexts.get(streamId);
@@ -420,15 +645,25 @@ export class TurnRunner {
         if (activeRun)
             activeRun.phase = 'running';
         const settledAt = { value: false };
-        let failed = false;
+        let succeeded = false;
+        let ownershipRefused = false;
+        let agentDispatched = false;
+        let streamWrites = Promise.resolve();
+        let streamWriteError;
         // TAL-424: steers an errored turn ended without taking, sent as one follow-up turn once this one is torn down.
         let followUp = [];
         try {
             if (!sidecar)
                 throw new SidecarError('The Agent sidecar is not running; chat is unavailable until it starts.', { condition: 'sidecar_unavailable' });
             // Python: budgets, reasoning config, personality and delivery context come from the profile's config.yaml; the
-            // system message carries the frozen session workspace, the WebUI guidance rides as the ephemeral prompt.
+            // system message carries the frozen session workspace, the WebUI guidance rides as the ephemeral prompt with this
+            // turn's workspace (TAL-710: never a tag on the user's message, which the Agent can persist from memory).
             const cfg = (await deps.profileConfig?.(s.profile ?? null)) ?? {};
+            const personality = str(s.personality);
+            // A selected personality never silently drops: an unknown catalog fails the turn (an absent name is the Agent's no-overlay).
+            const personalities = personality && deps.personalities
+                ? await deps.personalities(s.profile ?? null).catch((error) => { throw new SidecarError(`The session's personality "${personality}" could not be resolved: ${error instanceof Error ? error.message : String(error)}`, { condition: 'sidecar_error' }); })
+                : [];
             // The sidecar's per-session YOLO state does not survive its restarts: push the local flag (either way) before
             // the turn so a stale sidecar-side enable cannot auto-approve a session the UI reports as guarded.
             await this.withYoloLock(sessionId, async () => { try {
@@ -438,154 +673,226 @@ export class TurnRunner {
             // Last yield before `chat.start`: a cancel that landed during the attachment, config or YOLO awaits must not
             // start an Agent turn (the client also refuses an already-aborted signal outright).
             if (this.registry.cancelled.has(streamId)) {
-                this.finalizeCancelled(s, streamId, opts.ephemeral);
-                put('cancel', this.cancelFrame(sessionId));
+                await this.finalizeCancelled(s, streamId, opts.ephemeral);
+                await put('cancel', (await this.cancelFrame(sessionId)));
                 return;
             }
-            const frozenWorkspace = str(s.created_workspace) || str(s.workspace);
             const turnContext = {
-                system_message: workspaceSystemMessage(frozenWorkspace),
-                ephemeral_system_prompt: webuiEphemeralSystemPrompt({ config: cfg, personality: str(s.personality) || null, sessionId, profile: s.profile ?? null, workspace: frozenWorkspace, hermesHome: deps.profileHome(s.profile), homeDisplay: deps.homeDisplay?.() ?? '~/.hermes', background: turnOrigin(opts.source) === 'background' }),
-                max_iterations: processWakeupMaxIterations(maxIterationsFromConfig(cfg), opts.source ?? 'webui', deps.env ?? {}),
-                max_tokens: maxTokensFromConfig(cfg),
-                reasoning_config: reasoningConfigFromConfig(cfg, opts.modelProvider ?? s.model_provider ?? null),
+                system_message: workspaceSystemMessage(str(s.created_workspace) || str(s.workspace)),
+                ephemeral_system_prompt: webuiEphemeralSystemPrompt({ config: cfg, personalityPrompt: findPersonality(personalities, personality)?.prompt ?? null, sessionId, profile: s.profile ?? null, workspace: opts.workspace, hermesHome: deps.profileHome(s.profile), homeDisplay: deps.homeDisplay?.() ?? '~/.hermes', background: turnOrigin(opts.source) === 'background' }),
+                ...agentBuildParams(cfg, opts.source ?? 'webui', opts.modelProvider ?? s.model_provider ?? null, deps.env ?? {}),
             };
             // Python `_public_prefill_context_status`: the session-recall prefill hook was dropped with TAL-245, so the
             // context frame always reports the not-configured shape.
             const persistentBefore = persistentStateSnapshot(deps.profileHome(s.profile));
-            put('context_status', { session_id: sessionId, prefill: { status: 'not_configured', source: 'none', label: '', message_count: 0 } });
+            await put('context_status', { session_id: sessionId, prefill: { status: 'not_configured', source: 'none', label: '', message_count: 0 } });
+            // Setup awaits can outlive a gateway adoption; no await separates this read from dispatch.
+            deps.service().assertWritable(s);
+            agentDispatched = true;
             const result = await sidecar.call('chat.start', {
-                profile_home: deps.profileHome(s.profile), session_id: sessionId, stream_id: streamId, workspace: opts.workspace, model: opts.model ?? '', model_provider: opts.modelProvider,
+                profile_home: deps.profileHome(s.profile), session_id: sessionId, stream_id: streamId, ...this.agentWorkspace(s, opts.workspace), model: opts.model ?? '', model_provider: opts.modelProvider,
                 user_message: userMessage, ...turnContext, conversation_history: conversationHistory, enabled_toolsets: deps.toolsetsFor(s), ...(msgText ? { persist_user_message: msgText } : {}), ...(opts.ephemeral ? { ephemeral: true } : {}),
+                ...(opts.voiceTurn ? { voice_turn: true } : {}),
+                ...(opts.connectCard ? { connect_card: true } : {}),
+                ...(opts.skillSecret ? { skill_secret: true } : {}),
             }, {
                 signal: controller.signal,
                 timeoutMs: 0,
                 onStream: (frame) => {
-                    const data = isDict(frame.data) ? frame.data : {};
-                    switch (frame.event) {
-                        case 'token': {
-                            tokenSent = true;
-                            firstTokenAt ??= deps.now();
-                            // Visible text closes the reasoning segment: a stable title snapshot (or a clear) goes out first.
-                            const stable = titles.stableSnapshot(reasoningText.join(''));
-                            if (stable !== null)
-                                put('reasoning', { text: '', titles: stable });
-                            partialText.push(str(data.text));
-                            if (holdSilence && mayBecomeSilentReply(partialText.join(''))) {
-                                heldText += str(data.text);
-                                meterDelta();
+                    const receivedAt = deps.now();
+                    streamWrites = streamWrites.then(async () => {
+                        const publish = put;
+                        const emitFrame = (event, data, meta = {}) => publish(event, data, { ...meta, createdAt: receivedAt });
+                        const data = isDict(frame.data) ? frame.data : {};
+                        switch (frame.event) {
+                            case 'token': {
+                                tokenSent = true;
+                                firstTokenAt ??= receivedAt;
+                                // Visible text closes the reasoning segment: a stable title snapshot (or a clear) goes out first.
+                                const stable = titles.stableSnapshot(reasoningText.join(''));
+                                if (stable !== null)
+                                    (await emitFrame('reasoning', { text: '', titles: stable }));
+                                partialText.push(str(data.text));
+                                if (holdSilence && mayBecomeSilentReply(partialText.join(''))) {
+                                    heldText += str(data.text);
+                                    (await meterDelta(receivedAt));
+                                    return;
+                                }
+                                await emitFrame('token', { text: heldText + str(data.text) });
+                                heldText = '';
+                                await meterDelta(receivedAt);
                                 return;
                             }
-                            put('token', { text: heldText + str(data.text) });
-                            heldText = '';
-                            meterDelta();
-                            return;
-                        }
-                        case 'reasoning': {
-                            if (partialText.length && !reasoningText.length)
-                                titles.resetSegment();
-                            reasoningText.push(str(data.text));
-                            // Python `reasoning_event_payload`: bold or first-line titles ride on the delta they complete.
-                            put('reasoning', titles.prepare(reasoningEventPayload(str(data.text), reasoningText.join(''))));
-                            meterDelta();
-                            return;
-                        }
-                        case 'usage': {
-                            const usage = ChatUsageSchema.safeParse(data);
-                            if (usage.success) {
-                                applyAgentUsage(liveUsage, usage.data);
-                                this.liveUsage.set(streamId, liveUsage);
+                            case 'reasoning': {
+                                if (partialText.length && !reasoningText.length)
+                                    titles.resetSegment();
+                                reasoningText.push(str(data.text));
+                                // Python `reasoning_event_payload`: bold or first-line titles ride on the delta they complete.
+                                await emitFrame('reasoning', titles.prepare(reasoningEventPayload(str(data.text), reasoningText.join(''))));
+                                await meterDelta(receivedAt);
+                                return;
                             }
-                            putMetering(true);
-                            return;
-                        }
-                        case 'steer_pending':
-                            // TAL-424: while a withdraw or Send now rewrites the Agent's queue its pending text is ambiguous; it settles after.
-                            if (this.steerRewrites.has(streamId)) {
-                                this.deferredPendingSteer.set(streamId, str(data.text));
+                            case 'usage': {
+                                const usage = ChatUsageSchema.safeParse(data);
+                                if (usage.success) {
+                                    applyAgentUsage(liveUsage, usage.data);
+                                    this.liveUsage.set(streamId, liveUsage);
+                                }
+                                await putMetering(true, receivedAt);
                                 return;
                             }
-                            this.settleConsumedSteers(sessionId, streamId, str(data.text));
-                            return;
-                        case 'tool': {
-                            const call = { name: data.name, args: data.args ?? {}, tid: str(data.tid), done: false };
-                            liveToolCalls.push(call);
-                            toolIds.set(call, call.tid || mintToolId());
-                            toolStartedAt.set(call, deps.now());
-                            const redacted = deps.redactEnabled();
-                            put('tool', publicToolFrame(withToolId(data, toolIds.get(call)), redacted), { redacted });
-                            return;
-                        }
-                        case 'tool_complete': {
-                            // TAL-313: the server decides failure from the sidecar's raw result, which never leaves the server; TAL-315: so
-                            // are its display sections, which keep the stderr and exit code the flat preview drops.
-                            const { raw_result: rawResult, todo_result: todoResult, result_diff: resultDiff, ...complete } = data;
-                            const outcome = toolOutcome(rawResult);
-                            complete.is_error = outcome.is_error;
-                            if (rawResult !== undefined)
-                                complete.result_view = outcome.result_view;
-                            // TAL-448: a file edit's change, counted over the whole diff the sidecar forwards; redacted with the frame.
-                            const editDiff = toolEditDiff(data.name, { diff: resultDiff });
-                            if (editDiff)
-                                complete.edit_diff = editDiff;
-                            const tc = liveToolCalls[completedToolIndex(liveToolCalls, str(data.tid), data.name)];
-                            if (tc) {
-                                const startedAt = toolStartedAt.get(tc);
-                                if (startedAt !== undefined)
-                                    complete.duration = Math.round(Math.max(0, deps.now() - startedAt) * 1000) / 1000;
-                                // TAL-315: the decided view too, so a failed or cancelled turn's snapshot keeps the sections shown live.
-                                Object.assign(tc, { done: true, snippet: data.preview, is_error: complete.is_error, duration: complete.duration ?? null, ...(complete.result_view ? { result_view: complete.result_view } : {}), ...(editDiff ? { edit_diff: editDiff } : {}) });
+                            case 'steer_pending': {
+                                // TAL-424: while a withdraw or Send now rewrites the Agent's queue its pending text is ambiguous; it settles after.
+                                if (this.steerRewrites.has(streamId)) {
+                                    this.deferredPendingSteer.set(streamId, str(data.text));
+                                    return;
+                                }
+                                // Capture the causal position before the next frame; only persistence is asynchronous.
