turbollm 1.11.0 → 1.11.1

package/README.md

@@ -293,6 +293,9 @@ the detail:
   from Settings → Memory, and turning the toggle off stops new chats from seeing it immediately.
 - **Thinking-budget control** — a graduated slider, not just on/off: cap reasoning to a specific
   token count, disable it entirely, or leave it unlimited. Works in Chat and Code alike.
+- **Reasoning-effort control** for models whose own chat template supports it (e.g. Qwen3.8) —
+  an Off/Low/Medium/xhigh slider swaps in automatically in place of the thinking-budget one,
+  detected per model.
 
 </details>
 

package/dist/cli.js

@@ -626,6 +626,35 @@ function setModelProfile(cfg2, modelKey, engineId, profile) {
   ;
   (cfg2.modelProfiles[modelKey] ??= {})[engineId] = { profile, updatedAt: (/* @__PURE__ */ new Date()).toISOString() };
 }
+function migrateModelKey(cfg2, oldKey, newKey) {
+  if (oldKey === newKey) return false;
+  let moved = false;
+  if (cfg2.modelProfiles[oldKey] && !cfg2.modelProfiles[newKey]) {
+    cfg2.modelProfiles[newKey] = cfg2.modelProfiles[oldKey];
+    delete cfg2.modelProfiles[oldKey];
+    moved = true;
+  }
+  if (cfg2.modelPresets[oldKey] && !cfg2.modelPresets[newKey]) {
+    cfg2.modelPresets[newKey] = cfg2.modelPresets[oldKey];
+    delete cfg2.modelPresets[oldKey];
+    moved = true;
+  }
+  if (cfg2.lastPresetId[oldKey] && !cfg2.lastPresetId[newKey]) {
+    cfg2.lastPresetId[newKey] = cfg2.lastPresetId[oldKey];
+    delete cfg2.lastPresetId[oldKey];
+    moved = true;
+  }
+  if (cfg2.benchResults[oldKey] && !cfg2.benchResults[newKey]) {
+    cfg2.benchResults[newKey] = { ...cfg2.benchResults[oldKey], modelKey: newKey };
+    delete cfg2.benchResults[oldKey];
+    moved = true;
+  }
+  if (cfg2.lastLoaded.modelKey === oldKey) {
+    cfg2.lastLoaded = { ...cfg2.lastLoaded, modelKey: newKey };
+    moved = true;
+  }
+  return moved;
+}
 function deleteModelProfile(cfg2, modelKey, engineId) {
   const byEngine = cfg2.modelProfiles[modelKey];
   if (!byEngine) return;
@@ -4405,6 +4434,12 @@ function needsExtraPass(content) {
   return stripThinkingBlocks(content).trim() === "";
 }
 
+// src/chat/reasoning-effort.ts
+var VALID = /* @__PURE__ */ new Set(["off", "low", "medium", "xhigh"]);
+function parseReasoningEffort(value) {
+  return typeof value === "string" && VALID.has(value) ? value : void 0;
+}
+
 // src/tools/research-referee.ts
 var STOP_WORDS = /* @__PURE__ */ new Set([
   "a",
@@ -6509,6 +6544,7 @@ var SessionAuthRegistry = class {
   sessionToToken = /* @__PURE__ */ new Map();
   tokenToSession = /* @__PURE__ */ new Map();
   sessionThinkingBudget = /* @__PURE__ */ new Map();
+  sessionReasoningEffort = /* @__PURE__ */ new Map();
   /** Mint a token for a Code session — idempotent: returns the SAME token on every call for a
    *  given session (a reconnect/remount must never invalidate the CLI's already-running auth). */
   mint(codeSessionId) {
@@ -6547,6 +6583,20 @@ var SessionAuthRegistry = class {
     if (!sessionId) return null;
     return this.sessionThinkingBudget.get(sessionId) ?? null;
   }
+  /** Same override mechanism as setThinkingBudget, for reasoning_effort — see that method's
+   *  doc comment. Already validated by the caller (parseReasoningEffort); this registry never
+   *  re-validates, same as setThinkingBudget trusts its own caller's int coercion. */
+  setReasoningEffort(codeSessionId, effort) {
+    if (effort === null) this.sessionReasoningEffort.delete(codeSessionId);
+    else this.sessionReasoningEffort.set(codeSessionId, effort);
+  }
+  /** The reasoning-effort override for whichever session `token` belongs to, or null when the
+   *  token is unrecognized or no override has been set. */
+  getReasoningEffortForToken(token) {
+    const sessionId = this.tokenToSession.get(token);
+    if (!sessionId) return null;
+    return this.sessionReasoningEffort.get(sessionId) ?? null;
+  }
 };
 var sessionAuth = new SessionAuthRegistry();
 
@@ -9063,7 +9113,7 @@ ${content}` : b.docContext : content;
         ...images.map((url) => ({ type: "image_url", image_url: { url } }))
       ] : fullContent;
       engineMessages.push({ role: "user", content: userContent });
