@jmanuelcorral/openteam 0.24.1 → 0.26.0

package/dist/cli.js

@@ -424,6 +424,9 @@ function migrateConfig(raw) {
 var consoleDetailMessages = {
   follow: "Follow new activity",
   metadataOnly: "Lifecycle and telemetry stay metadata-only.",
+  logsDescription: "Current recorded conversations load automatically. Lifecycle and telemetry remain metadata-only; credentials are masked.",
+  credentialMasked: "[credential masked]",
+  offscreen: "Messages load automatically when this conversation is visible.",
   filter: "Filter sessions",
   incoming: "New activity available — show updates",
   refreshError: "Live updates unavailable; retrying.",
@@ -444,7 +447,7 @@ var consoleDetailMessages = {
   partialMetadata: "Some current session titles are unavailable or outside the bounded batch; fallback labels remain.",
   rootCurrent: "Current whole orchestrator session; may contain multiple batches. Not a historical or batch-only transcript.",
   recovered: "Read-only association verified from current opencode history; not an immutable lifecycle event.",
-  rootUnlinked: "Orchestrator session not recorded. Reveal to try a bounded, read-only relationship search.",
+  rootUnlinked: "Orchestrator session not recorded. A bounded, read-only relationship search is attempted when visible.",
   rootUnavailable: "Orchestrator session unavailable: no unique, verifiable relationship was found in local history.",
   rootAmbiguous: "Orchestrator session unavailable: conflicting or copied relationships are ambiguous.",
   searchIncomplete: "Orchestrator association search incomplete or exceeded its bounded budget. No session was guessed.",
@@ -457,9 +460,9 @@ var consoleDetailMessages = {
   arguments: "Arguments",
   result: "Result",
   loading: "Loading local conversation…",
-  current: "Present-day session, not a timeline reconstruction. May include later messages and tools. Refresh explicitly for updates.",
-  historical: "Historical cursor: conversation content is hidden to avoid showing future mutable text. Return to live before revealing this operation.",
-  disabled: "Content is disabled. Set console.sessionHistory.contentEnabled to true in .opencode/openteam.json and restart the Console. Then reveal content here.",
+  current: "Present-day whole session, not a timeline reconstruction or per-attempt transcript. Visible messages update automatically while live.",
+  historical: "Historical cursor: conversation content is hidden to avoid showing future mutable text. Return to live to read current messages.",
+  disabled: "Content is disabled by console.sessionHistory.contentEnabled: false. Lifecycle and telemetry remain metadata-only.",
   unavailable: "Conversation unavailable. Start an operator-owned local opencode serve against the same workspace and user history, configure console.sessionHistory.endpoint, then retry. No inference was requested.",
   empty: "No readable user messages, assistant output or tool calls remain in this session.",
   invalid: "This attempt has no recorded session in the selected run, or its history is incomplete.",
@@ -469,7 +472,7 @@ var consoleDetailMessages = {
   invalidEndpoint: "Expected a loopback HTTP origin without credentials, path, query or fragment.",
   readFailed: "Local session history could not be read.",
   oversized: "Conversation exceeds the local size limit (4 MiB). No content was loaded.",
-  truncated: "Conversation display is truncated by local safety limits; omitted content is not empty history.",
+  truncated: "Conversation display is truncated by local safety limits. The newest bounded message window is shown; older messages or oversized fields may be omitted, not empty history.",
   truncatedMarker: `
 [truncated]`,
   terminalGap: "Live collection was interrupted; events during the gap are unavailable. This is not complete session history.",
@@ -546,7 +549,7 @@ var ConsoleConfigSchema = z2.object({
   openBrowser: z2.boolean().default(false),
   remoteStorage: z2.boolean().default(false),
   sessionHistory: z2.object({
-    contentEnabled: z2.boolean().default(false),
+    contentEnabled: z2.boolean().default(true),
     endpoint: z2.string().refine(isLocalHistoryEndpoint, consoleDetailMessages.invalidEndpoint).optional()
   }).strict().optional(),
   terminal: z2.object({
@@ -684,7 +687,8 @@ var LocalRuntimeKindSchema = z2.enum([
   "ollama",
   "lmstudio",
   "foundry-local",
-  "lemonade"
+  "lemonade",
+  "llama-swap"
 ]);
 var LocalRuntimeSchema = z2.object({
   id: z2.string().min(1),
@@ -704,13 +708,26 @@ var LocalRuntimeSchema = z2.object({
     ctx.addIssue({
       code: z2.ZodIssueCode.custom,
       path: ["kind"],
-      message: `local runtime "${runtime.id}" has no resolvable kind: add "kind" ` + "(one of ollama, lmstudio, foundry-local, lemonade), since the id is " + "not itself a runtime kind."
+      message: `local runtime "${runtime.id}" has no resolvable kind: add "kind" ` + "(one of ollama, lmstudio, foundry-local, lemonade, llama-swap), " + "since the id is not itself a runtime kind."
     });
   }
+}).transform((runtime) => {
+  const resolvedKind = runtime.kind ?? runtime.id;
+  if (resolvedKind !== "llama-swap" || runtime.maxConcurrency !== undefined) {
+    return runtime;
+  }
+  return { ...runtime, maxConcurrency: 1 };
 });
 function localRuntimeKind(runtime) {
   return runtime.kind ?? runtime.id;
 }
+function effectiveMaxConcurrency(runtime) {
+  if (runtime.maxConcurrency !== undefined) {
+    return runtime.maxConcurrency;
+  }
+  const kind = runtime.kind ?? (runtime.id !== undefined && LocalRuntimeKindSchema.safeParse(runtime.id).success ? runtime.id : undefined);
+  return kind === "llama-swap" ? 1 : DEFAULT_LOCAL_MAX_CONCURRENCY;
+}
 var MemoryScopeSchema = z2.enum(["project", "user"]);
 var MemoryExtractionModeSchema = z2.enum([
   "local",
@@ -891,6 +908,27 @@ import { isAbsolute, normalize, parse, resolve, sep } from "node:path";
 import { createOpencodeClient } from "@opencode-ai/sdk";
 import { z as z4 } from "zod";
 
+// src/console/credentials.ts
+var URL_WITH_AUTHORITY = /https?:\/\/[^\s/?#]+/giu;
+function maskUrlCredentials(value) {
+  return value.replace(URL_WITH_AUTHORITY, (candidate) => {
+    const schemeEnd = candidate.indexOf("://") + 3;
+    const at = candidate.indexOf("@", schemeEnd);
+    if (at === -1)
+      return candidate;
+    const colon = candidate.indexOf(":", schemeEnd);
+    if (colon === -1 || colon <= schemeEnd || colon + 1 >= at || colon > at)
+      return candidate;
+    return `${candidate.slice(0, schemeEnd)}${consoleDetailMessages.credentialMasked}${candidate.slice(at)}`;
+  });
+}
+function maskConversationCredentials(value) {
+  return maskUrlCredentials(value.replace(/-----BEGIN (?:[A-Z]+ )?PRIVATE KEY-----[\s\S]*?(?:-----END (?:[A-Z]+ )?PRIVATE KEY-----|$)/gu, consoleDetailMessages.credentialMasked).replace(/\b(?:Bearer|Basic)\s+[A-Za-z0-9._~+/=-]+/giu, consoleDetailMessages.credentialMasked).replace(/\b(?:sk-[A-Za-z0-9_-]{16,}|gh[pousr]_[A-Za-z0-9]{16,}|github_pat_[A-Za-z0-9_]{16,}|AKIA[A-Z0-9]{16}|eyJ[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+)\b/gu, consoleDetailMessages.credentialMasked).replace(/(\b(?:password|passwd|secret|api[_-]?key|access[_-]?token|refresh[_-]?token|token|authorization|cookie|set-cookie)\b["']?\s*[:=]\s*)(?:"[^"]*"|'[^']*'|[^\s,;}\]]+)/giu, (_match, prefix) => prefix + consoleDetailMessages.credentialMasked));
+}
+function isCredentialKey(key) {
+  return /^(?:password|passwd|secret|api[_-]?key|access[_-]?token|refresh[_-]?token|token|authorization|cookie|set-cookie|private[_-]?key)$/iu.test(key);
+}
+
 // src/console/historyBoundary.ts
 function only(keys) {
   return (value) => {
@@ -1148,7 +1186,7 @@ class ConversationTooLargeError extends Error {
   }
 }
 var Info = z4.preprocess(only(["id", "sessionID", "role", "time"]), z4.object({
-  id: z4.string(),
+  id: z4.string().min(1).max(256),
   sessionID: z4.string(),
   role: z4.enum(["user", "assistant"]),
   time: z4.preprocess(only(["created", "completed"]), z4.object({
@@ -1168,7 +1206,7 @@ var Text = PartIdentity.extend({
 });
 var Tool = PartIdentity.extend({
   type: z4.literal("tool"),
-  callID: z4.string(),
+  callID: z4.string().min(1).max(256),
   tool: z4.string(),
   state: z4.preprocess(only(["status", "input", "output", "error", "time"]), z4.object({
     status: z4.enum(["pending", "running", "completed", "error"]),
@@ -1191,6 +1229,7 @@ function projectConversation(raw, sessionID) {
   let remaining = limits.totalChars;
   let toolsRemaining = limits.tools;
   function text(value) {
+    value = maskConversationCredentials(value);
     const length = Math.min(value.length, limits.fieldChars, remaining);
     remaining -= length;
     if (length === value.length)
@@ -1222,7 +1261,7 @@ function projectConversation(raw, sessionID) {
           truncated = true;
           break;
         }
-        const child = bounded(value[key], depth + 1);
+        const child = isCredentialKey(key) ? consoleDetailMessages.credentialMasked : bounded(value[key], depth + 1);
         if (Array.isArray(value))
           array.push(child);
         else
@@ -1233,7 +1272,7 @@ function projectConversation(raw, sessionID) {
     return text(JSON.stringify(bounded(input, 0), null, 2));
   }
   const messages = [];
-  for (const rawRow of raw.slice(0, limits.messages)) {
+  for (const rawRow of raw.slice(-limits.messages).reverse()) {
     if (remaining <= 0) {
       truncated = true;
       break;
@@ -1242,12 +1281,13 @@ function projectConversation(raw, sessionID) {
     if (row.info.sessionID !== sessionID)
       throw new Error(consoleDetailMessages.readFailed);
     const message = {
-      id: text(row.info.id),
+      id: row.info.id,
       role: row.info.role,
       createdAt: row.info.time.created,
       ...row.info.time.completed !== undefined ? { completedAt: row.info.time.completed } : {},
       text: [],
-      tools: []
+      tools: [],
+      order: []
     };
     if (row.parts.length > limits.parts)
       truncated = true;
@@ -1271,16 +1311,19 @@ function projectConversation(raw, sessionID) {
       if (part.sessionID !== sessionID || part.messageID !== row.info.id)
         throw new Error(consoleDetailMessages.readFailed);
       if (part.type === "text") {
-        if (!part.ignored)
+        if (!part.ignored) {
+          message.order?.push({ type: "text", index: message.text.length });
           message.text.push(text(part.text));
+        }
       } else {
         if (toolsRemaining-- <= 0) {
           truncated = true;
           continue;
         }
         const state = part.state;
+        message.order?.push({ type: "tool", index: message.tools.length });
         message.tools.push({
-          id: text(part.callID),
+          id: part.callID,
           name: text(part.tool),
           status: state.status,
           arguments: argumentsText(state.input),
@@ -1385,7 +1428,7 @@ function createConversationReader(deps) {
         if (workspaceKey(session.directory) !== workspaceKey(directory))
           throw new Error(consoleDetailMessages.readFailed);
         const time = z4.preprocess(only(["created"]), z4.object({ created: z4.number().finite().min(0).max(8640000000000000) }).strict()).safeParse(session.time);
-        const title = boundedTitle(session.title);
+        const title = boundedTitle(typeof session.title === "string" ? maskConversationCredentials(session.title) : session.title);
         return {
           id: session.id,
           directory: session.directory,
@@ -1435,10 +1478,13 @@ function createConversationReader(deps) {
             labels.delete(oldest);
           return session;
         },
-        messages: async (id) => {
+        messages: async (id, limit) => {
           if (!verified.has(id))
             throw new Error(consoleDetailMessages.readFailed);
-          const result = await client.session.messages({ path: { id } });
+          const result = await client.session.messages({
+            path: { id },
+            ...limit === undefined ? {} : { query: { limit } }
+          });
           if (!result.response.ok || result.error !== undefined)
             throw new Error(consoleDetailMessages.readFailed);
           return result.data;
@@ -1475,14 +1521,37 @@ function createConversationReader(deps) {
   const read = async (endpoint, sessionID, signal) => withHistory(endpoint, signal, async (port) => {
     const session = await port.get(sessionID);
     return {
-      ...projectConversation(await port.messages(sessionID), sessionID),
+      ...projectConversation(await port.messages(sessionID, CONVERSATION_LIMITS.messages + 1), sessionID),
       session: session.label
     };
   });
+  const roots = new Map;
   read.root = (endpoint, projection, signal) => withHistory(endpoint, signal, async (port) => {
-    const recovered = await recoverRootSession(port, projection);
+    const key = JSON.stringify([
+      endpoint,
+      projection.runID,
+      projection.rootOperationID,
+      projection.operations.flatMap((operation) => operation.attempts.map((attempt) => [
+        operation.operationID,
+        attempt.attemptID,
+        attempt.sessionID
+      ]))
+    ]);
+    const cached = roots.get(key);
+    const recovered = cached ? {
+      session: await port.get(cached),
+      association: "sdk-parent"
+    } : await recoverRootSession(port, projection);
+    roots.delete(key);
+    if (recovered.association === "sdk-parent")
+      roots.set(key, recovered.session.id);
+    if (roots.size > 32) {
+      const oldest = roots.keys().next().value;
+      if (oldest !== undefined)
+        roots.delete(oldest);
+    }
     return {
-      ...projectConversation(await port.messages(recovered.session.id), recovered.session.id),
+      ...projectConversation(await port.messages(recovered.session.id, CONVERSATION_LIMITS.messages + 1), recovered.session.id),
       session: recovered.session.label,
       association: recovered.association
     };
@@ -3356,6 +3425,23 @@ var executionSetupMessages = {
 `),
   localOnlyTitle: "Local execution",
   localRequiredError: "local execution requires an enabled local runtime",
+  llamaSwap: {
+    label: "llama-swap",
+    noLoadedModels: "llama-swap has no canonical models verified as ready with no active routing profile. Keep the intended model ready through the operator's existing runtime controls, then rerun setup; manual IDs cannot bypass admission.",
+    modelNotLoaded: (model) => `llama-swap model "${model}" is not positively verified as canonical and ready without an active routing profile. No substitute was selected.`,
+    probeRequired: "llama-swap setup requires a fresh read-only residency probe before writing configuration. Manual model IDs cannot bypass this requirement.",
+    probeFailed: "llama-swap readiness could not be verified. Check the configured /v1/models catalog and origin /running and /api/profiles endpoints; authenticated runtimes are not supported.",
+    endpointMismatch: "llama-swap endpoint/provider mapping differs from the selected runtime. Reconcile the configured origin or /v1 URL and provider; no configuration was written.",
+    modelIdentityUnknown: (model) => `llama-swap model "${model}" has no usable registered physical identity. Fix its model declaration; setup will not guess an alias.`,
+    declaredLimits: "llama-swap context and output limits are operator-declared metadata, not measured per-slot capacities. Missing output metadata uses openteam's client policy default, not a guessed model cap. Unknown context remains 0 and disables automatic compaction. Tool support remains unknown without evidence.",
+    declaredLimitsTitle: "Declared llama-swap limits",
+    outputReduction: (model, previous, ceiling) => `${model}: reducing limit.output from ${previous} to ${ceiling} to respect the declared runtime ceiling. Smaller operator limits are preserved.`,
+    outputReductionTitle: "Declared output ceiling",
+    outputChanged: (model, ceiling) => `${model}: fresh runtime metadata declares an output ceiling of ${ceiling} below the prepared configuration. Rerun setup to review the updated limit; no configuration was written.`
+  },
+  residencyV2Title: "Resident runtime transport compatibility",
+  residencyV2Restriction: (providers) => `Project policies disable opencode's separate native/core V2 transport for resident-only providers: ${providers.join(", ")}. Guarded V1 provider transport remains supported; this does not ban the public SDK v2 client library. Higher-precedence policies can override project restrictions. Readiness preflight is not an atomic no-autoload guarantee.`,
+  retainedResidencyProviderUnprotected: (providerID) => `Cannot safely retain resident-only provider "${providerID}": the effective configuration must retain its matching runtime, disabled when deselected. Reconcile the runtime/provider configuration; no configuration was written.`,
   lemonadeDetectedHint: (loaded, advertised) => `detected - ${loaded} confirmed loaded of ${advertised} advertised text model(s)`,
   lemonadeNoLoadedModels: "Lemonade has no text models verified as loaded and ready. Keep the intended model active in Lemonade and rerun setup; inactive or unknown models cannot be selected or entered manually.",
   lemonadeModelNotLoaded: (model) => `Lemonade model "${model}" is not positively verified as loaded and ready. Keep that exact text model active and rerun setup; no substitute was selected and no implicit loading is allowed.`,
@@ -3370,11 +3456,11 @@ var executionSetupMessages = {
   lemonadeV2InvalidPolicies: "Cannot safely preserve experimental.policies: experimental must be an object and policies must be an array of provider.use statements with a string resource and allow/deny effect. Fix the project configuration before rerunning setup.",
   lemonadeV2Title: "Lemonade transport compatibility",
   lemonadeV2Restriction: (providers) => `Project policies disable native V2 Lemonade access for: ${providers.join(", ")}. Ordinary guarded V1 configuration remains supported; V2 is not an openteam-hook execution path. Higher-precedence or global policies can override this project restriction; it is not a server-atomic no-autoload guarantee.`,
-  lemonadeInventory: (instanceID, providerID, models) => {
+  lemonadeInventory: (instanceID, providerID, models, runtime = "Lemonade") => {
     const inventory = models.map((model) => `${model.modelID} (${model.loaded === true ? "loaded" : model.loaded === false ? "inactive" : "residency unknown"})`).join(", ");
-    return `Lemonade inventory ${instanceID} (provider ${providerID}): ${inventory || "no advertised text models"}. Only confirmed loaded text models count as ready.`;
+    return `${runtime} inventory ${instanceID} (provider ${providerID}): ${inventory || "no advertised text models"}. Only confirmed loaded text models count as ready.`;
   },
-  lemonadeAgentResidency: (agent, model, state) => `Lemonade readiness: ${agent} (${model}): ${state === "not-loaded" ? "inactive (not ready)" : "residency unknown"}; advertised inventory is not loaded readiness. Keep this exact model active and verify residency before use; requests without verified readiness are blocked.`,
+  lemonadeAgentResidency: (agent, model, state, runtime = "Lemonade") => `${runtime} readiness: ${agent} (${model}): ${state === "not-loaded" ? "inactive (not ready)" : "residency unknown"}; advertised inventory is not loaded readiness. Keep this exact model active and verify residency before use; requests without verified readiness are blocked.`,
