@turmind/halo 1.3.3 → 1.3.4

package/dist/index.js

@@ -1627,83 +1627,90 @@ var init_agent_loop = __esm({
             return;
           const toolResults = [];
           let shouldEndTurn = false;
-          for (const tu of toolUseBlocks) {
-            if (options?.cancelSignal?.aborted)
-              return;
-            const toolDef = this.toolMap.get(tu.name);
-            if (toolDef?.forceEndTurn)
-              shouldEndTurn = true;
-            const startTime = Date.now();
-            let resultContent;
-            let resultText;
-            let isError = false;
-            if (!toolDef) {
-              resultContent = `${TOOL_ERROR_MARKER}
-Error: unknown tool "${tu.name}"`;
-              resultText = resultContent;
-              isError = true;
-            } else {
-              try {
-                const raw = await toolDef.callback(tu.input, options?.cancelSignal);
-                if (typeof raw === "string") {
-                  resultContent = raw;
-                  resultText = raw;
-                } else {
-                  resultContent = raw;
-                  resultText = raw.map((b) => b.type === "text" ? b.text : `[image ${b.source.media_type}]`).join("\n");
-                }
-              } catch (err) {
+          try {
+            for (const tu of toolUseBlocks) {
+              if (options?.cancelSignal?.aborted)
+                return;
+              const toolDef = this.toolMap.get(tu.name);
+              if (toolDef?.forceEndTurn)
+                shouldEndTurn = true;
+              const startTime = Date.now();
+              let resultContent;
+              let resultText;
+              let isError = false;
+              if (!toolDef) {
                 resultContent = `${TOOL_ERROR_MARKER}
-Error: ${err instanceof Error ? err.message : String(err)}`;
+Error: unknown tool "${tu.name}"`;
                 resultText = resultContent;
                 isError = true;
+              } else {
+                try {
+                  const raw = await toolDef.callback(tu.input, options?.cancelSignal);
+                  if (typeof raw === "string") {
+                    resultContent = raw;
+                    resultText = raw;
+                  } else {
+                    resultContent = raw;
+                    resultText = raw.map((b) => b.type === "text" ? b.text : `[image ${b.source.media_type}]`).join("\n");
+                  }
+                } catch (err) {
+                  resultContent = `${TOOL_ERROR_MARKER}
+Error: ${err instanceof Error ? err.message : String(err)}`;
+                  resultText = resultContent;
+                  isError = true;
+                }
               }
-            }
-            const uiCap = config.limits.toolResultUiMax;
-            const resultTextFull = resultText.length > uiCap ? resultText.slice(0, uiCap) + `
+              if (options?.cancelSignal?.aborted)
+                return;
+              const uiCap = config.limits.toolResultUiMax;
+              const resultTextFull = resultText.length > uiCap ? resultText.slice(0, uiCap) + `
 
 [Content truncated: ${resultText.length} chars total, showing first ${uiCap}. Use file_read for the complete content.]` : resultText;
-            const cap2 = tu.name === "activate_skill" ? Infinity : config.limits.toolResultMax;
-            const truncationNote = (origLen) => `
+              const cap2 = tu.name === "activate_skill" ? Infinity : config.limits.toolResultMax;
+              const truncationNote = (origLen) => `
 