-      await runGeneration(d, stream, { convId, conv, engineMessages, assistantMsg, ms, target, ac, thinkingBudget: b.thinkingBudget ?? -1, userText: content, isCodeAuthorized });
+      await runGeneration(d, stream, { convId, conv, engineMessages, assistantMsg, ms, target, ac, thinkingBudget: b.thinkingBudget ?? -1, reasoningEffort: parseReasoningEffort(b.reasoningEffort), userText: content, isCodeAuthorized });
     });
   });
   app2.post("/api/v1/conversations/:id/continue", async (c) => {
@@ -9102,7 +9152,7 @@ ${m.reasoning}
 ${m.content}` : m.content;
         engineMessages.push({ role: m.role, content });
       }
-      await runGeneration(d, stream, { convId, conv, engineMessages, assistantMsg, ms, target, ac, thinkingBudget: b.thinkingBudget ?? -1, isCodeAuthorized });
+      await runGeneration(d, stream, { convId, conv, engineMessages, assistantMsg, ms, target, ac, thinkingBudget: b.thinkingBudget ?? -1, reasoningEffort: parseReasoningEffort(b.reasoningEffort), isCodeAuthorized });
     });
   });
   app2.post("/api/v1/chat/stop", async (c) => {
@@ -9321,7 +9371,7 @@ function reportChatBenchResult(d, ms, stats) {
 }
 async function runGeneration(d, stream, ctx) {
   const { db: db2 } = d;
-  const { convId, conv, assistantMsg, ms, target, ac, thinkingBudget } = ctx;
+  const { convId, conv, assistantMsg, ms, target, ac, thinkingBudget, reasoningEffort } = ctx;
   const convS = conv.sampling ?? {};
   const engineKind = d.registry.active()?.kind ?? "";
   const repeatPenaltyKey = engineKind === "vllm" || engineKind === "sglang" || engineKind === "mlx-vlm" ? "repetition_penalty" : "repeat_penalty";
@@ -9427,6 +9477,12 @@ After reviewing the search results, include a confidence assessment on a line by
         reqBody.thinking_budget_tokens = thinkingBudget;
       }
       if (conv.preserveThinking) templateKwargs.preserve_thinking = true;
+      if (reasoningEffort === "off") {
+        reqBody.thinking_budget_tokens = 0;
+        templateKwargs.enable_thinking = false;
+      } else if (reasoningEffort) {
+        templateKwargs.reasoning_effort = reasoningEffort;
+      }
       if (Object.keys(templateKwargs).length) reqBody.chat_template_kwargs = templateKwargs;
       const toolsSupported = engineKind !== "vllm" && engineKind !== "sglang" && toolDefs.length > 0;
       if (toolsSupported) reqBody.tools = toolDefs;
@@ -9672,6 +9728,12 @@ After reviewing the search results, include a confidence assessment on a line by
         reqBody.thinking_budget_tokens = thinkingBudget;
       }
       if (conv.preserveThinking) templateKwargs.preserve_thinking = true;
+      if (reasoningEffort === "off") {
+        reqBody.thinking_budget_tokens = 0;
+        templateKwargs.enable_thinking = false;
+      } else if (reasoningEffort) {
+        templateKwargs.reasoning_effort = reasoningEffort;
+      }
       if (Object.keys(templateKwargs).length) reqBody.chat_template_kwargs = templateKwargs;
       const res = await fetch(`${target}/v1/chat/completions`, {
         method: "POST",
@@ -10755,7 +10817,7 @@ function pickPrefillProgress(slots) {
   return { processed, total, pct };
 }
 async function runCodeSession(params) {
-  const { d, convId, sessionId, repoRoot, mode, thinkingBudget, task: rawTask, signal, sink } = params;
+  const { d, convId, sessionId, repoRoot, mode, thinkingBudget, reasoningEffort, task: rawTask, signal, sink } = params;
   const ms = d.manager.status();
   if (ms.state !== "running" || !ms.model) throw new Error("model_not_loaded");
   const target = d.manager.target();
@@ -11175,6 +11237,20 @@ async function runCodeSession(params) {
           chat_template_kwargs: { ...payload.chat_template_kwargs ?? {}, enable_thinking: false }
         };
       }
+      if (reasoningEffort === "off") {
+        return {
+          ...payload,
+          thinking_budget_tokens: 0,
+          chat_template_kwargs: { ...payload.chat_template_kwargs ?? {}, enable_thinking: false }
+        };
+      }
+      if (reasoningEffort) {
+        return {
+          ...payload,
+          ...thinkingBudget > 0 ? { thinking_budget_tokens: thinkingBudget } : {},
+          chat_template_kwargs: { ...payload.chat_template_kwargs ?? {}, reasoning_effort: reasoningEffort }