   mixedPrimaryQuestion: "Primary coordinator model",
   mixedPrimaryLocal: (model) => `Use local thinking model ${model}`,
   mixedPrimaryFrontier: "Choose a frontier model",
@@ -3478,6 +3564,8 @@ var executionPolicyMessages = {
   primaryLocalMismatchRemedy: (sourceFile) => `      remedy: re-run \`openteam setup\` to regenerate ${sourceFile} and opencode.json from the same primary policy, then re-run \`openteam doctor\`.`
 };
 var doctorMessages = {
+  defaultRuntimeSlots: (slots) => ` · ${slots} slot(s) (default)`,
+  llamaSwapConcurrency: "    llama-swap does not report per-model slots; openteam uses the configured runtime-wide concurrency limit. No per-model capacity was inferred.",
   localModelLimits: {
     section: "  local model limits:",
     healthy: "    ✓ enabled local provider models declare valid output and context limits in opencode.json.",
@@ -3494,7 +3582,7 @@ var doctorMessages = {
     invalidInput: "invalid limit.input (expected a positive integer token count)",
     outputExceedsContext: "limit.output is greater than or equal to limit.context, so the configured context leaves no usable input window",
     outputExceedsInput: "limit.input is less than or equal to the reserved output budget, so compaction would have no usable input threshold",
-    outputRemedy: "    remedy: re-run `openteam setup` to write openteam's 8 192-token local output default, then re-run `openteam doctor`.",
+    outputRemedy: "    remedy: re-run `openteam setup` to write openteam's local output policy (at most 8 192 tokens, narrowed by declared runtime ceilings and smaller context/input budgets), then re-run `openteam doctor`.",
     unconfiguredLimitRemedy: "    remedy: the optional limit block may remain absent; to configure explicit limits, manually add a complete limit object in opencode.json with positive output and non-negative context (0 means unknown), because setup fills an absent limit block only for models the runtime currently discovers. Then re-run `openteam doctor`.",
     detectedContextRemedy: "    remedy: re-run `openteam setup` while the runtime is reachable to copy the detected usable context budget, then re-run `openteam doctor`.",
     unknownContextRemedy: "    remedy: re-run `openteam setup` to write the required context:0 sentinel; automatic compaction remains disabled until a positive context is configured.",
@@ -4482,234 +4570,286 @@ async function handleLifecycleHistoryRoute(request, response, url, deps) {
 // src/console/inspector.ts
 var CONVERSATION_CLIENT_JS = `
 function wireConversationInspector(){
+  if(document.__conversationsWired){return}document.__conversationsWired=true;
+  var msg=${JSON.stringify(consoleDetailMessages)},limits=${JSON.stringify(CONVERSATION_LIMITS)};
+  var records=new Map(),recoveries=new Map(),stopped=false,timer=null,scheduled=null,active=0,serial=0,retained=null;
   var mount=document.getElementById('lifecycle-timeline-mount');
-  if(!mount){return}
-  var msg=${JSON.stringify(consoleDetailMessages)};
-  var limits=${JSON.stringify(CONVERSATION_LIMITS)};
-  var panel=document.createElement('section');panel.className='lc-conversation';panel.hidden=true;panel.id='lc-attempt-conversation';
-  panel.setAttribute('data-conversation-inspector','');
-  panel.setAttribute('aria-labelledby','lc-conversation-heading');
-  mount.appendChild(panel);
-  var selection=null,request=null,epoch=0,retained=null;
-  function element(tag,text){var e=document.createElement(tag);if(text!==undefined){e.textContent=text}return e}
-  function selectedButton(){
-    if(!selection||mount.getAttribute('data-conversation-transition')==='1'||
-      mount.getAttribute('data-lifecycle-run')!==selection.run||
-      (mount.getAttribute('data-conversation-cursor')||null)!==selection.cursor){return null}
-    return Array.from(mount.querySelectorAll('[data-inspect-attempt],[data-inspect-root]')).find(function(button){
-      var attempt=button.closest('[data-lifecycle-attempt-id]');
-      var operation=button.closest('[data-lifecycle-operation-id]');
-      var session=button.closest(selection.target==='root'?'[data-root-session-id]':'[data-lifecycle-session-id]');
-      return !button.disabled&&operation&&session&&
-        (selection.target==='root'?button.hasAttribute('data-inspect-root'):attempt&&attempt.getAttribute('data-lifecycle-attempt-id')===selection.attempt&&
-        button.getAttribute('data-inspect-attempt')===selection.attempt)&&
-        operation.getAttribute('data-lifecycle-operation-id')===selection.operation&&
-        button.getAttribute('data-inspect-operation')===selection.operation&&
-        session.getAttribute(selection.target==='root'?'data-root-session-id':'data-lifecycle-session-id')===selection.session;
-    })||null;
-  }
-  function expanded(button,open){
-    if(button.getAttribute('aria-expanded')!==String(open)){button.setAttribute('aria-expanded',String(open))}
-    var label=open?msg.close:msg.inspectConversation;
-    if(button.textContent!==label){button.textContent=label}
-  }
-  function cancel(){
-    epoch++;if(request){request.abort();request=null}
-  }
-  function close(focus){
-    var button=selectedButton();
-    cancel();selection=null;retained=null;
-    mount.querySelectorAll('[data-inspect-attempt],[data-inspect-root]').forEach(function(control){expanded(control,false)});
-    panel.textContent='';panel.hidden=true;
-    ['data-conversation-kind','data-conversation-run','data-conversation-operation','data-conversation-attempt','data-conversation-session','data-conversation-target','aria-busy'].forEach(function(name){panel.removeAttribute(name)});
-    mount.appendChild(panel);
-    if(focus===true&&button){button.focus({preventScroll:true})}
-  }
-  // The renderer replaces metadata, not retained content. Restore only after
-  // the real run/cursor commits and the same attempt/session is still visible.
-  function retain(){
-    if(!selection||panel.hidden||!selectedButton()){return}
-    var scrolls=[],ancestor=panel.parentElement;
-    while(ancestor){scrolls.push([ancestor,ancestor.scrollTop,ancestor.scrollLeft]);ancestor=ancestor.parentElement}
-    panel.querySelectorAll('pre,[data-conversation-tool-list]').forEach(function(node){scrolls.push([node,node.scrollTop,node.scrollLeft])});
-    retained={scrolls:scrolls,top:panel.getBoundingClientRect().top,focus:panel.contains(document.activeElement)?document.activeElement:null};
-  }
-  function reanchor(){
-    if(!selection){return}
-    var button=selectedButton();
-    if(!button){close();return}
-    var attempt=selection.target==='root'?button.closest('.lct-operation-card'):button.closest('[data-lifecycle-attempt-id]');
-    if(panel.parentElement!==attempt){attempt.appendChild(panel)}
-    expanded(button,true);
-    if(retained){
-      retained.scrolls.forEach(function(item){if(item[0].isConnected){item[0].scrollTop=item[1];item[0].scrollLeft=item[2]}});
-      var tree=mount.querySelector('[data-lifecycle-tree-mount]');
-      if(tree){tree.scrollTop+=panel.getBoundingClientRect().top-retained.top}
-      if(retained.focus){retained.focus.focus({preventScroll:true})}
-      retained=null;
+  function el(tag,text){var node=document.createElement(tag);if(text!==undefined){node.textContent=text}return node}
+  function setText(node,text){if(node.textContent!==text){node.textContent=text}}
+  function visible(node){
+    if(document.hidden||!node.isConnected){return false}
+    var parent=node;
+    while(parent){if(parent.hidden){return false}parent=parent.parentElement}
+    var rect=node.getBoundingClientRect();
+    return rect.bottom>=0&&rect.top<(window.innerHeight||1000);
+  }
+  function cancel(record){
+    if(record.request){record.request.abort();record.request=null}
+    if(record.finish){record.finish();record.finish=null}
+    record.version++;
+  }
+  function discard(record){
+    cancel(record);record.panel.remove();record.body.textContent='';record.source.textContent='';record.rows.clear();records.delete(record.key);
+  }
+  function clear(scope){
+    Array.from(records.values()).forEach(function(record){
+      if(scope==='native'&&record.selection.target==='session'){return}
+      discard(record);
+    });
+    if(scope!=='native'){retained=null}
+  }
+  function recovery(record){
+    var saved=recoveries.get(record.key);
+    if(saved&&saved.until<=Date.now()){recoveries.delete(record.key);return null}
+    return saved||null;
+  }
+  function rootRecovery(record,data){
+    return record.selection.target==='root'&&!record.selection.session&&data.kind==='unavailable'&&
+      [msg.rootUnavailable,msg.rootAmbiguous,msg.searchIncomplete].includes(data.message)?
+      {message:data.message,until:Date.now()+30000}:null;
+  }
+  function rememberRecovery(record,saved){
+    if(!saved){recoveries.delete(record.key);return}
+    recoveries.delete(record.key);recoveries.set(record.key,saved);
+    if(recoveries.size>128){recoveries.delete(recoveries.keys().next().value)}
+  }
+  function desired(){
+    var result=[];
+    if(stopped||document.hidden){return result}
+    var explicitHistorical=!!((mount&&mount.getAttribute('data-conversation-cursor'))||new URLSearchParams(window.location.search).has('cursor'));
+    var transitioning=mount&&mount.getAttribute('data-conversation-transition')==='1';
+    if(mount){
+      var notice=mount.querySelector('[data-conversation-historical]');
+      if(explicitHistorical){
+        if(!notice){notice=el('p',msg.historical);notice.setAttribute('data-conversation-historical','');mount.appendChild(notice)}
+      }else if(notice){notice.remove()}
+    }
+    if(mount&&!explicitHistorical&&!transitioning){
+      mount.querySelectorAll('[data-inspect-root],[data-inspect-attempt]').forEach(function(button){
+        if(button.disabled){return}
+        var root=button.hasAttribute('data-inspect-root');
+        var session=button.closest(root?'[data-root-session-id]':'[data-lifecycle-session-id]');
+        var host=root?button.closest('.lct-operation-card'):button.closest('[data-lifecycle-attempt-id]');
+        if(!session||!host){return}
+        var selection={target:root?'root':'attempt',run:mount.getAttribute('data-lifecycle-run'),
+          operation:button.getAttribute('data-inspect-operation'),attempt:root?null:button.getAttribute('data-inspect-attempt'),
+          session:session.getAttribute(root?'data-root-session-id':'data-lifecycle-session-id')};
+        if(!selection.run||!selection.operation){return}
+        result.push({key:JSON.stringify(selection),selection:selection,host:host,button:button});
+      });
     }
+    document.querySelectorAll('[data-session-conversation]').forEach(function(host){
+      var selection={target:'session',session:host.getAttribute('data-session-conversation')};
+      if(explicitHistorical){
+        if(visible(host)){setText(host,msg.historical)}
+        return;
+      }
+      result.push({key:JSON.stringify(selection),selection:selection,host:host});
+    });
+    return result;
   }
-  function reset(){
-    panel.textContent='';
-    var header=element('header');header.className='lc-conversation-head';
-    var heading=element('h3',msg.title);heading.id='lc-conversation-heading';heading.tabIndex=-1;header.appendChild(heading);
-    var closeBtn=element('button',msg.close);closeBtn.type='button';
-    closeBtn.setAttribute('data-conversation-close','');
-    closeBtn.addEventListener('click',function(){close(true)});header.appendChild(closeBtn);panel.appendChild(header);
-  }
-  function showHeading(){
-    var heading=panel.querySelector('h3');
-    heading.focus({preventScroll:true});
-    heading.scrollIntoView({block:'start',inline:'nearest'});
-  }
-  function load(reveal){
-    if(!selectedButton()){return}
-    cancel();var current=epoch,selected=selection;
-    reset();
-    panel.setAttribute('data-conversation-kind','loading');
-    panel.setAttribute('aria-busy','true');
-    var status=element('p',msg.loading);status.setAttribute('role','status');panel.appendChild(status);
-    var opening=null;
-    if(reveal===true){
-      showHeading();
-      opening={focus:document.activeElement,scrolls:[mount.querySelector('[data-lifecycle-tree-mount]'),mount.closest('.scroll')].filter(Boolean).map(function(node){return [node,node.scrollTop,node.scrollLeft]})};
-    }
-    var controller=new AbortController();request=controller;
-    var query=new URLSearchParams({run:selected.run,operation:selected.operation,diagnostic:'1'});
-    if(selected.target==='root'){query.set('target','root')}else{query.set('attempt',selected.attempt)}
-    query.set('token',document.documentElement.getAttribute('data-token')||'');
-    if(selected.cursor!==null){query.set('cursor',selected.cursor)}
-    var timeout=setTimeout(function(){controller.abort()},15000);
-    fetch('/api/graph/conversation?'+query.toString(),{signal:controller.signal,cache:'no-store'})
-      .then(function(r){return r.json()}).then(function(data){
-        if(epoch!==current||selection!==selected||!selectedButton()){return}
-        var bringIntoView=opening&&document.activeElement===opening.focus&&opening.scrolls.every(function(item){return item[0].scrollTop===item[1]&&item[0].scrollLeft===item[2]});
-        if((data.kind==='conversation'||data.kind==='empty')&&(data.runID!==selected.run||data.operationID!==selected.operation||
-          (selected.target==='root'?(data.target!=='root'||data.attemptID!==undefined||!data.sessionID||(!selected.session&&!['sdk-parent','tool-result'].includes(data.association))):(data.target!=='attempt'||data.attemptID!==selected.attempt))||
-          (selected.session&&data.sessionID!==selected.session))){
-          data={kind:'unavailable',message:msg.unavailable};
-        }
-        status.textContent=data.message||msg.unavailable;
-        panel.setAttribute('data-conversation-kind',data.kind||'unavailable');
-        if(data.kind==='conversation'||data.kind==='empty'){
-          panel.setAttribute('data-conversation-session',data.sessionID);
-          var source=element('p');source.appendChild(element('strong',data.session&&data.session.title|| (selected.target==='root'?msg.orchestrator:msg.session)));
-          var created=consoleLabels.utcTime(data.session&&data.session.createdAt);
-          source.appendChild(element('small',' · '+(created===msg.timeUnknown?msg.timeUnknown:msg.sessionCreated+': '+created)));
-          source.appendChild(element('code',' · '+data.sessionID));panel.appendChild(source);
-          if(data.session&&data.session.id===data.sessionID){document.dispatchEvent(new CustomEvent('console-session-metadata',{detail:{session:data.session,runID:data.runID,target:selected.target}}))}
-        }
-        if(data.kind==='conversation'){
-          renderMessages(data.messages||[]);
-        }else if(data.kind==='empty'){
-          renderMessages([]);
+  function create(item){
+    var panel=el('section');panel.className='lc-conversation';panel.id='lc-conversation-'+(++serial);
+    panel.setAttribute('data-conversation-inspector','');
+    panel.setAttribute('data-conversation-target',item.selection.target);
+    panel.setAttribute('data-conversation-session',item.selection.session||'');
+    if(item.selection.run){panel.setAttribute('data-conversation-run',item.selection.run)}
+    if(item.selection.attempt){panel.setAttribute('data-conversation-attempt',item.selection.attempt)}
+    var heading=el('h3',msg.title);panel.appendChild(heading);
+    var status=el('p',msg.loading);status.setAttribute('role','status');panel.appendChild(status);
+    var source=el('p');panel.appendChild(source);
+    var body=el('div');body.setAttribute('data-conversation-messages','');panel.appendChild(body);
+    var refresh=el('button',msg.refresh);refresh.type='button';refresh.setAttribute('data-conversation-refresh','');panel.appendChild(refresh);
+    var record={key:item.key,selection:item.selection,host:item.host,panel:panel,status:status,source:source,body:body,rows:new Map(),
+      request:null,finish:null,version:0,due:0,button:item.button};
+    refresh.addEventListener('click',function(){record.due=0;recoveries.delete(record.key);sync()});
+    item.host.appendChild(panel);records.set(item.key,record);
+    return record;
+  }
+  function matches(record,data){
+    var s=record.selection;
+    return data.sessionID&&(s.session?data.sessionID===s.session:['sdk-parent','tool-result'].includes(data.association))&&
+      (s.target==='session'?data.target==='session'&&!data.runID:
+        data.runID===s.run&&data.operationID===s.operation&&data.target===s.target&&
+          (s.target==='root'?data.attemptID===undefined:data.attemptID===s.attempt));
+  }
+  function render(record,messages){
+    var remaining=limits.totalChars,tools=0,truncated=false,used=new Set();
+    function text(value){
+      value=String(value===undefined?'':value);
+      var length=Math.min(value.length,limits.fieldChars+msg.truncatedMarker.length,remaining);
+      remaining-=length;if(length<value.length){truncated=true}return value.slice(0,length);
+    }
+    if(messages.length>limits.messages){truncated=true}
+    messages.slice(0,limits.messages).forEach(function(message){
+      if(!message||!['user','assistant'].includes(message.role)||typeof message.id!=='string'||used.has(message.id)){return}
+      used.add(message.id);
+      var row=record.rows.get(message.id);
+      if(!row){
+        var article=el('article');article.setAttribute('data-conversation-role',message.role);
+        article.appendChild(el('h4',message.role==='user'?msg.prompt:msg.assistant));
+        var time=el('time');article.appendChild(time);
+        row={node:article,time:time,texts:[],tools:new Map()};record.rows.set(message.id,row);record.body.appendChild(article);
+      }
+      var selection=window.getSelection&&window.getSelection();
+      if(selection&&!selection.isCollapsed&&row.node.contains(selection.anchorNode)){return}
+      setText(row.time,consoleLabels.utcTime(message.createdAt));
+      var parts=(message.text||[]).slice(0,limits.parts);
+      parts.forEach(function(value,index){
+        if(!row.texts[index]){row.texts[index]=el('pre');row.node.insertBefore(row.texts[index],row.node.querySelector('details'))}
+        setText(row.texts[index],text(value));
+      });
+      while(row.texts.length>parts.length){row.texts.pop().remove()}
+      var seenTools=new Set();
+      (message.tools||[]).slice(0,limits.parts).forEach(function(tool){
+        if(tools++>=limits.tools||remaining<=0){truncated=true;return}
+        seenTools.add(tool.id);
+        var call=row.tools.get(tool.id);
+        if(!call){
+          var detail=el('details');detail.setAttribute('data-tool-call',tool.id);detail.open=true;
+          var summary=el('summary');detail.appendChild(summary);
+          detail.appendChild(el('h5',msg.arguments));var args=el('pre');detail.appendChild(args);
+          detail.appendChild(el('h5',msg.result));var result=el('pre');detail.appendChild(result);
+          call={node:detail,summary:summary,args:args,result:result};row.tools.set(tool.id,call);row.node.appendChild(detail);
         }
-        // The short loading card may not have enough scroll range. Complete
-        // the explicit reveal after layout grows, unless the reader moved on.
