openfox 2.0.118 → 2.0.120

package/CHANGELOG.md

@@ -1,5 +1,39 @@
 # Changelog
 
+## 2.0.120 - 2026-08-15
+
+### Bug Fixes
+
+- **DeepSeek default temperature matches the model card** — DeepSeek models now sample at 1.0 by default instead of 0.6
+- **Max tokens follow the session's model** — clamping and truncation retry budgets now use the active model's context window, not a global default
+- **Truncation retries no longer send negative max_tokens** — retry budget is floored at 256 tokens, preventing HTTP 400 errors
+
+## 2.0.119 - 2026-08-14
+
+### Features
+
+- **Per-model sampling controls** — pick which sampling parameters each backend receives, per model
+- **Read-only session status** — lightweight API and widget expose live state for external integrations
+
+### Enhancements
+
+- **Reviewer gets the full picture** — code reviews now factor in intended scope, verifier notes, and changed files
+- **No surprise navigation on In Progress** — moving a task keeps your place; open it via the card link
+- **Task editor shortcuts inverted** — Enter inserts a newline, Shift+Enter saves
+
+### Bug Fixes
+
+- **New-task autocomplete fixed** — slash/@ suggestions are no longer clipped and load on cold start
+- **Long chats safer against rejection** — max_tokens now keeps 2048 tokens of headroom
+- **Split panes keep dev state separate** — server status, logs, and config no longer bleed between workdirs
+- **ask_user prompts render as markdown** — no more raw text walls
+- **Submodule changes appear in the diff preview**
+- **Auto-config spots rejected reasoning** — providers that refuse it no longer receive reasoning history
+- **LM Studio context windows set by auto-config** — no longer left at zero
+- **No phantom password prompt** — the login modal no longer flashes while connecting
+- **Silent streams end their turn** — idle LLM connections time out cleanly instead of hanging
+- **Null token usage no longer crashes sessions**
+
 ## 2.0.118 - 2026-08-12
 
 ### Features

package/dist/CHANGELOG.md

@@ -1,5 +1,39 @@
 # Changelog
 
+## 2.0.120 - 2026-08-15
+
+### Bug Fixes
+
+- **DeepSeek default temperature matches the model card** — DeepSeek models now sample at 1.0 by default instead of 0.6
+- **Max tokens follow the session's model** — clamping and truncation retry budgets now use the active model's context window, not a global default
+- **Truncation retries no longer send negative max_tokens** — retry budget is floored at 256 tokens, preventing HTTP 400 errors
+
+## 2.0.119 - 2026-08-14
+
+### Features
+
+- **Per-model sampling controls** — pick which sampling parameters each backend receives, per model
+- **Read-only session status** — lightweight API and widget expose live state for external integrations
+
+### Enhancements
+
+- **Reviewer gets the full picture** — code reviews now factor in intended scope, verifier notes, and changed files
+- **No surprise navigation on In Progress** — moving a task keeps your place; open it via the card link
+- **Task editor shortcuts inverted** — Enter inserts a newline, Shift+Enter saves
+
+### Bug Fixes
+
+- **New-task autocomplete fixed** — slash/@ suggestions are no longer clipped and load on cold start
+- **Long chats safer against rejection** — max_tokens now keeps 2048 tokens of headroom
+- **Split panes keep dev state separate** — server status, logs, and config no longer bleed between workdirs
+- **ask_user prompts render as markdown** — no more raw text walls
+- **Submodule changes appear in the diff preview**
+- **Auto-config spots rejected reasoning** — providers that refuse it no longer receive reasoning history
+- **LM Studio context windows set by auto-config** — no longer left at zero
+- **No phantom password prompt** — the login modal no longer flashes while connecting
+- **Silent streams end their turn** — idle LLM connections time out cleanly instead of hanging
+- **Null token usage no longer crashes sessions**
+
 ## 2.0.118 - 2026-08-12
 
 ### Features

package/dist/agent-defaults/code-reviewer.agent.md

@@ -11,14 +11,19 @@ allowedTools:
   - session_metadata
 ---
 
-You are a code reviewer. Review the **git diff** of the modified files rather than reading the full files.
+You are a code reviewer. Your caller may provide situational context before the diff — **intended scope (criteria)**, **what the verifier already confirmed**, and the **modified files**. Review the **git diff** of the modified files rather than reading the full files.
+
+When context is provided, use it to judge the change against what was actually requested:
+
+- Read the intended scope (criteria) and check the diff delivers it faithfully — flag mismatches, partial implementations, or scope creep.
+- Treat the verifier's notes as already-confirmed ground truth. Don't re-verify what it covered; instead catch what it missed.
+- You're not looking for 100% perfection. Just something that feels good to use and doesn't bloat the software.
+
 You're mostly interested in:
 
 1. **For the user (UX interaction)** — does this feel good to use? Any rough edges, confusing behavior, or unnecessary friction?
 2. **For the project (code quality + guidelines)** — does this follow project conventions? Is it maintainable? Does it bloat the software?
 
