@mindstudio-ai/remy 0.1.312 → 0.1.314

package/dist/headless.js

@@ -434,6 +434,13 @@ var MODEL_SURFACES = {
     modelType: "text",
     userPickable: true
   },
+  research: {
+    default: "claude-5-sonnet",
+    label: "Research Agent",
+    description: "Researches using the web and reports back with citations.",
+    modelType: "text",
+    userPickable: true
+  },
   copyEditor: {
     default: "claude-5-sonnet",
     label: "Copy Agent",
@@ -486,37 +493,59 @@ var MODEL_SURFACES = {
     userPickable: false
   }
 };
+var DEFAULT_SUGGEST_COMPACT_AT = 3e5;
+var TEXT_MODELS = {
+  // Anthropic 1M-context, flat-priced.
+  "claude-5-opus": { forceCompactAt: 85e4 },
+  "claude-4-8-opus": { forceCompactAt: 85e4 },
+  "claude-4-7-opus": { forceCompactAt: 85e4 },
+  // Anthropic 1M-context with 2x long-context pricing above 200K input.
+  "claude-4-6-opus": { forceCompactAt: 85e4, suggestCompactAt: 18e4 },
+  "claude-4-6-sonnet": { forceCompactAt: 85e4, suggestCompactAt: 18e4 },
+  "claude-fable-5": { forceCompactAt: 85e4 },
+  "claude-fable-5-1": { forceCompactAt: 85e4 },
+  "claude-5-sonnet": { forceCompactAt: 85e4 },
+  // OpenAI gpt-5.5/5.6: ~1M window, but the usable input ceiling under
+  // `truncation: 'auto'` is ~794K (output + reasoning reserve), and all
+  // rates double above 272K input.
+  "gpt-5.5": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+  "gpt-5.6-sol": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+  "gpt-5.6-terra": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+  "gpt-5.6-luna": { forceCompactAt: 6e5, suggestCompactAt: 25e4 },
+  // Google ~1M-context; only 3.1-pro is tiered (higher rates above 200K).
+  "gemini-3-pro": { forceCompactAt: 85e4 },
+  "gemini-3.1-pro": { forceCompactAt: 85e4, suggestCompactAt: 18e4 },
+  "gemini-3-flash": { forceCompactAt: 85e4 },
+  "gemini-3.5-flash": { forceCompactAt: 85e4 },
+  "gemini-3.7-flash": { forceCompactAt: 85e4 },
+  // 256K window; its 200K pricing tier sits above the gate, so no nudge.
+  "grok-build-0.1": { forceCompactAt: 18e4 },
+  "grok-4.5": { forceCompactAt: 4e5 },
+  // 500K window
+  "grok-4.6": { forceCompactAt: 4e5 },
+  // 500K window
+  "glm-5.2": { forceCompactAt: 85e4 },
+  "muse-spark-1.1": { forceCompactAt: 85e4 },
+  "kimi-k2-7-code": { forceCompactAt: 2e5 },
+  // 262K window
+  "kimi-k3": { forceCompactAt: 85e4 },
+  "deepseek-v4-flash-0731": { forceCompactAt: 85e4 },
+  "qwen3.8-2.4t-a95b-deepinfra": { forceCompactAt: 2e5 },
+  // 262K window
+  "qwen3.8-27b-deepinfra": { forceCompactAt: 2e5 },
+  // 262K window
+  "minimax-m3": { forceCompactAt: 42e4 }
+  // 524K window
+};
+var DEFAULT_CONTEXT_LIMITS = { forceCompactAt: 85e4 };
+function getContextLimits(modelId) {
+  return TEXT_MODELS[modelId] ?? DEFAULT_CONTEXT_LIMITS;
+}
+function getSuggestCompactAt(modelId) {
+  return getContextLimits(modelId).suggestCompactAt ?? DEFAULT_SUGGEST_COMPACT_AT;
+}
 var ALLOWED_MODELS_BY_TYPE = {
-  text: [
-    "claude-5-opus",
-    "claude-4-8-opus",
-    "claude-4-7-opus",
-    "claude-4-6-opus",
-    "claude-4-6-sonnet",
-    "claude-fable-5",
-    "claude-fable-5-1",
-    "claude-5-sonnet",
-    "gpt-5.5",
-    "gpt-5.6-sol",
-    "gpt-5.6-terra",
-    "gpt-5.6-luna",
-    "gemini-3-pro",
-    "gemini-3.1-pro",
-    "gemini-3-flash",
-    "gemini-3.5-flash",
-    "gemini-3.7-flash",
-    "grok-build-0.1",
-    "grok-4.5",
-    "grok-4.6",
-    "glm-5.2",
-    "muse-spark-1.1",
-    "kimi-k2-7-code",
-    "kimi-k3",
-    "deepseek-v4-flash-0731",
-    "qwen3.8-2.4t-a95b-deepinfra",
-    "qwen3.8-27b-deepinfra",
-    "minimax-m3"
-  ]
+  text: Object.keys(TEXT_MODELS)
   // vision: undefined — unconstrained
   // image_generation: undefined — unconstrained
 };
@@ -559,6 +588,11 @@ function getEffectiveModelSurfaces() {
 function resolveModel(surfaceId, models, fallback) {
   return models?.[surfaceId] ?? fallback ?? orgDefaultModels[surfaceId] ?? MODEL_SURFACES[surfaceId].default;
 }
+function resolveParentModel(models, fallback, buildModel) {
+  const override = buildModel ? filterModelPicks({ parent: buildModel }).parent : void 0;
+  const baseline = resolveModel("parent", models, fallback);
+  return { baseline, effective: override ?? baseline };
+}
 
 // src/orgContext.ts
 var log3 = createLogger("orgContext");
@@ -1857,158 +1891,6 @@ var askMindStudioSdkTool = {
   }
 };
 