 [Content truncated: ${origLen} chars total, showing first ${cap2}. Re-run with narrower scope (e.g. grep / file_read with offset+limit) to see specific sections.]`;
-            if (typeof resultContent === "string" && resultContent.length > cap2) {
-              const orig = resultContent.length;
-              resultContent = resultContent.slice(0, cap2) + truncationNote(orig);
-              resultText = resultContent;
-            } else if (Array.isArray(resultContent)) {
-              let mutated = false;
-              resultContent = resultContent.map((b) => {
-                if (b.type === "text" && b.text.length > cap2) {
-                  mutated = true;
-                  return { type: "text", text: b.text.slice(0, cap2) + truncationNote(b.text.length) };
+              if (typeof resultContent === "string" && resultContent.length > cap2) {
+                const orig = resultContent.length;
+                resultContent = resultContent.slice(0, cap2) + truncationNote(orig);
+                resultText = resultContent;
+              } else if (Array.isArray(resultContent)) {
+                let mutated = false;
+                resultContent = resultContent.map((b) => {
+                  if (b.type === "text" && b.text.length > cap2) {
+                    mutated = true;
+                    return { type: "text", text: b.text.slice(0, cap2) + truncationNote(b.text.length) };
+                  }
+                  return b;
+                });
+                if (mutated) {
+                  resultText = resultContent.map((b) => b.type === "text" ? b.text : `[image ${b.source.media_type}]`).join("\n");
                 }
-                return b;
-              });
-              if (mutated) {
-                resultText = resultContent.map((b) => b.type === "text" ? b.text : `[image ${b.source.media_type}]`).join("\n");
               }
+              const durationMs = Date.now() - startTime;
+              toolResults.push({
+                type: "tool_result",
+                tool_use_id: tu.id,
+                content: resultContent,
+                // truncated — LLM-facing only
+                ...isError ? { is_error: true } : {}
+              });
+              yield {
+                type: "tool_result",
+                toolName: tu.name,
+                toolUseId: tu.id,
+                toolResult: resultText,
+                // truncated (LLM cap applied)
+                toolResultFull: resultTextFull,
+                // full — for UI display
+                durationMs
+              };
+            }
+          } finally {
+            if (toolResults.length > 0) {
+              this.messages.push({ role: "user", content: toolResults });
             }
-            const durationMs = Date.now() - startTime;
-            toolResults.push({
-              type: "tool_result",
-              tool_use_id: tu.id,
-              content: resultContent,
-              // truncated — LLM-facing only
-              ...isError ? { is_error: true } : {}
-            });
-            yield {
-              type: "tool_result",
-              toolName: tu.name,
-              toolUseId: tu.id,
-              toolResult: resultText,
-              // truncated (LLM cap applied)
-              toolResultFull: resultTextFull,
-              // full — for UI display
-              durationMs
-            };
           }
-          this.messages.push({ role: "user", content: toolResults });
           if (shouldEndTurn) {
             yield { type: "stop", stopReason: "end_turn" };
             return;
@@ -2490,7 +2497,7 @@ var init_init = __esm({
       path8.resolve(__dirname, "..", "bundled-docs"),
       path8.resolve(__dirname, "..", "..", "..", ".halo", "docs")
     ];
-    TEMPLATE_VERSION = 59;
+    TEMPLATE_VERSION = 60;
     VERSION_FILE = ".template-version";
     SKIP_NAMES = /* @__PURE__ */ new Set([".DS_Store", "schema.sql", "__pycache__", ".pytest_cache"]);
     BUILTIN_AGENT_IDS = /* @__PURE__ */ new Set([
@@ -2932,6 +2939,23 @@ ${skillInfo}` : ""
     queryAgentTool
   ];
 }
+function buildContinueTaskTool(sm, sessionId) {
+  return {
+    name: "continue_task",
+    description: "Call this when your turn was started by an interruption (a user or parent message arrived while you were working) and the task you were doing is NOT finished yet. After your current reply ends, the session automatically resumes the interrupted task. Do NOT call it if the task is done, if the user told you to stop, or in a normal (uninterrupted) turn \u2014 it does nothing there. The flag lasts one turn: if you get interrupted again before resuming, call it again.",