-        if(bringIntoView){panel.querySelector('h3').scrollIntoView({block:'start',inline:'nearest'})}
-      }).catch(function(){
-        if(epoch===current){status.textContent=msg.unavailable;panel.setAttribute('data-conversation-kind','unavailable')}
-      }).finally(function(){
-        clearTimeout(timeout);
-        if(epoch!==current){return}
-        request=null;
-        panel.setAttribute('aria-busy','false');
-        var refresh=element('button',msg.refresh);refresh.type='button';refresh.setAttribute('data-conversation-refresh','');
-        refresh.addEventListener('click',load);panel.appendChild(refresh);
+        setText(call.summary,text(tool.name)+' · '+text(tool.status));
+        setText(call.args,text(tool.arguments));setText(call.result,text(tool.result===undefined?msg.pending:tool.result));
       });
-  }
-  function renderMessages(messages){
-    var truncated=messages.length>limits.messages,remaining=limits.totalChars,toolCount=0;
-    messages=messages.slice(0,limits.messages);
-    var tools=element('details');tools.setAttribute('data-conversation-tools','');tools.open=true;
-    var count=messages.reduce(function(total,m){return total+(m.tools||[]).length},0);
-    tools.appendChild(element('summary',msg.tools.replace('{count}',String(count))));
-    var calls=element('div');calls.setAttribute('data-conversation-tool-list','');tools.appendChild(calls);
-    panel.appendChild(tools);
-    function text(value){
-      value=String(value);
-      var size=Math.min(value.length,limits.fieldChars+msg.truncatedMarker.length,remaining);
-      remaining-=size;
-      if(size<value.length){truncated=true}
-      return value.slice(0,size);
-    }
-    ['user','assistant'].forEach(function(role){
-      var section=element('section');section.setAttribute('data-conversation-role',role);
-      section.appendChild(element('h4',role==='user'?msg.prompt:msg.assistant));
-      var hasText=false;
-      messages.filter(function(m){return m.role===role}).forEach(function(m){
-        var item=element('article');item.appendChild(element('time',consoleLabels.utcTime(m.createdAt)));
-        if((m.text||[]).length>limits.parts){truncated=true}
-        (m.text||[]).slice(0,limits.parts).forEach(function(value){
-          if(!value){return}
-          hasText=true;
-          if(remaining>0){item.appendChild(element('pre',text(value)))}else{truncated=true}
+      row.tools.forEach(function(call,id){if(!seenTools.has(id)){call.node.remove();row.tools.delete(id)}});
+      if(Array.isArray(message.order)){
+        var previous=row.time;
+        message.order.slice(0,limits.parts).forEach(function(part){
+          var node=part.type==='text'?row.texts[part.index]:row.tools.get((message.tools[part.index]||{}).id);
+          node=node&&node.node||node;
+          if(node){if(previous.nextSibling!==node){row.node.insertBefore(node,previous.nextSibling||null)}previous=node}
         });
-        if(item.querySelector('pre')){section.appendChild(item)}
-      });panel.appendChild(section);
-      if(!hasText){section.appendChild(element('p',role==='user'?msg.noPrompts:msg.noOutput))}
+      }
+    });
+    record.rows.forEach(function(row,id){if(!used.has(id)){row.node.remove();record.rows.delete(id)}});
+    if(truncated){setText(record.status,msg.current+'\\n'+msg.truncated)}
+  }
+  function load(record){
+    var selected=record.selection,controller=new AbortController(),version=++record.version;
+    record.request=controller;active++;record.panel.setAttribute('aria-busy','true');
+    var finished=false;
+    record.finish=function(){if(!finished){finished=true;active--;clearTimeout(timeout)}};
+    var query=new URLSearchParams({token:document.documentElement.getAttribute('data-token')||''});
+    var path='/api/graph/conversation';
+    if(selected.target==='session'){path='/api/console/conversation';query.set('session',selected.session)}
+    else{query.set('run',selected.run);query.set('operation',selected.operation);
+      if(selected.target==='root'){query.set('target','root')}else{query.set('attempt',selected.attempt)}}
+    var timeout=setTimeout(function(){
+      if(record.version!==version){return}
+      cancel(record);record.due=Date.now()+30000;setText(record.status,msg.unavailable);
+      record.panel.setAttribute('aria-busy','false');schedule();
+    },15000);
+    fetch(path+'?'+query,{signal:controller.signal,cache:'no-store'}).then(function(response){return response.json()}).then(function(data){
+      if(record.version!==version||!records.has(record.key)||!visible(record.host)||controller.signal.aborted){return}
+      if((data.kind==='conversation'||data.kind==='empty')&&!matches(record,data)){data={kind:'unavailable',message:msg.unavailable}}
+      var negative=rootRecovery(record,data);
+      record.panel.setAttribute('data-conversation-kind',data.kind||'unavailable');
+      setText(record.status,data.message||((data.kind==='conversation'||data.kind==='empty')?msg.current:msg.unavailable));
+      if(data.kind==='conversation'||data.kind==='empty'){
+        recoveries.delete(record.key);
+        record.panel.setAttribute('data-conversation-session',data.sessionID);
+        setText(record.source,(data.session&&data.session.title||msg.session)+' · '+data.sessionID);
+        var focus=record.panel.contains(document.activeElement)?document.activeElement:null;
+        render(record,data.messages||[]);
+        if(focus&&focus.isConnected&&focus!==document.activeElement&&
+          (!document.activeElement||document.activeElement===document.body)){focus.focus({preventScroll:true})}
+        if(data.session&&data.session.id===data.sessionID){
+          document.dispatchEvent(new CustomEvent('console-session-metadata',{detail:{session:data.session,runID:data.runID,target:selected.target}}));
+        }
+      }else{rememberRecovery(record,negative);record.body.textContent='';record.source.textContent='';record.rows.clear()}
+      record.due=['disabled','unauthorized','historical','invalid','oversized'].includes(data.kind)?Infinity:Date.now()+(data.kind==='unavailable'?30000:2000);
+    }).catch(function(){
+      if(record.version===version&&!controller.signal.aborted){
+        recoveries.delete(record.key);
+        setText(record.status,msg.unavailable);record.panel.setAttribute('data-conversation-kind','unavailable');
+        record.body.textContent='';record.source.textContent='';record.rows.clear();record.due=Date.now()+30000;
+      }
+    }).finally(function(){
+      if(!finished){finished=true;clearTimeout(timeout);active--}
+      if(record.version===version){record.request=null;record.panel.setAttribute('aria-busy','false');if(controller.signal.aborted){record.due=Date.now()+30000;setText(record.status,msg.unavailable)}}
+      schedule();
+    });
+  }
+  function sync(){
+    var items=desired(),byKey=new Map(items.map(function(item){return [item.key,item]}));
+    Array.from(records.values()).forEach(function(record){
+      var item=byKey.get(record.key);
+      if(!item||!visible(item.host)){
+        if(item&&item.host.isConnected){item.host.style.minHeight=item.host.getBoundingClientRect().height+'px'}
+        discard(record);return;
+      }
+      record.host=item.host;
+      if(record.panel.parentElement!==item.host){item.host.appendChild(record.panel)}
+    });
+    items.forEach(function(item){
+      if(!visible(item.host)){return}
+      var record=records.get(item.key);
+      if(!record&&records.size<8){
+        if(item.selection.target==='session'){item.host.textContent=''}
+        record=create(item);
+      }
+      if(!record){return}
+      if(item.button){item.button.hidden=true;item.button.setAttribute('aria-controls',record.panel.id)}
+      var saved=recovery(record);
+      if(saved&&!record.rows.size&&!record.request){record.panel.setAttribute('data-conversation-kind','unavailable');setText(record.status,saved.message)}
+      if(!record.request&&active<2&&record.due<=Date.now()&&!saved){load(record)}
     });
-    messages.forEach(function(m){
-      if((m.tools||[]).length>limits.parts){truncated=true}
-      (m.tools||[]).slice(0,limits.parts).forEach(function(tool){
-        if(toolCount++>=limits.tools||remaining<=0){truncated=true;return}
-        var call=element('details');call.setAttribute('data-tool-call',tool.id);
-        call.appendChild(element('summary',text(tool.name)+' · '+text(tool.status)));
-        if(tool.startedAt!==undefined){call.appendChild(element('time',consoleLabels.utcTime(tool.startedAt)))}
-        call.appendChild(element('h5',msg.arguments));call.appendChild(element('pre',text(tool.arguments)));
-        call.appendChild(element('h5',msg.result));call.appendChild(element('pre',text(tool.result===undefined?msg.pending:tool.result)));
-        calls.appendChild(call);
+  }
+  function schedule(){
+    if(stopped||scheduled!==null){return}
+    scheduled=setTimeout(function(){scheduled=null;sync()},50);
+  }
+  document.addEventListener('conversation-selection',function(event){
+    clear(event&&event.detail&&event.detail.scope==='native'?'native':'all');
+    sync();
+  });
+  document.addEventListener('conversation-tree-replacing',function(){
+    var scroll=document.querySelector('.scroll'),tree=mount&&mount.querySelector('[data-lifecycle-tree-mount]');
+    retained={focus:document.activeElement,scroll:scroll,scrollTop:scroll&&scroll.scrollTop,treeTop:tree&&tree.scrollTop,positions:[]};
+    records.forEach(function(record){
+      [record.body].concat(Array.from(record.panel.querySelectorAll('pre'))).forEach(function(node){
+        retained.positions.push([node,node.scrollTop,node.scrollLeft]);
       });
     });
-    if(count===0){calls.appendChild(element('p',msg.noTools))}
-    if(truncated){panel.appendChild(element('p',msg.truncated))}
-  }
-  mount.addEventListener('click',function(event){
-    var button=event.target.closest('[data-inspect-attempt],[data-inspect-root]');
-    if(!button||button.disabled||mount.getAttribute('data-conversation-transition')==='1'){return}
-    if(button===selectedButton()){close(true);return}
-    close();
-    var root=button.hasAttribute('data-inspect-root');
-    var session=button.closest(root?'[data-root-session-id]':'[data-lifecycle-session-id]');
-    if(!session){return}
-    selection={
-      run:mount.getAttribute('data-lifecycle-run'),
-      operation:button.getAttribute('data-inspect-operation'),
-      attempt:button.getAttribute('data-inspect-attempt'),
-      target:root?'root':'attempt',
-      session:session.getAttribute(root?'data-root-session-id':'data-lifecycle-session-id'),
-      cursor:mount.getAttribute('data-conversation-cursor')||null
-    };
-    if(!selectedButton()){close();return}
-    panel.setAttribute('data-conversation-run',selection.run);
-    panel.setAttribute('data-conversation-operation',selection.operation);
-    if(selection.attempt){panel.setAttribute('data-conversation-attempt',selection.attempt)}
-    panel.setAttribute('data-conversation-target',selection.target);
-    panel.setAttribute('data-conversation-session',selection.session);
-    reanchor();
-    panel.hidden=false;
-    if(selection.cursor!==null){reset();panel.setAttribute('data-conversation-kind','historical');panel.appendChild(element('p',msg.historical));showHeading()}
-    else{load(true)}
+    records.forEach(function(record){if(record.selection.target!=='session'){record.panel.remove()}});
+  });
+  document.addEventListener('conversation-tree-committed',function(){
+    sync();
+    if(retained){
+      retained.positions.forEach(function(item){if(item[0].isConnected){item[0].scrollTop=item[1];item[0].scrollLeft=item[2]}});
+      var tree=mount&&mount.querySelector('[data-lifecycle-tree-mount]');
+      if(tree&&retained.treeTop!==null){tree.scrollTop=retained.treeTop}
+      if(retained.scroll){retained.scroll.scrollTop=retained.scrollTop}
+      if(retained.focus&&retained.focus.isConnected&&retained.focus!==document.activeElement&&
+        (!document.activeElement||document.activeElement===document.body||document.activeElement===document.documentElement)){
+        retained.focus.focus({preventScroll:true});
+      }
+      retained=null;
+    }
   });
-  document.addEventListener('conversation-selection',close);
-  document.addEventListener('conversation-tree-replacing',retain);
-  document.addEventListener('conversation-tree-committed',reanchor);
-  window.addEventListener('pagehide',close);
+  document.addEventListener('console-conversation-selection',sync);
+  document.addEventListener('visibilitychange',function(){if(document.hidden){clear()}else{sync()}});
+  document.addEventListener('scroll',schedule,true);
+  window.addEventListener('resize',schedule);
+  window.addEventListener('pagehide',function(){stopped=true;clear();clearInterval(timer);clearTimeout(scheduled);scheduled=null});
+  window.addEventListener('pageshow',function(event){if(event.persisted){stopped=false;timer=setInterval(sync,2000);sync()}});
+  timer=setInterval(sync,2000);sync();
 }
 `;
 var CONVERSATION_STYLE = `
-.lc-conversation{margin-top:12px;padding:12px;min-width:0;border:1px solid var(--cp-border);border-radius:16px;background:var(--cp-surface);color:var(--cp-text);overflow-anchor:none}
+.lc-conversation{margin-top:12px;padding:12px;min-width:0;border:1px solid var(--cp-border,var(--border));border-radius:12px;background:var(--cp-surface,var(--surface));color:var(--cp-text,var(--fg));overflow-anchor:none}
 .lc-conversation[hidden]{display:none}
-.lc-conversation-head{display:flex;flex-wrap:wrap;align-items:center;justify-content:space-between;gap:8px}
 .lc-conversation h3,.lc-conversation h4,.lc-conversation p{margin:8px 0}
-.lc-conversation h3{scroll-margin-top:12px}
-.lc-conversation pre{white-space:pre-wrap;overflow-wrap:anywhere;max-height:400px;overflow:auto;background:var(--cp-surface-soft);padding:12px;font-family:Consolas,"Courier New",Courier,monospace}
-.lc-conversation [data-conversation-role="user"] pre{max-height:180px}
-.lc-conversation [data-conversation-tool-list]{max-height:160px;overflow:auto}
-.lc-conversation details{margin-top:12px;padding:8px;border:1px solid var(--cp-border);border-radius:.625rem}
-.lc-conversation summary{cursor:pointer;color:var(--cp-accent)}
-.lc-conversation time{font-size:11px;color:var(--cp-text-muted)}
-.lc-conversation button,[data-inspect-attempt],[data-inspect-root]{padding:6px 10px;margin:6px;border:1px solid var(--cp-border);border-radius:.625rem;background:var(--cp-surface);color:var(--cp-accent)}
+.lc-conversation [data-conversation-messages]{max-height:520px;overflow:auto;overflow-anchor:none}
+.lc-conversation pre{white-space:pre-wrap;overflow-wrap:anywhere;max-height:400px;overflow:auto;background:var(--cp-surface-soft,var(--bg));padding:12px;font-family:var(--font-mono)}
+.lc-conversation details{margin:12px 0;padding:8px;border:1px solid var(--cp-border,var(--border));border-radius:8px}
+.lc-conversation summary{cursor:pointer;color:var(--cp-accent,var(--accent))}
+.lc-conversation time{font-size:11px;color:var(--cp-text-muted,var(--muted))}
+.lc-conversation button,[data-inspect-attempt],[data-inspect-root]{padding:6px 10px;margin:6px;border:1px solid var(--cp-border,var(--border));border-radius:8px;background:var(--cp-surface,var(--surface));color:var(--cp-accent,var(--accent))}
 .lct-operation-name{display:block;font-size:16px}
 .lct-operation-head details,.lct-session details{font-size:11px;color:var(--cp-text-muted)}
 .lct-session small,.session-btn small{display:block;color:var(--cp-text-muted)}
@@ -4743,7 +4883,7 @@ var CONSOLE_REFRESH_JS = `
       if(old.getAttribute('data-label-fallback')!==next.textContent){old.setAttribute('data-label-fallback',next.textContent)}
       return;
     }
-    if(old.matches('input,textarea,select,.term,.pty,.graph-drill,[data-config-mount]')){return}
+    if(old.matches('input,textarea,select,.term,.pty,.graph-drill,[data-config-mount],[data-session-conversation]')){return}
     Array.from(old.attributes).forEach(function(a){
       if(a.name!=='open'&&a.name!=='hidden'&&a.name!=='aria-selected'&&!next.hasAttribute(a.name)){old.removeAttribute(a.name)}
     });
@@ -4957,6 +5097,8 @@ svg.graph{display:block}
 .graph-drill .drilllog .tl.error{color:#f85149}
 `;
 var CONSOLE_CLIENT_JS = `
+${LABELS_CLIENT_JS}
+${CONVERSATION_CLIENT_JS}
 (function(){
   var detailMsg=${JSON.stringify(consoleDetailMessages)};
   var terminalCollectors=[];
@@ -5011,6 +5153,7 @@ var CONSOLE_CLIENT_JS = `
         if(show){matched=true;if(!screen.hidden){connectTerminal(cards[j],id)}}
       }
       updateTerminalLeases();
+      document.dispatchEvent(new CustomEvent('console-conversation-selection'));
       return matched;
     }
     if(!screen.__logsWired){
@@ -5071,6 +5214,7 @@ var CONSOLE_CLIENT_JS = `
     if(name==='config'&&activeScreen){loadConfig(activeScreen)}
     if(name==='logs'&&activeScreen){wireLogsScreen(activeScreen)}
     updateTerminalLeases();
+    document.dispatchEvent(new CustomEvent('console-conversation-selection'));
     try{localStorage.setItem('openteam.screen',name)}catch(e){}
     document.body.classList.remove('nav-open');
     return true;
@@ -5472,8 +5616,6 @@ var LIFECYCLE_THEME_BOOT = `
 })();
 `;
 var LIFECYCLE_CLIENT_JS = `
-${LABELS_CLIENT_JS}
-${CONVERSATION_CLIENT_JS}
 function wireLifecycleTimeline(){
   var mount=document.getElementById('lifecycle-timeline-mount');
   if(!mount){return}
@@ -5520,7 +5662,7 @@ function wireLifecycleTimeline(){
       treeMountEl.setAttribute('inert','');
       treeMountEl.querySelectorAll('[data-inspect-attempt],[data-inspect-root]').forEach(function(button){button.disabled=true});
     }
-    document.dispatchEvent(new CustomEvent('conversation-selection'));
+    document.dispatchEvent(new CustomEvent('conversation-selection',{detail:{scope:'native'}}));
   }
   function commitConversation(run,cursor){
     cursor=cursor===null?null:String(cursor);
@@ -5982,6 +6124,7 @@ function wireLifecycleTimeline(){
     inspect.setAttribute('aria-expanded','false');
     inspect.setAttribute('aria-controls','lc-attempt-conversation');
     inspect.setAttribute('data-inspect-attempt',attempt.attemptID);
+    inspect.hidden=true;
     inspect.setAttribute('data-inspect-operation',attempt.operationID);
     session.appendChild(inspect);
     if(attempt.parentSessionID){
@@ -6134,7 +6277,7 @@ function wireLifecycleTimeline(){