-// src/usageLedger.ts
-import fs12 from "fs";
-var LEDGER_FILE = ".logs/usage.ndjson";
-var fd = null;
-function nanoToDollars(nano) {
-  return typeof nano === "number" ? nano / 1e9 : void 0;
-}
-function recordUsage(entry) {
-  try {
-    if (fd === null) {
-      fs12.mkdirSync(".logs", { recursive: true });
-      fd = fs12.openSync(LEDGER_FILE, "a");
-    }
-    fs12.writeSync(fd, JSON.stringify(entry) + "\n");
-  } catch {
-  }
-}
-
-// src/subagents/common/runMindstudioCli.ts
-function stripFlags(args) {
-  const out = [];
-  for (let i = 0; i < args.length; i++) {
-    const arg = args[i];
-    if (arg === "--no-meta") {
-      continue;
-    }
-    if (arg === "--output-key") {
-      i++;
-      continue;
-    }
-    out.push(arg);
-  }
-  return out;
-}
-async function runMindstudioCliResult(args, options) {
-  const cleanArgs = stripFlags(args);
-  const cliAction = args[0];
-  const agentName = options?.caller ?? "mindstudio-cli";
-  const start = Date.now();
-  const res = await runCli("mindstudio", cleanArgs, options);
-  if (!res.ok) {
-    return { ok: false, value: formatCliResult(res) };
-  }
-  const truncNote = res.truncated ? "\n\n[output truncated]" : "";
-  let envelope;
-  try {
-    envelope = JSON.parse(res.output);
-  } catch {
-    return { ok: true, value: res.output + truncNote };
-  }
-  if (envelope && typeof envelope === "object" && Array.isArray(envelope.results)) {
-    const durationMs = Date.now() - start;
-    for (const step of envelope.results) {
-      if (typeof step?.billingCost === "number") {
-        recordUsage({
-          ts: Date.now(),
-          agentName,
-          cliAction: `${cliAction}:${step.stepType ?? "step"}`,
-          cost: nanoToDollars(step.billingCost),
-          inputTokens: 0,
-          outputTokens: 0,
-          durationMs,
-          toolNames: []
-        });
-      }
-    }
-    return { ok: true, value: JSON.stringify(envelope.results) + truncNote };
-  }
-  if (typeof envelope?.$billingCost === "number") {
-    recordUsage({
-      ts: Date.now(),
-      agentName,
-      cliAction,
-      cost: nanoToDollars(envelope.$billingCost),
-      billingEvents: envelope.$billingEvents,
-      // CLI billing isn't expressed as input/output tokens for most actions
-      // (image gen is per-image, scrape per-page, etc). `numUnits` inside each
-      // billingEvent carries the per-event unit count.
-      inputTokens: 0,
-      outputTokens: 0,
-      durationMs: Date.now() - start,
-      toolNames: []
-    });
-  }
-  if (options?.outputKey) {
-    const v = envelope?.[options.outputKey];
-    if (v === void 0 || v === null) {
-      return { ok: false, value: JSON.stringify(stripDollarKeys(envelope)) };
-    }
-    const value = typeof v === "string" ? v : JSON.stringify(v);
-    return { ok: true, value: value + truncNote };
-  }
-  return {
-    ok: true,
-    value: JSON.stringify(stripDollarKeys(envelope)) + truncNote
-  };
-}
-async function runMindstudioCli(args, options) {
-  return (await runMindstudioCliResult(args, options)).value;
-}
-function stripDollarKeys(envelope) {
-  if (!envelope || typeof envelope !== "object" || Array.isArray(envelope)) {
-    return envelope;
-  }
-  const out = {};
-  for (const [k, v] of Object.entries(envelope)) {
-    if (!k.startsWith("$")) {
-      out[k] = v;
-    }
-  }
-  return out;
-}
-
-// src/tools/common/searchGoogle.ts
-var FETCH_TOP_N = 5;
-var searchGoogleTool = {
-  definition: {
-    name: "searchGoogle",
-    description: "Search Google and return results. Use for research, finding documentation, looking up APIs, or any task where web search would help. The top results come back with their page content already included, so read those before reaching for scrapeWebUrl \u2014 you only need that for URLs this did not return, or for a result further down the list.",
-    inputSchema: {
-      type: "object",
-      properties: {
-        query: {
-          type: "string",
-          description: "The search query."
-        }
-      },
-      required: ["query"]
-    }
-  },
-  async execute(input, context) {
-    const query = input.query;
-    return runMindstudioCli(
-      [
-        "search-google",
-        "--query",
-        query,
-        "--export-type",
-        "json",
-        "--fetch-top-n",
-        String(FETCH_TOP_N)
-      ],
-      {
-        outputKey: "results",
-        maxBuffer: SEARCH_MAX_BUFFER,
-        onLog: context?.onLog,
-        caller: "parent"
-      }
-    );
-  }
-};
-
 // src/tools/common/setProjectMetadata.ts
 var setProjectMetadataTool = {
   definition: {
@@ -2133,7 +2015,7 @@ This reference lives at ${skill.path}. Re-read it with readFile if you need it a
 };
 
 // src/tools/code/readFile.ts
-import fs13 from "fs/promises";
+import fs12 from "fs/promises";
 var DEFAULT_WINDOW = 500;
 var MAX_BYTES = 64 * 1024;
 function isBinary(buffer) {
@@ -2174,7 +2056,7 @@ var readFileTool = {
   },
   async execute(input) {
     try {
-      const buffer = await fs13.readFile(input.path);
+      const buffer = await fs12.readFile(input.path);
       if (isBinary(buffer)) {
         const size = buffer.length;
         const unit = size > 1024 * 1024 ? `${(size / (1024 * 1024)).toFixed(1)}MB` : `${(size / 1024).toFixed(1)}KB`;
@@ -2246,7 +2128,7 @@ var readFileTool = {
 };
 
 // src/tools/code/writeFile.ts
-import fs14 from "fs/promises";
+import fs13 from "fs/promises";
 import path6 from "path";
 var writeFileTool = {
   definition: {
@@ -2283,7 +2165,7 @@ var writeFileTool = {
         lastNewlineCount = newlineCount;
         const lastNewline = partial.content.lastIndexOf("\n");
         const completeContent = partial.content.substring(0, lastNewline + 1);
-        const oldContent = await fs14.readFile(partial.path, "utf-8").catch(() => "");
+        const oldContent = await fs13.readFile(partial.path, "utf-8").catch(() => "");
         return `Writing ${partial.path} (${newlineCount} lines)
 ${unifiedDiff(partial.path, oldContent, completeContent)}`;
       }
@@ -2292,13 +2174,13 @@ ${unifiedDiff(partial.path, oldContent, completeContent)}`;
   async execute(input) {
     const release = await acquireFileLock(input.path);
     try {
-      await fs14.mkdir(path6.dirname(input.path), { recursive: true });
+      await fs13.mkdir(path6.dirname(input.path), { recursive: true });
       let oldContent = null;
       try {
-        oldContent = await fs14.readFile(input.path, "utf-8");
+        oldContent = await fs13.readFile(input.path, "utf-8");
       } catch {
       }
-      await fs14.writeFile(input.path, input.content, "utf-8");
+      await fs13.writeFile(input.path, input.content, "utf-8");
       const lineCount = input.content.split("\n").length;
       const label = oldContent !== null ? "Wrote" : "Created";
       return `${label} ${input.path} (${lineCount} lines)
@@ -2312,7 +2194,7 @@ ${unifiedDiff(input.path, oldContent ?? "", input.content)}`;
 };
 
