@robmclarty/volley 0.10.0 → 0.11.0

package/dist/index.d.ts

@@ -51,6 +51,7 @@ type VolleyConfig = {
   builder_max_steps?: number;
   builder_max_tool_calls?: number;
   builder_max_output_tokens?: number;
+  builder_trim_history?: number;
   allow_unsandboxed_builder?: boolean;
   critic_model?: string;
   critic_provider?: CriticProvider;
@@ -114,6 +115,11 @@ type ResolvedConfig = {
    * one step, not the session: fascicle feeds back that it hit the limit and
    * the loop goes on. Ignored by the `claude_cli` builder. */
   builder_max_output_tokens: number | null;
+  /** The estimated request size, in tokens, above which a tool-loop builder's
+   * turns send older tool results as one-line stubs (#7), or null to send
+   * every turn whole. Null by default, until a long local session is shown
+   * to run faster with it on. Ignored by the `claude_cli` builder. */
+  builder_trim_history: number | null;
   /** Operator opt-out permitting a non-`claude_cli` builder to run without a
    * sandbox — it gets a real host bash. Refused by default. */
   allow_unsandboxed_builder: boolean;
@@ -285,6 +291,40 @@ type PhaseRecord = {
    * another way. A thinking seat needs ai-sdk-ollama 4.3.0 or later, which
    * streams the reasoning even though volley never sets `think`. */
   throughput?: Throughput;
+  /** Each turn of a builder session, in order, so a long session's growth
+   * and where its time went can be read off the archive (#7). Set on a builder
+   * record whose provider reported steps (every tool-loop builder); absent on
+   * the critic and on a session with none. Additive. */
+  steps?: BuilderStepRecord[];
+};
+/** Where a builder turn's prefill and decode times came from.
+ * `ollama_reported`: Ollama's own `prompt_eval_duration` and `eval_duration`
+ * (native transport). `ollama_derived`: prefill as `total_duration` less
+ * `load_duration` and `eval_duration`, the ai_sdk transport's metadata, which
+ * leaves `prompt_eval_duration` out; slightly high, since the total includes
+ * Ollama's own overhead. `first_chunk`: prefill as the time to the first
+ * streamed chunk, decode as the rest of the round trip. `none`: no split. */
+type StepTimingBasis = 'ollama_reported' | 'ollama_derived' | 'first_chunk' | 'none';
+/** One turn of a builder session. */
+type BuilderStepRecord = {
+  index: number;
+  /** The whole request this turn sent, so the session's growth reads as the
+   * sequence of these. */
+  input_tokens: number;
+  output_tokens: number;
+  /** The turn's provider round trip, tool execution excluded. */
+  duration_ms: number;
+  prefill_ms?: number;
+  decode_ms?: number;
+  timing_basis: StepTimingBasis;
+  finish_reason: FinishReason;
+  truncated?: true;
+  /** Each tool call the turn made, and how many characters its result put
+   * into the transcript every later turn resends. */
+  tool_calls: Array<{
+    name: string;
+    output_chars: number;
+  }>;
 };
 type HaltReason = 'cost_cap' | 'gate_edit' | null;
 /** Carry-state threaded through the fascicle loop. */
@@ -486,10 +526,11 @@ export declare function create_renderer(options: RendererOptions): Renderer;
 //#endregion
 //#region src/phases.d.ts
 /** How the check runs when the caller sets it, as `volley acp` does. Left
- * out, the check inherits volley's environment and runs unasked, as on the
- * CLI. */
+ * out, the check runs unasked with volley's environment less the seats' keys,
+ * as on the CLI. */
 type CheckGate = {
-  /** The whole environment the check runs with, in place of volley's own. */
+  /** The whole environment the check runs with, in place of volley's own less
+   * the seats' keys. */
   env?: Readonly<Record<string, string>>;
   /** Asked on the check's card before each check, the way a model call asks
    * before a gated tool runs. Anything short of an allow keeps the check from