+                                const saving = this.settleConsumedSteers(sessionId, streamId, str(data.text));
+                                await saving;
+                                return;
                             }
-                            const id = (tc && toolIds.get(tc)) || str(data.tid) || mintToolId();
-                            this.lastCompletedTool.set(streamId, id);
-                            const redacted = deps.redactEnabled();
-                            put('tool_complete', publicToolFrame(withToolId(complete, id), redacted), { redacted });
-                            // TAL-397: the todo tool's result is the session's new list; the settled session derives the same one.
-                            const todos = data.name === 'todo' ? parseTodoToolResult(todoResult) : null;
-                            if (todos) {
-                                todos.ts = deps.now();
-                                this.liveTodos.set(streamId, todos);
-                                put('todo_state', redactNestedMessageContainers({ ...todos, session_id: sessionId, stream_id: streamId, source: 'live' }, redacted));
+                            case 'tool': {
+                                const call = { name: data.name, args: data.args ?? {}, tid: str(data.tid), done: false };
+                                liveToolCalls.push(call);
+                                toolIds.set(call, call.tid || mintToolId());
+                                toolStartedAt.set(call, receivedAt);
+                                const redacted = deps.redactEnabled();
+                                await emitFrame('tool', publicToolFrame(withToolId(data, toolIds.get(call)), redacted), { redacted });
+                                return;
                             }
-                            return;
-                        }
-                        // Python: the live chat frame carries the queue head plus depth, not the entry that just arrived.
-                        case 'approval': {
-                            deps.pending.submitApproval(sessionId, { ...data, session_id: sessionId });
-                            const { pending, pending_count } = deps.pending.approvalPending(sessionId);
-                            put('approval', { ...(pending ?? data), pending_count });
-                            return;
-                        }
-                        // TAL-514: the Agent stopped waiting on an approval (answer, timeout, or interrupt); withdraw its copy.
-                        case 'approval_resolved': {
-                            if (!str(data.approval_id))
+                            case 'tool_complete': {
+                                // TAL-313: the server decides failure from the sidecar's raw result, which never leaves the server; TAL-315: so
+                                // are its display sections, which keep the stderr and exit code the flat preview drops.
+                                const { raw_result: rawResult, todo_result: todoResult, result_diff: resultDiff, display_metadata: displayMetadata, ...complete } = data;
+                                const outcome = toolOutcome(rawResult);
+                                complete.is_error = outcome.is_error;
+                                if (rawResult !== undefined)
+                                    complete.result_view = outcome.result_view;
+                                // TAL-448: a file edit's change, counted over the whole diff the sidecar forwards; redacted with the frame. TAL-909:
+                                // else the diff in the metadata the Agent persisted on the result row, which a reload reads the same way.
+                                const editDiff = toolEditDiff(data.name, { diff: resultDiff }, displayMetadata);
+                                if (editDiff)
+                                    complete.edit_diff = editDiff;
+                                const tc = liveToolCalls[completedToolIndex(liveToolCalls, str(data.tid), data.name)];
+                                if (tc) {
+                                    const startedAt = toolStartedAt.get(tc);
+                                    if (startedAt !== undefined)
+                                        complete.duration = Math.round(Math.max(0, receivedAt - startedAt) * 1000) / 1000;
+                                    // TAL-315: the decided view too, so a failed or cancelled turn's snapshot keeps the sections shown live.
+                                    Object.assign(tc, { done: true, snippet: data.preview, is_error: complete.is_error, duration: complete.duration ?? null, ...(complete.result_view ? { result_view: complete.result_view } : {}), ...(editDiff ? { edit_diff: editDiff } : {}) });
+                                }
+                                const id = (tc && toolIds.get(tc)) || str(data.tid) || mintToolId();
+                                this.lastCompletedTool.set(streamId, id);
+                                const redacted = deps.redactEnabled();
+                                await emitFrame('tool_complete', publicToolFrame(withToolId(complete, id), redacted), { redacted });
+                                // TAL-397: the todo tool's result is the session's new list; the settled session derives the same one.
+                                // TAL-889: the sidecar attaches `todo_result` only to the Agent's todo tool, so no name check here.
+                                const todos = parseTodoToolResult(todoResult);
+                                if (todos) {
+                                    todos.ts = receivedAt;
+                                    this.liveTodos.set(streamId, todos);
+                                    await emitFrame('todo_state', redactNestedMessageContainers({ ...todos, session_id: sessionId, stream_id: streamId, source: 'live' }, redacted));
+                                }
                                 return;
-                            deps.pending.resolveApproval(sessionId, str(data.approval_id));
-                            put(...this.approvalHeadFrame(sessionId));
-                            return;
-                        }
-                        case 'clarify': {
-                            deps.pending.submitClarify(sessionId, { ...data, session_id: sessionId, kind: 'clarify' });
-                            const head = deps.pending.clarifyHeadFrame(sessionId);
-                            if (head)
-                                put('clarify', head);
-                            return;
-                        }
-                        case 'clarify_resolved': {
-                            const { head } = deps.pending.resolveClarify(sessionId, str(data.clarify_id));
-                            if (head)
-                                put('clarify', head);
-                            return;
+                            }
+                            // Python: the live chat frame carries the queue head plus depth, not the entry that just arrived.
+                            case 'approval': {
+                                deps.pending.submitApproval(sessionId, { ...data, session_id: sessionId });
+                                const { pending, pending_count } = deps.pending.approvalPending(sessionId);
+                                await emitFrame('approval', { ...(pending ?? data), pending_count });
+                                return;
+                            }
+                            // TAL-514: the Agent stopped waiting on an approval (answer, timeout, or interrupt); withdraw its copy.
+                            case 'approval_resolved': {
+                                if (!str(data.approval_id))
+                                    return;
+                                deps.pending.resolveApproval(sessionId, str(data.approval_id));
+                                await emitFrame(...this.approvalHeadFrame(sessionId));
+                                return;
+                            }
+                            case 'clarify': {
+                                deps.pending.submitClarify(sessionId, { ...data, session_id: sessionId });
+                                const head = deps.pending.clarifyHeadFrame(sessionId);
+                                if (head)
+                                    (await emitFrame('clarify', head));
+                                return;
+                            }
+                            case 'clarify_resolved': {
+                                const { head } = deps.pending.resolveClarify(sessionId, str(data.clarify_id));
+                                if (head)
+                                    (await emitFrame('clarify', head));
+                                return;
+                            }
+                            case 'provider_phase': {
+                                const phase = data.phase;
+                                if (phase !== null && phase?.phase !== 'prompt_reading')
+                                    return;
+                                const percent = phase?.progress_percent;
+                                providerPhase = phase === null ? null : { phase: 'prompt_reading', progress_percent: typeof percent === 'number' && Number.isFinite(percent) ? Math.max(0, Math.min(100, percent)) : null };
+                                await put('provider_phase', {}, { createdAt: receivedAt });
+                                return;
+                            }
+                            // TAL-821: the Agent sits out a provider's rate limit (`until`, epoch seconds: its reset, else the Agent's retry).
+                            case 'rate_limit_wait': {
+                                const until = typeof data.until === 'number' ? data.until : null;
+                                if (until === null && !waitDetail)
+                                    return;
+                                const provider = str(data.provider) || null;
+                                await emitFrame('rate_limit_wait', { session_id: sessionId, wait: until === null ? null : { provider, provider_label: provider && (deps.providerLabel?.(s.profile ?? null, provider) ?? displayName(provider)), until, until_kind: data.until_kind === 'reset' ? 'reset' : 'retry' } });
+                                return;
+                            }
+                            // TAL-818: a `manage_connections` card opened; clients poll and answer the operation by its `op_id`.
+                            case 'connection': {
+                                const snapshot = ConnectorOperationSnapshotSchema.safeParse(data);
+                                if (snapshot.success)
+                                    (await emitFrame('connect_card', connectorOperation(snapshot.data, catalogNames(deps.profileHome(s.profile)))));
+                                return;
+                            }
+                            case 'compressing':
+                                await emitFrame('compressing', data);
+                                return;
+                            case 'compression_finished':
+                                await closeCompression(receivedAt);
+                                return;
+                            case 'status':
+                                if (data.kind === 'terminal_error' && !capturedTerminalError)
+                                    capturedTerminalError = str(data.message);
+                                return;
+                            // TAL-803: the session's notice list changed; every client replaces its copy.
+                            case 'notice': {
+                                const notice = this.notices.add(sessionId, data);
+                                if (!notice)
+                                    return;
+                                await emitFrame('notices', { session_id: sessionId, notices: this.sessionNotices(sessionId) });
+                                // ponytail: the expiry reaches a client only while a turn streams; an idle tab drops it on its next session load.
+                                if (notice.expires_at !== null) {
+                                    const expiresAt = notice.expires_at;
+                                    const emitAfterExpiry = async () => {
+                                        // Timers can run before the wall clock expires the notice. Let SessionNotices decide, then retry.
+                                        if (this.notices.list(sessionId).includes(notice))
+                                            setTimeout(() => { void emitAfterExpiry().catch((error) => { deps.log(`[webui] notice expiry failed: ${String(error)}`); }); }, Math.max(1, Math.ceil((expiresAt - deps.now()) * 1000))).unref();
+                                        else
+                                            (await this.emitToSession(sessionId, 'notices', { session_id: sessionId, notices: this.sessionNotices(sessionId) }));
+                                    };
+                                    await emitAfterExpiry();
+                                }
+                                return;
+                            }
+                            case 'notice_clear':
+                            case 'notices_reset':
+                                if (frame.event === 'notice_clear' ? this.notices.clear(sessionId, str(data.key)) : this.notices.reset(sessionId))
+                                    (await emitFrame('notices', { session_id: sessionId, notices: this.sessionNotices(sessionId) }));
+                                return;
+                            default:
+                                await emitFrame(frame.event, data);
                         }
-                        case 'status':
-                            if (data.kind === 'terminal_error' && !capturedTerminalError)
-                                capturedTerminalError = str(data.message);
-                            return;
-                        default:
-                            put(frame.event, data);
-                    }
+                    }).catch((error) => { streamWriteError = error instanceof Error ? error : new Error(String(error)); });
                 },
             });
+            await streamWrites;
+            if (streamWriteError !== undefined)
+                throw streamWriteError;
             settledAt.value = true;
             // TAL-545: every transcript the Agent hands back settles with the history's images, not the request's projection.
             const withImages = (rows) => withNativeImagesRestored(rows, conversationHistory, apiHistory);
@@ -598,14 +905,14 @@ export class TurnRunner {
                 const interrupted = await this.interrupts.get(streamId);
                 applyAgentUsage(liveUsage, result.usage);
                 this.liveUsage.set(streamId, liveUsage);
-                this.finalizeCancelled(s, streamId, opts.ephemeral, interrupted?.checkpoint ? withImages(interrupted.checkpoint) : result.messages);
-                put('cancel', this.cancelFrame(sessionId));
+                await this.finalizeCancelled(s, streamId, opts.ephemeral, interrupted?.checkpoint ? withImages(interrupted.checkpoint) : result.messages);
+                await put('cancel', (await this.cancelFrame(sessionId)));
                 return;
             }
             // Writeback only while this stream still owns the session (Python `_stream_writeback_is_current`).
             let current;
             try {
-                current = deps.store.get(sessionId);
+                current = (await deps.store.get(sessionId));
             }
             catch {
                 current = s;
@@ -641,16 +948,15 @@ export class TurnRunner {
                 await this.steerRewrites.get(streamId);
                 const { events: steerEvents, leftovers } = this.finalizeSteers(streamId, str(result.pending_steer), 'followup');
                 followUp = leftovers;
-                this.persistError(s, streamId, classification.label, payload, activeTurnToken, resultMessages);
+                await this.persistError(s, streamId, classification.label, payload, activeTurnToken, resultMessages, result.provider);
                 // TAL-512: a btw error carries no session: its journaled frame must not keep a copy of the parent conversation.
                 if (!opts.ephemeral)
-                    payload.session = redactSessionData(this.terminalSessionPayload(s), deps.redactEnabled());
+                    payload.session = (await this.publicTerminalSession(s));
                 payload.session_id = s.session_id;
                 payload.old_session_id = sessionId;
                 for (const [event, data] of steerEvents)
-                    put(event, data);
-                put('apperror', payload);
-                failed = true;
+                    (await put(event, data));
+                await put('apperror', payload);
                 return;
             }
             // TAL-512: a btw turn passes the failure check above first, so its last assistant row is this turn's reply.
