fix: normalize OpenAI/GitHub finish_reason to Flynn stopReason conventions

OpenAI-compatible providers return 'stop' and 'tool_calls' as finish_reason
values, but Flynn's agent loop expects Anthropic-style 'end_turn' and
'tool_use'. This caused the agent to exit the tool loop prematurely when
falling back to GitHub Copilot (due to Anthropic API quota exhaustion).

- openai.ts: Map 'stop' → 'end_turn', 'length' → 'max_tokens', tool_calls
  with actual tools → 'tool_use', tool_calls without tools → 'end_turn'
- github.ts: Handle edge case where finish_reason is 'tool_calls' but no
  tools were parsed
- agent.ts: Accept both 'tool_use' and 'tool_calls' as valid stop reasons
  (belt-and-suspenders), extract toolCalls to local variable for TS narrowing
- openai.test.ts: Update expectations to match new normalized values

src/backends/native/agent.ts

@@ -174,8 +174,12 @@ export class NativeAgent {
         this._totalUsage.outputTokens += response.usage.outputTokens;
         this._callCount++;
 
-        // If the model didn't request tool use, we're done
-        if (response.stopReason !== 'tool_use' || !response.toolCalls?.length) {
+        // If the model didn't request tool use, we're done.
+        // Check both 'tool_use' (Anthropic) and 'tool_calls' (OpenAI-compatible) stop reasons,
+        // but always require actual toolCalls to be present.
+        const wantsToolUse = (response.stopReason === 'tool_use' || response.stopReason === 'tool_calls')
+          && response.toolCalls && response.toolCalls.length > 0;
+        if (!wantsToolUse) {
           let finalContent = response.content;
           if (response.thinkingContent) {
             finalContent = `<thinking>\n${response.thinkingContent}\n</thinking>\n\n${response.content}`;
@@ -185,8 +189,11 @@ export class NativeAgent {
           return finalContent;
         }
 
+        // Safe to assert non-null — wantsToolUse guarantees toolCalls exists and is non-empty
+        const toolCalls = response.toolCalls!;
+
         // Check for repeated tool calls — build a fingerprint from tool names + args
-        const fingerprint = response.toolCalls
+        const fingerprint = toolCalls
           .map(tc => `${tc.name}:${JSON.stringify(tc.args)}`)
           .sort()
           .join('|');
@@ -199,7 +206,7 @@ export class NativeAgent {
         }
 
         // Track consecutive calls to the same tool (by name, ignoring args)
-        const toolNames = response.toolCalls.map(tc => tc.name).sort().join(',');
+        const toolNames = toolCalls.map(tc => tc.name).sort().join(',');
         if (toolNames === lastToolName) {
           sameToolStreak++;
         } else {
@@ -213,7 +220,7 @@ export class NativeAgent {
         if (response.content) {
           assistantContent.push({ type: 'text', text: response.content });
         }
-        for (const tc of response.toolCalls) {
+        for (const tc of toolCalls) {
           assistantContent.push({
             type: 'tool_use',
             id: tc.id,
@@ -226,7 +233,7 @@ export class NativeAgent {
         // Execute each tool call and collect results
         const toolResultBlocks: unknown[] = [];
         lastToolResults = [];
-        for (const tc of response.toolCalls) {
+        for (const tc of toolCalls) {
           const internalName = this.toolRegistry!.getByApiName(tc.name)?.name ?? tc.name;
           this.onToolUse?.({ type: 'start', tool: internalName, args: tc.args });
 

src/models/github.ts

@@ -161,6 +161,9 @@ export class GitHubModelsClient implements ModelClient {
         stopReason = 'end_turn';
       } else if (reason === 'length') {
         stopReason = 'max_tokens';
+      } else if (reason === 'tool_calls') {
+        // Edge case: finish_reason says tool_calls but none were parsed
+        stopReason = 'end_turn';
       } else {
         stopReason = reason ?? 'end_turn';
       }

src/models/openai.test.ts

@@ -29,7 +29,7 @@ describe('OpenAIClient', () => {
     });
 
     expect(response.content).toBe('Hello from GPT!');
-    expect(response.stopReason).toBe('stop');
+    expect(response.stopReason).toBe('end_turn');
     expect(response.usage.inputTokens).toBe(10);
     expect(response.usage.outputTokens).toBe(5);
   });
@@ -66,7 +66,7 @@ describe('OpenAIClient tool use', () => {
       }],
     });
 
-    expect(response.stopReason).toBe('tool_calls');
+    expect(response.stopReason).toBe('tool_use');
     expect(response.toolCalls).toHaveLength(1);
     expect(response.toolCalls![0]).toEqual({
       id: 'call_1',

src/models/openai.ts

@@ -96,9 +96,27 @@ export class OpenAIClient implements ModelClient {
       args: JSON.parse(tc.function.arguments),
     })) ?? [];
 
+    // Map OpenAI finish reasons to Flynn's stop reasons
+    let stopReason: string;
+    if (toolCalls.length > 0) {
+      stopReason = 'tool_use';
+    } else {
+      const reason = choice?.finish_reason;
+      if (reason === 'stop') {
+        stopReason = 'end_turn';
+      } else if (reason === 'length') {
+        stopReason = 'max_tokens';
+      } else if (reason === 'tool_calls') {
+        // Edge case: finish_reason says tool_calls but none were parsed
+        stopReason = 'end_turn';
+      } else {
+        stopReason = reason ?? 'end_turn';
+      }
+    }
+
     return {
       content,
-      stopReason: choice?.finish_reason ?? 'stop',
+      stopReason,
       usage: {
         inputTokens: response.usage?.prompt_tokens ?? 0,
         outputTokens: response.usage?.completion_tokens ?? 0,