 // src/tools/code/editFile/index.ts
-import fs15 from "fs/promises";
+import fs14 from "fs/promises";
 var editFileTool = {
   definition: {
     name: "editFile",
@@ -2343,7 +2225,7 @@ var editFileTool = {
   async execute(input) {
     const release = await acquireFileLock(input.path);
     try {
-      const content = await fs15.readFile(input.path, "utf-8");
+      const content = await fs14.readFile(input.path, "utf-8");
       const { old_string, new_string, replace_all } = input;
       const occurrences = findOccurrences(content, old_string);
       if (replace_all) {
@@ -2359,7 +2241,7 @@ var editFileTool = {
             new_string
           );
         }
-        await fs15.writeFile(input.path, updated, "utf-8");
+        await fs14.writeFile(input.path, updated, "utf-8");
         return `Replaced ${occurrences.length} occurrence${occurrences.length > 1 ? "s" : ""} in ${input.path}
 ${unifiedDiff(input.path, content, updated)}`;
       }
@@ -2370,7 +2252,7 @@ ${unifiedDiff(input.path, content, updated)}`;
           old_string.length,
           new_string
         );
-        await fs15.writeFile(input.path, updated, "utf-8");
+        await fs14.writeFile(input.path, updated, "utf-8");
         return `Updated ${input.path}
 ${unifiedDiff(input.path, content, updated)}`;
       }
@@ -2386,7 +2268,7 @@ ${unifiedDiff(input.path, content, updated)}`;
           flex.matchedText.length,
           new_string
         );
-        await fs15.writeFile(input.path, updated, "utf-8");
+        await fs14.writeFile(input.path, updated, "utf-8");
         return `Updated ${input.path} (matched with flexible whitespace at line ${flex.line})
 ${unifiedDiff(input.path, content, updated)}`;
       }
@@ -2712,12 +2594,12 @@ var globTool = {
 };
 
 // src/tools/code/listDir.ts
-import fs16 from "fs/promises";
+import fs15 from "fs/promises";
 import path8 from "path";
 var EXCLUDE = /* @__PURE__ */ new Set([".git", "node_modules"]);
 var MAX_CHILDREN = 15;
 async function readAndSort(dirPath) {
-  const entries = await fs16.readdir(dirPath, { withFileTypes: true });
+  const entries = await fs15.readdir(dirPath, { withFileTypes: true });
   return entries.filter((e) => !EXCLUDE.has(e.name)).sort((a, b) => {
     if (a.isDirectory() && !b.isDirectory()) {
       return -1;
@@ -2758,7 +2640,7 @@ function formatSize(bytes) {
 }
 async function formatFile(dirPath, name, indent) {
   try {
-    const stat4 = await fs16.stat(path8.join(dirPath, name));
+    const stat4 = await fs15.stat(path8.join(dirPath, name));
     return `${indent}${name}${" ".repeat(Math.max(1, 30 - indent.length - name.length))}${formatSize(stat4.size)}`;
   } catch {
     return `${indent}${name}`;
@@ -3090,6 +2972,119 @@ var queryDatabaseTool = {
   }
 };
 
+// src/usageLedger.ts
+import fs16 from "fs";
+var LEDGER_FILE = ".logs/usage.ndjson";
+var fd = null;
+function nanoToDollars(nano) {
+  return typeof nano === "number" ? nano / 1e9 : void 0;
+}
+function recordUsage(entry) {
+  try {
+    if (fd === null) {
+      fs16.mkdirSync(".logs", { recursive: true });
+      fd = fs16.openSync(LEDGER_FILE, "a");
+    }
+    fs16.writeSync(fd, JSON.stringify(entry) + "\n");
+  } catch {
+  }
+}
+
+// src/subagents/common/runMindstudioCli.ts
+function stripFlags(args) {
+  const out = [];
+  for (let i = 0; i < args.length; i++) {
+    const arg = args[i];
+    if (arg === "--no-meta") {
+      continue;
+    }
+    if (arg === "--output-key") {
+      i++;
+      continue;
+    }
+    out.push(arg);
+  }
+  return out;
+}
+async function runMindstudioCliResult(args, options) {
+  const cleanArgs = stripFlags(args);
+  const cliAction = args[0];
+  const agentName = options?.caller ?? "mindstudio-cli";
+  const start = Date.now();
+  const res = await runCli("mindstudio", cleanArgs, options);
+  if (!res.ok) {
+    return { ok: false, value: formatCliResult(res) };
+  }
+  const truncNote = res.truncated ? "\n\n[output truncated]" : "";
+  let envelope;
+  try {
+    envelope = JSON.parse(res.output);
+  } catch {
+    return { ok: true, value: res.output + truncNote };
+  }
+  if (envelope && typeof envelope === "object" && Array.isArray(envelope.results)) {
+    const durationMs = Date.now() - start;
+    for (const step of envelope.results) {
+      if (typeof step?.billingCost === "number") {
+        recordUsage({
+          ts: Date.now(),
+          agentName,
+          cliAction: `${cliAction}:${step.stepType ?? "step"}`,
+          cost: nanoToDollars(step.billingCost),
+          inputTokens: 0,
+          outputTokens: 0,
+          durationMs,
+          toolNames: []
+        });
+      }
+    }
+    return { ok: true, value: JSON.stringify(envelope.results) + truncNote };
+  }
+  if (typeof envelope?.$billingCost === "number") {
+    recordUsage({
+      ts: Date.now(),
+      agentName,
+      cliAction,
+      cost: nanoToDollars(envelope.$billingCost),
+      billingEvents: envelope.$billingEvents,
+      // CLI billing isn't expressed as input/output tokens for most actions
+      // (image gen is per-image, scrape per-page, etc). `numUnits` inside each
+      // billingEvent carries the per-event unit count.
+      inputTokens: 0,
+      outputTokens: 0,
+      durationMs: Date.now() - start,
+      toolNames: []
+    });
+  }
+  if (options?.outputKey) {
+    const v = envelope?.[options.outputKey];
+    if (v === void 0 || v === null) {
+      return { ok: false, value: JSON.stringify(stripDollarKeys(envelope)) };
+    }
+    const value = typeof v === "string" ? v : JSON.stringify(v);
+    return { ok: true, value: value + truncNote };
+  }
+  return {
+    ok: true,
+    value: JSON.stringify(stripDollarKeys(envelope)) + truncNote
+  };
+}
+async function runMindstudioCli(args, options) {
+  return (await runMindstudioCliResult(args, options)).value;
+}
+function stripDollarKeys(envelope) {
+  if (!envelope || typeof envelope !== "object" || Array.isArray(envelope)) {
+    return envelope;
+  }
+  const out = {};
+  for (const [k, v] of Object.entries(envelope)) {
+    if (!k.startsWith("$")) {
+      out[k] = v;
+    }
+  }
+  return out;
+}
+
 // src/tools/_helpers/uploadImage.ts
 import { readFile, stat } from "fs/promises";
 import { basename, extname, resolve } from "path";
@@ -4568,7 +4563,7 @@ The first-party SDK (@mindstudio-ai/agent) provides access to 200+ AI models (Op
 ## What Remy apps are NOT good for
 