@@ -663,11 +969,12 @@ export class TurnRunner {
                         break;
                     }
                 }
-                opts.onDone?.(answer);
+                succeeded = true;
+                await opts.onDone?.(answer);
                 // Python `_ephemeral_session_payload`: only role and content leave the server for a btw turn.
-                put('done', { session: { session_id: sessionId, messages: resultMessages.map((m) => ({ role: m.role, content: m.content })) }, usage: { input_tokens: 0, output_tokens: 0 }, ephemeral: true, answer, terminal_state: answer.trim() ? 'completed' : 'no_response' });
+                await put('done', { session: { session_id: sessionId, messages: resultMessages.map((m) => ({ role: m.role, content: m.content })) }, usage: { input_tokens: 0, output_tokens: 0 }, ephemeral: true, answer, terminal_state: answer.trim() ? 'completed' : 'no_response' });
                 try {
-                    rmSync(deps.store.pathFor(sessionId), { force: true });
+                    await deps.store.delete(sessionId, { tombstone: false });
                 }
                 catch { /* ignore */ }
                 deps.store.sessions.delete(sessionId);
@@ -694,8 +1001,9 @@ export class TurnRunner {
             const { events: steerEvents, leftovers } = this.finalizeSteers(streamId, result.pending_steer, 'followup');
             s.messages = withoutToolImages(mergeDisplayMessagesAfterAgentResult(previousMessages, previousContext, resultMessages, msgText, { source: opts.source ?? 'webui', activeTurnToken, now: deps.now(), turnId: streamId, attachments: opts.attachments ?? [] }));
             s.context_messages = dedupeContext(context);
+            // The settled turn's own figures: a compaction it waited on (TAL-1005) installed an estimate its prompt now supersedes.
+            s.post_compression_context_tokens_estimate = result.compressed ? estimate : null;
             if (result.compressed) {
-                s.post_compression_context_tokens_estimate = estimate;
                 // TAL-540: the anchor is the last visible row before this turn (never its eager checkpoint); the merged transcript
                 // holds no compaction marker, so the summary is the model context's newest one.
                 const before = visibleMessagesForAnchor(previousMessages.filter((m) => m._turn_id !== streamId));
@@ -720,20 +1028,13 @@ export class TurnRunner {
                 s.threshold_tokens = result.context.threshold_tokens;
             if (typeof result.context.last_prompt_tokens === 'number')
                 s.last_prompt_tokens = result.context.last_prompt_tokens;
-            if (result.compressed) {
-                // TAL-540, Python `_live_usage_snapshot`: the counters and the ring's post-compression figures, so clients update before `done`.
-                const usage = {
-                    ...meteringUsage(s), context_length: s.context_length ?? 0, threshold_tokens: s.threshold_tokens ?? 0, last_prompt_tokens: s.last_prompt_tokens ?? 0,
-                    post_compression_context_tokens_estimate: s.post_compression_context_tokens_estimate, ...contextUsage(s, () => deps.service().deps.contextLengthFor(s.model, s.model_provider)),
-                };
-                put('compressed', { session_id: sessionId, old_session_id: sessionId, new_session_id: sessionId, continuation_session_id: sessionId, message: 'Compression finished', usage });
-            }
+            s.next_turn_threshold_tokens = result.context.next_turn_threshold_tokens ?? null;
+            if (result.compressed)
+                (await put('compressed', this.compressedFrame(s)));
             s.tool_calls = extractToolCallsFromMessages(s.messages, liveToolCalls, s.tool_calls);
-            s.active_stream_id = null;
-            s.pending_user_message = null;
-            s.pending_attachments = [];
-            s.pending_started_at = null;
-            s.pending_user_source = null;
+            this.clearPendingTurn(s);
+            // TAL-576: a turn that ran means the provider has usable credentials again.
+            s.process_wakeup_pause = null;
             const attachments = opts.attachments ?? [];
             if (attachments.length) {
                 for (let i = s.messages.length - 1; i >= 0; i -= 1) {
@@ -786,8 +1087,9 @@ export class TurnRunner {
             this.persistConsumedSteers(s, streamId, previousStartedAt(s, activeRun), now);
             // TAL-493: the rows this turn's Agent wrote to state.db are now in the transcript, so the merge must not replay them
             // (all of them, the dropped summary request too).
-            deps.service().settleStateDb(s, { turnId: streamId, previousContext: this.stopContexts.get(streamId)?.previousContext ?? previousContext, agentRows, startId: this.stopContexts.get(streamId)?.stateDbStartId ?? null });
-            deps.store.save(s);
+            this.settleStateDb(s, streamId, agentRows, { previousContext });
+            recordBackgroundReview(s, streamId);
+            await deps.store.save(s);
             deps.pending.clearApprovals(sessionId);
             deps.pending.clearClarifies(sessionId);
             deps.events.publish('session_done', { profile: s.profile, sessionId: s.session_id });
@@ -795,9 +1097,9 @@ export class TurnRunner {
             try {
                 const changes = persistentStateChanges(persistentBefore, persistentStateSnapshot(deps.profileHome(s.profile)));
                 if (changes.memory_saved)
-                    put('state_saved', { session_id: sessionId, kind: 'memory', action: 'saved' });
+                    (await put('state_saved', { session_id: sessionId, kind: 'memory', action: 'saved' }));
                 for (const change of changes.skills)
-                    put('state_saved', { session_id: sessionId, kind: 'skill', action: change.action, name: change.name });
+                    (await put('state_saved', { session_id: sessionId, kind: 'skill', action: change.action, name: change.name }));
             }
             catch (error) {
                 deps.log(`[webui] persistent state change detection failed for ${sessionId}: ${error.message}`);
@@ -813,7 +1115,7 @@ export class TurnRunner {
                 cache_hit_percent: cacheHit(cacheReadTokens, inputTokens), turn_cache_hit_percent: cacheHit(turnCacheRead, turnInput), duration_seconds: Math.round(duration * 1000) / 1000,
                 context_length: s.context_length ?? 0, threshold_tokens: s.threshold_tokens ?? 0, last_prompt_tokens: s.last_prompt_tokens ?? 0,
             };
-            const doneSession = redactSessionData(this.terminalSessionPayload(s), deps.redactEnabled());
+            const doneSession = (await this.publicTerminalSession(s));
             // TAL-299: the ring's figures match the terminal session's (and a detail reload's).
             for (const key of CONTEXT_USAGE_FIELDS)
                 doneUsage[key] = doneSession[key];
@@ -832,59 +1134,76 @@ export class TurnRunner {
             if (result.tool_limit_reached)
                 donePayload.terminal_reason = 'max_iterations';
             // TAL-397: the live meter ends on the persisted counters and turn rate.
-            put('metering', { session_id: sessionId, usage: meteringUsage(s), tps: doneUsage.tps ?? null, tps_available: doneUsage.tps !== undefined, estimated: false });
-            put('done', donePayload);
+            await put('metering', { session_id: sessionId, usage: meteringUsage(s), tps: doneUsage.tps ?? null, tps_available: doneUsage.tps !== undefined, estimated: false });
+            await put('done', donePayload);
             for (const [event, payload] of steerEvents)
-                put(event, payload);
+                (await put(event, payload));
             // The turn is over: release admission before the title work so a follow-up message is accepted while the
             // (up to two) title prompts run; the channel and journal stay open for the `title` events (Python retired the
             // worker before its daemon-thread title generation).
             this.registry.activeRuns.delete(streamId);
             this.settledStreams.add(streamId);
-            this.startSteerFollowUp(sessionId, leftovers);
+            // TAL-1005: a turn that settled at or above the next turn's threshold starts the compaction now, so the next message
+            // usually finds the session compacted; a follow-up turn admitted meanwhile joins it.
+            void deps.service().ensureCompacted(sessionId);
+            await this.startSteerFollowUp(sessionId, leftovers);
             // Before the title work, so a generated title's own state.db sync lands after this one's provisional title.
             await deps.syncUsage?.(s, usedModel || null);
             // Title work and the goal judge run side by side; `stream_end` follows both, so their frames reach the stream.
             const work = await Promise.allSettled([this.backgroundTitle(s, put), this.continueGoal(s, streamId, put)]);
-            put('stream_end', { session_id: sessionId });
+            await put('stream_end', { session_id: sessionId });
             for (const r of work)
                 if (r.status === 'rejected')
                     throw r.reason;
-            // Python: the last non-error assistant reply with content, else "(no answer produced)".
-            opts.onDone?.(lastAnswer(s.messages) || '(no answer produced)');
+            succeeded = true;
+            await opts.onDone?.(lastAnswer(s.messages));
         }
         catch (error) {
+            await streamWrites;
+            const readOnly = !agentDispatched ? (await deps.service().readOnlyReason(s)) : null;
+            if (readOnly) {
+                ownershipRefused = true;
+                if (this.registry.writebackOwners.get(sessionId) === streamId && s.active_stream_id === streamId) {
+                    // Cancellation persists the prompt and a Stop row; an unstarted, refused turn must add neither.
+                    s.messages = s.messages.filter((m) => !(m.role === 'user' && m._turn_id === streamId));
+                    this.clearPendingTurn(s);
+                    if (isDict(s.extra.completion_acceptances))
+                        s.extra.completion_acceptances = Object.fromEntries(Object.entries(s.extra.completion_acceptances).filter(([, value]) => value !== streamId));
+                    await deps.store.save(s, { touchUpdatedAt: false });
+                }
+                await put('apperror', { type: 'error', terminal_state: 'error', message: readOnly, session_id: sessionId, session: (await this.publicTerminalSession(s)) });
+                return;
+            }
             if (settledAt.value && !(error instanceof SidecarError)) {
                 deps.log(`[webui] ERROR settling turn ${streamId}\n${error instanceof Error ? (error.stack ?? error.message) : String(error)}`);
             }
-            failed = true;
             if (this.registry.cancelled.has(streamId)) {
-                this.finalizeCancelled(s, streamId, opts.ephemeral);
-                put('cancel', this.cancelFrame(sessionId));
+                await this.finalizeCancelled(s, streamId, opts.ephemeral);
+                await put('cancel', (await this.cancelFrame(sessionId)));
                 return;
             }
             const message = error instanceof Error ? error.message : String(error);
             const condition = error instanceof SidecarError ? error.condition : undefined;
-            const classification = classifyProviderError(message, { condition });
-            const payload = providerErrorPayload(message, classification.type, classification.hint, deps.redactEnabled());
-            if (condition === 'sidecar_unavailable' || condition === 'agent_runtime_stale') {
-                payload.type = condition === 'agent_runtime_stale' ? 'agent_runtime_stale' : 'sidecar_unavailable';
-                payload.condition = condition;
-            }
+            const runtime = error instanceof SidecarError && ['sidecar_unavailable', 'agent_runtime_stale', 'agent_incompatible'].includes(error.condition)
+                ? runtimeFailureDetails(error.condition, error.data) : null;
+            const classification = runtime ? { label: 'Error', type: runtime.type, hint: '' } : classifyProviderError(message, { condition });
+            const payload = providerErrorPayload(runtime?.message ?? message, classification.type, classification.hint, deps.redactEnabled());
+            if (runtime)
+                Object.assign(payload, runtime);
             try {
                 let current = s;
                 try {
-                    current = deps.store.get(sessionId);
+                    current = (await deps.store.get(sessionId));
                 }
                 catch {
                     current = s;
                 }
                 if (current.active_stream_id === streamId) {
                     for (const [event, data] of this.takeSteerEventsBefore(streamId, 'apperror'))
-                        put(event, data);
-                    this.persistError(current, streamId, classification.label, payload, activeTurnToken);
+                        (await put(event, data));
+                    await this.persistError(current, streamId, classification.label, payload, activeTurnToken);
                     if (!opts.ephemeral)
-                        payload.session = redactSessionData(this.terminalSessionPayload(current), deps.redactEnabled());
+                        payload.session = (await this.publicTerminalSession(current));
                 }
             }
             catch (persistError) {
@@ -892,15 +1211,20 @@ export class TurnRunner {
             }
             payload.session_id = sessionId;
             payload.old_session_id = sessionId;
-            put('apperror', payload);
+            await put('apperror', payload);
         }
         finally {
-            this.teardown(sessionId, streamId);
+            await this.teardown(sessionId, streamId);
             // After persistence and teardown: a background route's failure cleanup deletes the hidden session, which must
             // not race the error writeback above (a re-saved session would resurface in the sidebar).