+    inputSchema: { type: "object", properties: {} },
+    callback: async () => {
+      switch (sm.requestSelfKick(sessionId)) {
+        case "set":
+          return JSON.stringify({ code: 0, message: "Continue flag set. Finish your current reply \u2014 the interrupted task resumes automatically right after." });
+        case "not_interrupted":
+          return JSON.stringify({ code: 0, message: "This turn was not started by an interruption \u2014 nothing to resume. Just keep working normally; do not call this again this turn." });
+        case "no_turn":
+          return JSON.stringify({ code: 1, error: "Turn already ended (stopped) \u2014 flag not set." });
+      }
+    }
+  };
+}
 var init_session_tools = __esm({
   "packages/server/dist/agents/session-tools.js"() {
     "use strict";
@@ -7841,10 +7865,25 @@ var init_deepseek_agent = __esm({
         }
         return results;
       }
+      /** Block arrays become OpenAI content parts — text as-is, image blocks as
+       *  `image_url` data URLs (same shape as Kimi). Only `deepseek-flash`
+       *  actually sees images — the registry has `image: false` on v4-pro, so the
+       *  session manager strips them before they reach here for that model. */
       convertUserContent(content) {
         if (typeof content === "string")
           return content;
-        return content.filter((b) => b.type === "text").map((b) => b.text).join("\n");
+        const parts = [];
+        for (const block of content) {
+          if (block.type === "text") {
+            parts.push({ type: "text", text: block.text });
+          } else if (block.type === "image") {
+            parts.push({
+              type: "image_url",
+              image_url: { url: `data:${block.source.media_type};base64,${block.source.data}` }
+            });
+          }
+        }
+        return parts;
       }
       convertAssistantMessage(msg) {
         const results = [];
@@ -8921,8 +8960,8 @@ import { SignatureV4 } from "@smithy/signature-v4";
 import { Sha256 } from "@aws-crypto/sha256-js";
 import { defaultProvider } from "@aws-sdk/credential-provider-node";
 function extractRegionFromHost(hostname) {
-  const m = hostname.match(/bedrock-mantle\.([a-z0-9-]+)\.api\.aws/);
-  return m?.[1] ?? "us-east-2";
+  const m = hostname.match(/^bedrock-(?:mantle|runtime)\.([a-z0-9-]+)\./);
+  return m?.[1] ?? "us-east-1";
 }
 function safeParse6(json) {
   try {
@@ -9371,10 +9410,15 @@ function createModelRuntime(providerId, cfg) {
         maxTokens: cfg.maxTokens,
         thinking: cfg.thinking
       });
+    // aws-bedrock-openai is the same OpenAI Responses wire format on the
+    // bedrock-runtime host (/openai/v1, cross-Region profile ids only) —
+    // Grok 4.6 / Kimi K3 live there and not (or not everywhere) on Mantle.
+    // See aws-bedrock-openai.yaml.
     case "aws-bedrock-mantle":
+    case "aws-bedrock-openai":
       return new MantleAgent({
         modelId: cfg.modelId,
-        endpoint: cfg.endpoint ?? "https://bedrock-mantle.us-east-2.api.aws/openai/v1",
+        endpoint: cfg.endpoint ?? (providerId === "aws-bedrock-openai" ? "https://bedrock-runtime.us-east-1.amazonaws.com/openai/v1" : "https://bedrock-mantle.us-east-1.api.aws/openai/v1"),
         // No bearer token → MantleAgent falls back to SigV4 IAM auth using
         // these creds (or the SDK default chain when also empty).
         apiKey: cfg.apiKey ?? "",
@@ -10960,6 +11004,7 @@ var init_session_agent_builder = __esm({
         const nameSet = new Set(yamlConfig?.tools ?? []);
         const delegates = canDelegate(yamlConfig);
         const sessionTools = delegates ? this.host.createSessionTools(sessionId) : agentId === GOAL_AGENT_ID ? this.host.createGoalTools(sessionId) : [];
+        sessionTools.push(this.host.createContinueTaskTool(sessionId));
         let draftReset = null;
         if (nameSet.has("draft")) {
           const { tool, reset } = createDraftTool();
@@ -11442,7 +11487,7 @@ async function initObservability() {
     return true;
   }
   const { registerSdk: registerSdk2 } = await Promise.resolve().then(() => (init_otel_sdk(), otel_sdk_exports));
-  providers = registerSdk2(process.env.OTEL_SERVICE_NAME, "1.3.3");
+  providers = registerSdk2(process.env.OTEL_SERVICE_NAME, "1.3.4");
   console.log(`[Observability] OTLP export \u2192 ${process.env.OTEL_EXPORTER_OTLP_ENDPOINT} (service.name=${process.env.OTEL_SERVICE_NAME}, capture_content=${config.observability.captureContent})`);
   return true;
 }
@@ -11816,7 +11861,7 @@ function generateSessionId() {
   const rand = Math.random().toString(36).slice(2, 8);
   return `sid_${ts}_${rand}`;
 }