           var time=makeElement('small','',${JSON.stringify(consoleDetailMessages.timeUnknown)});time.setAttribute('data-session-time',rootID);rootSession.appendChild(time);
           var id=makeElement('code','',rootID);rootSession.appendChild(id);
         }else{rootSession.appendChild(makeElement('span','',${JSON.stringify(consoleDetailMessages.rootUnlinked)}))}
-        var inspect=makeElement('button','',message('inspect'));inspect.type='button';inspect.setAttribute('data-inspect-root','');inspect.setAttribute('data-inspect-operation',op.operationID);inspect.setAttribute('aria-expanded','false');inspect.setAttribute('aria-controls','lc-attempt-conversation');rootSession.appendChild(inspect);
+        var inspect=makeElement('button','',message('inspect'));inspect.type='button';inspect.hidden=true;inspect.setAttribute('data-inspect-root','');inspect.setAttribute('data-inspect-operation',op.operationID);rootSession.appendChild(inspect);
         card.insertBefore(rootSession,card.querySelector('.lct-attempts'));
       }
       var children=childrenByParent.get(op.operationID)||[];
@@ -6746,6 +6889,7 @@ function sessionDetailCard(tab, terminalEnabled, ptyEnabled, visible) {
   const idSuffix = domIdPart(tab.sessionID);
   const body = [
     sessionHeaderPanel(tab),
+    `<div data-session-conversation="${escapeHtml(tab.sessionID)}"></div>`,
     terminalEnabled ? terminalPanel(tab) : "",
     terminalEnabled && ptyEnabled ? ptyPanel(tab) : "",
     routesPanel(tab.recentRoutes, idSuffix),
@@ -7131,7 +7275,7 @@ function lcSessionHtml(attempt, role) {
     sessionNameHtml(attempt.sessionID, role),
     `<small data-session-time="${escapeHtml(attempt.sessionID)}">${escapeHtml(consoleDetailMessages.timeUnknown)}</small>`,
     `<details><summary>${escapeHtml(consoleDetailMessages.technicalID)}</summary><code>${escapeHtml(attempt.sessionID)}</code></details>`,
-    `<button type="button" aria-expanded="false" aria-controls="lc-attempt-conversation" data-inspect-operation="${escapeHtml(attempt.operationID)}" data-inspect-attempt="${escapeHtml(attempt.attemptID)}">${escapeHtml(consoleDetailMessages.inspectConversation)}</button>`,
+    `<button type="button" hidden data-inspect-operation="${escapeHtml(attempt.operationID)}" data-inspect-attempt="${escapeHtml(attempt.attemptID)}">${escapeHtml(consoleDetailMessages.inspectConversation)}</button>`,
     attempt.parentSessionID === undefined ? "" : `<span class="lct-parent-session">${escapeHtml(lifecycleMessages.parentSessionLabel)} <code>${escapeHtml(attempt.parentSessionID)}</code></span>`,
     "</div>"
   ].join("");
@@ -7218,7 +7362,7 @@ function lcOperationHtml(operation, operationIDs, childrenHtml, depth, rootSessi
     isRoot && operation.kind === "coordinator" ? [
       `<div class="lct-session" data-root-session-id="${escapeHtml(rootSessionID ?? "")}">`,
       rootSessionID ? `${sessionNameHtml(rootSessionID, consoleDetailMessages.orchestrator)} <small data-session-time="${escapeHtml(rootSessionID)}">${escapeHtml(consoleDetailMessages.timeUnknown)}</small><details><summary>${escapeHtml(consoleDetailMessages.technicalID)}</summary><code>${escapeHtml(rootSessionID)}</code></details>` : `<span>${escapeHtml(consoleDetailMessages.rootUnlinked)}</span>`,
-      `<button type="button" data-inspect-root data-inspect-operation="${escapeHtml(operation.operationID)}" aria-expanded="false" aria-controls="lc-attempt-conversation">${escapeHtml(consoleDetailMessages.inspectConversation)}</button>`,
+      `<button type="button" hidden data-inspect-root data-inspect-operation="${escapeHtml(operation.operationID)}">${escapeHtml(consoleDetailMessages.inspectConversation)}</button>`,
       "</div>"
     ].join("") : "",
     `<section class="lct-attempts" aria-label="${escapeHtml(lifecycleMessages.attemptsLabel)}">${attempts}</section>`,
@@ -7597,11 +7741,11 @@ function renderConsoleHtml(state, options = {}) {
   const body = [
     screenSection("tree", treeCopy.label, treeCopy.heading, treeCopy.description, treeBody, true),
     screenSection("routing", "Routing and telemetry", "Routing + telemetry", "Aggregate routing and cost telemetry from <code>GET /api/state</code> across all sessions.", routingScreenBody(state)),
-    screenSection("logs", "Logs and session activity", "Logs / activity", "Per-session activity. Prompts are referenced by hash only and never rendered in clear text.", `<div data-log-controls><label>${escapeHtml(consoleDetailMessages.filter)} <input type="search" data-log-filter></label> <label><input type="checkbox" data-log-follow> ${escapeHtml(consoleDetailMessages.follow)}</label> <button type="button" data-log-incoming hidden>${escapeHtml(consoleDetailMessages.incoming)}</button></div><div data-log-content>${logsScreenBody(state, terminalEnabled, ptyEnabled)}</div>`),
+    screenSection("logs", "Logs and session activity", "Logs / activity", consoleDetailMessages.logsDescription, `<div data-log-controls><label>${escapeHtml(consoleDetailMessages.filter)} <input type="search" data-log-filter></label> <label><input type="checkbox" data-log-follow> ${escapeHtml(consoleDetailMessages.follow)}</label> <button type="button" data-log-incoming hidden>${escapeHtml(consoleDetailMessages.incoming)}</button></div><div data-log-content>${logsScreenBody(state, terminalEnabled, ptyEnabled)}</div>`),
     screenSection("config", "Editable configuration", 'Configuration <span class="pill warn">editable</span>', "Edit the whole openteam JSON configuration. Roles are editable inside the JSON when present; unsupported surfaces are not shown.", configScreenBody(options.configEditor?.hasToken === true))
   ].join("");
   const session = state.session.id === undefined ? "" : ` · session ${escapeHtml(state.session.id)}`;
-  const tokenAttr = (terminalEnabled || graphDrilldown || configEnabled || lifecycleEnabled) && options.consoleToken !== undefined ? ` data-token="${escapeHtml(options.consoleToken)}"${terminalEnabled ? ' data-terminal="1"' : ""}${ptyEnabled ? ' data-pty="1"' : ""}${graphDrilldown ? ' data-drilldown="1"' : ""}${configEnabled ? ' data-config="1"' : ""}` : "";
+  const tokenAttr = options.consoleToken !== undefined ? ` data-token="${escapeHtml(options.consoleToken)}"${terminalEnabled ? ' data-terminal="1"' : ""}${terminalEnabled && ptyEnabled ? ' data-pty="1"' : ""}${graphDrilldown ? ' data-drilldown="1"' : ""}${configEnabled ? ' data-config="1"' : ""}` : "";
   return [
     "<!doctype html>",
     `<html lang="en" data-generated="${escapeHtml(state.generatedAt)}" data-refresh="${refreshMs}"${tokenAttr}>`,
@@ -7610,7 +7754,7 @@ function renderConsoleHtml(state, options = {}) {
     '<meta name="viewport" content="width=device-width,initial-scale=1">',
     "<title>openteam console</title>",
     lifecycleEnabled ? `<script>${LIFECYCLE_THEME_BOOT}</script>` : "",
-    `<style>${CONSOLE_STYLE}${graphOptionProvided ? GRAPH_STYLE : ""}${lifecycleEnabled ? LIFECYCLE_STYLE + CONVERSATION_STYLE : ""}</style>`,
+    `<style>${CONSOLE_STYLE}${CONVERSATION_STYLE}${graphOptionProvided ? GRAPH_STYLE : ""}${lifecycleEnabled ? LIFECYCLE_STYLE : ""}</style>`,
     "</head>",
     "<body>",
     '<div class="app">',
@@ -8387,7 +8531,8 @@ async function untilAborted(read, signal) {
 }
 async function handleConversationRoute(request, response, url, deps) {
   const labelsOnly = url.pathname === "/api/graph/session-labels";
-  if (url.pathname !== "/api/graph/conversation" && !labelsOnly)
+  const direct = url.pathname === "/api/console/conversation";
+  if (url.pathname !== "/api/graph/conversation" && !labelsOnly && !direct)
     return false;
   const send = (status, kind, message, extra = {}) => {
     response.writeHead(status, {
@@ -8407,11 +8552,60 @@ async function handleConversationRoute(request, response, url, deps) {
     send(405, "unavailable", consoleDetailMessages.unauthorized);
     return true;
   }
-  if (!deps.enabled || !deps.contentEnabled || url.searchParams.get("diagnostic") !== "1") {
+  if (!direct && !deps.enabled || !deps.contentEnabled) {
     send(403, "disabled", consoleDetailMessages.disabled);
     return true;
   }
   const params = url.searchParams;
+  if (direct) {
+    const sessionID = params.get("session");
+    if (!sessionID || sessionID.length > 256 || params.getAll("session").length !== 1 || ["run", "operation", "attempt", "target"].some((key) => params.has(key))) {
+      send(400, "invalid", consoleDetailMessages.invalid);
+      return true;
+    }
+    if (params.has("cursor")) {
+      send(409, "historical", consoleDetailMessages.historical);
+      return true;
+    }
+    const controller2 = new AbortController;
+    const cancel2 = () => controller2.abort();
+    response.on("close", cancel2);
+    request.on("aborted", cancel2);
+    const timeout2 = setTimeout(cancel2, SESSION_HISTORY_LIMITS.timeoutMs);
+    try {
+      const observed = await untilAborted(deps.observedSessionIDs?.() ?? Promise.resolve([]), controller2.signal);
+      if (!observed.includes(sessionID)) {
+        send(404, "invalid", consoleDetailMessages.invalid);
+        return true;
+      }
+      const endpoint = await untilAborted(deps.resolveEndpoint(sessionID), controller2.signal);
+      if (!endpoint) {
+        send(503, "unavailable", consoleDetailMessages.unavailable);
+        return true;
+      }
+      const conversation = await untilAborted(deps.readConversation(endpoint, sessionID, controller2.signal), controller2.signal);
+      send(200, conversation.messages.length || conversation.truncated ? "conversation" : "empty", `${consoleDetailMessages.current}${conversation.truncated ? `
+${consoleDetailMessages.truncated}` : conversation.messages.length ? "" : `
+${consoleDetailMessages.empty}`}`, {
+        scope: "present-day-session",
+        target: "session",
+        sessionID,
+        ...conversation
+      });
+    } catch (error) {
+      if (!response.destroyed)
+        send(error instanceof ConversationTooLargeError ? 413 : 503, error instanceof ConversationTooLargeError ? "oversized" : "unavailable", error instanceof ConversationTooLargeError ? consoleDetailMessages.oversized : consoleDetailMessages.unavailable);
+    } finally {
+      clearTimeout(timeout2);
+      response.off("close", cancel2);
+      request.off("aborted", cancel2);
+    }
+    return true;
+  }
+  if (!deps.reader) {
+    send(503, "unavailable", consoleDetailMessages.unavailable);
+    return true;
+  }
   const rootTarget = params.get("target") === "root";
   if ((labelsOnly ? params.has("run") ? ["run"] : [] : rootTarget ? ["run", "operation"] : ["run", "operation", "attempt"]).some((key) => params.getAll(key).length !== 1 || !params.get(key)) || params.has("session") || params.getAll("target").length > 1 || params.has("target") && !rootTarget && params.get("target") !== "attempt" || rootTarget && params.has("attempt") || labelsOnly && (params.has("operation") || params.has("attempt") || params.has("target"))) {
     send(400, "invalid", consoleDetailMessages.invalid);
@@ -8435,7 +8629,7 @@ async function handleConversationRoute(request, response, url, deps) {
     if (request.aborted || request.socket.destroyed)
       cancel2();
     const deadline = Date.now() + SESSION_HISTORY_LIMITS.timeoutMs;
-    const timeout = setTimeout(cancel2, SESSION_HISTORY_LIMITS.timeoutMs);
+    const timeout2 = setTimeout(cancel2, SESSION_HISTORY_LIMITS.timeoutMs);
     try {
       const history2 = selected?.kind === "view" ? selected : undefined;
       const selectedRoot = history2?.runs.find((run) => run.runID === history2.selection.runID)?.rootSessionID;
@@ -8501,7 +8695,7 @@ ${consoleDetailMessages.partialMetadata}` : consoleDetailMessages.currentMetadat
       if (!response.destroyed && !request.socket.destroyed)
         send(503, "unavailable", consoleDetailMessages.unavailable);
     } finally {
-      clearTimeout(timeout);
+      clearTimeout(timeout2);
       response.off("close", cancel2);
       request.off("aborted", cancel2);
       request.socket.off("close", cancel2);
@@ -8526,21 +8720,23 @@ ${consoleDetailMessages.partialMetadata}` : consoleDetailMessages.currentMetadat
   const controller = new AbortController;
   const cancel = () => controller.abort();
   response.on("close", cancel);
+  request.on("aborted", cancel);
+  const timeout = setTimeout(cancel, SESSION_HISTORY_LIMITS.timeoutMs);
   try {
     const run = history.runs.find((item) => item.runID === history.selection.runID);
     const recordedRoot = run?.rootSessionID;
     const sessionID = rootTarget ? recordedRoot : attempt?.sessionID;
     const endpointID = sessionID ?? history.projection.operations.flatMap((op) => op.attempts).filter((item) => item.sessionID).at(-1)?.sessionID ?? "";
-    const endpoint = await deps.resolveEndpoint(endpointID);
+    const endpoint = await untilAborted(deps.resolveEndpoint(endpointID), controller.signal);
     if (!endpoint) {
       send(503, "unavailable", consoleDetailMessages.unavailable);
       return true;
     }
-    const recovered = !sessionID && rootTarget && deps.readConversation.root ? await deps.readConversation.root(endpoint, history.projection, controller.signal) : undefined;
+    const recovered = !sessionID && rootTarget && deps.readConversation.root ? await untilAborted(deps.readConversation.root(endpoint, history.projection, controller.signal), controller.signal) : undefined;
     const resolvedID = sessionID ?? recovered?.session.id;
     if (!resolvedID)
       throw new AssociationError("unavailable");
-    const conversation = recovered ?? await deps.readConversation(endpoint, resolvedID, controller.signal);
+    const conversation = recovered ?? await untilAborted(deps.readConversation(endpoint, resolvedID, controller.signal), controller.signal);
     const current = rootTarget ? `${consoleDetailMessages.rootCurrent}
 ${consoleDetailMessages.current}${recovered ? `
 ${consoleDetailMessages.recovered}` : ""}` : consoleDetailMessages.current;
@@ -8556,10 +8752,12 @@ ${consoleDetailMessages.empty}`, {
       ...conversation
     });
   } catch (error) {
-    if (!controller.signal.aborted)
+    if (!response.destroyed)
       send(error instanceof ConversationTooLargeError ? 413 : 503, error instanceof ConversationTooLargeError ? "oversized" : "unavailable", error instanceof AssociationError ? error.message : error instanceof ConversationTooLargeError ? consoleDetailMessages.oversized : consoleDetailMessages.unavailable);
   } finally {
+    clearTimeout(timeout);
     response.off("close", cancel);
+    request.off("aborted", cancel);
   }
   return true;
 }
@@ -8839,15 +9037,19 @@ async function createConsoleServer(deps) {
   };
   const handle = async (request, response) => {
     const url2 = new URL(request.url ?? "/", "http://localhost");
-    if (["/api/graph/conversation", "/api/graph/session-labels"].includes(url2.pathname) && deps.storageApi && !readAuthorized2(request, url2, deps.storageApi.token)) {
+    if ([
+      "/api/graph/conversation",
+      "/api/graph/session-labels",
+      "/api/console/conversation"
+    ].includes(url2.pathname) && deps.storageApi && !readAuthorized2(request, url2, deps.storageApi.token)) {
       response.writeHead(401, { ...JSON_HEADERS5, "cache-control": "no-store" }).end(JSON.stringify({ version: 1, kind: "unauthorized" }));
       return;
     }
-    if (deps.console?.readConversation && deps.lifecycleHistoryApi && await handleConversationRoute(request, response, url2, {
+    if (deps.console?.readConversation && await handleConversationRoute(request, response, url2, {
       token: deps.console.token,
       enabled: deps.config.graphView?.enabled === true,
-      contentEnabled: deps.console.conversationContentEnabled === true,
-      reader: deps.lifecycleHistoryApi.reader,
+      contentEnabled: deps.console.conversationContentEnabled ?? deps.config.sessionHistory?.contentEnabled ?? true,
+      ...deps.lifecycleHistoryApi ? { reader: deps.lifecycleHistoryApi.reader } : {},
       resolveEndpoint: async (id) => deps.console?.historyEndpoint ?? await resolveEndpoint(id),
       observedSessionIDs: async () => buildConsoleState(await deps.readSnapshot()).sessions.map((session) => session.sessionID),
       readConversation: deps.console.readConversation
@@ -9014,6 +9216,7 @@ var PROVIDER_LABELS = {
   ollama: "Ollama",
   lmstudio: "LM Studio",
   lemonade: "Lemonade Server",
+  "llama-swap": executionSetupMessages.llamaSwap.label,
   "foundry-local": "Foundry Local"
 };
 function providerLabel(providerID) {
@@ -9235,7 +9438,8 @@ var KNOWN_LOCAL_PROVIDER_IDS = new Set([
   "ollama",
   "lmstudio",
   "foundry-local",
-  "lemonade"
+  "lemonade",
+  "llama-swap"
 ]);
 var KNOWN_FRONTIER_PROVIDER_IDS = new Set([
   "github-copilot",
@@ -11150,7 +11354,7 @@ async function runConsoleServe(deps, options) {
         ptyEnabled,
         createClient: makeClient,
         drilldownEnabled: graphViewEnabled,
-        conversationContentEnabled: config.console.sessionHistory?.contentEnabled === true,
+        conversationContentEnabled: config.console.sessionHistory?.contentEnabled !== false,
         ...config.console.sessionHistory?.endpoint !== undefined ? { historyEndpoint: config.console.sessionHistory.endpoint } : {},
         readConversation: createConversationReader({
           directory: deps.directory ?? process.cwd(),
@@ -16043,6 +16247,7 @@ var reasons = {
   unknown: "fresh residency metadata is unavailable, malformed, or ambiguous",
   inactive: "the requested model is not verified as resident and serving",
   identity: "the requested model identity can be rewritten or is ambiguous",
+  "authentication-unsupported": "authentication is required but unsupported by this integration",
   destination: "the request destination does not match its configured runtime",
   transport: "this provider execution path cannot enforce the residency guard",
   request: "this is not a supported, explicit model inference request",
@@ -16051,22 +16256,72 @@ var reasons = {
   configuration: "the residency guard configuration is invalid"
 };
 var modelResidencyMessages = {
-  blocked: (reason) => `[openteam] Lemonade model request blocked: ${reasons[reason]}.`,
-  providerDisabled: (providerID) => `[openteam] Lemonade provider "${providerID}" is disabled because its configured transport or runtime mapping cannot enforce loaded-only requests.`,
-  extractionRuntimeUnavailable: "[openteam] Memory extraction requires one enabled local runtime matching the configured extraction provider."