-            if (failed)
-                opts.onFailed?.();
-            this.startSteerFollowUp(sessionId, followUp);
+            if (!succeeded)
+                await opts.onFailed?.(failure);
+            // TAL-1005: a Stop or an error after the turn ran on a compaction's result left that job's estimate under a context
+            // its prompt and work may have grown; the idle trigger judges the context measured again.
+            if (!succeeded && !ownershipRefused && !opts.ephemeral)
+                void deps.service().remeasureContext(sessionId).then((measured) => { if (measured)
+                    void deps.service().ensureCompacted(sessionId); });
+            await this.startSteerFollowUp(sessionId, followUp);
         }
     }
     /**
@@ -941,9 +1265,8 @@ export class TurnRunner {
      * after the text (TAL-276), so an attachment-only turn still gives the model a request and its transcript row a
      * distinct identity.
      */
-    buildUserMessage(workspaceCtx, msgText, attachments, workspace, sessionId, imageMode) {
-        const text = workspaceCtx + msgText;
-        const withFiles = () => attachedFilesPrompt(text, attachments);
+    buildUserMessage(msgText, attachments, workspace, sessionId, imageMode) {
+        const withFiles = () => attachedFilesPrompt(msgText, attachments);
         const candidates = attachments.filter(isImageAttachment);
         if (!candidates.length || imageMode !== 'native')
             return withFiles();
@@ -981,18 +1304,23 @@ export class TurnRunner {
         }
         return images ? [{ type: 'text', text: withFiles() }, ...parts] : withFiles();
     }
+    /** The settled session as a terminal frame carries it: the display window, redacted. */
+    async publicTerminalSession(s) {
+        return redactSessionData((await this.terminalSessionPayload(s)), this.deps.redactEnabled());
+    }
     /**
      * Python `_session_payload_with_terminal_window`: the same visible-message window contract as `GET /api/session?
      * msg_limit=` (renderable rows, limited payload shape, windowed tool calls, todo state) with the full count.
      */
-    terminalSessionPayload(s) {
-        const payload = withSessionWireFlags(s.compact({ contextLengthFor: this.deps.service().deps.contextLengthFor, modelOptionFor: this.deps.service().deps.modelOptionFor }), this.registry.liveIds);
-        payload.assistant_name = this.deps.service().assistantName(s);
-        payload.workspace_name = this.deps.service().workspaceNames()(s);
+    async terminalSessionPayload(s) {
+        const payload = (await this.deps.service().wireRow(s));
+        payload.automatic_turns_held = (await this.automaticTurnsHeld(s.session_id));
+        payload.notices = this.sessionNotices(s.session_id);
+        payload.goal = this.sessionGoal(s.session_id);
         const service = this.deps.service();
-        const scened = service.backgroundLinked(s, hydrateAnchorActivityScenes(withToolCallOutcomes(withBackgroundUpdates(withMarkerKinds(withTurnIds(withAttachmentObjects(s.messages))), s), s.tool_calls, s.active_stream_id), s.anchor_activity_scenes, { activeTurnId: s.active_stream_id, clipToolResults: true }));
+        const scened = (await service.backgroundLinked(s, hydrateAnchorActivityScenes(withToolCallOutcomes((await service.backgroundUpdates(s, withMarkerKinds(withTurnIds(withAttachmentObjects(s.messages)), s))), s.tool_calls, s.active_stream_id), s.anchor_activity_scenes, { activeTurnId: s.active_stream_id, clipToolResults: true })));
         const [window, offset] = messageWindowForDisplay(scened, TERMINAL_SSE_VISIBLE_MESSAGE_LIMIT, null);
-        const limited = messagesForLimitedPayload(withBodyExcerpts(withDisplayMedia(window, service.mediaProjector(s)), s.active_stream_id));
+        const limited = messagesForLimitedPayload(withBodyExcerpts(withDisplayMedia(window, (await service.mediaProjector(s))), s.active_stream_id, newestSettledReply(scened, s.active_stream_id), offset));
         payload.messages = limited;
         payload.message_count = s.messages.length;
         payload._messages_offset = offset;
@@ -1002,16 +1330,43 @@ export class TurnRunner {
         payload.compression_reference = compressionReference(s, scened);
         return payload;
     }
-    /** Python `_materialize_pending_user_turn_before_error` + error message append + save. */
-    persistError(s, streamId, label, payload, activeTurnToken, agentRows = []) {
-        const startedAt = s.pending_started_at;
-        this.materializePendingUserTurn(s, activeTurnToken, streamId);
-        const duration = typeof startedAt === 'number' && startedAt > 0 ? Math.max(0, this.deps.now() - startedAt) : null;
+    /** TAL-540, Python `_live_usage_snapshot`: a finished compression's frame, with the counters and the ring's post-compression figures, so clients update before `done`. */
+    compressedFrame(s) {
+        const usage = {
+            ...meteringUsage(s), context_length: s.context_length ?? 0, threshold_tokens: s.threshold_tokens ?? 0, last_prompt_tokens: s.last_prompt_tokens ?? 0,
+            post_compression_context_tokens_estimate: s.post_compression_context_tokens_estimate, ...contextUsage(s, () => this.deps.service().deps.contextLengthFor(s.model, s.model_provider)),
+        };
+        const sid = s.session_id;
+        return { session_id: sid, old_session_id: sid, new_session_id: sid, continuation_session_id: sid, message: 'Compression finished', usage };
+    }
+    /** Release only the pending admission metadata; transcript settlement belongs to the caller. */
+    clearPendingTurn(s) {
         s.active_stream_id = null;
         s.pending_user_message = null;
         s.pending_attachments = [];
         s.pending_started_at = null;
         s.pending_user_source = null;
+    }
+    /** Python `_materialize_pending_user_turn_before_error` + error message append + save. */
+    /** `runtimeProvider`: the provider the Agent resolved for the turn, when it reported one. */
+    async persistError(s, streamId, label, payload, activeTurnToken, agentRows = [], runtimeProvider = '') {
+        const startedAt = s.pending_started_at;
+        // TAL-576: an automatic wakeup with no usable credentials pauses the next ones instead of failing each.
+        if (s.pending_user_source === 'process_wakeup' && payload.type === 'credential_pool_empty') {
+            // A profile whose home now escapes the profiles root (`profileHome` throws) records no pause; the error still persists.
+            let home = null;
+            try {
+                home = this.deps.profileHome(s.profile);
+            }
+            catch {
+                home = null;
+            }
+            if (home !== null)
+                recordWakeupPause(s, home, this.deps.now(), runtimeProvider, this.deps.configPath?.(home));
+        }
+        this.materializePendingUserTurn(s, activeTurnToken, streamId);
+        const duration = typeof startedAt === 'number' && startedAt > 0 ? Math.max(0, this.deps.now() - startedAt) : null;
+        this.clearPendingTurn(s);
         this.appendPartialSnapshot(s, streamId);
         const hint = str(payload.hint);
         const errorMessage = { role: 'assistant', content: `**${label}:** ${str(payload.message) || label}${hint ? `\n\n*${hint}*` : ''}`, timestamp: Math.trunc(this.deps.now()), _error: true, _turn_id: streamId };
@@ -1035,10 +1390,15 @@ export class TurnRunner {
         s.messages.push(errorMessage);
         this.keepLiveState(s, streamId);
         this.persistConsumedSteers(s, streamId, startedAt, this.deps.now());
-        const stop = this.stopContexts.get(streamId);
-        this.deps.service().settleStateDb(s, { turnId: streamId, previousContext: stop?.previousContext ?? [], agentRows, startId: stop?.stateDbStartId ?? null });
+        // TAL-921: the next prompt remembers the turn's work; an exhausted compression's work would only overflow again. One
+        // read builds the context and settles, so a row committed in between is never covered without reaching either.
+        const read = this.deps.service().stateDbRead(s);
+        const context = payload.type === 'compression_exhausted' ? null : this.unsettledContext(streamId, agentRows, read);
+        if (context)
+            s.context_messages = context;
+        this.settleStateDb(s, streamId, agentRows, { read });
         try {
-            this.deps.store.save(s);
+            (await this.deps.store.save(s));
         }
         catch (error) {
             this.deps.log(`[webui] WARNING: failed to save error turn for ${s.session_id}: ${error.message}`);
@@ -1049,6 +1409,24 @@ export class TurnRunner {
         this.deps.pending.clearClarifies(s.session_id);
         this.deps.events.publish('session_error', { profile: s.profile, sessionId: s.session_id });
     }
+    /** Every terminal path carries the same starting read's id and generation into settlement. */
+    settleStateDb(s, streamId, agentRows, opts = {}) {
+        const stop = this.stopContexts.get(streamId);
+        const service = this.deps.service();
+        service.settleStateDb(s, { turnId: streamId, previousContext: stop?.previousContext ?? opts.previousContext ?? [], agentRows, startId: stop?.stateDbStartId ?? null, startGeneration: stop?.stateDbStartGeneration ?? null, read: opts.read ?? service.stateDbRead(s) });
+    }
+    /**
+     * TAL-921: the model context a turn that ends without its settled result leaves: the work its Agent reported, else the
+     * rows it committed past its starting read in `read`, the read its settlement uses.
+     */
+    unsettledContext(streamId, agentRows, read) {
+        const stop = this.stopContexts.get(streamId);
+        if (!stop)
+            return null;
+        const after = stop.stateDbStartGeneration !== null && read.generation !== null && stop.stateDbStartGeneration !== read.generation ? -1 : stop.stateDbStartId;
+        const stateRows = after === null ? [] : read.rows.filter((m) => typeof m._state_db_row_id === 'number' && m._state_db_row_id > after);
+        return this.deps.service().unsettledTurnContext({ ...stop, streamedText: (this.registry.partialText.get(streamId) ?? []).join(''), agentRows, stateRows });
+    }
     materializePendingUserTurn(s, activeTurnToken, turnId) {
         const pendingText = str(s.pending_user_message);
         if (!s.hasPendingPrompt)
@@ -1094,11 +1472,11 @@ export class TurnRunner {
      * Python `_emit_cancel_event`: the terminal frame carries the settled session so tabs can render the partial without a
      * reload. TAL-364: no copy — clients show their one localized status for the `cancelled` outcome.