+        };
+      }
       if (thinkingBudget > 0) {
         return { ...payload, thinking_budget_tokens: thinkingBudget };
       }
@@ -11913,7 +11989,7 @@ var CodeRunManager = class {
   enqueue(sessionId, params) {
     const s = this.ensure(sessionId, params.convId, params.repoRoot);
     const willQueue = s.active !== null;
-    s.queue.push({ task: params.task, userMsgId: params.userMsgId, thinkingBudget: params.thinkingBudget ?? -1, kind: params.kind ?? "followUp" });
+    s.queue.push({ task: params.task, userMsgId: params.userMsgId, thinkingBudget: params.thinkingBudget ?? -1, reasoningEffort: params.reasoningEffort, kind: params.kind ?? "followUp" });
     this.emitQueue(sessionId);
     void this.pump(sessionId);
     return { queued: willQueue };
@@ -12043,6 +12119,7 @@ var CodeRunManager = class {
         repoRoot: s.repoRoot,
         mode,
         thinkingBudget: turn.thinkingBudget,
+        reasoningEffort: turn.reasoningEffort,
         task: turn.task,
         signal: ac.signal,
         sink,
@@ -12747,7 +12824,8 @@ function parseGguf(path) {
       else if (key.endsWith(".nextn_predict_layers")) m.nextnLayers = readNumberOrMax(r, t);
       else if (key === "tokenizer.chat_template") {
         m.hasChatTemplate = true;
-        skipValue(r, t);
+        const template = t === T_STRING ? String(readScalar(r, t)) : (skipValue(r, t), "");
+        if (template.includes("reasoning_effort")) m.reasoningEffort = true;
       } else {
         skipValue(r, t);
       }
@@ -12765,6 +12843,7 @@ function parseGguf(path) {
       expertCount: m.expertCount ?? 0,
       nextnLayers: m.nextnLayers ?? 0,
       hasChatTemplate: m.hasChatTemplate ?? false,
+      reasoningEffort: m.reasoningEffort ?? false,
       slidingWindow: m.slidingWindow ?? 0,
       slidingWindowPattern: m.slidingWindowPattern ?? [],
       headDimSwa: keyLengthSwa ?? valueLengthSwa ?? 0,
@@ -12779,15 +12858,15 @@ function parseGguf(path) {
   }
 }
 function quantFromName(filename) {
-  const m = filename.match(/(I?Q\d[A-Z0-9_]*|BF16|F16|F32)/i);
-  if (m) return m[1].toUpperCase();
+  const matches = [...filename.matchAll(/(?:^|[-_. ])(I?Q\d(?:_[A-Za-z0-9]+)*|BF16|F16|F32)(?=$|[-_. ])/gi)];
+  if (matches.length > 0) return matches[matches.length - 1][1].toUpperCase();
   const apex = filename.match(/APEX(?:-[A-Za-z0-9]+){0,3}?-(I-)?(NANO|MINI|COMPACT|BALANCED|QUALITY)(?=\.|$|-)/i);
   if (apex) return `APEX-${apex[1] ? "I-" : ""}${apex[2].toUpperCase()}`;
   return "?";
 }
 
 // src/models/scanner.ts
-var CACHE_VERSION = 5;
+var CACHE_VERSION = 6;
 function attentionLayoutOf(meta) {
   const a = meta ?? {};
   const out = {};
@@ -12854,6 +12933,9 @@ var Scanner = class {
   lastScanAt = "";
   cache = /* @__PURE__ */ new Map();
   cachePath;
+  // [legacyKey, key] pairs discovered during the CURRENT rescan — see the comment in `entryFor`
+  // for why these are batched into one config write instead of applied as they're found.
+  pendingKeyMigrations = [];
   list() {
     return { models: this.entries, scanning: this.scanning, lastScanAt: this.lastScanAt };
   }
@@ -12912,11 +12994,18 @@ var Scanner = class {
         if (existsSync19(d)) walk(d, scan);
         await tick();
       }
+      this.pendingKeyMigrations = [];
       const gguf = await this.build(scan.ggufs);
       const mlx = scan.mlxDirs.map((dir) => mlxEntryFor(dir));
       this.entries = [...gguf, ...mlx].sort((a, b) => a.name.localeCompare(b.name));
       this.lastScanAt = (/* @__PURE__ */ new Date()).toISOString();
       this.saveCache();
+      if (this.pendingKeyMigrations.length > 0) {
+        const pairs = this.pendingKeyMigrations;
+        this.store.update((cfg2) => {
+          for (const [oldKey, newKey] of pairs) migrateModelKey(cfg2, oldKey, newKey);
+        });
+      }
     } finally {
       this.scanning = false;
     }
@@ -12986,12 +13075,16 @@ var Scanner = class {
       }
     }
     const fileName = basename2(path);
-    const quant = meta?.quant || quantFromName(fileName);
+    const quant = quantFromName(fileName) !== "?" ? quantFromName(fileName) : meta?.quant || "?";
