fluxflow-cli 2.2.8 → 2.2.10

package/ARCHITECTURE.md

@@ -22,9 +22,9 @@ The execution flow of a single user prompt follows this loop:
    - Each tool is dispatched to its respective handler in `src/tools/`.
    - Tool outputs are collected and appended to the context as `[TOOL RESULT]: ...`.
 4. **Security Governance**: During tool execution, the loop enforces security checks (e.g., blocking `exec_command` from accessing system root drives if "External Workspace Access" is off) and pauses for Human-in-the-Loop (HITL) approval if necessary.
-5. **Turn Management & Continuation**: The model is instructed to append `[turn: finish]` if its goal is complete, or `[turn: continue]` if it expects tool results.
+5. **Turn Management & Continuation**: The model is instructed to append `[[END]]` or `[turn: finish]` if its goal is complete, or `[turn: continue]` if it expects tool results.
    - If tools were called or `[turn: continue]` is present, the loop increments and re-prompts the model with the newly gathered `[TOOL RESULT]` data.
-   - If `[turn: finish]` is detected and no further tools were called, the main loop terminates, passing the final synthesized context to the background Janitor process.
+   - If `[[END]]` or `[turn: finish]` is detected and no further tools were called, the main loop terminates, passing the final synthesized context to the background Janitor process.
 6. **Loop Limits & Resilience**: To prevent infinite loops or excessive API usage, **Flux mode** is capped at 70 iterations per user prompt, while **Flow mode** is capped at 7.
    - **Multi-Stage Failover**: The loop features a sophisticated 16-attempt retry engine with exponential backoff (1s - 32s).
    - **Critical Fallback Pivot**: If the primary model fails 14 consecutive times, the agent surgically pivots to a lighter, high-concurrency fallback model (`gemini-3.1-flash-lite`) for the final 3 attempts to ensure session navigation through API congestion.