      */
-    cancelFrame(sessionId) {
+    async cancelFrame(sessionId) {
         let snapshot = null;
         if (sessionId)
             try {
-                snapshot = redactSessionData(this.terminalSessionPayload(this.deps.store.get(sessionId)), this.deps.redactEnabled());
+                snapshot = (await this.publicTerminalSession((await this.deps.store.get(sessionId))));
             }
             catch {
                 snapshot = null;
@@ -1110,12 +1488,12 @@ export class TurnRunner {
      * turn's model context keeps what the Agent had captured (`checkpoint`, its canonical transcript) and the prose that
      * streamed, so the next request continues from it; the terminal row carries the outcome and no copy.
      */
-    finalizeCancelled(s, streamId, ephemeral, checkpoint = null) {
+    async finalizeCancelled(s, streamId, ephemeral, checkpoint = null) {
         if (this.registry.writebackOwners.get(s.session_id) !== streamId)
             return false;
         let current = s;
         try {
-            current = this.deps.store.get(s.session_id);
+            current = (await this.deps.store.get(s.session_id));
         }
         catch {
             return false;
@@ -1124,7 +1502,7 @@ export class TurnRunner {
             return false;
         if (ephemeral) {
             try {
-                rmSync(this.deps.store.pathFor(current.session_id), { force: true });
+                await this.deps.store.delete(current.session_id, { tombstone: false });
             }
             catch { /* ignore */ }
             this.deps.store.sessions.delete(current.session_id);
@@ -1136,9 +1514,9 @@ export class TurnRunner {
         if (quiet && current.active_stream_id === null) {
             // TAL-493: rows its worker committed while unwinding are covered by what it reports; others are kept.
             if (checkpoint) {
-                this.deps.service().settleStateDb(current, { turnId: streamId, previousContext: this.stopContexts.get(streamId)?.previousContext ?? [], agentRows: checkpoint, startId: this.stopContexts.get(streamId)?.stateDbStartId ?? null });
+                this.settleStateDb(current, streamId, checkpoint);
                 try {
-                    this.deps.store.save(current);
+                    (await this.deps.store.save(current));
                 }
                 catch {
                     return false;
@@ -1148,26 +1526,28 @@ export class TurnRunner {
         }
         const stop = this.stopContexts.get(streamId);
         const canonical = stop !== undefined && checkpoint !== null && checkpointTurnStart(checkpoint, stop.msgText, stop.historyLength) !== null;
-        const settle = () => {
+        // TAL-921: one state.db read builds the context and settles the turn.
+        const settle = (read) => {
             if (!stop)
                 return null;
             if (canonical)
                 stop.checkpointed = true;
-            return stoppedTurnContext(stop.previousContext, checkpoint, stop.prompt, stop.msgText, (this.registry.partialText.get(streamId) ?? []).join(''), stop.historyLength);
+            return this.unsettledContext(streamId, checkpoint, read);
         };
         if (current.messages.some((m) => isCancelMarker(m)) && current.active_stream_id === null && !current.pending_user_message) {
             // The worker's canonical result can arrive after cancel() settled without a checkpoint (the interrupt reply failed
             // or timed out): it replaces this stream's context and never adds a second Stop row. A settlement that already had
             // one keeps it, so work the Agent finished after the Stop stays out.
             const ownStop = current.messages.some((m) => isCancelMarker(m) && m._turn_id === streamId);
-            const late = canonical && !stop.checkpointed && ownStop ? settle() : null;
-            if (late)
-                current.context_messages = dedupeContext(late);
             // TAL-493: the worker's rows committed after the first settlement are covered by what it reports; others are kept.
-            if (late || (ownStop && checkpoint)) {
-                this.deps.service().settleStateDb(current, { turnId: streamId, previousContext: stop?.previousContext ?? [], agentRows: checkpoint, startId: stop?.stateDbStartId ?? null });
+            if (ownStop && checkpoint) {
+                const read = this.deps.service().stateDbRead(current);
+                const late = canonical && !stop.checkpointed ? settle(read) : null;
+                if (late)
+                    current.context_messages = dedupeContext(late);
+                this.settleStateDb(current, streamId, checkpoint, { read });
                 try {
-                    this.deps.store.save(current);
+                    (await this.deps.store.save(current));
                 }
                 catch {
                     return false;
@@ -1177,23 +1557,20 @@ export class TurnRunner {
         }
         const startedAt = current.pending_started_at;
         this.materializePendingUserTurn(current, buildActiveTurnToken(streamId, current.pending_started_at), streamId);
-        const context = settle();
+        const read = this.deps.service().stateDbRead(current);
+        const context = settle(read);
         if (context)
             current.context_messages = dedupeContext(context);
-        current.active_stream_id = null;
-        current.pending_user_message = null;
-        current.pending_attachments = [];
-        current.pending_started_at = null;
-        current.pending_user_source = null;
+        this.clearPendingTurn(current);
         if (!quiet) {
             this.appendPartialSnapshot(current, streamId);
             current.messages.push({ role: 'assistant', content: '', _error: true, _terminal_state: 'cancelled', timestamp: Math.trunc(this.deps.now()), _turn_id: streamId });
         }
         this.keepLiveState(current, streamId);
         this.persistConsumedSteers(current, streamId, startedAt, this.deps.now());
-        this.deps.service().settleStateDb(current, { turnId: streamId, previousContext: stop?.previousContext ?? [], agentRows: checkpoint, startId: stop?.stateDbStartId ?? null });
+        this.settleStateDb(current, streamId, checkpoint, { read });
         try {
-            this.deps.store.save(current);
+            (await this.deps.store.save(current));
         }
         catch {
             return false;
@@ -1205,8 +1582,8 @@ export class TurnRunner {
         return true;
     }
     /** Push a frame onto the session's live chat stream from outside the turn (Python gateway notify callbacks). */
-    emitToSession(sessionId, event, data) {
-        this.sessionPuts.get(sessionId)?.(event, data);
+    async emitToSession(sessionId, event, data) {
+        await this.sessionPuts.get(sessionId)?.(event, data);
     }
     /** TAL-397: a turn that ends without its result keeps what the live stream already showed: counters and todo list. */
     keepLiveState(s, streamId) {
@@ -1217,13 +1594,15 @@ export class TurnRunner {
         if (todos)
             s.extra[UNSETTLED_TODO_KEY] = todos;
     }
-    teardown(sessionId, streamId) {
+    async teardown(sessionId, streamId) {
+        this.runStatuses.delete(streamId);
         this.liveUsage.delete(streamId);
+        this.compressionSpans.delete(streamId);
         this.liveTodos.delete(streamId);
         const writer = this.writers.get(streamId);
         if (writer) {
             try {
-                writer.close();
+                (await writer.close());
             }
             catch { /* ignore */ }
             this.writers.delete(streamId);
@@ -1252,21 +1631,27 @@ export class TurnRunner {
         catch { /* best effort */ }
     }
     /** Journal files with a live writer (retention must leave them alone). */
-    activeJournalPaths() {
-        return new Set([...this.writers.values()].map((w) => this.deps.journal.pathFor(w.sessionId, w.runId)));
-    }
     // ── title ────────────────────────────────────────────────────────────────
     /** Python `_generate_llm_session_title_via_aux`: aux-route statuses (`llm_aux`, `llm_aux_retry`, `llm_*_aux`). */
     async llmTitle(s, userText, assistantText) {
         const sidecar = this.deps.sidecar();
         if (!sidecar)
             return { title: '', status: 'llm_error_aux', rawPreview: '' };
+        const service = this.deps.service();
+        await service.deps.warmSessionModelRepair?.(s.profile);
+        const [model, provider] = service.turnModel(s);
         const [qa, prompts] = titlePrompts(userText, assistantText);
         let status = 'llm_error_aux';
         for (const [idx, prompt] of prompts.entries()) {
+            if ((await this.deps.service().isReadOnly(s)))
+                return { title: '', status: 'read_only', rawPreview: '' };
             let raw = '';
             try {
-                const result = await sidecar.call('aux.complete', { profile_home: this.deps.profileHome(s.profile), task: 'title_generation', messages: [{ role: 'system', content: prompt }, { role: 'user', content: qa }], max_tokens: 64 }, { timeoutMs: 15_000 });
+                const result = await sidecar.call('aux.complete', {
+                    profile_home: this.deps.profileHome(s.profile), task: 'title_generation',
+                    messages: [{ role: 'system', content: prompt }, { role: 'user', content: qa }], max_tokens: 64,
+                    main_runtime: { model: model ?? '', provider }, main_fallback: true,
+                });
                 raw = str(result.text);
             }
             catch (error) {
@@ -1287,6 +1672,42 @@ export class TurnRunner {
         }
         return { title: '', status, rawPreview: '' };
     }
+    /**
+     * TAL-902: `/refine [focus]` starts the Agent's memory/skill review of the chat's model context on its agent
+     * (`chat.refine`, the agent a turn would build); no turn starts. Without a chat, while a turn runs, or before anything
+     * was said it answers with the Agent's wording instead.
+     */
+    async refine(s, focus) {
+        if (!s)
+            return 'Nothing to refine yet — send a message first.';
+        if (s.active_stream_id)
+            return 'Agent is running — wait for the turn to finish, then /refine.';
+        const messages = sanitizeMessagesForApi(this.deps.service().modelContext(s), { preserveApiContent: true, issuedReasoningSince: s.reasoning_replay_from });
+        if (!messages.length)
+            return 'Nothing to refine yet — the conversation is empty.';
+        // Like a turn: no agent is built in a profile whose home is going away.
+        if (this.deps.profileDeleting?.(s.profile ?? null))
+            throw new HttpFailure(409, `Profile '${str(s.profile)}' is being deleted.`);
+        const sidecar = this.deps.sidecar();
+        if (!sidecar)
+            throw new SidecarError('The Agent sidecar is not running; /refine is unavailable until it starts.', { condition: 'sidecar_unavailable' });
+        const [model, provider] = this.deps.service().turnModel(s);
+        const cfg = (await this.deps.profileConfig?.(s.profile ?? null)) ?? {};
+        this.deps.service().assertWritable(s);
+        return (await sidecar.call('chat.refine', {
+            profile_home: this.deps.profileHome(s.profile), session_id: s.session_id, messages, focus, model: model ?? '', model_provider: provider,
+            enabled_toolsets: this.deps.toolsetsFor(s), ...agentBuildParams(cfg, 'webui', provider, this.deps.env ?? {}),
+            ...this.agentWorkspace(s, (await this.deps.service().turnWorkspace(s))),
+        }, { timeoutMs: 120_000 })).output;
+    }
+    /**
+     * TAL-903: the workspace a session's agent is built at, and whether it is a host path (the profile's terminal runs
+     * locally). Only then does the Agent anchor AGENTS.md discovery and its memory provider there; `chat.start` and
+     * `chat.refine` send the same pair, so the agent `/refine` builds is the one the next turn reuses.
+     */
+    agentWorkspace(s, workspace) {
+        return { workspace, workspace_local: this.deps.workspaces.profileSupportsLocalIo(s.profile) };
+    }
     /**
      * Python `generate_session_title_for_session`: on-demand title from the persisted transcript; never touches
      * `llm_title_generated`. An explicit regenerate runs even with `auxiliary.title_generation.enabled` off (TAL-531).