     const name = meta?.name || cleanName(fileName);
     const vision = mmprojPath !== null;
     const arch2 = meta?.arch ?? "unknown";
+    const key = `${name.toLowerCase()}|${quant}|${sizeBytes}`;
+    const legacyQuant = meta?.quant || quantFromName(fileName);
+    const legacyKey = `${name.toLowerCase()}|${legacyQuant}|${sizeBytes}`;
+    if (legacyKey !== key) this.pendingKeyMigrations.push([legacyKey, key]);
     return {
-      key: `${name.toLowerCase()}|${quant}|${sizeBytes}`,
+      key,
       name,
       path,
       dir,
@@ -13015,6 +13108,7 @@ var Scanner = class {
       mmprojPath: vision ? mmprojPath : null,
       mmprojSizeBytes: vision ? mmprojSizeBytes : 0,
       hasChatTemplate: meta?.hasChatTemplate ?? false,
+      reasoningEffort: meta?.reasoningEffort ?? false,
       embedding: isEmbeddingModel(arch2, fileName),
       incomplete,
       parseError,
@@ -13107,6 +13201,7 @@ function mlxEntryFor(dir) {
   let sizeBytes = 0;
   let mtimeMs = 0;
   let hasChatTemplate = false;
+  let reasoningEffort = false;
   try {
     for (const n of readdirSync7(dir)) {
       const lower = n.toLowerCase();
@@ -13117,7 +13212,11 @@ function mlxEntryFor(dir) {
       }
     }
     const tc = join23(dir, "tokenizer_config.json");
-    hasChatTemplate = existsSync19(tc) && readFileSync8(tc, "utf8").includes("chat_template") || existsSync19(join23(dir, "chat_template.jinja"));
+    const tcText = existsSync19(tc) ? readFileSync8(tc, "utf8") : "";
+    const jinjaPath = join23(dir, "chat_template.jinja");
+    const jinjaText = existsSync19(jinjaPath) ? readFileSync8(jinjaPath, "utf8") : "";
+    hasChatTemplate = tcText.includes("chat_template") || jinjaText !== "";
+    reasoningEffort = tcText.includes("reasoning_effort") || jinjaText.includes("reasoning_effort");
   } catch {
   }
   let incomplete = false;
@@ -13162,6 +13261,7 @@ function mlxEntryFor(dir) {
     mmprojPath: null,
     mmprojSizeBytes: 0,
     hasChatTemplate,
+    reasoningEffort,
     embedding: isEmbeddingModel(arch2, basename2(dir)),
     incomplete,
     parseError,
@@ -15868,9 +15968,20 @@ function spillMb(baselineSharedMb, loadedSharedMb) {
   if (baselineSharedMb === null || loadedSharedMb === null) return null;
   return Math.max(0, loadedSharedMb - baselineSharedMb);
 }
-function isSpilling(spill) {
+var SPILL_PROXIMITY_MB = 1024;
+var SPILL_FLOOR_CAP_MB = 2048;
+function isSpilling(spill, vramAbsMb, budgetMb, floorMb = null, pooled = false) {
   if (spill === null) return false;
-  return spill > SPILL_TOLERANCE_MB;
+  const floor = pooled ? 0 : Math.min(floorMb ?? 0, SPILL_FLOOR_CAP_MB);
+  if (spill - floor <= SPILL_TOLERANCE_MB) return false;
+  if (!pooled && vramAbsMb !== null && budgetMb > 0 && vramAbsMb < budgetMb - SPILL_PROXIMITY_MB) return false;
+  return true;
+}
+function spillFloor(floorMb, spill, vramAbsMb, budgetMb, pooled = false) {
+  if (pooled || spill === null || vramAbsMb === null || budgetMb <= 0) return floorMb;
+  if (vramAbsMb >= budgetMb - SPILL_PROXIMITY_MB) return floorMb;
+  const learned = floorMb === null ? spill : Math.min(floorMb, spill);
+  return Math.min(learned, SPILL_FLOOR_CAP_MB);
 }
 
 // src/api/path-utils.ts
@@ -17145,6 +17256,8 @@ var BenchRunner = class {
     if (base2.nglFit) return this.benchAt(entry, sys, base2, caps, results, "auto-fit");
     let bestNgl = 0;
     const budgetMb = gpuBudgetMb(sys, base2);
+    let spillFloorMb = null;
+    const pooled = sys.gpus.length > 1 && base2.gpu.splitMode !== "none";
     if (sys.gpus.length > 0) {
       const hi0 = entry.blockCount > 0 ? entry.blockCount : 99;
       let lo = 0, hi = hi0;
@@ -17155,7 +17268,8 @@ var BenchRunner = class {
         const probe2 = await this.probeVram(entry, sys, { ...base2, ngl: mid }, caps);
         this.pushProbe(results, base2, "ngl", mid, probe2);
         await this.settleGpu(sys);
-        const verdict = probeVerdict(probe2, budgetMb, headroomMb);
+        spillFloorMb = spillFloor(spillFloorMb, probe2.spillMb, probe2.vramAbsMb, budgetMb, pooled);
+        const verdict = probeVerdict(probe2, budgetMb, headroomMb, spillFloorMb, pooled);
         if (verdict.decision === "fits") {
           bestNgl = mid;