+  blocked: (reason, runtimeLabel = "Lemonade", detail) => `[openteam] ${runtimeLabel} model request blocked: ${reasons[reason]}${detail === undefined ? "" : ` (${detail})`}.`,
+  providerDisabled: (providerID, runtimeLabel = "Lemonade") => `[openteam] ${runtimeLabel} provider "${providerID}" is disabled because its configured transport or runtime mapping cannot enforce loaded-only requests.`,
+  extractionRuntimeUnavailable: "[openteam] Memory extraction requires one enabled local runtime matching the configured extraction provider.",
+  extractionRuntimeUnavailableFor: (runtimeLabel) => `[openteam] Memory extraction requires one enabled ${runtimeLabel} local runtime matching the configured extraction provider.`
+};
+var llamaSwapResidencyMessages = {
+  catalogModelDetail: (id, detail) => `catalog model "${id}" ${detail}`,
+  invalidCatalogField: (field) => `declares an invalid ${field}`,
+  catalogEntryMustExposeStringId: "catalog entries must expose a string id",
+  catalogModelIdsMustBeNonEmptyTrimmedStrings: "catalog model ids must be non-empty, trimmed strings",
+  expectedDataArray: (path) => `GET ${path} must return an object with a data array`,
+  duplicateCanonicalModelId: (modelID, path) => `duplicate canonical model id "${modelID}" in ${path}`,
+  httpFailure: (path, status) => `GET ${path} failed with HTTP ${status}`,
+  unverifiableResponse: (path) => `GET ${path} failed before a verifiable response`,
+  expectedProfilesArray: (path) => `GET ${path} must return an object with a profiles array`,
+  activeProfileRewrite: (profile) => `active profile "${profile}" can rewrite canonical model routing`,
+  invalidActiveProfileValue: (path) => `GET ${path} reported an invalid active profile value`,
+  expectedRunningArray: (path) => `GET ${path} must return an object with a running array`,
+  malformedRunningEvidence: (path) => `GET ${path} returned malformed model/state evidence`,
+  duplicateRunningEvidence: (path, modelID) => `GET ${path} returned duplicate evidence for "${modelID}"`
 };
 
-// src/local/lemonadeResidency.ts
+// src/local/residency.ts
 class ModelResidencyError extends Error {
   code;
   statusCode = 400;
   isRetryable = false;
-  constructor(code) {
-    super(modelResidencyMessages.blocked(code));
+  constructor(code, runtimeLabel = "Lemonade", detail) {
+    super(modelResidencyMessages.blocked(code, runtimeLabel, detail));
     this.code = code;
     this.name = "ModelResidencyError";
   }
 }
+async function withResidencyDeadline(operation, options) {
+  const timeoutMs = options.timeoutMs ?? 5000;
+  const runtimeLabel = options.runtimeLabel;
+  if (!Number.isFinite(timeoutMs) || timeoutMs <= 0) {
+    throw new ModelResidencyError("configuration", runtimeLabel);
+  }
+  const controller = new AbortController;
+  const cancel = () => controller.abort(new ModelResidencyError("cancelled", runtimeLabel));
+  if (options.signal?.aborted)
+    throw new ModelResidencyError("cancelled", runtimeLabel);
+  options.signal?.addEventListener("abort", cancel, { once: true });
+  const schedule = options.scheduleTimeout ?? ((callback, ms2) => {
+    const timer = setTimeout(callback, ms2);
+    return () => clearTimeout(timer);
+  });
+  let onAbort = () => {};
+  const aborted = new Promise((_, reject) => {
+    onAbort = () => reject(controller.signal.reason);
+    controller.signal.addEventListener("abort", onAbort, { once: true });
+  });
+  const clearTimer = schedule(() => controller.abort(new ModelResidencyError("timeout", runtimeLabel)), timeoutMs);
+  try {
+    controller.signal.throwIfAborted();
+    return await Promise.race([operation(controller.signal), aborted]);
+  } finally {
+    clearTimer();
+    options.signal?.removeEventListener("abort", cancel);
+    controller.signal.removeEventListener("abort", onAbort);
+    controller.abort();
+  }
+}
+
+// src/local/lemonadeResidency.ts
 var modelName = z22.string().min(1).refine((name) => name === name.trim());
 var LoadedModelSchema = z22.object({
   model_name: modelName,
@@ -16203,36 +16458,6 @@ function lemonadeBaseURL(value) {
   }
   return `${url.origin}${url.pathname.replace(/\/$/, "")}`;
 }
-async function withResidencyDeadline(operation, options) {
-  const timeoutMs = options.timeoutMs ?? 5000;
-  if (!Number.isFinite(timeoutMs) || timeoutMs <= 0) {
-    throw new ModelResidencyError("configuration");
-  }
-  const controller = new AbortController;
-  const cancel = () => controller.abort(new ModelResidencyError("cancelled"));
-  if (options.signal?.aborted)
-    throw new ModelResidencyError("cancelled");
-  options.signal?.addEventListener("abort", cancel, { once: true });
-  const schedule = options.scheduleTimeout ?? ((callback, ms2) => {
-    const timer = setTimeout(callback, ms2);
-    return () => clearTimeout(timer);
-  });
-  let onAbort = () => {};
-  const aborted = new Promise((_, reject) => {
-    onAbort = () => reject(controller.signal.reason);
-    controller.signal.addEventListener("abort", onAbort, { once: true });
-  });
-  const clearTimer = schedule(() => controller.abort(new ModelResidencyError("timeout")), timeoutMs);
-  try {
-    controller.signal.throwIfAborted();
-    return await Promise.race([operation(controller.signal), aborted]);
-  } finally {
-    clearTimer();
-    options.signal?.removeEventListener("abort", cancel);
-    controller.signal.removeEventListener("abort", onAbort);
-    controller.abort();
-  }
-}
 async function metadata(url, fetch2, signal, headers) {
   signal.throwIfAborted();
   let payload;
@@ -16536,9 +16761,314 @@ function isObject2(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 
+// src/local/llamaSwapResidency.ts
+function isObject3(value) {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+function readPositiveSafeInteger(value) {
+  return typeof value === "number" && Number.isSafeInteger(value) && value > 0 ? value : undefined;
+}
+function readSupportsTools2(model) {
+  const direct = readBoolean2(model, [
+    "tool_call",
+    "tool_calls",
+    "toolCalling",
+    "supportsToolCalling",
+    "supports_tools",
+    "supportsTools"
+  ]);
+  if (direct !== undefined) {
+    return direct;
+  }
+  const capabilities = readCapabilityBoolean2(model.capabilities);
+  if (capabilities !== undefined) {
+    return capabilities;
+  }
+  const features = readCapabilityBoolean2(model.features);
+  if (features !== undefined) {
+    return features;
+  }
+  const labels = readCapabilityBoolean2(model.labels);
+  if (labels !== undefined) {
+    return labels;
+  }
+  return "unknown";
+}
+function readCapabilityBoolean2(value) {
+  if (Array.isArray(value)) {
+    if (value.some((entry) => typeof entry === "string" && ["tools", "tool_call", "tool-calling", "function_calling"].includes(entry))) {
+      return true;
+    }
+    return;
+  }
+  if (!isObject3(value)) {
+    return;
+  }
+  return readBoolean2(value, [
+    "tools",
+    "tool_call",
+    "tool_calls",
+    "toolCalling",
+    "function_calling",
+    "supportsToolCalling"
+  ]);
+}
+function readBoolean2(value, keys) {
+  for (const key of keys) {
+    if (typeof value[key] === "boolean") {
+      return value[key];
+    }
+  }
+  return;
+}
+function modelDetail(id, detail) {
+  return llamaSwapResidencyMessages.catalogModelDetail(id, detail);
+}
+function llamaSwapBaseURL(value) {
+  let url;
+  try {
+    url = new URL(value);
+  } catch {
+    throw new ModelResidencyError("destination", "llama-swap");
+  }
+  if (url.protocol !== "http:" && url.protocol !== "https:" || url.username !== "" || url.password !== "" || url.search !== "" || url.hash !== "") {
+    throw new ModelResidencyError("destination", "llama-swap");
+  }
+  if (url.pathname === "/" || url.pathname === "") {
+    return `${url.origin}/v1`;
+  }
+  if (url.pathname === "/v1" || url.pathname === "/v1/") {
+    return `${url.origin}/v1`;
+  }
+  throw new ModelResidencyError("destination", "llama-swap");
+}
+function parseCanonicalType(entry, id) {
+  if (!isObject3(entry.meta) || !isObject3(entry.meta.llamaswap)) {
+    return;
+  }
+  const type = entry.meta.llamaswap.type;
+  if (type === undefined) {
+    return;
+  }
+  if (typeof type !== "string" || type.length === 0 || type !== type.trim()) {
+    throw new ModelResidencyError("unknown", "llama-swap", modelDetail(id, llamaSwapResidencyMessages.invalidCatalogField("meta.llamaswap.type")));
+  }
+  return type === "model" ? "model" : undefined;
+}
+function readContextCandidates(entry) {
+  const values = [
+    readPositiveSafeInteger(entry.context_length),
+    readPositiveSafeInteger(entry.context_window),
+    isObject3(entry.meta) ? readPositiveSafeInteger(entry.meta.n_ctx) : undefined
+  ];
+  return values.filter((value) => value !== undefined);
+}
+function readDeclaredMaxOutput(entry, id) {
+  if (!isObject3(entry.meta) || !isObject3(entry.meta.llamaswap)) {
+    return;
+  }
+  if (!Object.hasOwn(entry.meta.llamaswap, "max_output_tokens")) {
+    return;
+  }
+  const maxOutputTokens = readPositiveSafeInteger(entry.meta.llamaswap.max_output_tokens);
+  if (maxOutputTokens === undefined) {
+    throw new ModelResidencyError("unknown", "llama-swap", modelDetail(id, llamaSwapResidencyMessages.invalidCatalogField("meta.llamaswap.max_output_tokens")));
+  }
+  return maxOutputTokens;
+}
+function parseCatalogModel(value) {
+  if (!isObject3(value) || typeof value.id !== "string") {
+    throw new ModelResidencyError("unknown", "llama-swap", llamaSwapResidencyMessages.catalogEntryMustExposeStringId);
+  }
+  const id = value.id;
+  if (id.length === 0 || id !== id.trim()) {
+    throw new ModelResidencyError("unknown", "llama-swap", llamaSwapResidencyMessages.catalogModelIdsMustBeNonEmptyTrimmedStrings);
+  }
+  if (parseCanonicalType(value, id) !== "model") {
+    return;
+  }
+  const model = {
+    modelID: id,
+    supportsTools: readSupportsTools2(value)
+  };
+  const contexts = readContextCandidates(value);
+  if (contexts.length > 0) {
+    model.contextWindow = Math.min(...contexts);
+    model.contextWindowProvenance = "declared";
+  }
+  const maxOutputTokens = readDeclaredMaxOutput(value, id);
+  if (maxOutputTokens !== undefined) {
+    model.maxOutputTokens = maxOutputTokens;
+    model.maxOutputTokensProvenance = "declared";
+  }
+  return model;
+}
+function parseCatalog(payload) {
+  if (!isObject3(payload) || !Array.isArray(payload.data)) {
+    throw new ModelResidencyError("unknown", "llama-swap", llamaSwapResidencyMessages.expectedDataArray("/v1/models"));
+  }
+  const seen = new Set;
+  const models = [];
+  for (const entry of payload.data) {
+    const model = parseCatalogModel(entry);
+    if (model === undefined) {
+      continue;
+    }
+    if (seen.has(model.modelID)) {
+      throw new ModelResidencyError("unknown", "llama-swap", llamaSwapResidencyMessages.duplicateCanonicalModelId(model.modelID, "/v1/models"));
+    }
+    seen.add(model.modelID);
+    models.push(model);
+  }
+  return models;
+}
+async function metadata2(url, fetch2, signal, headers) {
+  signal.throwIfAborted();
+  try {
+    const response = await fetch2(url, {
+      method: "GET",
+      headers,
+      signal,
+      redirect: "error",
+      cache: "no-store"
+    });
+    signal.throwIfAborted();
+    if (response.status === 401 || response.status === 403) {
+      throw new ModelResidencyError("authentication-unsupported", "llama-swap");
+    }
+    if (!response.ok || response.redirected || response.url !== "" && response.url !== url) {
+      throw new ModelResidencyError("unknown", "llama-swap", llamaSwapResidencyMessages.httpFailure(new URL(url).pathname, response.status));
+    }
+    return await response.json();
+  } catch (error) {
+    signal.throwIfAborted();
+    if (error instanceof ModelResidencyError) {
+      throw error;
+    }
+    throw new ModelResidencyError("unknown", "llama-swap", llamaSwapResidencyMessages.unverifiableResponse(new URL(url).pathname));
+  }
+}
+function readProfiles(payload) {
+  if (!isObject3(payload) || !Array.isArray(payload.profiles)) {
+    throw new ModelResidencyError("unknown", "llama-swap", llamaSwapResidencyMessages.expectedProfilesArray("/api/profiles"));
+  }
+  if (payload.active === null) {
+    return;
+  }
+  if (typeof payload.active === "string" && payload.active.length > 0 && payload.active === payload.active.trim()) {
+    throw new ModelResidencyError("identity", "llama-swap", llamaSwapResidencyMessages.activeProfileRewrite(payload.active));
+  }
+  throw new ModelResidencyError("unknown", "llama-swap", llamaSwapResidencyMessages.invalidActiveProfileValue("/api/profiles"));
+}
+function readRunning(payload, modelIDs) {
+  if (!isObject3(payload) || !Array.isArray(payload.running)) {
+    throw new ModelResidencyError("unknown", "llama-swap", llamaSwapResidencyMessages.expectedRunningArray("/running"));
+  }
+  const result = new Map;
+  for (const modelID of modelIDs) {
+    result.set(modelID, { loaded: false });
+  }
+  const seen = new Set;
+  for (const entry of payload.running) {
+    if (!isObject3(entry) || typeof entry.model !== "string" || entry.model.length === 0 || entry.model !== entry.model.trim() || typeof entry.state !== "string" || entry.state.length === 0 || entry.state !== entry.state.trim()) {
+      throw new ModelResidencyError("unknown", "llama-swap", llamaSwapResidencyMessages.malformedRunningEvidence("/running"));
+    }
+    if (seen.has(entry.model)) {
+      throw new ModelResidencyError("unknown", "llama-swap", llamaSwapResidencyMessages.duplicateRunningEvidence("/running", entry.model));
+    }
+    seen.add(entry.model);
+    if (!modelIDs.has(entry.model)) {
+      continue;
+    }
+    result.set(entry.model, { loaded: entry.state === "ready" });
+  }
+  return result;
+}
+function requestHeaders(headers) {
+  const result = new Headers(headers);
+  result.delete("content-type");
+  result.delete("content-length");
+  result.set("accept", "application/json");
+  result.set("cache-control", "no-cache, no-store");
+  result.delete("if-none-match");
+  result.delete("if-modified-since");
+  return result;
+}
+function llamaSwapResidencyOptions(options) {
+  return { ...options, runtimeLabel: "llama-swap" };
+}
+async function readLlamaSwapCatalog(baseURL, fetch2, options = {}) {
+  const base = llamaSwapBaseURL(baseURL);
+  const headers = requestHeaders(options.headers);
+  return withResidencyDeadline(async (signal) => parseCatalog(await metadata2(`${base}/models`, fetch2, signal, headers)), llamaSwapResidencyOptions(options));
+}
+async function readLlamaSwapResidency(baseURL, fetch2, options = {}) {
+  const base = llamaSwapBaseURL(baseURL);
+  const origin = new URL(base).origin;
+  const headers = requestHeaders(options.headers);
+  return withResidencyDeadline(async (signal) => {
+    const models = await readLlamaSwapCatalog(base, fetch2, {
+      ...llamaSwapResidencyOptions(options),
+      signal
+    });
+    readProfiles(await metadata2(`${origin}/api/profiles`, fetch2, signal, headers));
+    return readRunning(await metadata2(`${origin}/running`, fetch2, signal, headers), new Set(models.map((model) => model.modelID)));
+  }, llamaSwapResidencyOptions(options));
+}
+
+// src/local/llamaSwap.ts
+var LLAMA_SWAP_DEFAULT_BASE_URL = "http://localhost:8080/v1";
+function withLoadedState(models, residency) {
+  return models.map((model) => {
+    const loaded = residency.get(model.modelID)?.loaded;
+    return loaded === undefined ? model : { ...model, loaded };
+  });
+}
+function createLlamaSwapAdapter() {
+  const listModels = async (options) => {
+    const baseURL = llamaSwapBaseURL(options.baseURL ?? LLAMA_SWAP_DEFAULT_BASE_URL);
+    const models = await readLlamaSwapCatalog(baseURL, options.fetch);
+    try {
+      return withLoadedState(models, await readLlamaSwapResidency(baseURL, options.fetch));
+    } catch (error) {
+      if (error instanceof ModelResidencyError && (error.code === "identity" || error.code === "unknown")) {
+        return models;
+      }
+      throw error;
+    }
+  };
+  return {
+    id: "llama-swap",
+    defaultBaseURL: LLAMA_SWAP_DEFAULT_BASE_URL,
+    listModels,
+    async probe(options) {
+      const baseURL = llamaSwapBaseURL(options.baseURL ?? LLAMA_SWAP_DEFAULT_BASE_URL);
+      try {
+        return {
+          id: "llama-swap",
+          baseURL,
+          reachable: true,