@@ -1300,6 +1721,7 @@ export class TurnRunner {
         let rawPreview = '';
         if (assistantText)
             ({ title: next, status: llmStatus, rawPreview } = await this.llmTitle(s, userText, assistantText));
+        this.deps.service().assertWritable(s);
         if (next)
             return { title: next, status: llmStatus, rawPreview };
         const fallback = fallbackTitleFromExchange(userText, assistantText);
@@ -1307,6 +1729,15 @@ export class TurnRunner {
             return { title: fallback, status: `local_summary:${llmStatus}`, rawPreview };
         return { title: null, status: llmStatus || 'empty_title', rawPreview };
     }
+    /**
+     * TAL-591 (Python `_queue_generated_title_for_imported_session`): a CLI session a mutation claims while it still
+     * carries its default title gets the background title job at once instead of after its first WebUI turn.
+     */
+    async titleClaimed(s) {
+        if (!looksLikeDefaultCliTitle({ title: s.title, source_tag: s.source_tag, raw_source: s.raw_source, session_source: s.session_source, source_label: s.source_label }))
+            return;
+        await this.backgroundTitle(s, () => Promise.resolve()).catch((error) => { this.deps.log(`[webui] imported session title failed for ${s.session_id}: ${error.message}`); });
+    }
     /** Python `_run_background_title_update` (spawned only when `_background_title_generation_inputs` is eligible). */
     async backgroundTitle(s, put) {
         const sessionId = s.session_id;
@@ -1317,27 +1748,34 @@ export class TurnRunner {
         const eligible = ['Untitled', 'New Chat', ''].includes(placeholder) || placeholder === titleFrom(s.messages, '') || invalidExisting || defaultTitle(s);
         const [userText, assistantText] = firstExchangeSnippets(s.messages);
         // Python `_put_title_status`: empty reason/title/raw_preview keys are omitted.
-        const status = (status, reason = '', title = '', rawPreview = '') => {
-            put('title_status', { session_id: sessionId, status, ...(reason ? { reason } : {}), ...(title ? { title } : {}), ...(rawPreview ? { raw_preview: rawPreview } : {}) });
+        const status = async (status, reason = '', title = '', rawPreview = '') => {
+            await put('title_status', { session_id: sessionId, status, ...(reason ? { reason } : {}), ...(title ? { title } : {}), ...(rawPreview ? { raw_preview: rawPreview } : {}) });
         };
-        if (!eligible || (s.llm_title_generated && !invalidExisting) || !userText || !assistantText)
+        // Python ran the adaptive refresh only when the first-title job was not eligible.
+        if (!eligible || (s.llm_title_generated && !invalidExisting) || !userText || !assistantText) {
+            await this.refreshTitle(s, put, status);
             return;
+        }
         if (s.manual_title) {
-            status('skipped', 'manual_title', placeholder);
+            (await status('skipped', 'manual_title', placeholder));
             return;
         }
         // A missing config.yaml reads as `{}` (enabled); an unreadable one cannot confirm the user has not opted out.
         const cfg = await this.deps.profileConfig?.(s.profile ?? null);
         if (!cfg) {
-            status('skipped', 'config_unavailable', placeholder);
+            (await status('skipped', 'config_unavailable', placeholder));
             return;
         }
         if (!titleGenerationEnabled(cfg)) {
-            status('skipped', 'title_generation_disabled', placeholder);
+            (await status('skipped', 'title_generation_disabled', placeholder));
             return;
         }
         const generated = await this.llmTitle(s, userText, assistantText);
         const { status: llmStatus, rawPreview } = generated;
+        if (llmStatus === 'read_only') {
+            (await status('skipped', 'read_only', placeholder));
+            return;
+        }
         let next = generated.title;
         let source = llmStatus;
         if (!next) {
@@ -1350,23 +1788,30 @@ export class TurnRunner {
         const fallbackReason = source === 'fallback' && llmStatus ? `local_summary:${llmStatus}` : 'local_summary';
         let current = s;
         try {
-            current = this.deps.store.get(sessionId);
+            current = (await this.deps.store.get(sessionId));
         }
         catch {
+            // A session deleted while the title prompt ran stays deleted.
+            if ((await this.deps.store.wasDeleted(sessionId)))
+                return;
             current = s;
         }
+        if ((await this.deps.service().isReadOnly(current))) {
+            (await status('skipped', 'read_only', str(current.title).trim()));
+            return;
+        }
         let effective = str(current.title).trim();
         let wrote = false;
         if (next) {
             const stillAuto = effective === placeholder || ['Untitled', 'New Chat', ''].includes(effective) || effective === titleFrom(current.messages, '') || looksInvalidGeneratedTitle(current.title) || defaultTitle(current);
             if (current.manual_title || !stillAuto) {
-                status('skipped', 'manual_title', effective);
+                (await status('skipped', 'manual_title', effective));
                 return;
             }
             if (next !== effective) {
                 current.title = next;
                 markSessionTitleGenerated(current);
-                this.deps.store.save(current, { touchUpdatedAt: false });
+                await this.deps.store.save(current, { touchUpdatedAt: false });
                 // Python `sync_session_title` after generation: the state.db row follows when `sync_to_insights` is on.
                 await this.deps.syncTitle?.(current);
                 this.deps.events.publish('title', { profile: current.profile, sessionId });
@@ -1375,11 +1820,80 @@ export class TurnRunner {
             }
         }
         if (wrote) {
-            status(source, source === 'fallback' ? fallbackReason : llmStatus, effective, rawPreview);
-            put('title', { session_id: sessionId, title: effective });
+            await status(source, source === 'fallback' ? fallbackReason : llmStatus, effective, rawPreview);
+            await put('title', { session_id: sessionId, title: effective });
         }
         else
-            status('skipped', source || 'unchanged', effective, rawPreview);
+            (await status('skipped', source || 'unchanged', effective, rawPreview));
+    }
+    /**
+     * Python `_maybe_schedule_title_refresh` + `_run_background_title_refresh`: every N exchanges a generated title is
+     * regenerated from the latest exchange; a manual title is never touched.
+     */
+    async refreshTitle(s, put, status) {
+        const every = this.deps.titleRefreshEvery?.() ?? 0;
+        const exchanges = countExchanges(s.messages);
+        if (every <= 0 || exchanges <= 0 || exchanges % every !== 0)
+            return;
+        const [userText, assistantText] = latestExchangeSnippets(s.messages);
+        if (!userText || !assistantText)
+            return;
+        // Admission is released before title work, so an earlier turn's refresh may still be running: refreshes apply in
+        // turn order and each reads the title its predecessor wrote.
+        const sessionId = s.session_id;
+        const run = (this.titleRefreshes.get(sessionId) ?? Promise.resolve()).catch(() => undefined).then(async () => { await this.applyTitleRefresh(sessionId, s, userText, assistantText, put, status); });
+        this.titleRefreshes.set(sessionId, run);
+        try {
+            await run;
+        }
+        finally {
+            if (this.titleRefreshes.get(sessionId) === run)
+                this.titleRefreshes.delete(sessionId);
+        }
+    }
+    async applyTitleRefresh(sessionId, s, userText, assistantText, put, status) {
+        const read = async () => { try {
+            return (await this.deps.store.get(sessionId));
+        }
+        catch {
+            return s;
+        } };
+        const latest = (await read());
+        const placeholder = str(latest.title).trim();
+        if (['Untitled', 'New Chat', ''].includes(placeholder) || latest.manual_title || !latest.llm_title_generated)
+            return;
+        const cfg = await this.deps.profileConfig?.(latest.profile ?? null);
+        if (!cfg || !titleGenerationEnabled(cfg)) {
+            (await status('refresh_skipped', cfg ? 'title_generation_disabled' : 'config_unavailable', placeholder));
+            return;
+        }
+        const { title: next, status: llmStatus, rawPreview } = await this.llmTitle(latest, userText, assistantText);
+        if (!next) {
+            (await status('refresh_skipped', llmStatus || 'empty', placeholder, rawPreview));
+            return;
+        }
+        const normalize = (value) => value.replace(/\s+/g, ' ').trim().toLowerCase();
+        if (normalize(next) === normalize(placeholder)) {
+            (await status('refresh_skipped', 'same_title', placeholder, rawPreview));
+            return;
+        }
+        const current = (await read());
+        // The user may have renamed the session while the title prompt ran.
+        if (current.manual_title || str(current.title).trim() !== placeholder) {
+            (await status('skipped', 'manual_title', str(current.title).trim()));
+            return;
+        }
+        if ((await this.deps.service().isReadOnly(current))) {
+            (await status('refresh_skipped', 'read_only', str(current.title).trim()));
+            return;
+        }
+        current.title = next;
+        markSessionTitleGenerated(current);
+        await this.deps.store.save(current, { touchUpdatedAt: false });
+        await this.deps.syncTitle?.(current);
+        this.deps.events.publish('title', { profile: current.profile, sessionId });
+        await status('refreshed', llmStatus, next, rawPreview);
+        await put('title', { session_id: sessionId, title: next });
     }
     // ── cancel / steer ───────────────────────────────────────────────────────
     /**
@@ -1387,6 +1901,26 @@ export class TurnRunner {
      * `steer_withdrawn` (stopped) this Stop emitted, for the cancel response (TAL-426).
      */
     async cancel(streamId) {
+        const ownerCancelled = await this.deps.service().owner?.cancel(streamId);
+        if (ownerCancelled !== null && ownerCancelled !== undefined)
+            return { cancelled: ownerCancelled, withdrawn: [] };
+        // TAL-899: the user's Stop holds automatic turns, even one that lands after `done`, unless a newer turn already
+        // took the session: a late or retried Stop for a turn the user followed holds nothing.
+        const owner = this.registry.ownerSessionId(streamId) ?? this.registry.activeRuns.get(streamId)?.session_id;
+        if (owner && this.registry.writebackOwners.get(owner) === streamId) {
+            this.stopHeld.add(owner);
+            try {
+                const session = (await this.deps.store.get(owner));
+                session.automatic_turns_stop_held = true;
+                await this.deps.store.save(session, { touchUpdatedAt: false });
+            }
+            catch (error) {
+                this.deps.log(`[webui] WARNING: failed to persist Stop hold for session ${owner}; interrupting with the in-memory hold: ${error.message}`);
+            }
+        }
+        return (await this.stop(streamId));
+    }
+    async stop(streamId) {
         const channel = this.registry.peek(streamId);
         const run = this.registry.activeRuns.get(streamId);
         if (!channel && !run)
@@ -1421,24 +1955,24 @@ export class TurnRunner {
         if (sessionId) {
             let current = null;
             try {
-                current = this.deps.store.get(sessionId);
+                current = (await this.deps.store.get(sessionId));
             }
             catch {
                 current = null;
             }
             if (current?.active_stream_id === streamId) {
-                this.finalizeCancelled(current, streamId, run?.ephemeral, checkpoint);
+                await this.finalizeCancelled(current, streamId, run?.ephemeral, checkpoint);
             }
         }
         // The worker may already have written its terminal row while the interrupt was in flight; the stream is then
         // gone from the registry and this path must not add a second one.
         if (channel && this.registry.streams.has(streamId)) {
             const writer = this.writers.get(streamId) ?? null;
-            const emit = (event, data) => {
+            const emit = async (event, data) => {
                 let eventId = null;
                 if (writer) {
                     try {
-                        eventId = writer.appendSseEvent(event, data).event_id;
+                        eventId = (await writer.appendSseEvent(event, data)).event_id;
                     }
                     catch {
                         eventId = null;
@@ -1447,8 +1981,10 @@ export class TurnRunner {
                 channel.put([event, data, eventId]);
             };
             for (const [event, data] of steerEvents)
-                emit(event, data);
-            emit('cancel', this.cancelFrame(sessionId));
+                (await emit(event, data));
+            this.runStatuses.delete(streamId);
+            await emit('run_status', { session_id: sessionId, status: null });
+            await emit('cancel', (await this.cancelFrame(sessionId)));
             this.registry.streams.delete(streamId);
             this.registry.liveIds.delete(streamId);
         }
@@ -1464,7 +2000,7 @@ export class TurnRunner {
         const streamId = this.registry.activeRunStreamForSession(sessionId);
         if (!streamId || this.registry.activeRuns.get(streamId)?.origin !== 'background')
             return false;
-        await this.cancel(streamId);
+        await this.stop(streamId);
         let timer;
         await Promise.race([this.workers.get(streamId), new Promise((resolve) => { timer = setTimeout(resolve, BACKGROUND_UNWIND_WAIT_MS); })]);
         clearTimeout(timer);
@@ -1473,7 +2009,7 @@ export class TurnRunner {
     async steer(sessionId, text, displayText, steerId) {
         let s = null;
         try {
-            s = this.deps.store.get(sessionId, { metadataOnly: true });
+            s = (await this.deps.store.get(sessionId, { metadataOnly: true }));
         }
         catch {
             s = null;
@@ -1492,7 +2028,7 @@ export class TurnRunner {
             return { accepted: false, fallback: 'no_cached_agent', stream_id: null };
         // Register provisionally before the RPC: the Agent may apply the steer and finish the turn before the steer
         // reply arrives, and the turn's finalisation must then already see the record. A rejection removes it again.