 - Native mobile apps (iOS/Android). Mobile-responsive web apps are fine.
-- Fast-twitch multiplayer and live co-editing (shared cursors, 60fps sync) \u2014 client\u2192server is always a method invoke, so sub-100ms bidirectional interaction isn't a fit.
+- Real-time action games (server-authoritative simulation, guaranteed-order sync) \u2014 event delivery is at-most-once and ordering is the app's job, so there is no server tick to build one on.
 </platform_brief>`;
 }
 
@@ -4857,56 +4852,155 @@ var screenshotTool = {
   execute: (input, context) => executeScreenshot(input, context?.onLog, context)
 };
 
-// src/subagents/designExpert/tools/searchGoogle.ts
-var searchGoogle_exports = {};
-__export(searchGoogle_exports, {
-  definition: () => definition,
-  execute: () => execute
-});
-var FETCH_TOP_N2 = 5;
-var definition = {
+// src/subagents/research/tools.ts
+var searchGoogleDefinition = {
   name: "searchGoogle",
-  description: 'Search Google for web results. Reserch modern design trends in industries or verticals, "best [domain] apps 2026", ui patterns, or find something specific if the the user has an explicit reference. Searching for and reading case studies is a great way to get information and context about a project\'s domain. Prioritize authoritative sources like Figma and other design leaders, avoid random blog spam. The top results come back with their page content already included, so read those directly \u2014 only use `scrapeWebUrl` for a result further down the list, or for a URL this did not return.',
+  description: "Search Google. Returns ~25 results (title, description, URL) \u2014 fast (~5s) when `fetchTopN` is omitted. Set `fetchTopN` (3\u20135) to also get the top results' page content inline, capped at ~4,000 chars per page \u2014 one call delivers excerpts, but it is much slower (~30\u201360s). Default pattern: fire SERP-only searches in parallel, then scrape the specific pages worth reading in full.",
   inputSchema: {
     type: "object",
     properties: {
       query: {
         type: "string",
         description: "The search query."
+      },
+      fetchTopN: {
+        type: "number",
+        description: "Also fetch page content (capped ~4,000 chars each) for this many top results. Omit for a fast SERP-only search."
       }
     },
     required: ["query"]
   }
 };
-async function execute(input, onLog) {
+var scrapeWebUrlDefinition = {
+  name: "scrapeWebUrl",
+  description: "Fetch a web page as markdown. Renders JavaScript, so client-rendered pages come back complete. Returns the full page \u2014 long docs can run tens of thousands of tokens, so fetch only pages you have chosen to read, and fire independent fetches in parallel. For code hosted on GitHub/npm, prefer cloning with bash over scraping the repo page.",
+  inputSchema: {
+    type: "object",
+    properties: {
+      url: {
+        type: "string",
+        description: "The URL to fetch."
+      }
+    },
+    required: ["url"]
+  }
+};
+var RESEARCH_TOOLS = [
+  ...COMMON_READ_TOOLS,
+  searchGoogleDefinition,
+  scrapeWebUrlDefinition,
+  bashTool.definition
+];
+async function executeSearchGoogle(input, onLog, caller) {
+  const fetchTopN = Math.max(0, Math.round(Number(input.fetchTopN) || 0));
   return runMindstudioCli(
     [
       "search-google",
       "--query",
-      input.query,
+      String(input.query),
       "--export-type",
       "json",
       "--fetch-top-n",
-      String(FETCH_TOP_N2)
+      String(fetchTopN)
     ],
     {
       outputKey: "results",
+      maxBuffer: SEARCH_MAX_BUFFER,
       onLog,
-      caller: "designExpert",
-      maxBuffer: SEARCH_MAX_BUFFER
+      caller
     }
   );
 }
+async function executeScrapeWebUrl(input, onLog, caller) {
+  return runMindstudioCli(
+    [
+      "scrape-url",
+      "--url",
+      String(input.url),
+      "--page-options",
+      JSON.stringify({ onlyMainContent: true })
+    ],
+    {
+      maxBuffer: SCRAPE_MAX_BUFFER,
+      onLog,
+      caller
+    }
+  );
+}
+
+// src/subagents/research/index.ts
+var BASE_PROMPT2 = readAsset("subagents/research", "prompt.md");
+async function runResearch(task, context) {
+  const specIndex = loadSpecIndex();
+  const parts = [BASE_PROMPT2, loadPlatformBrief()];
+  parts.push("<!-- cache_breakpoint -->");
+  if (specIndex) {
+    parts.push(specIndex);
+  }
+  const system = parts.join("\n\n");
+  const result = await runSubAgent({
+    system,
+    task,
+    tools: RESEARCH_TOOLS,
+    externalTools: /* @__PURE__ */ new Set(),
+    executeTool: (name, toolInput, toolCallId, onLog, sams) => {
+      const childCtx = toolCallId ? {
+        ...deriveContext(context, toolCallId, onLog),
+        subAgentMessages: sams
+      } : context;
+      if (name === "searchGoogle") {
+        return executeSearchGoogle(toolInput, childCtx.onLog, "research");
+      }
+      if (name === "scrapeWebUrl") {
+        return executeScrapeWebUrl(toolInput, childCtx.onLog, "research");
+      }
+      return executeTool(name, toolInput, childCtx);
+    },
+    apiConfig: context.apiConfig,
+    model: resolveModel("research", context.models, context.model),
+    subAgentId: "research",
+    signal: context.signal,
+    parentToolId: context.toolCallId,
+    requestId: context.requestId,
+    onEvent: context.onEvent,
+    resolveExternalTool: context.resolveExternalTool,
+    toolRegistry: context.toolRegistry
+  });
+  context.subAgentMessages?.set(context.toolCallId, result.messages);
+  return result.text;
+}
+var researchTool = {
+  definition: {
+    name: "research",
+    description: "Your researcher. Hand it a question and it searches the web, reads the pages and source code that matter, and returns a distilled, citation-backed report. Use it for both objective research like third-party APIs and services, as well as for subjective things like current trends, patterns, and approaches (from architecture to UI to even just high-level framings and ideas). Brief it neutrally: state the question and the relevant project context, not the answer you expect.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        task: {
+          type: "string",
+          description: "What you need to find out, in natural language, plus any project context it cannot get from reading the spec. Phrase it as a question to answer, not a conclusion to confirm."
+        }
+      },
+      required: ["task"]
+    }
+  },
+  async execute(input, context) {
+    if (!context) {
+      return "Error: research requires execution context";
+    }
+    return runResearch(input.task, context);
+  }
+};
 