+          models: await listModels({ ...options, baseURL }),
+          probedAt: options.probedAt
+        };
+      } catch (error) {
+        return unavailableSnapshot({
+          id: "llama-swap",
+          baseURL,
+          probedAt: options.probedAt,
+          error
+        });
+      }
+    }
+  };
+}
+
 // src/local/modelResidency.ts
+function runtimeRequiresModelResidency(kind) {
+  return kind === "lemonade" || kind === "llama-swap";
+}
 function hasRequiredModelResidency(kind, model) {
-  return kind !== "lemonade" || model?.loaded === true;
+  return !runtimeRequiresModelResidency(kind) || model?.loaded === true;
 }
 
 // src/local/lmstudio.ts
@@ -16593,6 +17123,7 @@ class RuntimeRegistry {
       lmstudio: createLMStudioAdapter(),
       "foundry-local": createFoundryLocalAdapter(),
       lemonade: createLemonadeAdapter(),
+      "llama-swap": createLlamaSwapAdapter(),
       ...options.adapters
     };
     this.clock = options.clock;
@@ -16730,6 +17261,12 @@ var DETECTION_RUNTIMES = [
     baseURL: LEMONADE_DEFAULT_BASE_URL,
     defaultModel: { providerID: "lemonade", modelID: "probe" }
   },
+  {
+    id: "llama-swap",
+    enabled: true,
+    baseURL: LLAMA_SWAP_DEFAULT_BASE_URL,
+    defaultModel: { providerID: "llama-swap", modelID: "probe" }
+  },
   {
     id: "foundry-local",
     enabled: true,
@@ -17228,7 +17765,7 @@ function runtimeLine(snapshot, runtime) {
   const mark = snapshot.reachable ? "✓" : "✗";
   const detail = snapshot.reachable ? `${snapshot.models.length} model(s)` : snapshot.error ?? "unreachable";
   const declared = runtime?.maxConcurrency;
-  const slots = declared === undefined ? ` · ${DEFAULT_LOCAL_MAX_CONCURRENCY} slot(s) (default)` : ` · ${declared} slot(s) (declared)`;
+  const slots = declared === undefined ? doctorMessages.defaultRuntimeSlots(effectiveMaxConcurrency(runtime ?? {})) : ` · ${declared} slot(s) (declared)`;
   return `  ${mark} ${snapshot.id.padEnd(14)} ${snapshot.baseURL || "(no baseURL)"} — ${detail}${slots}`;
 }
 function runtimeModelSlotsLine(runtime, snapshot) {
@@ -17240,6 +17777,8 @@ function runtimeModelSlotsLine(runtime, snapshot) {
   if (runtime === undefined || snapshot === undefined || !snapshot.reachable || snapshot.models.length === 0) {
     return [];
   }
+  if (localRuntimeKind(runtime) === "llama-swap")
+    return [doctorMessages.llamaSwapConcurrency];
   const reported = snapshot.models.flatMap((model) => model.slots === undefined ? [] : [`${model.modelID} ${model.slots} slot(s)`]);
   if (reported.length > 0) {
     return [
@@ -17643,7 +18182,7 @@ function renderDoctor(input) {
       continue;
     }
     const runtime = enabledRuntimes.find((r) => r.id === snapshot.id);
-    const cap = runtime?.maxConcurrency ?? DEFAULT_LOCAL_MAX_CONCURRENCY;
+    const cap = effectiveMaxConcurrency(runtime ?? {});
     lines.push(`  ⚠ ${snapshot.id} is reached over the network: its ${cap}-slot cap is enforced per openteam process, so`, "    two processes at once (an opencode session plus 'openteam console') can together exceed it.");
   }
   lines.push(`  telemetry: ${input.telemetryPath} — ${input.telemetryRecords} record(s)`);
@@ -18046,7 +18585,7 @@ import { posix as pathPosix, win32 as pathWin32 } from "node:path";
 // package.json
 var package_default = {
   name: "@jmanuelcorral/openteam",
-  version: "0.24.1",
+  version: "0.26.0",
   packageManager: "bun@1.3.14",
   description: "Cost-aware, local-first routing plugin for opencode with cheapest-capable frontier fallback and multi-agent orchestration.",
   license: "MIT",
@@ -18365,14 +18904,25 @@ function frontierModelsForProvider(profiles, providerID) {
 var OPENCODE_GITIGNORE_PATH = ".opencode/.gitignore";
 var DEFAULT_LOCAL_OUTPUT_TOKENS = 8192;
 var MIN_USABLE_INPUT_TOKENS = 1024;
-function computeModelLimit(contextWindow, inputBudget) {
+function computeModelLimit(contextWindow, inputBudget, maxOutputTokens) {
   const context = contextWindow ?? 0;
   const contextBudget = context > 0 ? context : undefined;
   const validInputBudget = inputBudget !== undefined && inputBudget > 0 ? inputBudget : undefined;
   const effectiveBudget = contextBudget === undefined ? validInputBudget : validInputBudget === undefined ? contextBudget : Math.min(contextBudget, validInputBudget);
-  const output = effectiveBudget !== undefined && effectiveBudget - DEFAULT_LOCAL_OUTPUT_TOKENS < MIN_USABLE_INPUT_TOKENS ? Math.max(1, Math.floor(effectiveBudget / 2)) : DEFAULT_LOCAL_OUTPUT_TOKENS;
+  const desired = Math.min(DEFAULT_LOCAL_OUTPUT_TOKENS, positiveInteger3(maxOutputTokens) ?? DEFAULT_LOCAL_OUTPUT_TOKENS);
+  const output = Math.min(desired, effectiveBudget !== undefined && effectiveBudget - desired < MIN_USABLE_INPUT_TOKENS ? Math.max(1, Math.floor(effectiveBudget / 2)) : desired);
   return { context, output };
 }
+function collectMaxOutputTokens(modelInfo, usedModelIDs) {
+  const used = new Set(usedModelIDs);
+  const ceilings = {};
+  for (const model of modelInfo ?? []) {
+    const ceiling = positiveInteger3(model.maxOutputTokens);
+    if (ceiling !== undefined && used.has(model.modelID))
+      ceilings[model.modelID] = ceiling;
+  }
+  return Object.keys(ceilings).length > 0 ? ceilings : undefined;
+}
 function collectContextWindows(modelInfo, usedModelIDs) {
   if (modelInfo === undefined) {
     return;
@@ -18489,6 +19039,12 @@ var KNOWN_RUNTIMES = [
     defaultBaseURL: LEMONADE_DEFAULT_BASE_URL,
     fallbackModelID: "Qwen3-Coder-30B-A3B-Instruct-GGUF"
   },
+  {
+    id: "llama-swap",
+    label: executionSetupMessages.llamaSwap.label,
+    defaultBaseURL: LLAMA_SWAP_DEFAULT_BASE_URL,
+    fallbackModelID: ""
+  },
   {
     id: "foundry-local",
     label: "Foundry Local",
@@ -18496,7 +19052,7 @@ var KNOWN_RUNTIMES = [
     dynamicPort: true
   }
 ];
-var LEMONADE_PROJECT_CONFIG_CANDIDATES = [
+var RESIDENT_PROJECT_CONFIG_CANDIDATES = [
   ".opencode/opencode.jsonc",
   ".opencode/opencode.json",
   "opencode.jsonc",
@@ -18518,6 +19074,19 @@ function choiceKind(choice) {
 function choiceProviderID(choice) {
   return choice.providerID ?? choice.id;
 }
+function residencySetupMessages(kind) {
+  return kind === "llama-swap" ? executionSetupMessages.llamaSwap : {
+    noLoadedModels: executionSetupMessages.lemonadeNoLoadedModels,
+    modelNotLoaded: executionSetupMessages.lemonadeModelNotLoaded,
+    probeRequired: executionSetupMessages.lemonadeProbeRequired,
+    probeFailed: executionSetupMessages.lemonadeProbeFailed,
+    endpointMismatch: executionSetupMessages.lemonadeEndpointMismatch,
+    modelIdentityUnknown: executionSetupMessages.lemonadeModelIdentityUnknown
+  };
+}
+function residencyBaseURL(kind, value) {
+  return kind === "llama-swap" ? llamaSwapBaseURL(value ?? LLAMA_SWAP_DEFAULT_BASE_URL) : lemonadeBaseURL(value ?? LEMONADE_DEFAULT_BASE_URL);
+}
 function selectableModelIDs(models, modelInfo, kind) {
   const byID = new Map(modelInfo?.map((model) => [model.modelID, model]));
   return models.filter((id) => hasRequiredModelResidency(kind, byID.get(id)));
@@ -18535,7 +19104,7 @@ function assertRuntimeResidency(choice, primary) {
       continue;
     const info = choice.models.includes(id) ? choice.modelInfo?.find((model) => model.modelID === id) : undefined;
     if (!hasRequiredModelResidency(kind, info))
-      throw new Error(executionSetupMessages.lemonadeModelNotLoaded(`${choiceProviderID(choice)}/${id}`));
+      throw new Error(residencySetupMessages(kind).modelNotLoaded(`${choiceProviderID(choice)}/${id}`));
   }
 }
 function assertSetupResidency(answers) {
@@ -18543,11 +19112,11 @@ function assertSetupResidency(answers) {
     if (choice.enabled) {
       assertRuntimeResidency(choice, answers.primaryModel);
     } else if (choiceProviderID(choice) === answers.primaryModel.providerID && !hasRequiredModelResidency(localRuntimeKind(choice), undefined)) {
-      throw new Error(executionSetupMessages.lemonadeModelNotLoaded(`${answers.primaryModel.providerID}/${answers.primaryModel.modelID}`));
+      throw new Error(residencySetupMessages(localRuntimeKind(choice)).modelNotLoaded(`${answers.primaryModel.providerID}/${answers.primaryModel.modelID}`));
     }
   }
 }
-function lemonadeV2Policies(providers) {
+function residentV2Policies(providers) {
   return [...new Set(providers)].sort().map((providerID) => {
     if (providerID.length === 0 || /[*?\\/\s]/.test(providerID))
       throw new Error(executionSetupMessages.lemonadeV2InvalidProvider(providerID));
@@ -18564,7 +19133,7 @@ function buildOpenTeamConfig(answers) {
         providerID: choiceProviderID(choice),
         modelID: choice.defaultModelID
       },
-      maxConcurrency: choice.maxConcurrency ?? DEFAULT_LOCAL_MAX_CONCURRENCY
+      maxConcurrency: effectiveMaxConcurrency(choice)
     };
     if (choice.modelConcurrency !== undefined && Object.keys(choice.modelConcurrency).length > 0) {
       runtime.modelConcurrency = { ...choice.modelConcurrency };
@@ -18579,7 +19148,7 @@ function buildOpenTeamConfig(answers) {
       runtime.kind = choice.kind;
     }
     if (choice.baseURL !== undefined) {
-      runtime.baseURL = choice.baseURL;
+      runtime.baseURL = localRuntimeKind(choice) === "llama-swap" ? llamaSwapBaseURL(choice.baseURL) : choice.baseURL;
     } else if (knownRuntime(choiceKind(choice)).dynamicPort === true) {
       runtime.discovery = "cli";
     }
@@ -18625,7 +19194,7 @@ function orchestratorAgentModel(answers) {
 }
 function buildOpencodeConfig(answers) {
   assertSetupResidency(answers);
-  const policies = lemonadeV2Policies(answers.runtimes.filter((choice) => choice.enabled && choiceKind(choice) === "lemonade").map(choiceProviderID));
+  const policies = residentV2Policies(answers.runtimes.filter((choice) => choice.enabled && runtimeRequiresModelResidency(localRuntimeKind(choice))).map(choiceProviderID));
   const provider = {};
   const providerOwner = new Map;
   for (const choice of answers.runtimes) {
@@ -18633,7 +19202,8 @@ function buildOpencodeConfig(answers) {
       continue;
     }
     const meta = knownRuntime(choiceKind(choice));
-    const baseURL = choice.baseURL ?? meta.defaultBaseURL;
+    const configuredBaseURL = choice.baseURL ?? meta.defaultBaseURL;
+    const baseURL = meta.id === "llama-swap" ? llamaSwapBaseURL(configuredBaseURL ?? LLAMA_SWAP_DEFAULT_BASE_URL) : configuredBaseURL;
     if (baseURL === undefined) {
       continue;
     }
@@ -18652,7 +19222,7 @@ function buildOpencodeConfig(answers) {
     const models = {};
     for (const modelID of modelIDs) {
       const contextWindow = choice.modelContextWindows?.[modelID];
-      const limit = computeModelLimit(contextWindow);
+      const limit = computeModelLimit(contextWindow, undefined, choice.modelMaxOutputTokens?.[modelID]);
       models[modelID] = { name: modelID, limit };
     }
     const owner = providerOwner.get(providerID);
@@ -18833,8 +19403,8 @@ function preferredModelID(models, modelInfo, kind) {
 async function chooseRuntimeModel(prompt, models, meta, modelInfo) {
   const offered = selectableModelIDs(models, modelInfo, meta.id);
   if (offered.length === 0) {
-    if (meta.id === "lemonade")
-      throw new Error(executionSetupMessages.lemonadeNoLoadedModels);
+    if (runtimeRequiresModelResidency(meta.id))
+      throw new Error(residencySetupMessages(meta.id).noLoadedModels);
     return prompt.text({
       message: `Default model for ${meta.label} (no models detected)`,
       placeholder: meta.fallbackModelID,
@@ -18850,8 +19420,8 @@ async function chooseRuntimeModel(prompt, models, meta, modelInfo) {
     opts.initial = first;
   }
   const selected = await prompt.select(opts);
-  if (meta.id === "lemonade" && !offered.includes(selected))
-    throw new Error(executionSetupMessages.lemonadeModelNotLoaded(selected));
+  if (runtimeRequiresModelResidency(meta.id) && !offered.includes(selected))
+    throw new Error(residencySetupMessages(meta.id).modelNotLoaded(selected));
   return selected;
 }
 function remoteBaseURLExample(meta) {
@@ -18907,28 +19477,31 @@ async function resolveRuntimeConnection(deps, meta, info) {
     initial: RUNTIME_LOCATION_LOCAL
   });
   if (location === RUNTIME_LOCATION_LOCAL) {
-    if (meta.id === "lemonade" && info?.reachable !== true)
-      throw new Error(executionSetupMessages.lemonadeProbeFailed);
+    if (runtimeRequiresModelResidency(meta.id) && info?.reachable !== true)
+      throw new Error(info?.error ?? residencySetupMessages(meta.id).probeFailed);
     const baseURL2 = info?.baseURL !== undefined && info.baseURL.length > 0 ? info.baseURL : meta.defaultBaseURL;
-    return { baseURL: baseURL2, ...connectionOf(info) };
+    return {
+      baseURL: meta.id === "llama-swap" ? llamaSwapBaseURL(baseURL2) : baseURL2,
+      ...connectionOf(info)
+    };
   }
   const entered = await deps.prompt.text({
     message: `${meta.label} URL on the network (OpenAI-compatible endpoint)`,
     placeholder: remoteBaseURLExample(meta)
   });
-  const baseURL = normalizeRemoteBaseURL(entered, meta.defaultBaseURL);
+  const baseURL = meta.id === "llama-swap" ? llamaSwapBaseURL(entered) : normalizeRemoteBaseURL(entered, meta.defaultBaseURL);
   let probedRuntime;
-  if (meta.id === "lemonade" && deps.probeRuntime === undefined)
-    throw new Error(executionSetupMessages.lemonadeProbeRequired);
+  if (runtimeRequiresModelResidency(meta.id) && deps.probeRuntime === undefined)
+    throw new Error(residencySetupMessages(meta.id).probeRequired);
   if (deps.probeRuntime !== undefined) {
     try {
       probedRuntime = await deps.probeRuntime(meta.id, baseURL);
     } catch (cause) {
-      if (meta.id === "lemonade")
-        throw new Error(executionSetupMessages.lemonadeProbeFailed, { cause });
+      if (runtimeRequiresModelResidency(meta.id))
+        throw new Error(residencySetupMessages(meta.id).probeFailed, { cause });
     }
-    if (meta.id === "lemonade" && probedRuntime?.reachable !== true)
-      throw new Error(executionSetupMessages.lemonadeProbeFailed);
+    if (runtimeRequiresModelResidency(meta.id) && probedRuntime?.reachable !== true)
+      throw new Error(probedRuntime?.error ?? residencySetupMessages(meta.id).probeFailed);
   }
   return { baseURL, ...connectionOf(probedRuntime) };
 }
@@ -18936,8 +19509,8 @@ async function chooseClassModel(prompt, models, meta, defaultModelID, modelClass
   const label = modelClass === "thinking" ? `Thinking model for ${meta.label} (design, analysis, review reasoning)` : `Coder model for ${meta.label} (writing and editing code)`;
   const offered = selectableModelIDs(models, modelInfo, meta.id);
   if (offered.length === 0) {
-    if (meta.id === "lemonade")
-      throw new Error(executionSetupMessages.lemonadeNoLoadedModels);
+    if (runtimeRequiresModelResidency(meta.id))
+      throw new Error(residencySetupMessages(meta.id).noLoadedModels);
     const raw = await prompt.text({
       message: `${label} — enter to reuse the default`,
       placeholder: defaultModelID,
@@ -18953,8 +19526,8 @@ async function chooseClassModel(prompt, models, meta, defaultModelID, modelClass
     })),
     initial: defaultModelID
   });
-  if (meta.id === "lemonade" && !offered.includes(selected))
-    throw new Error(executionSetupMessages.lemonadeModelNotLoaded(selected));
+  if (runtimeRequiresModelResidency(meta.id) && !offered.includes(selected))
+    throw new Error(residencySetupMessages(meta.id).modelNotLoaded(selected));
   return selected;
 }
 async function resolveEnabledRuntime(deps, meta, info) {
@@ -18991,6 +19564,16 @@ async function resolveEnabledRuntime(deps, meta, info) {
   if (modelContextWindows !== undefined) {
     choice.modelContextWindows = modelContextWindows;
   }
+  const modelMaxOutputTokens = collectMaxOutputTokens(connection.modelInfo, [
+    defaultModelID,
+    thinkingModelID,
+    coderModelID,
+    ...connection.models
+  ]);
+  if (modelMaxOutputTokens !== undefined)
+    choice.modelMaxOutputTokens = modelMaxOutputTokens;
+  if (meta.id === "llama-swap")
+    deps.prompt.note(executionSetupMessages.llamaSwap.declaredLimits, executionSetupMessages.llamaSwap.declaredLimitsTitle);
   return choice;
 }
 function reportedModelSlots(modelInfo, usedModelIDs) {
@@ -19017,11 +19600,11 @@ async function promptSlotCount(prompt, meta, reported) {
   }
   const raw = await prompt.text({
     message: `How many concurrent generation slots does ${meta.label} have? (match its real capacity; leaving one free keeps interactive work responsive)`,