-        return this.withSteerChain(activeStreamId, () => this.steerInto(sidecar, sessionId, activeStreamId, text, displayText, steerId));
+        return this.withSteerChain(activeStreamId, sessionId, () => this.steerInto(sidecar, sessionId, activeStreamId, text, displayText, steerId));
     }
     async steerInto(sidecar, sessionId, activeStreamId, text, displayText, steerId) {
         const record = { steer_id: steerId, session_id: sessionId, stream_id: activeStreamId, text, display_text: displayText || text, created_at: this.deps.now(), state: 'pending', can_redirect: false };
@@ -1524,7 +2060,7 @@ export class TurnRunner {
             record.can_redirect = result.can_redirect === true;
             // TAL-424: every client shows it until the Agent takes it (unless it already did while the reply was in flight).
             if (this.steers.get(activeStreamId)?.includes(record))
-                this.emitToSession(sessionId, 'steer_pending', this.pendingSteerView(record));
+                (await this.emitToSession(sessionId, 'steer_pending', this.pendingSteerView(record)));
         }
         return { accepted: result.accepted, fallback: result.accepted ? null : (result.fallback ?? 'not_running'), stream_id: activeStreamId, steer_id: steerId };
     }
@@ -1593,10 +2129,10 @@ export class TurnRunner {
         return boundary;
     }
     /** A steer is saved as it enters the stream, so a reload mid-turn already has it; the terminal save re-places it. */
-    saveConsumedSteers(sessionId, streamId) {
+    async saveConsumedSteers(sessionId, streamId) {
         let s;
         try {
-            s = this.deps.store.get(sessionId);
+            s = (await this.deps.store.get(sessionId));
         }
         catch {
             return;
@@ -1604,7 +2140,7 @@ export class TurnRunner {
         if (s.active_stream_id !== streamId || this.placeConsumedSteers(s, streamId, s.pending_started_at) === null)
             return;
         try {
-            this.deps.store.save(s, { touchUpdatedAt: false });
+            (await this.deps.store.save(s, { touchUpdatedAt: false }));
         }
         catch (error) {
             this.deps.log(`[webui] WARNING: failed to save consumed steer for ${sessionId}: ${error.message}`);
@@ -1658,30 +2194,24 @@ export class TurnRunner {
             return { events, leftovers: [...(unmatched ? [unmatched] : []), ...left.map((r) => r.text)] };
         if (unmatched)
             events.push(['steer_withdrawn', { steer_id: null, reason: 'stopped', text: unmatched }]);
-        // ponytail: legacy `pending_steer_leftover` keeps today's Web and App from losing a stopped steer (they requeue it);
-        // drop it once TAL-425 / TAL-426 put `steer_withdrawn` text back in the composer.
-        for (const r of left)
-            events.push(['pending_steer_leftover', { steer_id: r.steer_id, stream_id: streamId, text: r.text, display_text: r.display_text, leftover_at: this.deps.now() }]);
-        if (unmatched)
-            events.push(['pending_steer_leftover', { stream_id: streamId, text: unmatched }]);
         return { events, leftovers: [] };
     }
     /**
      * TAL-424: the steers a finished turn never took go on as one follow-up turn, sent once by the server for every client.
      * Should another turn already hold the session, they steer into that one instead, so no text is lost.
      */
-    startSteerFollowUp(sessionId, texts) {
+    async startSteerFollowUp(sessionId, texts) {
         if (!texts.length)
             return;
         let s;
         try {
-            s = this.deps.store.get(sessionId);
+            s = (await this.deps.store.get(sessionId));
         }
         catch {
             return;
         }
         const msg = texts.join('\n');
-        const started = this.start(s, { msg, attachments: [], workspace: s.workspace, model: s.model, modelProvider: s.model_provider });
+        const started = (await this.startServerTurn(s, { msg }));
         if (!started._status || started._status < 400)
             return;
         if (started._status === 409 && started.active_stream_id) {
@@ -1709,61 +2239,83 @@ export class TurnRunner {
         // Only a known goal turn shows progress up front; any other turn shows nothing unless a goal turns out to be active.
         const goalTurn = this.registry.goalRelated.has(streamId);
         if (goalTurn)
-            put('goal', { session_id: sessionId, state: 'evaluating', message: 'Evaluating goal progress…', message_key: 'goal_evaluating_progress' });
+            (await put('goal', { session_id: sessionId, state: 'evaluating', message: 'Evaluating goal progress…', message_key: 'goal_evaluating_progress' }));
         let decision;
         try {
-            decision = await sidecar.call('goals.evaluate', { session_id: sessionId, profile_home: this.deps.profileHome(s.profile), last_response: lastResponse, user_initiated: true });
+            // The workspace the turn ran in, unrecovered: the Agent pauses the goal rather than run its gates anywhere else.
+            this.deps.service().assertWritable(s);
+            decision = await sidecar.call('goals.evaluate', { session_id: sessionId, profile_home: this.deps.profileHome(s.profile), last_response: lastResponse, user_initiated: true, workspace: str(s.workspace) || undefined });
         }
         catch (error) {
             this.deps.log(`[webui] WARNING: goal evaluation failed for ${sessionId}: ${error.message}`);
             if (goalTurn)
-                put('goal', { session_id: sessionId, state: 'idle', decision: 'error', message: '' });
+                (await put('goal', { session_id: sessionId, state: 'idle', decision: 'error', message: '' }));
             return;
         }
+        // The judgement moved the goal on (turns, park, pause, done); an inactive one only matters if a goal still shows.
+        if (decision.verdict !== 'inactive' || this.goals.view(sessionId))
+            await this.refreshGoal(s, put);
         if (!goalTurn && decision.verdict === 'inactive')
             return;
         const prompt = str(decision.continuation_prompt).trim();
         const proceed = decision.should_continue && Boolean(prompt);
         const status = { message: decision.message, message_key: decision.message_key, message_args: decision.message_args ?? [], decision: decision.verdict };
-        put('goal', { session_id: sessionId, state: proceed ? 'continuing' : 'idle', ...status });
-        if (!proceed)
+        if (!proceed) {
+            (await put('goal', { session_id: sessionId, state: 'idle', ...status }));
             return;
+        }
         let current;
         try {
-            current = this.deps.store.get(sessionId);
+            current = (await this.deps.store.get(sessionId));
         }
         catch {
             return;
         }
-        const started = this.start(current, { msg: prompt, attachments: [], workspace: current.workspace, model: current.model, modelProvider: current.model_provider, goalRelated: true });
+        const started = (await this.startServerTurn(current, { msg: prompt, goalRelated: true }));
+        // TAL-899: a Stop, even one after the answer, holds the continuation; the user's next turn is judged instead.
+        if (started.held) {
+            (await put('goal', { session_id: sessionId, state: 'idle', decision: 'held', message: '' }));
+            return;
+        }
+        await put('goal', { session_id: sessionId, state: 'continuing', ...status });
         if (started._status === 409 && started.active_stream_id) {
             this.registry.goalRelated.add(started.active_stream_id);
             return;
         }
         if (started._status !== undefined && started._status >= 400) {
             this.deps.log(`[webui] WARNING: goal continuation for ${sessionId} not started (${String(started._status)}): ${str(started.error)}`);
-            put('goal', { session_id: sessionId, state: 'idle', decision: 'error', message: '' });
+            await put('goal', { session_id: sessionId, state: 'idle', decision: 'error', message: '' });
             return;
         }
-        put('goal_continue', { session_id: sessionId, state: 'continuing', continuation_prompt: prompt, text: prompt, stream_id: str(started.stream_id), ...status });
+        await put('goal_continue', { session_id: sessionId, state: 'continuing', continuation_prompt: prompt, text: prompt, stream_id: str(started.stream_id), ...status });
     }
     /** The Agent's pending steer text settles which steers it has taken: those become `steer_consumed` and their rows. */
-    settleConsumedSteers(sessionId, streamId, pendingText) {
+    async settleConsumedSteers(sessionId, streamId, pendingText) {
         for (const record of this.takeConsumedSteers(streamId, pendingText, { keepLeftovers: true }))
-            this.emitToSession(sessionId, 'steer_consumed', record);
-        this.saveConsumedSteers(sessionId, streamId);
+            (await this.emitToSession(sessionId, 'steer_consumed', record));
+        await this.saveConsumedSteers(sessionId, streamId);
     }
     /** What every client shows for a pending steer; the server decides which actions it offers. */
     pendingSteerView(r) {
         const open = r.state === 'pending';
         return { steer_id: r.steer_id, text: r.display_text, submitted_at: r.created_at, state: r.state, actions: { edit: open, cancel: open, send_now: open && r.can_redirect } };
     }
+    runStatus(streamId) {
+        return this.runStatuses.get(streamId) ?? null;
+    }
+    /** TAL-1006: the start (epoch seconds) of the stream's open compression span, else null. */
+    compressionStartedAt(streamId) {
+        return this.compressionSpans.get(streamId) ?? null;
+    }
     /** TAL-424: the active stream's pending steers, oldest first (session detail `pending_steers`). */
     pendingSteers(streamId) {
         return (this.steers.get(streamId) ?? []).map((r) => this.pendingSteerView(r));
     }
-    withSteerChain(streamId, fn) {
-        const run = (this.steerChains.get(streamId) ?? Promise.resolve()).then(fn);
+    withSteerChain(streamId, sessionId, fn) {
+        const run = (this.steerChains.get(streamId) ?? Promise.resolve()).then(async () => {
+            await this.deps.service().getForMutation(sessionId);
+            return fn();
+        });
         const settled = run.catch(() => undefined).then(() => { if (this.steerChains.get(streamId) === settled)
             this.steerChains.delete(streamId); });
         this.steerChains.set(streamId, settled);
@@ -1777,7 +2329,7 @@ export class TurnRunner {
     async rewriteSteer(sessionId, steerId, fallback, rewrite) {
         let s = null;
         try {
-            s = this.deps.store.get(sessionId, { metadataOnly: true });
+            s = (await this.deps.store.get(sessionId, { metadataOnly: true }));
         }
         catch {
             s = null;
@@ -1785,21 +2337,21 @@ export class TurnRunner {
         const streamId = s?.active_stream_id;
         const sidecar = this.deps.sidecar();
         if (!streamId || !sidecar)
-            return fallback;
-        return this.withSteerChain(streamId, async () => {
+            return fallback('not_running');
+        return this.withSteerChain(streamId, sessionId, async () => {
             const records = this.steers.get(streamId) ?? [];
             const index = records.findIndex((r) => r.steer_id === steerId);
             if (index < 0 || records[index].state !== 'pending')
-                return fallback;
+                return fallback('taken');
             // In flight before the sidecar call, so a pending-text report racing its reply is deferred too.