 // src/subagents/designExpert/tools/scrapeWebUrl.ts
 var scrapeWebUrl_exports = {};
 __export(scrapeWebUrl_exports, {
-  definition: () => definition2,
-  execute: () => execute2
+  definition: () => definition,
+  execute: () => execute
 });
-var definition2 = {
+var definition = {
   name: "scrapeWebUrl",
-  description: "Fetch the content of a web page as markdown. Use when reading sites from search results or specific things the user wants to incorporate.",
+  description: "Fetch the content of a web page as markdown. Use for reading a specific URL \u2014 a site the user referenced, a brand to match, a page the researcher cited that you want in full.",
   inputSchema: {
     type: "object",
     properties: {
@@ -4918,7 +5012,7 @@ var definition2 = {
     required: ["url"]
   }
 };
-async function execute2(input, onLog) {
+async function execute(input, onLog) {
   const pageOptions = { onlyMainContent: true };
   return runMindstudioCli(
     [
@@ -4935,8 +5029,8 @@ async function execute2(input, onLog) {
 // src/subagents/designExpert/tools/analyzeDesign.ts
 var analyzeDesign_exports = {};
 __export(analyzeDesign_exports, {
-  definition: () => definition4,
-  execute: () => execute4
+  definition: () => definition3,
+  execute: () => execute3
 });
 import { readFile as readFile2 } from "fs/promises";
 import { extname as extname2, join as join2 } from "path";
@@ -4945,8 +5039,8 @@ import { extname as extname2, join as join2 } from "path";
 var renderImage_exports = {};
 __export(renderImage_exports, {
   RENDER_ANALYZE_PROMPT: () => RENDER_ANALYZE_PROMPT,
-  definition: () => definition3,
-  execute: () => execute3
+  definition: () => definition2,
+  execute: () => execute2
 });
 import { mkdir, unlink, writeFile } from "fs/promises";
 import { tmpdir } from "os";
@@ -4955,7 +5049,7 @@ import { randomUUID } from "crypto";
 var MIN_DIMENSION = 16;
 var MAX_DIMENSION = 4096;
 var RENDER_ANALYZE_PROMPT = 'You are reviewing a browser-rendered graphic (composed from HTML/CSS by a designer) for fidelity. Report: whether the composition fills the full canvas or leaves unintended gaps at any edge, any clipped or overflowing text, whether custom webfonts appear to have loaded (distinctive letterforms vs generic fallback serif/sans), any misalignment or uneven spacing, any unintended scrollbars or default-styling artifacts, and \u2014 if the background is transparent \u2014 any fringing or stray opaque pixels at the edges. Then briefly describe the overall composition and how polished it looks. Be concise and practical. Respond only with your analysis as Markdown (starting with the title "Render Review") and absolutely no other text. Do not use emojis - use unicode if you need symbols.';
-var definition3 = {
+var definition2 = {
   name: "renderImage",
   description: "Render a self-contained HTML document in a real browser and capture it as a hosted PNG at exact pixel dimensions, with a fidelity review included. Deterministic \u2014 exact hex colors, real loaded webfonts, precise geometry \u2014 unlike generateImages, which is an image model. Use for token-exact graphics: Open Graph share cards, wordmarks, flat/geometric icon tiles, badges, any composition where letterforms and spacing carry the design. Compose with HTML/CSS (link webfonts from CDNs \u2014 the renderer waits for them to load); inline existing SVG markup when needed, but never hand-write new SVG path data.",
   inputSchema: {
@@ -4989,7 +5083,7 @@ var definition3 = {
     required: ["html", "width", "height"]
   }
 };
-async function execute3(input, onLog, context) {
+async function execute2(input, onLog, context) {
   const html = typeof input.html === "string" ? input.html : "";
   const width = Math.round(Number(input.width));
   const height = Math.round(Number(input.height));
@@ -5098,7 +5192,7 @@ Identify the specific design moves that make this page interesting and unique, d
 
 Respond only with your analysis as Markdown and absolutely no other text. Do not use emojis - use unicode if you need symbols.
 `;
-var definition4 = {
+var definition3 = {
   name: "analyzeDesign",
   description: "Analyze the visual design of a website, an image (URL or file on disk), or an HTML document on disk. Websites are screenshotted first; an HTML file is rendered in a real browser first and the render is analyzed, which is how you look at a wireframe you authored. Provides static image analysis only, will not capture animations or video. With no prompt, a website or image gets a full design reference analysis (mood, color, typography, layout, distinctiveness) and a rendered HTML document gets a fidelity review (clipping, overflow, webfont loading, alignment). Provide a custom prompt to ask a specific design question instead. Use a bulleted list to ask many questions at once.",
   inputSchema: {
@@ -5138,7 +5232,7 @@ function splitFrontmatter(raw) {
     html: raw.slice(match[0].length)
   };
 }
-async function execute4(input, onLog, context) {
+async function execute3(input, onLog, context) {
   const url = String(input.url ?? "").trim();
   if (!url) {
     return "Error: url is required.";
@@ -5246,10 +5340,10 @@ ${basePrompt}` : basePrompt;
 // src/subagents/designExpert/tools/analyzeImage.ts
 var analyzeImage_exports = {};
 __export(analyzeImage_exports, {
-  definition: () => definition5,
-  execute: () => execute5
+  definition: () => definition4,
+  execute: () => execute4
 });
-var definition5 = {
+var definition4 = {
   name: "analyzeImage",
   description: "Analyze an image using a vision model. Provides static image analysis only, will not capture animations or video. Returns an objective description of what is visible \u2014 shapes, colors, layout, text, artifacts. Use for factual inventory of image contents, not for subjective design judgment - the vision model providing the analysis has no sense of design. You are the design expert - use the analysis tool for factual inventory, then apply your own expertise for quality and suitability assessments. Optionally provide specific questions about what you're looking for. Use a bulleted list to ask many questions at once. If you are analyzing a screenshot of the app preview, you can reuse the same screenshot URL multiple times to ask multiple questions.",
   inputSchema: {
@@ -5270,7 +5364,7 @@ var definition5 = {
     required: ["imageUrl"]
   }
 };
-async function execute5(input, onLog, context) {
+async function execute4(input, onLog, context) {
   const prompt = buildScreenshotAnalysisPrompt({
     prompt: input.prompt
   });
@@ -5287,8 +5381,8 @@ async function execute5(input, onLog, context) {
 // src/subagents/designExpert/tools/images/generateImages.ts
 var generateImages_exports = {};
 __export(generateImages_exports, {
-  definition: () => definition6,
-  execute: () => execute6
+  definition: () => definition5,
+  execute: () => execute5
 });
 