-var SessionManager;
+var CONTINUE_TASK_KICK, SessionManager;
 var init_session_manager = __esm({
   "packages/server/dist/agents/session-manager.js"() {
     "use strict";
@@ -11847,6 +11892,7 @@ var init_session_manager = __esm({
     init_session_skill_commands();
     init_session_state_store();
     init_genai_spans();
+    CONTINUE_TASK_KICK = `[System] You called continue_task: the task interrupted earlier is not finished. Resume it now from where it stopped. Any tool call marked "[tool execution interrupted \u2014 no result]" did not run \u2014 re-issue it if you still need the result (this overrides the marker's do-not-retry note \u2014 you explicitly asked to continue).`;
     SessionManager = class _SessionManager {
       // The fields below are marked `internal:` because they're consumed by
       // helpers in this directory (session-tools.ts). Treat them as
@@ -12063,6 +12109,11 @@ var init_session_manager = __esm({
       createSessionTools(sessionId) {
         return buildSessionTools(this, sessionId);
       }
+      /** Built-in continue_task tool (every agent, no yaml opt-in) — see
+       *  session-agent-builder / requestSelfKick / drainQueue. */
+      createContinueTaskTool(sessionId) {
+        return buildContinueTaskTool(this, sessionId);
+      }
       /** Goal-mode tool set — injected only for the `goal` agent (see
        *  session-agent-builder / goal-mode.ts). */
       createGoalTools(sessionId) {
@@ -12104,6 +12155,8 @@ var init_session_manager = __esm({
           warnedToolHashes: /* @__PURE__ */ new Set(),
           turnStartTime: 0,
           interruptRequested: false,
+          selfKick: false,
+          resumedAfterInterrupt: false,
           isCompacting: false,
           compactAbortController: null,
           compactedThisTurn: false,
@@ -12707,6 +12760,8 @@ var init_session_manager = __esm({
             session.toolCallLog = [];
             session.warnedToolHashes.clear();
             session.interruptRequested = false;
+            session.resumedAfterInterrupt = false;
+            session.selfKick = false;
             result = await this.runAgentTurn(session, message);
             console.debug(`[SessionManager] runSession ${sessionId} first turn done \u2014 result: ${result.slice(0, 150)}`);
           }
@@ -12847,6 +12902,9 @@ ${list}`;
       async drainQueue(session) {
         while (session.messageQueue.length > 0) {
           const batch = session.messageQueue.splice(0);
+          session.resumedAfterInterrupt = session.interruptRequested;
+          const kickFlagReset = session.selfKick;
+          session.selfKick = false;
           session.interruptRequested = false;
           session.toolCallLog = [];
           session.warnedToolHashes.clear();
@@ -12861,7 +12919,8 @@ ${q.text}` : q.text).join("\n\n");
           if (suffix && session.parentId === null) {
             this.emitEvent(session.id, { type: "system", text: suffix.trim(), agentName: session.agentName });
           }
-          const input = this.buildInput(merged + suffix, images.length > 0 ? images : void 0, session.supportsImage);
+          const kickNote = kickFlagReset ? "\n\n[System] The continue_task flag you set last turn was reset by this interrupt. If the interrupted task is still unfinished after you answer, call continue_task again." : "";
+          const input = this.buildInput(merged + suffix + kickNote, images.length > 0 ? images : void 0, session.supportsImage);
           try {
             await this.runAgentTurn(session, input);
           } catch (err) {
@@ -12869,6 +12928,15 @@ ${q.text}` : q.text).join("\n\n");
             console.error(`[SessionManager] Drain run error for ${session.id}: ${errMsg}`);
             throw err;
           }
+          if (session.messageQueue.length === 0) {
+            if (session.selfKick && !session.interruptRequested) {
+              session.messageQueue.push({ text: CONTINUE_TASK_KICK });
+              const taskId = session.parentId !== null ? session.id : void 0;
+              this.emitEvent(session.id, { type: "user", text: CONTINUE_TASK_KICK, agentName: "user", report: true, taskId });
+              console.debug(`[SessionManager] continue_task kick for ${session.id}`);
+            }
+            session.selfKick = false;
+          }
           if (session.parentId === null && session.messageQueue.length > 0) {
             this.emitEvent(session.id, { type: "complete", batchBoundary: true });
           }
@@ -12944,6 +13012,18 @@ ${q.text}` : q.text).join("\n\n");
         this.markPendingToolCallsInterrupted(sessionId);
         console.debug(`[SessionManager] Interrupted session ${sessionId}`);