-    placeholder: String(DEFAULT_LOCAL_MAX_CONCURRENCY),
-    initial: String(DEFAULT_LOCAL_MAX_CONCURRENCY)
+    placeholder: String(effectiveMaxConcurrency({ id: meta.id })),
+    initial: String(effectiveMaxConcurrency({ id: meta.id }))
   });
   const parsed = Number.parseInt(raw.trim(), 10);
-  return Number.isInteger(parsed) && parsed > 0 ? parsed : DEFAULT_LOCAL_MAX_CONCURRENCY;
+  return Number.isInteger(parsed) && parsed > 0 ? parsed : effectiveMaxConcurrency({ id: meta.id });
 }
 function reportedSlotSummary(reported) {
   const entries = Object.entries(reported ?? {});
@@ -19080,6 +19663,29 @@ async function resolveRuntimeInstances(deps, meta, info, usedIds) {
 function isRecord6(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value);
 }
+function setupOutputCeilings(answers, existing) {
+  const result = {};
+  const providers = recordAt2(existing, "provider");
+  for (const choice of answers.runtimes.filter((runtime) => runtime.enabled)) {
+    const physical = choice.modelMaxOutputTokens;
+    if (physical === undefined)
+      continue;
+    const providerID = choiceProviderID(choice);
+    const ceilings = { ...physical };
+    const models = recordAt2(recordAt2(providers, providerID), "models");
+    for (const [alias, value] of Object.entries(models ?? {})) {
+      if (!isRecord6(value))
+        continue;
+      const id = typeof value.id === "string" ? value.id : alias;
+      delete ceilings[alias];
+      const ceiling = physical[id];
+      if (ceiling !== undefined)
+        ceilings[alias] = ceiling;
+    }
+    result[providerID] = ceilings;
+  }
+  return result;
+}
 var LIMIT_FIELDS = [
   "context",
   "input",
@@ -19569,7 +20175,7 @@ function readModelLimitValues(limit) {
 function mergeModelLimitForRerun(input) {
   const target = readModelLimitValues(input.targetLimit);
   const effective = readModelLimitValues(input.effectiveLimit);
-  const computed = computeModelLimit(effective.context ?? input.generatedLimit?.context, effective.input);
+  const computed = computeModelLimit(effective.context ?? input.generatedLimit?.context, effective.input, input.maxOutputTokens);
   const merged = { ...input.targetLimit ?? {} };
   const targetOutputIsGeneratedDefault = target.output === DEFAULT_LOCAL_OUTPUT_TOKENS;
   if (targetOutputIsGeneratedDefault && effective.output !== undefined && effective.output !== DEFAULT_LOCAL_OUTPUT_TOKENS) {
@@ -19583,6 +20189,8 @@ function mergeModelLimitForRerun(input) {
   } else {
     merged.output = computed.output;
   }
+  if (input.maxOutputTokens !== undefined && typeof merged.output === "number")
+    merged.output = Math.min(merged.output, input.maxOutputTokens);
   if (target.context !== undefined) {
     merged.context = target.context;
   } else if (effective.context !== undefined) {
@@ -19602,7 +20210,8 @@ function completePhysicalModelLimit(input) {
   const target = readModelLimitValues(input.targetLimit);
   const effective = readModelLimitValues(input.effectiveLimit);
   const context = target.context ?? effective.context ?? input.generatedLimit?.context ?? 0;
-  const output = target.output ?? effective.output ?? computeModelLimit(context, effective.input).output;
+  const requestedOutput = target.output ?? effective.output ?? computeModelLimit(context, effective.input, input.maxOutputTokens).output;
+  const output = input.maxOutputTokens === undefined ? requestedOutput : Math.min(requestedOutput, input.maxOutputTokens);
   return { ...input.targetLimit, context, output };
 }
 function repairManagedModelLimits(input) {
@@ -19623,16 +20232,18 @@ function repairManagedModelLimits(input) {
     let providerChanged = false;
     const repairedModels = { ...targetModels };
     for (const [modelID, modelValue] of Object.entries(targetModels)) {
-      if (!isRecord6(modelValue) || !hasOwn2(modelValue, "limit")) {
+      const ceiling = input.modelOutputCeilings?.[providerID]?.[modelID];
+      if (!isRecord6(modelValue) || !hasOwn2(modelValue, "limit") && ceiling === undefined) {
         continue;
       }
-      const targetLimit = recordAt2(modelValue, "limit");
+      const targetLimit = recordAt2(modelValue, "limit") ?? (ceiling === undefined ? undefined : {});
       if (targetLimit === undefined) {
         continue;
       }
       const effectiveLimit = recordAt2(recordAt2(effectiveModels, modelID), "limit");
       const generatedLimit = generatedModelLimit(input.generated, providerID, modelID);
       const repairedLimit = completePhysicalModelLimit({
+        maxOutputTokens: ceiling,
         ...generatedLimit !== undefined ? { generatedLimit } : {},
         targetLimit,
         ...effectiveLimit !== undefined ? { effectiveLimit } : {}
@@ -19691,7 +20302,7 @@ function highestExistingPinVersion(files) {
   }
   return highest;
 }
-function reconcileLemonadeV2Policies(raw, denials) {
+function reconcileResidentV2Policies(raw, denials) {
   if (denials.length === 0)
     return raw;
   const experimental = raw.experimental;
@@ -19750,6 +20361,7 @@ function mergeOpencodeConfigForRerun(generated, targetRaw, options) {
       const targetLimit = recordAt2(targetModel, "limit");
       const effectiveLimit = recordAt2(effectiveModel, "limit");
       const mergedLimit = mergeModelLimitForRerun({
+        maxOutputTokens: options.modelOutputCeilings?.[providerID]?.[modelID],
         ...modelConfig.limit !== undefined ? { generatedLimit: modelConfig.limit } : {},
         ...targetLimit !== undefined ? { targetLimit } : {},
         ...effectiveLimit !== undefined ? { effectiveLimit } : {}
@@ -19792,7 +20404,7 @@ function mergeOpencodeConfigForRerun(generated, targetRaw, options) {
   if (managesV2Policies && hasOwn2(targetRaw, "experimental")) {
     result.experimental = targetRaw.experimental;
   }
-  return reconcileLemonadeV2Policies(result, generated.experimental?.policies ?? []);
+  return reconcileResidentV2Policies(result, generated.experimental?.policies ?? []);
 }
 function mergeOpenTeamConfigForRerun(generated, existing) {
   const existingOrch = existing.orchestrator;
@@ -19842,9 +20454,9 @@ function mergeOpenTeamConfigForRerun(generated, existing) {
   assertEffectiveConfigModelDomains(merged);
   return merged;
 }
-function preserveRetainedLemonadeRuntimes(generated, previous, opencode) {
+function preserveRetainedResidentRuntimes(generated, previous, opencode) {
   const providers = recordAt2(opencode, "provider") ?? {};
-  const retained = (previous?.local.runtimes ?? []).filter((runtime) => localRuntimeKind(runtime) === "lemonade" && hasOwn2(providers, runtime.defaultModel.providerID));
+  const retained = (previous?.local.runtimes ?? []).filter((runtime) => runtimeRequiresModelResidency(localRuntimeKind(runtime)) && hasOwn2(providers, runtime.defaultModel.providerID));
   const deselected = retained.filter((runtime) => !generated.local.runtimes.some((selected) => selected.defaultModel.providerID === runtime.defaultModel.providerID)).map((runtime) => ({ ...runtime, enabled: false }));
   return {
     config: deselected.length === 0 ? generated : OpenTeamConfigSchema.parse({
@@ -19884,31 +20496,33 @@ async function confirmJsonNormalization(prompt, path2, format) {
     initial: false
   });
 }
-function assertPreviousLemonadePrimary(answers, previousConfig) {
-  const priorLemonadePrimary = previousConfig?.local.runtimes.some((runtime) => runtime.defaultModel.providerID === answers.primaryModel.providerID && localRuntimeKind(runtime) === "lemonade");
-  if (priorLemonadePrimary && !answers.runtimes.some((choice) => choice.enabled && choiceProviderID(choice) === answers.primaryModel.providerID)) {
-    throw new Error(executionSetupMessages.lemonadeModelNotLoaded(`${answers.primaryModel.providerID}/${answers.primaryModel.modelID}`));
+function assertPreviousResidentPrimary(answers, previousConfig) {
+  const priorResidentPrimary = previousConfig?.local.runtimes.find((runtime) => runtime.defaultModel.providerID === answers.primaryModel.providerID && runtimeRequiresModelResidency(localRuntimeKind(runtime)));
+  if (priorResidentPrimary && !answers.runtimes.some((choice) => choice.enabled && choiceProviderID(choice) === answers.primaryModel.providerID)) {
+    throw new Error(residencySetupMessages(localRuntimeKind(priorResidentPrimary)).modelNotLoaded(`${answers.primaryModel.providerID}/${answers.primaryModel.modelID}`));
   }
 }
-async function verifyLemonadeBeforeWrite(deps, answers, config, opencode) {
+async function verifyResidencyBeforeWrite(deps, answers, config, opencode) {
   for (const choice of answers.runtimes) {
-    if (choice.enabled && localRuntimeKind(choice) === "lemonade" && !config.local.runtimes.some((runtime) => runtime.enabled && runtime.defaultModel.providerID === choiceProviderID(choice) && localRuntimeKind(runtime) === "lemonade"))
-      throw new Error(executionSetupMessages.lemonadeEndpointMismatch);
+    if (choice.enabled && runtimeRequiresModelResidency(localRuntimeKind(choice)) && !config.local.runtimes.some((runtime) => runtime.enabled && runtime.defaultModel.providerID === choiceProviderID(choice) && localRuntimeKind(runtime) === localRuntimeKind(choice)))
+      throw new Error(residencySetupMessages(localRuntimeKind(choice)).endpointMismatch);
   }
   for (const runtime of config.local.runtimes) {
-    if (localRuntimeKind(runtime) !== "lemonade")
+    const kind = localRuntimeKind(runtime);
+    if (!runtimeRequiresModelResidency(kind))
       continue;
+    const messages = residencySetupMessages(kind);
     const providerID = runtime.defaultModel.providerID;
     const provider = isRecord6(opencode.provider) ? opencode.provider[providerID] : undefined;
     if (!runtime.enabled && provider === undefined)
       continue;
     const endpoint = isRecord6(provider) ? (isRecord6(provider.options) ? provider.options.baseURL : undefined) ?? provider.api : undefined;
-    if (typeof endpoint !== "string" || lemonadeBaseURL(endpoint) !== lemonadeBaseURL(runtime.baseURL ?? LEMONADE_DEFAULT_BASE_URL) || config.local.runtimes.filter((candidate) => candidate.defaultModel.providerID === providerID).length !== 1)
-      throw new Error(executionSetupMessages.lemonadeEndpointMismatch);
+    if (typeof endpoint !== "string" || residencyBaseURL(kind, endpoint) !== residencyBaseURL(kind, runtime.baseURL) || config.local.runtimes.filter((candidate) => candidate.defaultModel.providerID === providerID).length !== 1)
+      throw new Error(messages.endpointMismatch);
     if (!runtime.enabled)
       continue;
     if (deps.probeRuntime === undefined)
-      throw new Error(executionSetupMessages.lemonadeProbeRequired);
+      throw new Error(messages.probeRequired);
     const references2 = new Set([
       runtime.defaultModel.modelID,
       runtime.thinkingModel,
@@ -19946,21 +20560,30 @@ async function verifyLemonadeBeforeWrite(deps, answers, config, opencode) {
       const declaration = recordAt2(models, alias);
       const wireID = declaration !== undefined && hasOwn2(declaration, "id") ? declaration.id : alias;
       if (declaration === undefined || typeof wireID !== "string" || wireID.length === 0 || wireID !== wireID.trim())
-        throw new Error(executionSetupMessages.lemonadeModelIdentityUnknown(`${providerID}/${alias}`));
+        throw new Error(messages.modelIdentityUnknown(`${providerID}/${alias}`));
       wireIDs.add(wireID);
     }
     let fresh;
     try {
       fresh = await deps.probeRuntime(localRuntimeKind(runtime), endpoint);
     } catch (cause) {
-      throw new Error(executionSetupMessages.lemonadeProbeFailed, { cause });
+      throw new Error(messages.probeFailed, { cause });
     }
-    if (!fresh.reachable || fresh.id !== localRuntimeKind(runtime) || fresh.baseURL.replace(/\/+$/, "") !== endpoint.replace(/\/+$/, ""))
-      throw new Error(executionSetupMessages.lemonadeProbeFailed);
+    if (!fresh.reachable || fresh.id !== localRuntimeKind(runtime) || residencyBaseURL(kind, fresh.baseURL) !== residencyBaseURL(kind, endpoint))
+      throw new Error(fresh.error ?? messages.probeFailed);
     for (const wireID of wireIDs) {
       const info = fresh.models.includes(wireID) ? fresh.modelInfo?.find((model) => model.modelID === wireID) : undefined;
-      if (!hasRequiredModelResidency("lemonade", info))
-        throw new Error(executionSetupMessages.lemonadeModelNotLoaded(`${providerID}/${wireID}`));
+      if (!hasRequiredModelResidency(kind, info))
+        throw new Error(messages.modelNotLoaded(`${providerID}/${wireID}`));
+    }
+    for (const [alias, declaration] of Object.entries(models ?? {})) {
+      if (!isRecord6(declaration))
+        continue;
+      const wireID = typeof declaration.id === "string" ? declaration.id : alias;
+      const ceiling = fresh.modelInfo?.find((model) => model.modelID === wireID)?.maxOutputTokens;
+      const output = positiveInteger3(recordAt2(declaration, "limit")?.output);
+      if (ceiling !== undefined && output !== undefined && output > ceiling)
+        throw new Error(executionSetupMessages.llamaSwap.outputChanged(`${providerID}/${alias}`, ceiling));
     }
   }
 }
@@ -19975,7 +20598,7 @@ async function runSetup(deps) {
       return {
         value: meta.id,
         label: meta.label,
-        hint: info?.reachable ? meta.id === "lemonade" ? executionSetupMessages.lemonadeDetectedHint(selectableModelIDs(info.models, info.modelInfo, meta.id).length, info.modelInfo?.length ?? info.models.length) : `detected · ${info.models.length} model(s)` : "not detected (you can point it to another machine)",
+        hint: info?.reachable ? runtimeRequiresModelResidency(meta.id) ? executionSetupMessages.lemonadeDetectedHint(selectableModelIDs(info.models, info.modelInfo, meta.id).length, info.modelInfo?.length ?? info.models.length) : `detected · ${info.models.length} model(s)` : "not detected (you can point it to another machine)",
         selected: info?.reachable ?? false
       };
     });
@@ -20092,7 +20715,7 @@ async function runSetup(deps) {
     };
     const openTeamConfig = buildOpenTeamConfig(answers);
     const opencodeConfig = buildOpencodeConfig(answers);
-    const opencodeConfigPaths = LEMONADE_PROJECT_CONFIG_CANDIDATES;
+    const opencodeConfigPaths = RESIDENT_PROJECT_CONFIG_CANDIDATES;
     const orchestratorAgent = buildOrchestratorAgent(orchestratorAgentModel(answers), { yolo });
     const slashCommands = buildOpenteamCommands();
     const openTeamPath = DEFAULT_CONFIG_PATH;
@@ -20113,29 +20736,44 @@ async function runSetup(deps) {
       }
       throw error;
     }
-    assertPreviousLemonadePrimary(answers, existingOpenTeamConfig);
+    assertPreviousResidentPrimary(answers, existingOpenTeamConfig);
     const effectiveExistingOpencode = mergeExistingOpencodeConfigFiles(existingOpencodeFiles, opencodeConfigPaths);
+    const modelOutputCeilings = setupOutputCeilings(answers, effectiveExistingOpencode);
+    const reductions = new Set;
+    for (const file of existingOpencodeFiles) {
+      for (const [providerID, ceilings] of Object.entries(modelOutputCeilings)) {
+        const models = recordAt2(recordAt2(recordAt2(file.raw, "provider"), providerID), "models");
+        for (const [modelID, ceiling] of Object.entries(ceilings)) {
+          const output = positiveInteger3(recordAt2(recordAt2(models, modelID), "limit")?.output);
+          if (output !== undefined && output > ceiling)
+            reductions.add(executionSetupMessages.llamaSwap.outputReduction(`${providerID}/${modelID}`, output, ceiling));
+        }
+      }
+    }
+    if (reductions.size > 0)
+      prompt.note([...reductions].join(`
+`), executionSetupMessages.llamaSwap.outputReductionTitle);
     const {
       config: finalOpenTeamConfig,
-      providerIDs: retainedLemonadeProviderIDs
-    } = preserveRetainedLemonadeRuntimes(existingOpenTeamConfig !== undefined ? mergeOpenTeamConfigForRerun(openTeamConfig, existingOpenTeamConfig) : openTeamConfig, existingOpenTeamConfig, effectiveExistingOpencode);
+      providerIDs: retainedResidentProviderIDs
+    } = preserveRetainedResidentRuntimes(existingOpenTeamConfig !== undefined ? mergeOpenTeamConfigForRerun(openTeamConfig, existingOpenTeamConfig) : openTeamConfig, existingOpenTeamConfig, effectiveExistingOpencode);
     assertEffectiveConfigModelDomains(finalOpenTeamConfig);
