@robmclarty/volley 0.9.0 → 0.10.0

package/dist/index.d.ts

@@ -50,6 +50,7 @@ type VolleyConfig = {
   builder_url?: string;
   builder_max_steps?: number;
   builder_max_tool_calls?: number;
+  builder_max_output_tokens?: number;
   allow_unsandboxed_builder?: boolean;
   critic_model?: string;
   critic_provider?: CriticProvider;
@@ -98,11 +99,21 @@ type ResolvedConfig = {
    * lmstudio, openrouter) takes one. */
   builder_url: string | null;
   builder_max_steps: number;
-  /** How many tool calls a tool-loop builder may make in one turn; fascicle
-   * runs the first that many, in order, and drops the rest. 1 by default,
-   * since some model and runtime pairs return parallel calls badly; raise it
-   * for a pair that handles them. Ignored by the `claude_cli` builder. */
+  /** How many tool calls a tool-loop builder may make in one turn, reads
+   * aside (`read_file`, `search_files`, `list_files` are `read_only` and
+   * don't count); fascicle runs a turn's calls in order up to the first one
+   * past the cap, drops the rest, and tells the model which it dropped. 1 by
+   * default, since some model and runtime pairs return parallel calls badly;
+   * raise it for a pair that handles them. Ignored by the `claude_cli`
+   * builder. */
   builder_max_tool_calls: number;
+  /** The most output tokens a tool-loop builder may generate in one turn,
+   * reasoning included, or null to leave each turn to the provider's own
+   * limit. 16384 by default on a local builder (ollama, lmstudio), null on
+   * openrouter. A turn cut off at it before it made a tool call costs that
+   * 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;
   /** 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;
@@ -176,6 +187,9 @@ type CheckResult = {
   detail: CheckArtifact[];
   /** Raw checkride summary JSON (when the runner is checkride). */
   summary?: unknown;
+  /** checkride's capped excerpt of the failing slots (`.check/digest.md`),
+   * when the check failed and checkride wrote one. */
+  digest?: string;
   /** Captured combined output (when the runner is a shell command). */
   log?: string;
 };
@@ -232,6 +246,10 @@ type PhaseRecord = {
    * counts behind the salvage-rate health metric. */
   tool_calls: number;
   salvaged_tool_calls: number;
+  /** Turns the provider cut off at the output cap before they made a tool
+   * call (`--builder-max-output-tokens`), each of which cost a step and went
+   * on. Set on a builder record when any was; absent otherwise. Additive. */
+  truncated_turns?: number;
   /** Provider-error retries that preceded this phase's successful call
    * — the local-critic degradation ladder's first rung: a stochastic
    * Ollama stream death is retried once before falling back. Set on the critic
@@ -288,6 +306,9 @@ type LoopState = {
   builder_cost_usd: number;
   critic_cost_usd: number;
   check_duration_ms: number;
+  /** How many of the run's builder sessions ended on the step limit
+   * (`max_steps`) rather than calling finish. */
+  builder_max_steps_hits: number;
   iteration_cost_usd: number;
   halt: HaltReason;
   cost_warned: boolean;
@@ -304,6 +325,12 @@ type RunResult = {
   builder_cost_usd: number;
   critic_cost_usd: number;
   check_duration_ms: number;
+  /** How many of the run's builder sessions hit the step limit before they
+   * called finish: work handed to the check and critic unverified. The
+   * warning a run prints for each goes to stderr, which `--json` silences, so
+   * a harness reads it here. Always 0 for a `claude_cli` builder, which has
+   * no step limit, and for a critique. */
+  builder_max_steps_hits: number;
   final_verdict: Verdict | null;
   /** The run branch teardown left behind because this run's work was never
    * integrated: under `--worktree` without `--git` a converged run has no