-You're not looking for 100% perfection. Just something that feels good to use and doesn't bloat the software.
-
 ## Managing Findings
 
 Use `session_metadata` to track your findings:

package/dist/{auto-config-AAUO4PKW.js → auto-config-YXNCBYEY.js}

@@ -1,3 +1,6 @@
+import {
+  findLmStudioModel
+} from "./chunk-K2QKX4ZJ.js";
 import {
   ensureVersionPrefix
 } from "./chunk-HNCM3D7Y.js";
@@ -98,14 +101,10 @@ async function detectLmstudioInfo(baseUrl, modelId) {
   const nativeUrl = `${base.replace(/\/v\d+\/?$/, "")}/api/v1/models`;
   const response = await fetch(nativeUrl, { signal: AbortSignal.timeout(5e3) });
   if (!response.ok) throw new Error(`HTTP ${response.status}`);
-  const data = await response.json();
-  const model = data.find((m) => (m.key ?? m.id) === modelId);
+  const model = findLmStudioModel(await response.json(), modelId);
   if (!model) throw new Error(`Model ${modelId} not found in LM Studio`);
-  const loadedContext = model.loaded_instances?.[0]?.config?.context_length;
-  const ctxLen = loadedContext ?? model.max_context_length;
-  const supportsVision = model.capabilities?.vision ?? false;
-  if (ctxLen) {
-    return { contextWindow: ctxLen, source: "backend", supportsVision };
+  if (model.contextWindow) {
+    return { contextWindow: model.contextWindow, source: "backend", supportsVision: model.supportsVision };
   }
   throw new Error("No context_length in LM Studio response");
 }
@@ -157,6 +156,117 @@ async function probeCombos(baseUrl, apiKey, model, combos) {
   logger.debug("Auto-config: no working combo found", { model });
   return null;
 }
+async function probeReasoningInMessages(baseUrl, apiKey, model) {
+  const headers = { "Content-Type": "application/json" };
+  if (apiKey) headers["Authorization"] = `Bearer ${apiKey}`;
+  try {
+    const response = await fetch(`${ensureVersionPrefix(baseUrl)}/chat/completions`, {
+      method: "POST",
+      headers,
+      body: JSON.stringify({
+        model,
+        messages: [
+          { role: "user", content: "say hi in one word" },
+          { role: "assistant", content: "hi", reasoning: "probe" },
+          { role: "user", content: "say hi in one word" }
+        ],
+        max_tokens: 1
+      }),
+      signal: AbortSignal.timeout(15e3)
+    });
+    return response.ok;
+  } catch {
+    return void 0;
+  }
+}
+var STANDARD_PARAMS = ["temperature", "top_p", "max_tokens", "top_k", "reasoning_effort"];
+var PARAM_VALUES = {
+  temperature: 0.7,
+  top_p: 0.9,
+  max_tokens: 50,
+  top_k: 40,
+  reasoning_effort: "high"
+};
+var escapeRegExp = (value) => value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+var REJECTED_PARAM_PATTERNS = STANDARD_PARAMS.map((param) => ({
+  param,
+  pattern: new RegExp(`\\b${escapeRegExp(param)}\\b`, "i")
+}));
+function extractRejectedParam(errorText) {
+  for (const { param, pattern } of REJECTED_PARAM_PATTERNS) {
+    if (pattern.test(errorText)) return param;
+  }
+  return void 0;
+}
+var DUMMY_TOOL = {
+  type: "function",
+  function: {
+    name: "noop",
+    description: "No-op tool for probing",
+    parameters: { type: "object", properties: {}, required: [] }
+  }
+};
+async function probeChatCompletions(baseUrl, headers, model, samplingParams) {
+  const body = {
+    model,
+    messages: [{ role: "user", content: "say hi in one word" }],
+    tools: [DUMMY_TOOL],
+    tool_choice: "auto"
+  };
+  for (const param of samplingParams) {
+    body[param] = PARAM_VALUES[param];
+  }
+  try {
+    const response = await fetch(`${ensureVersionPrefix(baseUrl)}/chat/completions`, {
+      method: "POST",
+      headers,
+      body: JSON.stringify(body),
+      signal: AbortSignal.timeout(15e3)
+    });
+    if (response.ok) return { ok: true, status: response.status, errorText: "" };
+    return { ok: false, status: response.status, errorText: await response.text() };
+  } catch {
+    return { ok: false, status: 0, errorText: "" };
+  }
+}
+async function probeRejectedParams(baseUrl, apiKey, model, backend) {
+  const headers = { "Content-Type": "application/json" };
+  if (apiKey) headers["Authorization"] = `Bearer ${apiKey}`;
+  const candidates = [...STANDARD_PARAMS];
+  if (["openai", "anthropic", "ollama"].includes(backend)) {
+    const idx = candidates.indexOf("top_k");
+    if (idx !== -1) candidates.splice(idx, 1);
+  }
+  const control = await probeChatCompletions(baseUrl, headers, model, []);
+  if (!control.ok) {
+    logger.debug("Auto-config: rejection-probe control failed, skipping param detection", {
+      model,
+      status: control.status
+    });
+    return [];
+  }
+  const rejected = [];
+  const remaining = [...candidates];
+  while (remaining.length > 0) {
+    const probe = await probeChatCompletions(baseUrl, headers, model, remaining);
+    if (probe.ok) break;
+    if (probe.status === 400) {
+      const rejectedParam = extractRejectedParam(probe.errorText);
+      if (rejectedParam && remaining.includes(rejectedParam)) {
+        rejected.push(rejectedParam);
+        remaining.splice(remaining.indexOf(rejectedParam), 1);
+        logger.debug("Auto-config: param rejected, retrying without", { model, rejectedParam });
+        continue;
+      }
+      break;
+    }
+    break;
+  }
+  if (rejected.length > 0) {
+    logger.info("Auto-config: detected rejected params", { model, rejected });
+  }
+  return rejected;
+}
 async function autoConfig(input) {
   const { url, apiKey, backend, models } = input;
   const baseUrl = url.replace(/\/+$/, "");
@@ -168,17 +278,21 @@ async function autoConfig(input) {
       source: contextSource,
       supportsVision
     } = await detectModelInfo(baseUrl, apiKey, backend, model.id);
-    const [thinkingConfig, nonThinkingConfig] = await Promise.all([
+    const [thinkingConfig, nonThinkingConfig, rejectedParams] = await Promise.all([
       probeCombos(baseUrl, apiKey, model.id, THINKING_COMBOS),
-      probeCombos(baseUrl, apiKey, model.id, NON_THINKING_COMBOS)
+      probeCombos(baseUrl, apiKey, model.id, NON_THINKING_COMBOS),
+      probeRejectedParams(baseUrl, apiKey, model.id, backend)
     ]);
+    const sendReasoningInMessages = thinkingConfig ? await probeReasoningInMessages(baseUrl, apiKey, model.id) : void 0;
     results.push({
       id: model.id,
       contextWindow,
       contextSource,
       supportsVision,
       thinkingConfig,
-      nonThinkingConfig
+      nonThinkingConfig,
+      ...sendReasoningInMessages !== void 0 ? { sendReasoningInMessages } : {},
+      ...rejectedParams.length > 0 ? { rejectedParams } : {}
     });
   }
   return { models: results };
@@ -186,4 +300,4 @@ async function autoConfig(input) {
 export {
   autoConfig
 };
-//# sourceMappingURL=auto-config-AAUO4PKW.js.map
+//# sourceMappingURL=auto-config-YXNCBYEY.js.map

package/dist/{chat-handler-KNBP6WVI.js → chat-handler-TIR5CMWE.js}

@@ -2,14 +2,14 @@ import {
   buildRunChatTurnParams,
   finalizeTurnCompletion,
   generateSessionNameForSession
-} from "./chunk-MTFLMNLN.js";
+} from "./chunk-USNG72WA.js";
 import {
   runChatTurn
-} from "./chunk-CX7UJZA5.js";
-import "./chunk-IDUXGSQV.js";
-import "./chunk-5E5G3DLI.js";
+} from "./chunk-HWHWXLM5.js";
+import "./chunk-JDWJ2PJJ.js";
+import "./chunk-IW46BUOP.js";
 import "./chunk-2FCLCUVR.js";
-import "./chunk-NP2SV2CX.js";
+import "./chunk-IQFPRQWM.js";
 import "./chunk-27NABDXB.js";
 import "./chunk-M6X6AIEG.js";
 import "./chunk-ZTOG6LJD.js";
@@ -18,7 +18,7 @@ import "./chunk-JE7LW7Y6.js";
 import "./chunk-OAN4BXEW.js";
 import "./chunk-O6DU226D.js";
 import "./chunk-J7GAZGZT.js";
-import "./chunk-HZSKKJHQ.js";
+import "./chunk-AS7KAH66.js";
 import "./chunk-EQ5CGNOT.js";
 import "./chunk-XZGQZ42D.js";
 import {
@@ -38,13 +38,14 @@ import "./chunk-LCLH6ZUL.js";
 import "./chunk-LNNBLB6G.js";
 import "./chunk-OSXPQSLW.js";
 import "./chunk-3PG65AXG.js";
-import "./chunk-GZTMCTUS.js";
+import "./chunk-WN76RDRU.js";
+import "./chunk-K2QKX4ZJ.js";
 import "./chunk-J2GP3J3X.js";
-import "./chunk-MKWS7TQH.js";
-import "./chunk-W7VIYHPQ.js";
+import "./chunk-KFQKBXLO.js";
+import "./chunk-GJL3QSMP.js";
 import "./chunk-AZHWLLPP.js";
 import "./chunk-HNCM3D7Y.js";
-import "./chunk-TOKQ2UVG.js";
+import "./chunk-MIPQOHWS.js";
 import "./chunk-LTYL6ZGH.js";
 import "./chunk-HD6Y77RO.js";
 import "./chunk-5PT4UMRR.js";
@@ -173,4 +174,4 @@ export {
   startChatSession,
   stopSessionExecution
 };
-//# sourceMappingURL=chat-handler-KNBP6WVI.js.map
+//# sourceMappingURL=chat-handler-TIR5CMWE.js.map