             let finished = () => undefined;
             this.steerRewrites.set(streamId, new Promise((resolve) => { finished = resolve; }));
-            let outcome = { result: fallback, changed: false };
+            let outcome = { result: fallback('failed'), changed: false };
             try {
                 outcome = await rewrite(sidecar, streamId, records, index);
             }
             catch {
-                outcome = { result: fallback, changed: false };
+                outcome = { result: fallback('failed'), changed: false };
             }
             finally {
                 this.steerRewrites.delete(streamId);
@@ -1807,7 +2359,7 @@ export class TurnRunner {
                 const deferred = this.deferredPendingSteer.get(streamId);
                 this.deferredPendingSteer.delete(streamId);
                 if (deferred !== undefined && !outcome.changed)
-                    this.settleConsumedSteers(sessionId, streamId, deferred);
+                    (await this.settleConsumedSteers(sessionId, streamId, deferred));
             }
             return outcome.result;
         });
@@ -1822,54 +2374,61 @@ export class TurnRunner {
      * TAL-424 Edit and Cancel: take a pending steer back before the Agent takes it. `withdrawn: false` (and nothing
      * emitted) when it is unknown, already taken or being sent; a taken one then settles as consumed.
      */
-    withdrawSteer(sessionId, steerId, reason) {
-        return this.rewriteSteer(sessionId, steerId, { withdrawn: false }, async (sidecar, streamId, records, index) => {
+    async withdrawSteer(sessionId, steerId, reason) {
+        return (await this.rewriteSteer(sessionId, steerId, () => ({ withdrawn: false }), async (sidecar, streamId, records, index) => {
             const record = records[index];
             const reply = await sidecar.call('chat.steer_withdraw', { stream_id: streamId, pending: records.map((r) => r.text), index });
             if (!reply.withdrawn)
                 return { result: { withdrawn: false }, changed: false };
             this.dropSteer(streamId, record);
-            this.emitToSession(sessionId, 'steer_withdrawn', { steer_id: record.steer_id, reason, text: record.display_text });
+            await this.emitToSession(sessionId, 'steer_withdrawn', { steer_id: record.steer_id, reason, text: record.display_text });
             return { result: { withdrawn: true, text: record.display_text }, changed: true };
-        });
+        }));
     }
     /**
      * TAL-424 Send now: deliver a pending steer with the Agent's redirect. A model request restarts with it (consumed at
      * once); during tools it goes last on the Agent's queue and stays `sending_now` until taken; with nothing live it
      * stays pending.
      */
-    sendSteerNow(sessionId, steerId) {
-        return this.rewriteSteer(sessionId, steerId, { redirected: false }, async (sidecar, streamId, records, index) => {
+    async sendSteerNow(sessionId, steerId) {
+        return (await this.rewriteSteer(sessionId, steerId, (outcome) => ({ redirected: false, outcome }), async (sidecar, streamId, records, index) => {
             const record = records[index];
             if (!record.can_redirect)
-                return { result: { redirected: false }, changed: false };
+                return { result: { redirected: false, outcome: 'failed' }, changed: false };
             const reply = await sidecar.call('chat.steer_now', { stream_id: streamId, pending: records.map((r) => r.text), index });
             if (!reply.withdrawn)
-                return { result: { redirected: false }, changed: false };
+                return { result: { redirected: false, outcome: reply.reason === 'not_running' ? 'not_running' : 'taken' }, changed: false };
             if (!reply.redirected) {
                 // The Agent's queue changed meanwhile, so it went back last: clients follow the new order.
                 if (reply.requeued === 'last') {
                     this.dropSteer(streamId, record);
                     records.push(record);
-                    this.emitToSession(sessionId, 'steer_pending', this.pendingSteerView(record));
+                    (await this.emitToSession(sessionId, 'steer_pending', this.pendingSteerView(record)));
                 }
-                return { result: { redirected: false }, changed: true };
+                return { result: { redirected: false, outcome: 'next_step' }, changed: true };
             }
             this.dropSteer(streamId, record);
             if (reply.delivery === 'steer') {
                 record.state = 'sending_now';
                 records.push(record);
-                this.emitToSession(sessionId, 'steer_pending', this.pendingSteerView(record));
+                await this.emitToSession(sessionId, 'steer_pending', this.pendingSteerView(record));
             }
             else {
-                this.emitToSession(sessionId, 'steer_consumed', this.consumedSteerPayload(record, { redirected: true }));
-                this.saveConsumedSteers(sessionId, streamId);
+                await this.emitToSession(sessionId, 'steer_consumed', this.consumedSteerPayload(record, { redirected: true }));
+                await this.saveConsumedSteers(sessionId, streamId);
             }
-            return { result: { redirected: true }, changed: true };
-        });
+            return { result: { redirected: true, outcome: 'delivered' }, changed: true };
+        }));
     }
     // ── approvals / clarify ──────────────────────────────────────────────────
     async respondApproval(sessionId, choice, approvalId, enableYolo) {
+        if (this.deps.service().owner?.isOwned(sessionId)) {
+            if (enableYolo)
+                throw new HttpFailure(403, 'Change approval policy through the owning Agent');
+            await this.deps.service().owner.respond(sessionId, 'approval', approvalId, choice);
+            return { ok: true, choice };
+        }
+        sessionId = (await this.deps.service().idForMutation(sessionId));
         // Python `_enable_session_yolo_and_release_pending`: a stale card id while another approval is parked
         // must not commit YOLO for the whole session — fail closed with the relay envelope instead.
         if (enableYolo && approvalId && !this.deps.pending.hasApprovalId(sessionId, approvalId) && this.deps.pending.hasPendingApproval(sessionId)) {
@@ -1882,7 +2441,7 @@ export class TurnRunner {
         const relayFailure = (message) => ({ ok: false, choice, error: `The Agent sidecar did not accept the approval (${message}); retry in a moment.`, _status: 503 });
         let s = null;
         try {
-            s = this.deps.store.get(sessionId, { metadataOnly: true });
+            s = (await this.deps.store.get(sessionId, { metadataOnly: true }));
         }
         catch {
             s = null;
@@ -1890,7 +2449,9 @@ export class TurnRunner {
         const profileHome = this.deps.profileHome(s?.profile ?? null);
         let resolved = false;
         let withdrawn = false;
-        if (entry) {
+        if (entry?._injected === true)
+            resolved = Boolean(this.deps.pending.resolveApproval(sessionId, str(entry.approval_id)).entry);
+        else if (entry) {
             if (!sidecar)
                 return relayFailure('not running');
             try {
@@ -1922,7 +2483,7 @@ export class TurnRunner {
                         return relayFailure(str(error.message));
                     }
                 }
-                this.deps.service().setYolo(sessionId, true);
+                await this.deps.service().setYolo(sessionId, true);
                 return null;
             });
             if (failure)
@@ -1956,17 +2517,21 @@ export class TurnRunner {
             return;
         }
         if (this.deps.pending.dropApprovals(sessionId, new Set(before.filter((id) => !live.has(id)))).length)
-            this.emitToSession(sessionId, ...this.approvalHeadFrame(sessionId));
+            (await this.emitToSession(sessionId, ...this.approvalHeadFrame(sessionId)));
     }
     withYoloLock(sessionId, fn) {
-        const run = (this.yoloChains.get(sessionId) ?? Promise.resolve()).then(fn);
+        const run = (this.yoloChains.get(sessionId) ?? Promise.resolve()).then(async () => {
+            sessionId = (await this.deps.service().idForMutation(sessionId));
+            return fn();
+        });
         const settled = run.catch(() => undefined).then(() => { if (this.yoloChains.get(sessionId) === settled)
             this.yoloChains.delete(sessionId); });
         this.yoloChains.set(sessionId, settled);
         return run;
     }
     /** Python `set_session_yolo_enabled(False)`: the sidecar must drop its per-session YOLO state before the local flag clears. */
-    disableYolo(sessionId) {
+    async disableYolo(sessionId) {
+        sessionId = (await this.deps.service().idForMutation(sessionId));
         return this.withYoloLock(sessionId, async () => {
             const sidecar = this.deps.sidecar();
             if (sidecar) {
@@ -1977,7 +2542,7 @@ export class TurnRunner {
                     return { ok: false, error: `The Agent sidecar did not disable YOLO (${str(error.message)}); retry in a moment.`, yolo_enabled: this.deps.service().yolo(sessionId).yolo_enabled === true, _status: 503 };
                 }
             }
-            return this.deps.service().setYolo(sessionId, false);
+            return (await this.deps.service().setYolo(sessionId, false));
         });
     }
     /**
@@ -1986,12 +2551,26 @@ export class TurnRunner {
      * did not acknowledge (prompt retained).
      */
     async respondClarify(sessionId, clarifyId, reply) {
+        if (this.deps.service().owner?.isOwned(sessionId)) {
+            await this.deps.service().owner.respond(sessionId, 'clarify', clarifyId, reply);
+            return { ok: true };
+        }
+        sessionId = (await this.deps.service().idForMutation(sessionId));
         const entry = this.deps.pending.peekClarify(sessionId, clarifyId);
         if (!entry)
             return { ok: false, stale: true };
-        const response = typeof reply === 'string' ? reply : clarifyReply(entry, reply);
+        // TAL-887: a password-manager answer goes to the Agent and nowhere else; `null` (a decline) is its empty reply.
+        const vault = isVaultPrompt(entry);
+        if (vault && typeof reply === 'string')
+            return { ok: false, invalid: true };
+        const response = vault ? vaultReply(entry, reply) : reply === null ? null : clarifyReply(entry, reply);
         if (response === null)
             return { ok: false, invalid: true };
+        // A test-hook prompt has no parked Agent waiter: it is answered here.
+        if (entry._injected === true) {
+            this.deps.pending.resolveClarify(sessionId, str(entry.clarify_id));
+            return vault ? { ok: true } : { ok: true, response };
+        }
         const sidecar = this.deps.sidecar();
         if (!sidecar)
             return { ok: false, error: 'The Agent sidecar is not running; retry in a moment.' };
@@ -2000,14 +2579,16 @@ export class TurnRunner {
             ok = (await sidecar.call('clarify.respond', { session_id: sessionId, clarify_id: str(entry.clarify_id), response })).ok;
         }
         catch (error) {
+            if (vault)
+                return { ok: false, error: 'The Agent sidecar did not accept the secure answer; retry in a moment.' };
             return { ok: false, error: `The Agent sidecar did not accept the answer (${str(error.message)}); retry in a moment.` };
         }
         if (!ok)
             return { ok: false, stale: true };
         const { head } = this.deps.pending.resolveClarify(sessionId, str(entry.clarify_id));
         if (head)
-            this.emitToSession(sessionId, 'clarify', head);
-        return { ok: true, response };
+            (await this.emitToSession(sessionId, 'clarify', head));
+        return vault ? { ok: true } : { ok: true, response };
     }
 }
 /** The turn's answer: the last non-error assistant reply with content. */