@blockrun/franklin 3.15.35 → 3.15.36

package/dist/agent/llm.js

@@ -8,6 +8,21 @@ import { USER_AGENT } from '../config.js';
 import { routeRequest, parseRoutingProfile } from '../router/index.js';
 import { ThinkTagStripper } from './think-tag-stripper.js';
 import { isNemotronProseModel, stripNemotronProse } from './nemotron-prose-stripper.js';
+// Reasoning-tier models the gateway routes to that reject `tool_choice`
+// outright. Pattern: OpenAI o1/o3 family + DeepSeek's reasoner variant.
+// Add new entries as their 400 errors appear in real sessions; this is
+// a known-bad allowlist, not a guess. Wildcard substring match keeps it
+// resilient to model-revision suffixes (`o1-mini`, `o3-2026-04`, etc.).
+const MODELS_WITHOUT_TOOL_CHOICE_SUBSTR = [
+    'deepseek-reasoner',
+    'openai/o1',
+    'openai/o3',
+];
+function modelDoesNotSupportToolChoice(model) {
+    if (!model)
+        return false;
+    return MODELS_WITHOUT_TOOL_CHOICE_SUBSTR.some(s => model.includes(s));
+}
 function parseTimeoutEnv(name) {
     const raw = process.env[name];
     const parsed = raw ? Number.parseInt(raw, 10) : NaN;
@@ -309,6 +324,17 @@ export class ModelClient {
             (!Array.isArray(requestPayload['tools']) || requestPayload['tools'].length === 0)) {
             delete requestPayload['tool_choice'];
         }
+        // Models that don't support `tool_choice` (reasoning-only families).
+        // Verified 2026-05-04 from a real session: grounding-retry forced
+        // tool_choice on a request that ended up on deepseek-reasoner, which
+        // returned `400 Invalid request: deepseek-reasoner does not support
+        // this tool_choice`. Same shape applies to OpenAI o1 / o3 and
+        // similar restricted reasoning models. Strip silently — the agent
+        // loop's grounding-retry contract already tolerates the field
+        // disappearing (it'll re-evaluate next turn).
+        if (requestPayload['tool_choice'] !== undefined && modelDoesNotSupportToolChoice(request.model)) {
+            delete requestPayload['tool_choice'];
+        }
         // ── GLM-specific optimizations ───────────────────────────────────────────
         // GLM models work best with temperature=0.8 per official zai spec.
         // Enable thinking mode only for explicit reasoning variants (-thinking-).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.15.35",
+  "version": "3.15.36",
   "description": "Franklin — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {