orchestrator: recover from context overflow on fallback

src/backends/native/orchestrator.test.ts

@@ -605,6 +605,94 @@ describe('AgentOrchestrator', () => {
     });
   });
 
+  describe('process()', () => {
+    it('rolls back tool-loop provider errors, hard-trims on context overflow, and retries once', async () => {
+      let callCount = 0;
+      const mockClient: ModelClient = {
+        chat: vi.fn().mockImplementation(async () => {
+          callCount++;
+          if (callCount === 1) {
+            return {
+              content: '',
+              stopReason: 'tool_use',
+              usage: { inputTokens: 10, outputTokens: 5 },
+              toolCalls: [{ id: 'call_1', name: 'test.echo', args: { text: 'hi' } }],
+            } as ChatResponse;
+          }
+          if (callCount === 2) {
+            // Simulate llama.cpp context overflow buried inside an aggregated router error.
+            throw new Error(
+              'llama-server error (400): {"error":{"type":"exceedcontextsizeerror","nprompttokens":9183,"nctx":4096}}',
+            );
+          }
+          return {
+            content: 'ok',
+            stopReason: 'end_turn',
+            usage: { inputTokens: 10, outputTokens: 5 },
+          } as ChatResponse;
+        }),
+      };
+
+      const router = new ModelRouter({
+        default: mockClient,
+        fallbackChain: [],
+      });
+
+      // Minimal Session stub that supports rollback via replaceHistory().
+      const history: any[] = [];
+      const session = {
+        id: 'test',
+        addMessage: vi.fn((m: any) => { history.push(m); }),
+        getHistory: vi.fn(() => [...history]),
+        clear: vi.fn(() => { history.length = 0; }),
+        replaceHistory: vi.fn((msgs: any[]) => {
+          history.length = 0;
+          history.push(...msgs);
+        }),
+        getConfig: vi.fn(() => undefined),
+        setConfig: vi.fn(),
+        deleteConfig: vi.fn(),
+      } as any;
+
+      const registry = new ToolRegistry();
+      registry.register({
+        name: 'test.echo',
+        description: 'echo',
+        inputSchema: { type: 'object', properties: { text: { type: 'string' } }, required: ['text'] },
+        execute: async (args: any) => ({ success: true, output: String(args.text ?? '') }),
+      });
+
+      const hooks = new HookEngine({ confirm: [], log: [], silent: [] });
+      const executor = new ToolExecutor(registry, hooks);
+
+      const orchestrator = new AgentOrchestrator({
+        modelRouter: router,
+        systemPrompt: 'You are helpful.',
+        session,
+        toolRegistry: registry,
+        toolExecutor: executor,
+        primaryTier: 'default',
+        delegation: {
+          compaction: 'fast',
+          memory_extraction: 'default',
+          classification: 'complex',
+          tool_summarisation: 'default',
+          complex_reasoning: 'complex',
+        },
+        maxDelegationDepth: 3,
+      });
+
+      const res = await orchestrator.process('hello');
+      expect(res).toBe('ok');
+
+      // Ensure we didn't persist the low-level error string in history.
+      const textHistory = history
+        .map(m => (typeof m.content === 'string' ? m.content : ''))
+        .join('\n');
+      expect(textHistory).not.toContain('Error in tool loop');
+    });
+  });
+
   describe('setModelTier()', () => {
     it('sets model tier on primary agent', () => {
       const orchestrator = new AgentOrchestrator({

src/backends/native/orchestrator.ts

@@ -9,11 +9,12 @@ import type { Attachment } from '../../channels/types.js';
 import { NativeAgent } from './agent.js';
 import type { ToolUseEvent } from './agent.js';
 import type { OutboundAttachmentCollector } from './attachments.js';
-import { shouldCompact } from '../../context/tokens.js';
+import { estimateMessageTokens, shouldCompact } from '../../context/tokens.js';
 import { compactHistory, type CompactionConfig, type CompactionResult, DEFAULT_COMPACTION_CONFIG } from '../../context/compaction.js';
 import { estimateCost } from '../../models/costs.js';
 import { auditLogger } from '../../audit/index.js';
 import { buildAdaptiveMemoryContext, buildRecentMemoryContext } from '../../memory/adaptive.js';
+import { buildUserMessage } from '../../models/media.js';
 