           lo = mid + 1;
@@ -17169,7 +17283,7 @@ var BenchRunner = class {
     }
     if (bestNgl === null) return null;
     let found = await this.benchAt(entry, sys, { ...base2, ngl: bestNgl }, caps, results, `ngl=${bestNgl}`);
-    if (found && isSpilling(found.cand.spillMb) && bestNgl > 0) {
+    if (found && isSpilling(found.cand.spillMb, found.cand.vramAbsMb, budgetMb, spillFloorMb, pooled) && bestNgl > 0) {
       this.emit(`ngl=${bestNgl} spilled ${Math.round(found.cand.spillMb)} MB once actually generating (the load-only probe could not see this) \u2014 retrying at ngl=${bestNgl - 1}`);
       const safer = await this.benchAt(entry, sys, { ...base2, ngl: bestNgl - 1 }, caps, results, `ngl=${bestNgl - 1} (spill backoff)`);
       if (safer) {
@@ -17204,13 +17318,16 @@ var BenchRunner = class {
     const budgetMb = gpuBudgetMb(sys, base2);
     let lo = 0, hi = maxN;
     let bestN = null;
+    let spillFloorMb = null;
+    const pooled = sys.gpus.length > 1 && base2.gpu.splitMode !== "none";
     while (lo <= hi && !this.cancelled && Date.now() <= this.deadline) {
       const mid = Math.floor((lo + hi) / 2);
       this.state = { ...this.state, step: `KV ${base2.kvTypeK}: probing nCpuMoe=${mid} (range ${lo}\u2013${hi})\u2026`, candidates: results };
       const probe2 = await this.probeVram(entry, sys, { ...base2, nCpuMoe: mid }, caps);
       this.pushProbe(results, base2, "nCpuMoe", mid, probe2);
       await this.settleGpu(sys);
-      const verdict = probeVerdict(probe2, budgetMb, headroomMb);
+      spillFloorMb = spillFloor(spillFloorMb, probe2.spillMb, probe2.vramAbsMb, budgetMb, pooled);
+      const verdict = probeVerdict(probe2, budgetMb, headroomMb, spillFloorMb, pooled);
       if (verdict.decision === "fits") {
         bestN = mid;
         hi = mid - 1;
@@ -17238,7 +17355,7 @@ var BenchRunner = class {
       }
       if (n < maxN) this.emit(`nCpuMoe=${n} failed to bench (${results[results.length - 1]?.outcome ?? "crash"}) \u2014 retrying at nCpuMoe=${n + 1}`);
     }
-    if (found && isSpilling(found.cand.spillMb) && foundN < maxN) {
+    if (found && isSpilling(found.cand.spillMb, found.cand.vramAbsMb, budgetMb, spillFloorMb, pooled) && foundN < maxN) {
       this.emit(`nCpuMoe=${foundN} spilled ${Math.round(found.cand.spillMb)} MB once actually generating (the load-only probe could not see this) \u2014 retrying at nCpuMoe=${foundN + 1}`);
       const safer = await this.benchAt(entry, sys, { ...base2, nCpuMoe: foundN + 1 }, caps, results, `nCpuMoe=${foundN + 1} (spill backoff)`);
       if (safer) {
@@ -17929,9 +18046,9 @@ function pickKvQuants(baseKv, kvTypes) {
   for (const t of [baseKv, ...QUALITY_KV]) if (supported(t) && !out.includes(t)) out.push(t);
   return out;
 }
-function probeVerdict(probe2, budgetMb, headroomMb) {
+function probeVerdict(probe2, budgetMb, headroomMb, spillFloorMb = null, pooled = false) {
   if (probe2.outcome === "oom") return { decision: "offload-more", reason: "oom" };
-  if (isSpilling(probe2.spillMb)) {
+  if (isSpilling(probe2.spillMb, probe2.vramAbsMb, budgetMb, spillFloorMb, pooled)) {
     return { decision: "offload-more", reason: "spill", shortfallMb: probe2.spillMb };
   }
   if (overHeadroom(probe2.vramAbsMb, budgetMb, headroomMb)) {
@@ -24682,6 +24799,15 @@ function registerCodeRoutes(app2, d, codeRuns) {
     sessionAuth.setThinkingBudget(run.id, tokens === -1 ? null : tokens);
     return c.json({ ok: true, tokens });
   });
+  app2.patch("/api/v1/code/sessions/:id/reasoning-effort", async (c) => {
+    const id = c.req.param("id");
+    const run = db2.getAgentRun(id);
+    if (!run) return err8(c, 404, "not_found", "Session not found.");
+    const b = await body8(c);
+    const effort = parseReasoningEffort(b.effort) ?? null;
+    sessionAuth.setReasoningEffort(run.id, effort);
+    return c.json({ ok: true, effort });
+  });
   app2.get("/api/v1/code/sessions/:id/last-usage", (c) => {
     const id = c.req.param("id");
     const run = db2.getAgentRun(id);
@@ -24930,7 +25056,7 @@ function registerCodeRoutes(app2, d, codeRuns) {
       userMsgId = lastUser.id;
       task = contextFilesBlock(lastUser.textAttachments) + lastUser.content;
     }