       }
+      /** continue_task tool backend. Arms the one-turn selfKick flag that drainQueue
+       *  turns into a synthetic resume turn — but only in a turn that itself started
+       *  after an interrupt (resumedAfterInterrupt), so kicks ≤ interrupts. */
+      requestSelfKick(sessionId) {
+        const session = this.sessions.get(sessionId);
+        if (!session || session.abortController === null)
+          return "no_turn";
+        if (!session.resumedAfterInterrupt)
+          return "not_interrupted";
+        session.selfKick = true;
+        return "set";
+      }
       /** Append text as a user turn to a session's agent.messages, coalescing into
        *  a trailing user message if one exists. Two invariants drive this:
        *   1. Anthropic rejects consecutive same-role messages, and the next
@@ -12981,6 +13061,7 @@ ${q.text}` : q.text);
             }
             session.messageQueue = [];
             session.interruptRequested = false;
+            session.selfKick = false;
             if (session.abortController) {
               session.abortController.abort(abortReason("stop"));
               session.abortController = null;
@@ -13702,6 +13783,7 @@ ${result.summary}`, agentName: "user", report: true, taskId });
             session.abortController.abort(abortReason("delete"));
             session.abortController = null;
           }
+          session.selfKick = false;
           if (session.promise) {
             try {
               await session.promise;
@@ -13718,6 +13800,7 @@ ${result.summary}`, agentName: "user", report: true, taskId });
               childSession.abortController.abort(abortReason("delete"));
               childSession.abortController = null;
             }
+            childSession.selfKick = false;
             this.sessions.delete(id);
           }
         }
@@ -13762,6 +13845,7 @@ ${result.summary}`, agentName: "user", report: true, taskId });
             session.abortController.abort(abortReason("archive"));
             session.abortController = null;
           }
+          session.selfKick = false;
           if (session.promise) {
             try {
               await session.promise;
@@ -13848,8 +13932,10 @@ ${result.summary}`, agentName: "user", report: true, taskId });
        *  message), and requests a SOFT interrupt so a busy turn yields after its
        *  current tool — same effect as sendUserMessage's busy branch, minus the
        *  idle/run path (the caller already established the session is busy or
-       *  compacting). When compacting, the flag is harmless (no live turn) and
-       *  endCompact drains the queue. */
+       *  compacting). When compacting there is no live turn to interrupt, so the
+       *  flag is NOT set (mirrors sendUserMessage's compacting branch): endCompact's
+       *  drain would snapshot it into resumedAfterInterrupt and let continue_task
+       *  arm a kick in a turn that interrupted nothing. endCompact drains the queue. */
       async enqueueUserMessage(sessionId, text6, images) {
         const session = this.sessions.get(sessionId);
         if (!session)
@@ -13859,7 +13945,8 @@ ${result.summary}`, agentName: "user", report: true, taskId });
           this.emitEvent(sessionId, { type: "system", text: `\u26A0 ${w}` });
         }
         session.messageQueue.push({ text: scoped.text, images });
-        session.interruptRequested = true;
+        if (!session.isCompacting)
+          session.interruptRequested = true;
         console.debug(`[SessionManager] User message enqueued for ${sessionId} (${session.messageQueue.length} in queue)`);
       }
       /** Hard stop — abort immediately (for explicit user stop button).
@@ -13880,6 +13967,7 @@ ${q.text}` : q.text);
         }
         session.messageQueue = [];
         session.interruptRequested = false;
+        session.selfKick = false;
         if (session.abortController) {
           session.abortController.abort();
           session.abortController = null;
@@ -31078,7 +31166,7 @@ var init_dist2 = __esm({
     init_git_ssh();
     __dirname3 = path68.dirname(fileURLToPath3(import.meta.url));
     PORT = config.server.port;
-    HALO_VERSION = "1.3.3";
+    HALO_VERSION = "1.3.4";
     PROJECT_ROOT = findProjectRoot();
     HALO_HOME4 = path68.join(homedir24(), ".halo");
     GIT_SHA = readGitSha();
@@ -32013,7 +32101,7 @@ import path69 from "node:path";
 import fs63 from "node:fs";
 import { homedir as homedir25 } from "node:os";
 import { parseArgs as parseArgs2 } from "node:util";
-var VERSION = "1.3.3";
+var VERSION = "1.3.4";
 var HELP_TOP = `Usage: halo <command> [options]
 