 // src/subagents/designExpert/tools/images/enhancePrompt.ts
@@ -5515,7 +5609,7 @@ async function generateImageAssets(opts) {
 }
 
 // src/subagents/designExpert/tools/images/generateImages.ts
-var definition6 = {
+var definition5 = {
   name: "generateImages",
   description: "Generate images. Returns CDN URLs with a quality analysis for each image. Produces high-quality results for everything from photorealistic images and abstract/creative visuals. Pass multiple prompts to generate in parallel. No need to analyze images separately after generating \u2014 the analysis is included.",
   inputSchema: {
@@ -5548,7 +5642,7 @@ var definition6 = {
     required: ["prompts"]
   }
 };
-async function execute6(input, onLog, context) {
+async function execute5(input, onLog, context) {
   return generateImageAssets({
     prompts: input.prompts,
     width: input.width,
@@ -5579,10 +5673,10 @@ async function execute6(input, onLog, context) {
 // src/subagents/designExpert/tools/images/editImages.ts
 var editImages_exports = {};
 __export(editImages_exports, {
-  definition: () => definition7,
-  execute: () => execute7
+  definition: () => definition6,
+  execute: () => execute6
 });
-var definition7 = {
+var definition6 = {
   name: "editImages",
   description: "Edit or transform existing images. Provide one or more source image URLs as reference and a prompt describing the desired edit. Use for compositing, style transfer, subject transformation, blending multiple references, or incorporating one or more references into something new. Returns CDN URLs with analysis.",
   inputSchema: {
@@ -5618,7 +5712,7 @@ var definition7 = {
     required: ["prompts", "sourceImages"]
   }
 };
-async function execute7(input, onLog, context) {
+async function execute6(input, onLog, context) {
   return generateImageAssets({
     prompts: input.prompts,
     sourceImages: input.sourceImages,
@@ -5649,15 +5743,15 @@ async function execute7(input, onLog, context) {
 // src/subagents/designExpert/tools/polishCopy.ts
 var polishCopy_exports = {};
 __export(polishCopy_exports, {
-  definition: () => definition8,
-  execute: () => execute8
+  definition: () => definition7,
+  execute: () => execute7
 });
 
 // src/subagents/copyEditor/tools.ts
 var COPY_EDITOR_TOOLS = [...COMMON_READ_TOOLS];
 
 // src/subagents/copyEditor/index.ts
-var BASE_PROMPT2 = readAsset("subagents/copyEditor", "prompt.md");
+var BASE_PROMPT3 = readAsset("subagents/copyEditor", "prompt.md");
 var copyEditorTool = {
   definition: {
     name: "copyEditor",
@@ -5678,7 +5772,7 @@ var copyEditorTool = {
       return "Error: copy editor requires execution context";
     }
     const specIndex = loadSpecIndex();
-    const parts = [BASE_PROMPT2];
+    const parts = [BASE_PROMPT3];
     parts.push("<!-- cache_breakpoint -->");
     if (specIndex) {
       parts.push(specIndex);
@@ -5712,7 +5806,7 @@ var copyEditorTool = {
 };
 
 // src/subagents/designExpert/tools/polishCopy.ts
-var definition8 = {
+var definition7 = {
   name: "polishCopy",
   description: "Hand off any user-facing copy you've written \u2014 headlines, captions, labels, body text \u2014 and get back a sharper version: better built for its audience and free of the fingerprints that make writing read as AI. It elevates how the copy communicates without inventing facts or claims you didn't give it. Give it the text plus what it's for (where it appears, the audience).",
   inputSchema: {
@@ -5726,15 +5820,15 @@ var definition8 = {
     required: ["task"]
   }
 };
-async function execute8(input, _onLog, context) {
+async function execute7(input, _onLog, context) {
   return copyEditorTool.execute(input, context);
 }
 
 // src/subagents/designExpert/tools/loadSkill.ts
 var loadSkill_exports = {};
 __export(loadSkill_exports, {
-  definition: () => definition9,
-  execute: () => execute9
+  definition: () => definition8,
+  execute: () => execute8
 });
 
 // src/subagents/designExpert/skills/_catalog.ts
@@ -5748,7 +5842,7 @@ var designSkillCatalog = buildSkillCatalog({
 });
 