     const effectiveOpenTeamConfig = await effectiveSetupConfig(finalOpenTeamConfig, existingOpencodeFiles, deps);
     assertEffectiveConfigModelDomains(effectiveOpenTeamConfig);
-    for (const providerID of retainedLemonadeProviderIDs) {
+    for (const providerID of retainedResidentProviderIDs) {
       const matches = effectiveOpenTeamConfig.local.runtimes.filter((runtime) => runtime.defaultModel.providerID === providerID);
-      if (matches.length !== 1 || matches[0] === undefined || localRuntimeKind(matches[0]) !== "lemonade")
-        throw new Error(executionSetupMessages.lemonadeRetainedProviderUnprotected(providerID));
+      if (matches.length !== 1 || matches[0] === undefined || !runtimeRequiresModelResidency(localRuntimeKind(matches[0])))
+        throw new Error(finalOpenTeamConfig.local.runtimes.some((runtime) => runtime.defaultModel.providerID === providerID && localRuntimeKind(runtime) === "lemonade") ? executionSetupMessages.lemonadeRetainedProviderUnprotected(providerID) : executionSetupMessages.retainedResidencyProviderUnprotected(providerID));
     }
-    const v2Denials = lemonadeV2Policies([
+    const v2Denials = residentV2Policies([
       ...opencodeConfig.experimental?.policies.map((policy) => policy.resource) ?? [],
-      ...effectiveOpenTeamConfig.local.runtimes.filter((runtime) => localRuntimeKind(runtime) === "lemonade").map((runtime) => runtime.defaultModel.providerID)
+      ...effectiveOpenTeamConfig.local.runtimes.filter((runtime) => runtimeRequiresModelResidency(localRuntimeKind(runtime))).map((runtime) => runtime.defaultModel.providerID)
     ]);
     if (v2Denials.length > 0)
       opencodeConfig.experimental = { policies: v2Denials };
-    const v2Notice = v2Denials.length > 0 ? executionSetupMessages.lemonadeV2Restriction(v2Denials.map((policy) => policy.resource)) : undefined;
+    const v2Notice = v2Denials.length > 0 ? executionSetupMessages.residencyV2Restriction(v2Denials.map((policy) => policy.resource)) : undefined;
     if (v2Notice !== undefined)
-      prompt.note(v2Notice, executionSetupMessages.lemonadeV2Title);
+      prompt.note(v2Notice, executionSetupMessages.residencyV2Title);
     const relevantProviderIDs = relevantLocalProviderIDs(opencodeConfig, existingOpenTeamConfig);
     for (const file of existingOpencodeFiles) {
       assertRelevantLimitNumbers(file, relevantProviderIDs);
@@ -20166,12 +20804,14 @@ async function runSetup(deps) {
     const includePluginInDataFile = pluginOwnerPath === opencodeDataPath;
     const dataTargetRaw = opencodeFileAt(existingOpencodeFiles, opencodeDataPath)?.raw ?? {};
     const mergedOpencodeRaw = mergeOpencodeConfigForRerun(opencodeConfig, dataTargetRaw, {
+      modelOutputCeilings,
       ...effectiveExistingOpencode !== undefined ? { effectiveRaw: effectiveExistingOpencode } : {},
       includePlugin: includePluginInDataFile,
       ...includePluginInDataFile ? { pluginEntries: pluginPlan.plugins } : {},
       yolo
     });
     const repairedDataTarget = repairManagedModelLimits({
+      modelOutputCeilings,
       raw: mergedOpencodeRaw,
       effectiveRaw: effectiveExistingOpencode,
       generated: opencodeConfig,
@@ -20185,12 +20825,13 @@ async function runSetup(deps) {
         continue;
       }
       const repaired = repairManagedModelLimits({
+        modelOutputCeilings,
         raw: file.raw,
         effectiveRaw: effectiveExistingOpencode,
         generated: opencodeConfig,
         providerIDs: relevantProviderIDs
       });
-      let plannedRaw = reconcileLemonadeV2Policies(repaired.raw, v2Denials);
+      let plannedRaw = reconcileResidentV2Policies(repaired.raw, v2Denials);
       let changed = repaired.changed || plannedRaw !== repaired.raw;
       if (file.path === pluginOwnerPath && !includePluginInDataFile && pluginPlan.changed) {
         plannedRaw = { ...plannedRaw, plugin: [...pluginPlan.plugins] };
@@ -20264,7 +20905,7 @@ async function runSetup(deps) {
     const effectivePlannedOpencode = mergeExistingOpencodeConfigFiles(existingOpencodeFiles, opencodeConfigPaths, opencodeWritePlan);
     if (effectivePlannedOpencode === undefined)
       throw new Error(executionSetupMessages.lemonadeEndpointMismatch);
-    await verifyLemonadeBeforeWrite(deps, answers, effectiveOpenTeamConfig, effectivePlannedOpencode);
+    await verifyResidencyBeforeWrite(deps, answers, effectiveOpenTeamConfig, effectivePlannedOpencode);
     await deps.writeFile(openTeamPath, serializeOpenTeamConfig(finalOpenTeamConfig));
     for (const [path2, contents] of opencodeWritePlan) {
       await deps.writeFile(path2, serializeOpencodeConfig(contents));
@@ -22647,7 +23288,7 @@ function liveModelsFromSnapshots(snapshots, config) {
     map.set(runtime.defaultModel.providerID, {
       reachable: snapshot.reachable,
       modelIDs: new Set(snapshot.models.filter((model) => hasRequiredModelResidency(kind, model)).map((model) => model.modelID)),
-      ...kind === "lemonade" ? {
+      ...runtimeRequiresModelResidency(kind) ? {
         residency: new Map(snapshot.models.map((model) => [model.modelID, model.loaded]))
       } : {}
     });
@@ -22977,9 +23618,9 @@ async function runCli(argv, deps) {
       const inventoryNotes = [];
       const readySnapshots = snapshots.map((snapshot) => {
         const runtime = config.local.runtimes.find((runtime2) => runtime2.id === snapshot.id);
-        if (runtime === undefined || localRuntimeKind(runtime) !== "lemonade")
+        if (runtime === undefined || !runtimeRequiresModelResidency(localRuntimeKind(runtime)))
           return snapshot;
-        inventoryNotes.push(executionSetupMessages.lemonadeInventory(runtime.id, runtime.defaultModel.providerID, snapshot.models));
+        inventoryNotes.push(executionSetupMessages.lemonadeInventory(runtime.id, runtime.defaultModel.providerID, snapshot.models, localRuntimeKind(runtime) === "llama-swap" ? "llama-swap" : "Lemonade"));
         return {
           ...snapshot,
           models: snapshot.models.filter((model) => hasRequiredModelResidency(localRuntimeKind(runtime), model))
@@ -22990,7 +23631,7 @@ async function runCli(argv, deps) {
         if (agent.model === undefined || live !== "not-loaded" && live !== "residency-unknown")
           return [];
         return [
-          executionSetupMessages.lemonadeAgentResidency(agent.name, `${agent.model.providerID}/${agent.model.modelID}`, live)
+          executionSetupMessages.lemonadeAgentResidency(agent.name, `${agent.model.providerID}/${agent.model.modelID}`, live, config.local.runtimes.some((runtime) => runtime.defaultModel.providerID === agent.model?.providerID && localRuntimeKind(runtime) === "llama-swap") ? "llama-swap" : "Lemonade")
         ];
       });
       return {
@@ -24812,6 +25453,159 @@ var realCacheAdapter = {
   isReparsePath
 };
 
+// src/local/lemonadeRequestGuard.ts
+var inferencePaths = new Set([
+  "/chat/completions",
+  "/completions",
+  "/embeddings",
+  "/responses"
+]);
+function modelID(payload) {
+  if (typeof payload !== "object" || payload === null || !("model" in payload) || typeof payload.model !== "string" || payload.model.length === 0 || payload.model !== payload.model.trim()) {
+    throw new ModelResidencyError("request");
+  }
+  if (payload.model.endsWith(":latest"))
+    throw new ModelResidencyError("identity");
+  return payload.model;
+}
+function createLemonadeGuardedFetch(options) {
+  const base = lemonadeBaseURL(options.baseURL);
+  const upstream = options.fetch;
+  const healthFetch = options.healthFetch ?? upstream;
+  return Object.assign(async (input, init) => {
+    const request = new Request(input, init);
+    const url = new URL(request.url);
+    if (url.username !== "" || url.password !== "" || url.search !== "" || url.hash !== "" || !request.url.startsWith(`${base}/`)) {
+      throw new ModelResidencyError("destination");
+    }
+    const path4 = request.url.slice(base.length);
+    const metadataOnly = request.method === "GET" && (path4 === "/health" || /^\/models(?:\/[^/]+)?$/.test(path4));
+    if (!metadataOnly && (request.method !== "POST" || !inferencePaths.has(path4))) {
+      throw new ModelResidencyError("request");
+    }
+    if (!metadataOnly) {
+      await withResidencyDeadline(async (signal) => {
+        signal.throwIfAborted();
+        let payload;
+        try {
+          payload = await request.clone().json();
+        } catch {
+          signal.throwIfAborted();
+          throw new ModelResidencyError("request");
+        }
+        signal.throwIfAborted();
+        const model = modelID(payload);
+        const modelIDs = typeof options.modelIDs === "function" ? await options.modelIDs() : options.modelIDs;
+        signal.throwIfAborted();
+        if (modelIDs !== undefined && !modelIDs.has(model)) {
+          throw new ModelResidencyError("identity");
+        }
+        const residency = await readLemonadeResidency(base, healthFetch, {
+          ...options,
+          signal,
+          headers: request.headers
+        });
+        signal.throwIfAborted();
+        const entry = residency.get(model);
+        if (entry?.loaded !== true) {
+          throw new ModelResidencyError(entry?.loaded === undefined && entry !== undefined ? "unknown" : "inactive");
+        }
+      }, { ...options, signal: request.signal });
+    }
+    if (request.signal.aborted)
+      throw new ModelResidencyError("cancelled");
+    const {
+      body: _body,
+      headers: _headers,
+      method: _method,
+      signal: _signal,
+      redirect: _redirect,
+      ...transportOptions
+    } = init ?? {};
+    return upstream(request, { ...transportOptions, redirect: "error" });
+  }, {
+    preconnect: (...args) => upstream.preconnect?.(...args)
+  });
+}
+
+// src/local/llamaSwapRequestGuard.ts
+var inferencePaths2 = new Set([
+  "/chat/completions",
+  "/completions",
+  "/embeddings",
+  "/responses"
+]);
+async function validatedModelID(payload, modelIDs) {
+  if (typeof payload !== "object" || payload === null || !("model" in payload) || typeof payload.model !== "string" || payload.model.length === 0 || payload.model !== payload.model.trim()) {
+    throw new ModelResidencyError("request", "llama-swap");
+  }
+  if (modelIDs === undefined) {
+    return payload.model;
+  }
+  const ids = typeof modelIDs === "function" ? await modelIDs() : modelIDs;
+  if (ids.has(payload.model)) {
+    return payload.model;
+  }
+  throw new ModelResidencyError("identity", "llama-swap");
+}
+function createLlamaSwapGuardedFetch(options) {
+  const base = llamaSwapBaseURL(options.baseURL);
+  const upstream = options.fetch;
+  const healthFetch = options.healthFetch ?? upstream;
+  return Object.assign(async (input, init) => {
+    const request = new Request(input, init);
+    const url = new URL(request.url);
+    if (url.username !== "" || url.password !== "" || url.search !== "" || url.hash !== "" || !request.url.startsWith(`${base}/`)) {
+      throw new ModelResidencyError("destination", "llama-swap");
+    }
+    const path4 = request.url.slice(base.length);
+    const metadataOnly = request.method === "GET" && /^\/models(?:\/[^/]+)?$/.test(path4);
+    if (!metadataOnly && (request.method !== "POST" || !inferencePaths2.has(path4))) {
+      throw new ModelResidencyError("request", "llama-swap");
+    }
+    if (!metadataOnly) {
+      await withResidencyDeadline(async (signal) => {
+        signal.throwIfAborted();
+        let payload;
+        try {
+          payload = await request.clone().json();
+        } catch {
+          signal.throwIfAborted();
+          throw new ModelResidencyError("request", "llama-swap");
+        }
+        signal.throwIfAborted();
+        const model = await validatedModelID(payload, options.modelIDs);
+        const residency = await readLlamaSwapResidency(base, healthFetch, {
+          ...options,
+          signal,
+          headers: request.headers
+        });
+        signal.throwIfAborted();
+        if (!residency.has(model)) {
+          throw new ModelResidencyError("identity", "llama-swap");
+        }
+        if (residency.get(model)?.loaded !== true) {
+          throw new ModelResidencyError("inactive", "llama-swap");
+        }
+      }, { ...options, signal: request.signal, runtimeLabel: "llama-swap" });
+    }
+    if (request.signal.aborted) {
+      throw new ModelResidencyError("cancelled", "llama-swap");
+    }
+    const {
+      body: _body,
+      headers: _headers,
+      method: _method,
+      signal: _signal,
+      redirect: _redirect,
+      ...transportOptions
+    } = init ?? {};
+    return upstream(request, { ...transportOptions, redirect: "error" });
+  }, {
+    preconnect: (...args) => upstream.preconnect?.(...args)
+  });
+}
+
 // src/memory/types.ts
 import { z as z28 } from "zod";
 var SHARED_OWNER_KEY = "*";
@@ -25476,6 +26270,13 @@ function createLocalEmbedder(request, deps) {
   return (texts) => embedTexts({ ...request, texts }, deps);
 }
 
+// src/messages/memoryRuntime.ts
+var memoryRuntimeMessages = {
+  providerUnavailable: "[openteam] Semantic memory requires exactly one enabled local runtime matching the configured embeddings provider. No embedding request was sent.",
+  unavailable: "[openteam] Semantic memory initialization failed; recall and memory injection are unavailable.",
+  rejectedLines: (count) => `[openteam] memory log rejected ${count} line(s); semantic memory may be incomplete.`
+};
+
 // src/memory/embedText.ts
 function normalize3(parts) {
   return parts.map((part) => part.trim()).filter((part) => part.length > 0).join(" ");
@@ -25553,25 +26354,39 @@ async function buildMemoryRuntimeFromConfig(config, deps, options = {}) {
   if (options.requireInjectionEnabled === true && !semantic.injection.enabled) {
     return;
   }
-  const runtime = config.local.runtimes.find((rt) => rt.enabled && rt.baseURL !== undefined);
+  const warn = deps.warn ?? console.warn;
+  const configuredModel = semantic.embeddings.model;
+  const candidates = configuredModel == null ? config.local.runtimes.filter((rt) => rt.enabled && rt.baseURL !== undefined) : config.local.runtimes.filter((rt) => rt.defaultModel.providerID === configuredModel.providerID);
+  const runtime = candidates[0];
+  if (configuredModel != null && (candidates.length !== 1 || runtime?.enabled !== true)) {
+    warn(memoryRuntimeMessages.providerUnavailable);
+    return;
+  }
   if (runtime?.baseURL === undefined) {
     return;
   }
-  const modelID = semantic.embeddings.model?.modelID ?? runtime.defaultModel.modelID;
+  const modelID2 = semantic.embeddings.model?.modelID ?? runtime.defaultModel.modelID;
   const openDatabase = deps.openDatabase ?? createBunSqliteDatabase;
   const makeEmbedder = deps.createEmbedder ?? createLocalEmbedder;
   const fetchImpl = deps.fetch ?? globalThis.fetch;
   const now = deps.now ?? (() => Date.now());
-  const warn = deps.warn ?? console.warn;
+  const kind = localRuntimeKind(runtime);
   try {
+    const baseURL = kind === "llama-swap" ? llamaSwapBaseURL(runtime.baseURL) : runtime.baseURL;
+    const guardedFetch = runtimeRequiresModelResidency(kind) ? (kind === "llama-swap" ? createLlamaSwapGuardedFetch : createLemonadeGuardedFetch)({
+      baseURL,
+      fetch: fetchImpl,
+      healthFetch: fetchImpl,
+      modelIDs: new Set([modelID2])
+    }) : fetchImpl;
     const db = await openDatabase(semantic.indexPath);
     const index = createMemoryIndex(db, warn);
     const rebuildResult = await rebuildMemoryIndexFromStorage(index, semantic.logPath, { storage: deps.storage });
     if (rebuildResult.rejectedLines > 0) {
-      warn(`[openteam] memory log rejected ${rebuildResult.rejectedLines} line(s); semantic memory may be incomplete.`);
+      warn(memoryRuntimeMessages.rejectedLines(rebuildResult.rejectedLines));
     }
-    const embedder = makeEmbedder({ baseURL: runtime.baseURL, modelID }, { fetch: fetchImpl });
-    await ensureMemoryEmbeddings({ index, embedder, model: modelID, warn });
+    const embedder = makeEmbedder({ baseURL, modelID: modelID2 }, { fetch: guardedFetch });
+    await ensureMemoryEmbeddings({ index, embedder, model: modelID2, warn });
     return {
       index,
       embedder,
@@ -25580,6 +26395,7 @@ async function buildMemoryRuntimeFromConfig(config, deps, options = {}) {
       semantic
     };
   } catch {
+    warn(memoryRuntimeMessages.unavailable);
     return;
   }
 }