 Commands:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@turmind/halo",
-  "version": "1.3.3",
+  "version": "1.3.4",
   "description": "Halo — multi-agent collaborative workspace. CLI + TUI + HTTP/WS server in one package.",
   "type": "module",
   "license": "MIT",

package/templates/models/aws-bedrock-claude-invoke.yaml

@@ -245,6 +245,35 @@ models:
           - { value: xhigh, label: XHigh }
           - { value: max, label: Max }
 
+  - id: global.anthropic.claude-opus-5-5
+    displayName: Claude Opus 5.5
+    maxOutputTokens: 128000
+    contextWindow: 1000000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 1h
+        ttlPresets:
+          - { value: 5m, label: 5min }
+          - { value: 1h, label: 1hour }
+      thinking:
+        # Opus 5.5 (Bedrock GA 2026-09-22). Live-probed 2026-09-24 on invoke:
+        # invalid-value probe lists low / medium / high / xhigh / max;
+        # max_tokens ceiling 128000; smoke call (xhigh) returned normally.
+        # Model card: 1M context, adaptive thinking always-on (can't be
+        # disabled), image input.
+        mode: adaptive
+        defaultEnabled: true
+        default: high       # flagship — high by default; xhigh/max on demand
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
+          - { value: max, label: Max }
+
   - id: global.anthropic.claude-haiku-4-5-20251001-v1:0
     displayName: Claude Haiku 4.5
     maxOutputTokens: 64000