 // src/subagents/designExpert/tools/loadSkill.ts
-var definition9 = {
+var definition8 = {
   name: "loadSkill",
   description: "Load the full craft reference for a design surface that isn't in your prompt. The available skills and the trigger for each are listed in <available_skills>. Load one before designing in its area, not after \u2014 these are hard-won technique recipes, and the defaulted version of these surfaces is exactly what they exist to prevent. Calling this is cheap and expected \u2014 if you're unsure whether you need it, load it.",
   inputSchema: {
@@ -5766,7 +5860,7 @@ var definition9 = {
     required: ["skill"]
   }
 };
-async function execute9(input) {
+async function execute8(input) {
   const id = String(input.skill ?? "");
   const skill = designSkillCatalog.get(id);
   if (!skill) {
@@ -5787,15 +5881,15 @@ This reference lives at ${skill.path}. Re-read it with readFile if you need it a
 var createWireframe_exports = {};
 __export(createWireframe_exports, {
   WIREFRAMES_DIR: () => WIREFRAMES_DIR,
-  definition: () => definition10,
-  execute: () => execute10
+  definition: () => definition9,
+  execute: () => execute9
 });
 import { mkdir as mkdir2, stat as stat2, writeFile as writeFile2 } from "fs/promises";
 import { join as join3 } from "path";
 var log9 = createLogger("createWireframe");
 var WIREFRAMES_DIR = "src/.wireframes";
 var UPLOAD_TIMEOUT_MS2 = 3e4;
-var definition10 = {
+var definition9 = {
   name: "createWireframe",
   description: "Generate a wireframe from self-contained HTML+CSS you author and write it to disk as a design artifact. This is how a wireframe comes to exist \u2014 the way generateImages is how an image comes to exist \u2014 and the developer builds from the file it creates. The result also hands back the reference line that embeds the wireframe in your response and in specs; paste it wherever the wireframe belongs and it renders as a live preview. Calling again with the same slug revises the wireframe in place, so existing references stay current \u2014 a revision is a call to this tool, never a prose description of changes to an earlier wireframe.",
   inputSchema: {
@@ -5889,7 +5983,7 @@ async function uploadMirror(context, slug, content) {
     return { ok: false, note: err.message };
   }
 }
-async function execute10(input, onLog, context) {
+async function execute9(input, onLog, context) {
   if (!context) {
     return "Error: createWireframe requires execution context";
   }
@@ -5937,8 +6031,29 @@ async function execute10(input, onLog, context) {
 }
 
 // src/subagents/designExpert/tools/index.ts
+var research = {
+  definition: {
+    name: "research",
+    description: "Your researcher, for design questions your built-in catalogs and expertise do not settle: how leading products present a specific kind of data or interaction, current conventions or best practices for a pattern, general-purpose ideas and inspiration. It searches the web, reads the sources that matter, and returns a distilled report with citations and concrete specifics. Brief it neutrally: the question or topic, not the answer you expect.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        task: {
+          type: "string",
+          description: "What you need to find out, in natural language, with enough product context to make the findings relevant."
+        }
+      },
+      required: ["task"]
+    }
+  },
+  execute: (input, _onLog, context) => {
+    if (!context) {
+      return Promise.resolve("Error: research requires execution context");
+    }
+    return runResearch(input.task, context);
+  }
+};
 var tools = {
-  searchGoogle: searchGoogle_exports,
   scrapeWebUrl: scrapeWebUrl_exports,
   analyzeDesign: analyzeDesign_exports,
   analyzeImage: analyzeImage_exports,
@@ -5952,7 +6067,8 @@ var tools = {
   polishCopy: polishCopy_exports,
   loadSkill: loadSkill_exports,
   // Appended last: tool order is part of the subagent's prompt-cache prefix.
-  createWireframe: createWireframe_exports
+  createWireframe: createWireframe_exports,
+  research
 };
 var DESIGN_EXPERT_TOOLS = [
   ...COMMON_READ_TOOLS,
@@ -6510,11 +6626,11 @@ ${result.text}`;
 }
 
 // src/subagents/productVision/prompt.ts
-var BASE_PROMPT3 = readAsset("subagents/productVision", "prompt.md");
+var BASE_PROMPT4 = readAsset("subagents/productVision", "prompt.md");
 function getProductVisionPrompt() {
   const specIndex = loadSpecIndex();
   const roadmapIndex = loadRoadmapIndex();
-  const parts = [BASE_PROMPT3, loadPlatformBrief()];
+  const parts = [BASE_PROMPT4, loadPlatformBrief()];
   parts.push("<!-- cache_breakpoint -->");
   if (specIndex) {
     parts.push(specIndex);
@@ -6638,11 +6754,26 @@ var SANITY_CHECK_TOOLS = [
       },
       required: ["command"]
     }
+  },
+  // Appended last: tool order is part of the subagent's prompt-cache prefix.
+  {
+    name: "research",
+    description: "Deep researcher. When the plan hinges on an unfamiliar third-party service, API, or claim you cannot settle with a quick search, hand it the question \u2014 it researches properly (multiple sources, official docs, reading real source code) and returns a distilled, citation-backed report. Quick package-liveness checks stay on searchGoogle; use this when being wrong would be expensive. Brief it neutrally: the question, not the answer you expect.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        task: {
+          type: "string",
+          description: "What you need to find out, in natural language."
+        }
+      },
+      required: ["task"]
+    }
   }
 ];
 
 // src/subagents/codeSanityCheck/index.ts
-var BASE_PROMPT4 = readAsset("subagents/codeSanityCheck", "prompt.md");
+var BASE_PROMPT5 = readAsset("subagents/codeSanityCheck", "prompt.md");
 var codeSanityCheckTool = {
   definition: {
     name: "codeSanityCheck",
@@ -6663,7 +6794,7 @@ var codeSanityCheckTool = {
       return "Error: code sanity check requires execution context";
     }
     const specIndex = loadSpecIndex();
-    const parts = [BASE_PROMPT4, loadPlatformBrief()];
+    const parts = [BASE_PROMPT5, loadPlatformBrief()];
     parts.push("<!-- cache_breakpoint -->");
     if (specIndex) {
       parts.push(specIndex);
@@ -6679,6 +6810,16 @@ var codeSanityCheckTool = {
           ...deriveContext(context, toolCallId, onLog),
           subAgentMessages: sams
         } : context;
+        if (name === "searchGoogle") {
+          return executeSearchGoogle(
+            { ...toolInput, fetchTopN: 5 },
+            childCtx.onLog,
+            "codeSanityCheck"
+          );
+        }
+        if (name === "research") {
+          return runResearch(String(toolInput.task ?? ""), childCtx);
+        }
         return executeTool(name, toolInput, childCtx);
       },
       apiConfig: context.apiConfig,
@@ -6832,7 +6973,7 @@ function acquireSpecSyncLock() {
 }
 