-    const enqueueParams = { convId: run.convId, repoRoot: agentCwd(run), task, userMsgId, thinkingBudget: b.thinkingBudget };
+    const enqueueParams = { convId: run.convId, repoRoot: agentCwd(run), task, userMsgId, thinkingBudget: b.thinkingBudget, reasoningEffort: parseReasoningEffort(b.reasoningEffort) };
     if (kind === "steer") {
       const { steered, queued: queued2 } = await runs.steer(id, enqueueParams);
       return c.json({ ok: true, queued: queued2, steered, userMessageId: userMsgId }, 202);
@@ -25856,6 +25982,15 @@ function registerGateway(app2, d, opts = {}) {
     const requestOrigin = new URL(c.req.url).origin;
     const guidance = req.tools?.length ? applyAgentGuidance(req, requestOrigin) : null;
     const oaiBody = mapToOpenAI(req);
+    if (anthropicCodeSessionId) {
+      const effortOverride = sessionAuth.getReasoningEffortForToken(anthropicToken);
+      if (effortOverride === "off") {
+        oaiBody.thinking_budget_tokens = 0;
+        oaiBody.chat_template_kwargs = { ...oaiBody.chat_template_kwargs ?? {}, enable_thinking: false };
+      } else if (effortOverride !== null) {
+        oaiBody.chat_template_kwargs = { ...oaiBody.chat_template_kwargs ?? {}, reasoning_effort: effortOverride };
+      }
+    }
     if (guidance?.forceTextOnly) oaiBody.tool_choice = "none";
     const oaiAlias = engineModelAlias(d.registry.active()?.kind ?? "", d.manager.currentOpts()?.modelPath);
     if (oaiAlias) oaiBody.model = oaiAlias;
@@ -26102,6 +26237,13 @@ function registerGateway(app2, d, opts = {}) {
             if (override > 0) parsedBody.thinking_budget_tokens = override;
             else delete parsedBody.thinking_budget_tokens;
           }
+          const effortOverride = sessionAuth.getReasoningEffortForToken(chatToken);
+          if (effortOverride === "off") {
+            parsedBody.thinking_budget_tokens = 0;
+            parsedBody.chat_template_kwargs = { ...parsedBody.chat_template_kwargs ?? {}, enable_thinking: false };
+          } else if (effortOverride !== null) {
+            parsedBody.chat_template_kwargs = { ...parsedBody.chat_template_kwargs ?? {}, reasoning_effort: effortOverride };
+          }
         }
         if (parsedBody?.stream === true && parsedBody.stream_options === void 0) {
           parsedBody.stream_options = { include_usage: true };

package/dist/webdist/assets/AgentEditPage-D5ccnu5s.js

@@ -0,0 +1 @@
+import{j as e}from"./radix-B0YBjihE.js";import{b as W,u as $,r as b}from"./vendor-C1I6euDx.js";import{u as G,d as H,e as Y,f as J,g as U,a as z,B as C,t as j,c as q,b as h,A as _}from"./index-Djd745uq.js";import{f as V}from"./chat-api-B7cm_Kjj.js";import{f as X,a as Z}from"./agent-api-DyPUCnyd.js";import{P as B}from"./personas-Bdrl3tL6.js";import{C as ee}from"./chevron-left-BQ3TyvUR.js";import{R as te}from"./rotate-ccw-Dl2CDY8F.js";import{T as se}from"./trash-2-C5UxcfNn.js";import{C as ne}from"./chevron-right-C1TzfDRc.js";import"./dropdown-menu-Co-9JBK7.js";const ae=()=>({name:"",description:"",systemPrompt:"",skillIds:[],tools:[]});function oe(s){const d=[],x=new Map;for(const n of s){const a=/^mcp__([^_]+(?:_[^_]+)*)__/.exec(n.name);if(!a){d.push(n);continue}const i=a[1],p=/^\[([^\]]+)\]/.exec(n.description),c=p?p[1]:i.replace(/_/g," ");x.has(i)||x.set(i,{key:i,label:c,tools:[]}),x.get(i).tools.push(n)}return{builtin:d,groups:[...x.values()].sort((n,a)=>n.label.localeCompare(a.label))}}function le(s){return s.replace(/^mcp__[^_]+(?:_[^_]+)*__/,"")}function re(s){return s.replace(/^\[[^\]]+\]\s*/,"")}function D({tool:s,checked:d,onToggle:x}){return e.jsxs("label",{className:"flex cursor-pointer items-start gap-2 text-[12px]",children:[e.jsx("input",{type:"checkbox",className:"mt-0.5",checked:d,onChange:x}),e.jsxs("span",{className:"flex flex-col",children:[e.jsx("span",{className:"font-mono text-ink",children:le(s.name)}),s.description&&e.jsx("span",{className:"text-faint",children:re(s.description)})]})]})}function ie({group:s,selected:d,onToggleAll:x,onToggleOne:n}){const[a,i]=b.useState(!1),p=s.tools.map(l=>l.name),c=p.filter(l=>d.has(l)).length,o=c===p.length,g=c>0&&!o,u=b.useRef(null);return b.useEffect(()=>{u.current&&(u.current.indeterminate=g)},[g]),e.jsxs("div",{className:"rounded-md