 // ── Public types ──────────────────────────────────────────────────────
 
@@ -227,7 +228,48 @@ export class AgentOrchestrator {
   async process(userMessage: string, attachments?: Attachment[]): Promise<string> {
     this._injectMemoryContext(userMessage);
     await this.compactIfNeeded();
-    return this._agent.process(userMessage, attachments);
+
+    // Snapshot history so we can rollback if the underlying tool loop returns an error message.
+    // This avoids persisting low-level provider errors to the user-visible conversation state.
+    const before = this.getHistory();
+
+    const result = await this._agent.process(userMessage, attachments);
+
+    // NativeAgent currently converts tool-loop exceptions into a user-visible error string.
+    // Intercept a few common cases here to self-heal (context overflow) and/or degrade gracefully.
+    if (this._isToolLoopErrorMessage(result)) {
+      // Roll back the user message + error message inserted by the agent.
+      this._restoreHistory(before);
+
+      const underlying = this._stripToolLoopErrorPrefix(result);
+      const ctx = this._extractContextWindowFromError(underlying);
+      if (ctx) {
+        // Attempt: compact + hard-trim to fit the discovered context window, then retry once.
+        await this._compactAndTrimToFit(ctx);
+        const retry = await this._agent.process(userMessage, attachments);
+        if (!this._isToolLoopErrorMessage(retry)) {
+          return retry;
+        }
+        // If we still failed, roll back again so we don't persist the error string.
+        this._restoreHistory(before);
+      }
+
+      // Persist a short, user-friendly failure message (without provider internals).
+      const friendly =
+        [
+          'I ran into a model/provider error while processing that message.',
+          '',
+          'Try again. If it keeps happening:',
+          '1. Run `/compact` or `/reset` to shrink the conversation context.',
+          '2. Switch to a different model tier (e.g. `/model local`).',
+        ].join('\n');
+
+      // Re-add the user message so the conversation state matches what the user sent.
+      this._appendUserAndAssistant(userMessage, attachments, friendly);
+      return friendly;
+    }
+
+    return result;
   }
 
   /**
@@ -445,6 +487,77 @@ export class AgentOrchestrator {
     await this.compact();
   }
 
+  private _isToolLoopErrorMessage(text: string): boolean {
+    return text.startsWith('Error in tool loop (iteration ');
+  }
+
+  private _stripToolLoopErrorPrefix(text: string): string {
+    const m = text.match(/^Error in tool loop \(iteration \d+\):\s*(.*)$/s);
+    return m ? m[1] : text;
+  }
+
+  private _restoreHistory(messages: Message[]): void {
+    if (this._session) {
+      this._session.replaceHistory(messages);
+      return;
+    }
+    // No session available; nothing safe to do here.
+  }
+
+  private _appendUserAndAssistant(userMessage: string, attachments: Attachment[] | undefined, assistantText: string): void {
+    if (!this._session) {
+      return;
+    }
+    const userMsg = buildUserMessage(userMessage, attachments);
+    this._session.addMessage(userMsg);
+    this._session.addMessage({ role: 'assistant', content: assistantText });
+  }
+
+  private _extractContextWindowFromError(errorText: string): number | undefined {
+    // Try a few common patterns and pick the smallest plausible context window.
+    // Example llama.cpp error:
+    //   exceeds the available context size (4096 tokens) ... "nctx":4096
+    const candidates: number[] = [];
+    const jsonNctx = errorText.match(/"nctx"\s*:\s*(\d{3,7})/);
+    if (jsonNctx) { candidates.push(Number(jsonNctx[1])); }
+    const paren = errorText.match(/context size \((\d{3,7}) tokens\)/);
+    if (paren) { candidates.push(Number(paren[1])); }
+    const maxContext = errorText.match(/maximum context length is (\d{3,7}) tokens/i);
+    if (maxContext) { candidates.push(Number(maxContext[1])); }
+
+    const valid = candidates.filter(n => Number.isFinite(n) && n >= 256);
+    if (valid.length === 0) {
+      return undefined;
+    }
+    return Math.min(...valid);
+  }
+
+  private async _compactAndTrimToFit(contextWindow: number): Promise<void> {
+    // Compaction is best-effort; if it fails (e.g., providers down), fall back to a hard trim.
+    try {
+      await this.compact();
+    } catch (error) {
+      console.warn('[Flynn:compact] Emergency compaction failed:', error);
+    }
+
+    if (!this._session) {
+      return;
+    }
+
+    const threshold = Math.floor((this._compactionConfig?.thresholdPct ?? 80) / 100 * contextWindow);
+    let messages = this.getHistory();
+    let estimated = estimateMessageTokens(messages);
+
+    // Drop oldest messages until we're under budget.
+    // This is intentionally blunt; it only triggers after a real provider context overflow.
+    while (messages.length > 1 && estimated > threshold) {
+      messages = messages.slice(1);
+      estimated = estimateMessageTokens(messages);
+    }
+
+    this._session.replaceHistory(messages);
+  }
+
   /** Accumulate usage stats for a given tier. */
   private _trackUsage(tier: ModelTier, usage: TokenUsage): void {
     const existing = this._usageByTier.get(tier);

