Method recordToolFailure

src/agent/loop.ts:136–150

Record a tool failure to episodic memory (best-effort, opt-in). Only fires when * failure-driven learning is enabled; the pattern miner later decides what's worth * learning. Fire-and-forget — never blocks or throws into the loop.

(tool: string, content: string)

  private currentTurn = 0;
  /**
   * Mid-task steering queue. The UI pushes `/btw …` notes here while a run is in
   * flight; the run loop drains them at the top of each iteration and injects them
   * into the conversation so the model can adjust course without being stopped.
   * Single-threaded JS event loop ⇒ no lock needed.
   */
  private steerQueue: string[] = [];
  /** Session ledger of files the model has demonstrably read (read-before-write gate). */
  private readLedger = new ReadLedger();
  /** Skills already auto-injected this session — never inject the same one twice. */
  private autoInjectedSkills = new Set<string>();
  /** Monotonic union of tool names shipped this session. The relevance gate only ever
   *  ADDS to this — never drops — so the tools block stays a byte-stable cache prefix
   *  across turns (a sliding per-turn set would flip and invalidate the prompt cache). */
  private sessionToolNames = new Set<string>();
  /** Where the active model is actually served from ('ollama'/'lmstudio'/'anthropic'/…) and
   *  its LIVE context window — set each iteration, forwarded on budget_update for the UI. */
  private lastModelSource = '';
  private lastEffectiveCtxWindow = 0;