 // src/subagents/specSync/index.ts
-var BASE_PROMPT5 = readAsset("subagents/specSync", "prompt.md");
+var BASE_PROMPT6 = readAsset("subagents/specSync", "prompt.md");
 var MSFM_DOCS = `<mindstudio_flavored_markdown_spec_docs>
 ${readAsset(
   "prompt",
@@ -6871,7 +7012,7 @@ var specSyncTool = {
     }
     const specIndex = loadSpecIndex();
     const parts = [
-      BASE_PROMPT5,
+      BASE_PROMPT6,
       loadPlatformBrief(),
       MSFM_DOCS,
       "<!-- cache_breakpoint -->"
@@ -6986,7 +7127,6 @@ var ALL_TOOLS = [
   confirmDestructiveActionTool,
   askMindStudioSdkTool,
   scrapeWebUrlTool,
-  searchGoogleTool,
   setProjectMetadataTool,
   designExpertTool,
   productVisionTool,
@@ -7025,7 +7165,8 @@ var ALL_TOOLS = [
   // Appended rather than grouped: position is part of the cache prefix, so a
   // new tool goes at the end to leave every existing session's prefix intact.
   loadSkillTool,
-  testJewelTool
+  testJewelTool,
+  researchTool
 ];
 var SUBAGENT_TOOL_NAMES = /* @__PURE__ */ new Set([
   "visualDesignExpert",
@@ -7034,7 +7175,8 @@ var SUBAGENT_TOOL_NAMES = /* @__PURE__ */ new Set([
   "copyEditor",
   "specSync",
   "runAutomatedBrowserTest",
-  "askMindStudioSdk"
+  "askMindStudioSdk",
+  "research"
 ]);
 function getToolDefinitions() {
   return ALL_TOOLS.map((t) => t.definition);
@@ -8616,6 +8758,12 @@ function classify(line, matches) {
     tailClean = SEPARATORS_ONLY.test(gap);
   }
 }
+function chipCase(label) {
+  if (label.startsWith("`")) {
+    return label;
+  }
+  return label.charAt(0).toUpperCase() + label.slice(1);
+}
 function parseSuggestions(raw) {
   if (!raw.includes(SUGGEST_MARKER)) {
     return { text: raw, suggestions: [] };
@@ -8645,7 +8793,7 @@ function parseSuggestions(raw) {
       const key = `${m.label}::${m.message}`;
       if (!seen.has(key)) {
         seen.add(key);
-        suggestions.push({ label: m.label, message: m.message });
+        suggestions.push({ label: chipCase(m.label), message: m.message });
       }
     }
     let rebuilt = "";
@@ -8745,10 +8893,12 @@ async function runTurn(params) {
     onBackgroundComplete
   } = params;
   const tools2 = getToolDefinitions();
-  const buildModelOverride = buildModel ? filterModelPicks({ parent: buildModel }).parent : void 0;
-  const baseline = resolveModel("parent", state.models, model);
-  const parentModel = buildModelOverride ?? baseline;
-  const modelOverride = buildModelOverride && buildModelOverride !== baseline ? { from: baseline } : void 0;
+  const { baseline, effective: parentModel } = resolveParentModel(
+    state.models,
+    model,
+    buildModel
+  );
+  const modelOverride = parentModel !== baseline ? { from: baseline } : void 0;
   const totalAttachments = entries.reduce(
     (n, e) => n + (e.attachments?.length ?? 0),
     0
@@ -8756,7 +8906,7 @@ async function runTurn(params) {
   log15.info("Turn started", {
     requestId,
     model,
-    buildModel: buildModelOverride,
+    buildModel: modelOverride ? parentModel : void 0,
     toolCount: tools2.length,
     ...entries.length > 1 && { entryCount: entries.length },
     ...totalAttachments > 0 && { attachmentCount: totalAttachments }
@@ -9591,12 +9741,13 @@ function loadPassiveResults() {
   }
   return [];
 }
-function writeStats(stats, queue, passiveResults) {
+function writeStats(stats, queue, passiveResults, suggestCompactAt) {
   try {
     writeFileAtomicSync(
       STATS_FILE,
       JSON.stringify({
         ...stats,
+        suggestCompactAt,
         queue,
         passiveResults
       })
@@ -9798,7 +9949,6 @@ var USER_FACING_TOOLS = /* @__PURE__ */ new Set([
   "confirmDestructiveAction",
   "presentPublishPlan"
 ]);
-var FORCED_COMPACTION_THRESHOLD_TOKENS = 85e4;
 var HeadlessSession = class {
   // Configuration
   opts;
@@ -10071,7 +10221,15 @@ var HeadlessSession = class {
   /** Persist sessionStats + queue snapshot + passive pen to .remy-stats.json. */
   persistStats() {
     this.sessionStats.updatedAt = Date.now();
-    writeStats(this.sessionStats, this.queue.snapshot(), this.passivePen);
+    const suggestCompactAt = getSuggestCompactAt(
+      resolveParentModel(this.state.models, this.opts.model).effective
+    );
+    writeStats(
+      this.sessionStats,
+      this.queue.snapshot(),
+      this.passivePen,
+      suggestCompactAt
+    );
   }
   //////////////////////////////////////////////////////////////////////////////
   // Background completions (tool-block mutation; message delivery via queue)
@@ -10101,8 +10259,9 @@ var HeadlessSession = class {
     saveSession(this.state);
   }
   /**
-   * Forced compaction gate. If lastContextSize exceeds the threshold, compact
-   * before letting the upcoming turn run. Coalesces with any in-flight
+   * Forced compaction gate. If lastContextSize exceeds the upcoming turn's
+   * model threshold (per-model — see ModelContextLimits in models/surfaces),
+   * compact before letting the turn run. Coalesces with any in-flight
    * compaction (e.g., one already started by /compact or a tool call). No
    * timeout — compaction takes as long as it takes.
    *
@@ -10113,13 +10272,15 @@ var HeadlessSession = class {
    * On compaction failure we don't bail — the turn proceeds and surfaces any
    * downstream overflow through the existing "prompt is too long" path.
    */
-  async runForcedCompactionIfNeeded(requestId) {
-    if (this.sessionStats.lastContextSize <= FORCED_COMPACTION_THRESHOLD_TOKENS) {
+  async runForcedCompactionIfNeeded(requestId, parentModel) {
+    const threshold = getContextLimits(parentModel).forceCompactAt;
+    if (this.sessionStats.lastContextSize <= threshold) {
       return;
     }
     log17.info("Forced compaction gate triggered", {
       contextSize: this.sessionStats.lastContextSize,
-      threshold: FORCED_COMPACTION_THRESHOLD_TOKENS,
+      threshold,
+      model: parentModel,
       requestId
     });
     try {
@@ -10601,7 +10762,12 @@ var HeadlessSession = class {
       }
       return steered;
     };
-    await this.runForcedCompactionIfNeeded(requestId);
+    const parentModel = resolveParentModel(
+      this.state.models,
+      this.opts.model,
+      params.buildModel
+    ).effective;
+    await this.runForcedCompactionIfNeeded(requestId, parentModel);
     try {
       await runTurn({
         state: this.state,