src/daemon/routing.ts

@@ -79,25 +79,31 @@ export function createMessageRouter(deps: {
     const baseSid = agentConfigName
       ? `${channel}:${senderId}:${agentConfigName}`
       : `${channel}:${senderId}`;
-    const sessionId = tierFromMetadata ? `${baseSid}:${tierFromMetadata}` : baseSid;
+    const session = deps.sessionManager.getSession(channel, senderId);
+
+    // Read per-session model tier override (persisted in SQLite)
+    const sessionTierOverride = session.getConfig('modelTier') as ModelTier | undefined;
+
+    // Resolution chain: metadata (cron) → session override → agent config → global default
+    const effectiveTier = tierFromMetadata
+      ?? sessionTierOverride
+      ?? agentConfig?.modelTier
+      ?? deps.config.agents.primary_tier
+      ?? 'default';
+
+    // Cache agents by tier too so switching tiers updates context-window heuristics.
+    const sessionId = `${baseSid}:${effectiveTier}`;
 
     let entry = agents.get(sessionId);
     if (!entry) {
-      const session = deps.sessionManager.getSession(channel, senderId);
-
-      // Read per-session model tier override (persisted in SQLite)
-      const sessionTierOverride = session.getConfig('modelTier') as ModelTier | undefined;
-
-      // Resolution chain: metadata (cron) → session override → agent config → global default
-      const effectiveTier = tierFromMetadata
-        ?? sessionTierOverride
-        ?? agentConfig?.modelTier
-        ?? deps.config.agents.primary_tier
-        ?? 'default';
-
       // Use agent config overrides where available, falling back to global config
       const effectiveSystemPrompt = agentConfig?.systemPrompt ?? deps.systemPrompt;
-      const effectiveProvider = deps.config.models.default.provider;
+
+      const modelsConfig = deps.config.models as Record<string, { provider?: string; model?: string; context_window?: number } | undefined>;
+      const tierConfig = modelsConfig[effectiveTier] ?? deps.config.models.default;
+      const effectiveProvider = tierConfig?.provider ?? deps.config.models.default.provider;
+      const effectiveModelName = tierConfig?.model ?? deps.config.models.default.model;
+      const effectiveContextWindow = tierConfig?.context_window ?? deps.config.models.default.context_window;
 
       const delegationConfig: DelegationConfig = {
         compaction: deps.config.agents.delegation.compaction ?? 'fast',
@@ -181,8 +187,8 @@ export function createMessageRouter(deps: {
           summaryMaxTokens: deps.config.compaction.summary_max_tokens,
           importanceThreshold: deps.config.compaction.importance_threshold,
         } : undefined,
-        modelName: deps.config.models.default.model,
-        contextWindow: deps.config.models.default.context_window,
+        modelName: effectiveModelName,
+        contextWindow: effectiveContextWindow,
         memoryStore: deps.memoryStore,
         memoryInjectionStrategy: deps.config.memory?.injection_strategy,
         memoryMaxInjectionTokens: deps.config.memory?.max_injection_tokens,