package/templates/models/aws-bedrock-mantle.yaml

@@ -1,21 +1,35 @@
 id: aws-bedrock-mantle
 displayName: AWS Bedrock Mantle (OpenAI GPT-6 / 5.6)
 displayName_zh: AWS Bedrock Mantle（OpenAI GPT-6 / 5.6）
-description: OpenAI GPT-6 Astra and GPT-5.6 (Sol / Terra / Luna) on Amazon Bedrock via the bedrock-mantle endpoint (OpenAI Responses API only — not Converse/InvokeModel). Astra is served from the us-west-2 endpoint only.
-description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-6 Astra 与 GPT-5.6（Sol / Terra / Luna）（仅 OpenAI Responses API，不支持 Converse/InvokeModel）。Astra 仅 us-west-2 端点提供。
+description: OpenAI GPT-6 (Astra / Sol / Luna) and GPT-5.6 (Sol / Terra / Luna) on Amazon Bedrock via the bedrock-mantle endpoint (OpenAI Responses API only — not Converse/InvokeModel). Region-split — GPT-6 Sol / Luna are on us-east-1 only, Astra on us-west-2 only.
+description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-6（Astra / Sol / Luna）与 GPT-5.6（Sol / Terra / Luna）（仅 OpenAI Responses API，不支持 Converse/InvokeModel）。按区域分布：GPT-6 Sol / Luna 仅 us-east-1，Astra 仅 us-west-2。
 # Mantle is a distinct endpoint from bedrock-runtime; the OpenAI surface lives
-# under /openai/v1. Region split (live-probed 2026-09-13, both directions 404
-# on the wrong endpoint): us-east-2 serves every 5.6 model; us-west-2 serves
-# GPT-6 Astra + Terra / Luna but NOT Sol. No model is on both endpoints
-# except Terra / Luna, so the endpoint is a per-agent pick — the default
-# stays us-east-2 (existing Sol configs), Astra users select the us-west-2
-# preset in the agent form.
-defaultEndpoint: https://bedrock-mantle.us-east-2.api.aws/openai/v1
+# under /openai/v1. Region split (live-probed 2026-09-24, 404 on the wrong
+# endpoint):
+#
+#                us-east-1   us-east-2   us-west-2
+#   gpt-6-astra      -           -           ✓
+#   gpt-6-sol        ✓           -           -
+#   gpt-6-luna       ✓           -           -
+#   gpt-5.6-sol      ✓           ✓           -
+#   gpt-5.6-terra    ✓           ✓           ✓
+#   gpt-5.6-luna     ✓           ✓           ✓
+#
+# us-east-1 serves everything but Astra, so it is the default; existing
+# us-east-2 configs (5.6-era) keep working; Astra users pick us-west-2 in the
+# agent form. The endpoint is a per-agent pick.
+#
+# The same six models are also in aws-bedrock-openai.yaml (bedrock-runtime
+# host, `global.openai.*` ids, no region split) — AWS's recommended endpoint
+# for new agents. Stay here only for Mantle-only features (server-side
+# web_search, background inference), which the web-search skill relies on.
+defaultEndpoint: https://bedrock-mantle.us-east-1.api.aws/openai/v1
 endpointPresets:
+  - https://bedrock-mantle.us-east-1.api.aws/openai/v1
   - https://bedrock-mantle.us-east-2.api.aws/openai/v1
   - https://bedrock-mantle.us-west-2.api.aws/openai/v1
 
-defaultModelId: openai.gpt-5.6-sol
+defaultModelId: openai.gpt-6-sol
 
 # Auth: a Bedrock long-term API key passed as a Bearer token. Generate at
 # console → Bedrock → API keys → long-term. The default placeholder reads the
@@ -87,6 +101,72 @@ models:
           - { value: medium, label: Medium }
           - { value: high, label: High }
 
+  # GPT-6 Sol / Luna (Bedrock GA 2026-09-22) — us-east-1 endpoint ONLY (404 on
+  # us-east-2 / us-west-2). Same 1,050,000 context / 128,000 max output as
+  # Astra per the OpenAI model pages; live-probed 2026-09-24 on us-east-1 with
+  # invalid-value requests: reasoning.effort accepts none/low/medium/high/
+  # xhigh/max, text.verbosity accepts low/medium/high — identical surface, so
+  # MantleAgent needs no change. Sol ($2/$10) is the coding/agentic tier and
+  # the new default; Luna ($0.10/$0.50) is the high-volume tier. The 272K→2x
+  # input-price step noted on Astra applies here too.
+  - id: openai.gpt-6-sol
+    displayName: GPT-6 Sol
+    maxOutputTokens: 128000
+    contextWindow: 1050000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
+          - { value: max, label: Max }
+      verbosity:
+        default: low
+        presets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+
+  - id: openai.gpt-6-luna
+    displayName: GPT-6 Luna
+    maxOutputTokens: 128000
+    contextWindow: 1050000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
+          - { value: max, label: Max }
+      verbosity:
+        default: low
+        presets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+
   # GPT-5.6 flagship.
   - id: openai.gpt-5.6-sol
     displayName: GPT-5.6 Sol