border border-border",children:[e.jsxs("div",{className:"flex items-center gap-2 px-2.5 py-2",children:[e.jsx("input",{ref:u,type:"checkbox",checked:o,onChange:()=>x(p,!o)}),e.jsxs("button",{type:"button",onClick:()=>{j("agents","toggle_agent_tool_group"),i(l=>!l)},className:"flex flex-1 items-center gap-1.5 text-left text-[13px] text-ink",children:[e.jsx(ne,{size:12,className:q("shrink-0 text-faint transition-transform",a&&"rotate-90")}),s.label,e.jsxs("span",{className:"text-[11px] text-faint",children:["(",c,"/",p.length,")"]})]})]}),a&&e.jsx("div",{className:"flex flex-col gap-1.5 border-t border-border px-3 py-2 pl-9",children:s.tools.map(l=>e.jsx(D,{tool:l,checked:d.has(l.name),onToggle:()=>n(l.name)},l.name))})]})}function ce({tools:s,selected:d,onChange:x}){const n=new Set(d),{builtin:a,groups:i}=oe(s),p=o=>x(n.has(o)?d.filter(g=>g!==o):[...d,o]),c=(o,g)=>{const u=d.filter(l=>!o.includes(l));x(g?[...u,...o]:u)};return s.length===0?e.jsx("p",{className:"text-[12px] text-faint",children:"No tools available yet."}):e.jsxs("div",{className:"flex flex-col gap-2",children:[a.length>0&&e.jsx("div",{className:"flex flex-col gap-2 rounded-md border border-border bg-panel-2 p-3",children:a.map(o=>e.jsx(D,{tool:o,checked:n.has(o.name),onToggle:()=>p(o.name)},o.name))}),i.map(o=>e.jsx(ie,{group:o,selected:n,onToggleAll:c,onToggleOne:p},o.key))]})}function je(){const{agentId:s="new"}=W(),d=$(),x=G(),n=s==="new",a=!n&&B.some(t=>t.id===s),i=a?B.find(t=>t.id===s):void 0,p=H(),c=!n&&!a?(p.data??[]).find(t=>t.id===s):void 0,o=Y(),g=J(),u=U(),l=a?g.data?.[s]:void 0,I=z({queryKey:Z.list(),queryFn:X,staleTime:0}),k=I.data??[],w=z({queryKey:["available-tools"],queryFn:V}),N=w.data??[],[m,f]=b.useState(ae()),[S,T]=b.useState(!1),[A,P]=b.useState(!1),[L,R]=b.useState(!1),[E,Q]=b.useState("skills");n&&!S&&(k.length||N.length)&&(f(t=>({...t,skillIds:k.map(r=>r.id),tools:N.map(r=>r.name)})),T(!0)),!n&&!a&&!S&&c&&(f({name:c.name,description:c.description,systemPrompt:c.systemPrompt,skillIds:c.skillIds,tools:c.tools}),T(!0)),a&&!S&&i&&g.isSuccess&&(k.length||N.length)&&(f({name:l?.name??i.name,description:l?.description??i.description,systemPrompt:l?.systemPrompt??i.systemPrompt,skillIds:l?.skillIds??k.map(t=>t.id),tools:l?.tools??N.map(t=>t.name)}),T(!0));const y=()=>{j("agents","back_to_agents"),d("/customize")},O=t=>f(r=>({...r,skillIds:r.skillIds.includes(t)?r.skillIds.filter(v=>v!==t):[...r.skillIds,t]})),F=()=>{if(!m.name.trim()){h.error("Name is required.");return}P(!0);const t={name:m.name.trim(),description:m.description.trim(),systemPrompt:m.systemPrompt,skillIds:m.skillIds,tools:m.tools};if(a){u.save.mutate({id:s,override:t},{onSuccess:()=>{h.success("Agent saved."),y()},onError:v=>{h.error(v instanceof _?v.message:"Could not save agent."),P(!1)}});return}const r={onSuccess:()=>{x.invalidateQueries({queryKey:["chat-agents"]}),h.success(n?"Agent created.":"Agent saved."),y()},onError:v=>{h.error(v instanceof _?v.message:"Could not save agent."),P(!1)}};n?o.add.mutate(t,r):o.update.mutate({id:s,patch:t},r)},K=()=>{u.reset.mutate(s,{onSuccess:()=>{h.success("Reset to default."),i&&f({name:i.name,description:i.description,systemPrompt:i.systemPrompt,skillIds:k.map(t=>t.id),tools:N.map(t=>t.name)})},onError:t=>h.error(t instanceof _?t.message:"Could not reset agent.")})},M=()=>{o.remove.mutate(s,{onSuccess:()=>{h.success("Agent deleted."),y()},onError:t=>h.error(t instanceof _?t.message:"Could not delete agent.")})};return!n&&!a&&p.isSuccess&&!c?e.jsxs("div",{className:"flex w-full flex-col items-center gap-3 px-8 py-16",children:[e.jsx("p",{className:"text-[14px] text-muted",children:"That agent no longer exists."