feat(FN-882): add loop detection recovery with compact-and-resume

- Add ContextLimitDetector to detect agent loops via repeated tool call patterns
- Implement compact-and-resume strategy: summarize conversation and restart agent from current step
- Add loop recovery to StuckTaskDetector with configurable attempt tracking and retry limits
- Extend executor with automatic loop recovery on context limit detection
- Add loop recovery support to pi executor with same compact-and-resume pattern
- Add comprehensive tests for context-limit-detector, stuck-task-detector loop detection, executor, and pi recovery
- Add changeset for patch bump to @gsxdsm/fusion
- Update README with loop detection and recovery documentation

packages/engine/src/context-limit-detector.ts

@@ -0,0 +1,60 @@
+/**
+ * Context limit error detection.
+ *
+ * Classifies errors from LLM providers that indicate the conversation context
+ * has grown too large for the model's window. Used by the executor to trigger
+ * compact-and-resume recovery before falling back to kill/requeue.
+ *
+ * Patterns are intentionally conservative — we only match errors that
+ * explicitly reference context/token overflow, NOT generic rate limits or
+ * server errors (those are handled by usage-limit-detector and transient-error-detector).
+ */
+
+/** Patterns that indicate a context-window overflow from the LLM provider. */
+const CONTEXT_OVERFLOW_PATTERNS: RegExp[] = [
+  // Anthropic: "prompt is too long: X tokens > Y maximum"
+  /prompt is too long/i,
+  // OpenAI (Completions & Responses): "exceeds the context window"
+  /exceeds?\s+the\s+context\s+window/i,
+  // Google Gemini: "input token count exceeds the maximum"
+  /input token count exceeds/i,
+  // xAI (Grok): "maximum prompt length is X but request contains Y"
+  /maximum prompt length/i,
+  // Groq: "reduce the length of the messages"
+  /reduce the length of the messages/i,
+  // Mistral: "too large for model with Y maximum context length"
+  /too large for model with.*maximum context length/i,
+  // OpenRouter (all backends): "maximum context length is X tokens"
+  /maximum context length is \d+ tokens/i,
+  // llama.cpp: "exceeds the available context size"
+  /exceeds?\s+the\s+available\s+context\s+size/i,
+  // LM Studio: "greater than the context length"
+  /greater than the context length/i,
+  // Kimi: "exceeded model token limit"
+  /exceeded model token limit/i,
+  // Generic catch-all: "context length exceeded" / "context window exceeded"
+  /context (?:length|window|size) exceeded/i,
+  // Token limit patterns with context keywords
+  /token limit.*context/i,
+  /too many tokens/i,
+  // Anthropic variant: "messages with that many tokens would exceed"
+  /tokens? would exceed/i,
+];
+
+/**
+ * Check if an error message indicates a context-window overflow.
+ *
+ * Returns true only when the message explicitly references context overflow
+ * from a known LLM provider pattern. Returns false for:
+ * - Rate limit errors (handled by usage-limit-detector)
+ * - Transient network errors (handled by transient-error-detector)
+ * - Generic "limit exceeded" without context keywords (false positive prevention)
+ * - "Aborted" errors without context signal
+ *
+ * @param message — The error message string to classify
+ * @returns true if the message indicates a context overflow
+ */
+export function isContextLimitError(message: string): boolean {
+  if (!message) return false;
+  return CONTEXT_OVERFLOW_PATTERNS.some((pattern) => pattern.test(message));
+}