}),e.jsx(C,{size:"sm",variant:"outline",onClick:y,children:"Back to agents"})]}):e.jsxs("div",{className:"flex w-full max-w-4xl flex-col gap-5 px-4 py-6 md:px-8",children:[e.jsxs("div",{className:"flex flex-wrap items-center gap-2",children:[e.jsx("button",{type:"button",onClick:y,className:"text-faint hover:text-ink",title:"Back to agents",children:e.jsx(ee,{size:18})}),e.jsx("span",{className:"text-[15px] font-medium text-ink",children:n?"New agent":a?l?"Edit agent (modified built-in)":"Edit agent (built-in)":"Edit agent"}),e.jsxs("div",{className:"ml-auto flex items-center gap-2",children:[a&&e.jsxs(C,{size:"sm",variant:"outline",onClick:()=>{j("agents","reset_agent_to_default"),K()},disabled:u.reset.isPending,children:[e.jsx(te,{size:13})," Reset to default"]}),e.jsx(C,{size:"sm",variant:"outline",onClick:y,children:"Cancel"}),e.jsx(C,{size:"sm",onClick:()=>{j("agents","save_agent"),F()},disabled:A||!m.name.trim(),children:A?"Saving…":"Save"})]})]}),a&&e.jsx("p",{className:"rounded-md border border-border bg-panel px-3 py-2 text-[12px] text-muted",children:"Built-in agents can be edited in place — Reset to default discards your changes and restores TurboLLM's original prompt and access."}),e.jsxs("div",{className:"flex flex-col gap-1.5",children:[e.jsx("label",{className:"text-[12px] font-medium text-muted",children:"Name"}),e.jsx("input",{className:"rounded-md border border-border bg-bg px-3 py-1.5 text-[13px] text-ink outline-none focus:border-accent placeholder:text-faint",placeholder:"Release Notes Writer",value:m.name,onChange:t=>f(r=>({...r,name:t.target.value}))})]}),e.jsxs("div",{className:"flex flex-col gap-1.5",children:[e.jsx("label",{className:"text-[12px] font-medium text-muted",children:"Description"}),e.jsx("input",{className:"rounded-md border border-border bg-bg px-3 py-1.5 text-[13px] text-ink outline-none focus:border-accent placeholder:text-faint",placeholder:"What this agent is for, in one line",value:m.description,onChange:t=>f(r=>({...r,description:t.target.value}))})]}),e.jsxs("div",{className:"flex flex-col gap-1.5",children:[e.jsx("label",{className:"text-[12px] font-medium text-muted",children:"System prompt"}),e.jsx("textarea",{className:"min-h-[160px] resize-y rounded-md border border-border bg-bg px-3 py-2 font-mono text-[12px] leading-relaxed text-ink outline-none focus:border-accent placeholder:text-faint",placeholder:"You are…",value:m.systemPrompt,onChange:t=>f(r=>({...r,systemPrompt:t.target.value}))})]}),e.jsxs("div",{className:"flex flex-col gap-2.5",children:[e.jsx("div",{className:"inline-flex w-fit rounded-md border border-border p-0.5",children:["skills","tools"].map(t=>e.jsx("button",{type:"button",onClick:()=>{j("agents","switch_agent_form_tab"),Q(t)},className:q("rounded px-3 py-1.5 text-[13px] font-medium transition-colors",E===t?"bg-accent/12 text-accent":"text-muted hover:text-ink"),children:t==="skills"?`Skills (${m.skillIds.length})`:`Tools (${m.tools.length})`},t))}),E==="skills"?I.isLoading?e.jsx("p",{className:"text-[12px] text-faint",children:"Loading…"}):k.length===0?e.jsx("p",{className:"text-[12px] text-faint",children:"No skills yet — create one under Customize → Skills."}):e.jsx("div",{className:"flex flex-col gap-2 rounded-md border border-border bg-panel-2 p-3",children:k.map(t=>e.jsxs("label",{className:"flex cursor-pointer items-start gap-2 text-[13px]",children:[e.jsx("input",{type:"checkbox",className:"mt-0.5",checked:m.skillIds.includes(t.id),onChange:()=>O(t.id)}),e.jsxs("span",{className:"flex flex-col",children:[e.jsx("span",{className:"text-ink",children:t.name}),t.description&&e.jsx("span",{className:"text-[12px] text-muted",children:t.description})]})]},t.id))}):w.isLoading?e.jsx("p",{className:"text-[12px] text-faint",children:"Loading…"}):e.jsx(ce,{tools:N,selected:m.tools,onChange:t=>f(r=>({...r,tools:t}))})]}),!n&&!a&&(L?e.jsxs("div",{className:"flex items-center gap-2 self-start rounded-md border border-border bg-panel p-2 text-[12px]",children:[e.jsx("span",{className:"flex-1 text-muted",children:"Delete this agent?"}),e.jsx("button",{type:"button",onClick:()=>{j("agents","delete_agent"),M()},className:"rounded px-2 py-1 text-[color:var(--err)] hover:bg-[color:color-mix(in_srgb,var(--err)_12%,transparent)]",children:"Delete"}),e.jsx("button",{type:"button",onClick:()=>{j("agents","cancel_delete_agent"),R(!1)},className:"rounded px-2 py-1 text-faint hover:text-ink",children:"Cancel"})]}):e.jsxs("button",{type:"button",onClick:()=>R(!0),className:"flex items-center gap-1.5 self-start text-[12px] text-faint hover:text-[color:var(--err)]",children:[e.jsx(se,{size:13})," Delete agent"]}))]})}export{je as AgentEditPage};