@cat-factory/executor-harness 1.131.0 → 1.132.3

package/dist/pi.d.ts

@@ -314,6 +314,21 @@ export interface HarnessCallMetric {
      * turn. `CliInlineLanguageModel`'s step-level row is the same idea on the inline path.
      */
     standsForJob?: boolean;
+    /**
+     * This row carries only TOKENS and stands for no model call of its own, so the backend counts it
+     * in every token sum and in no `calls` figure.
+     *
+     * Separate from {@link standsForJob}, which is only about the missing TURN ordinal. The
+     * shortfall row occupies no turn either way, but when the CLI narrated NO turns at all it is the
+     * job's only record and IS the call — nothing else recorded it, and a count excluding it would
+     * report a step that burned tokens across zero calls. The producer is the only layer that can
+     * tell the two apart: the backend sees one BATCH of a job's calls at a time (the live drain
+     * splits them across polls), so it cannot ask whether any turn was narrated.
+     *
+     * Absent on every real turn. `CliInlineLanguageModel.fileUnaccounted` makes the same split on
+     * the inline path, off its own `reported > 0`.
+     */
+    spendOnly?: boolean;
 }
 /**
  * Publish one captured model call: append it to the run's list (which becomes the terminal

package/dist/usage-attribution.d.ts

@@ -49,6 +49,14 @@ export declare function claudeUsage(raw: unknown): {
  *
  * Clamped at 0 per side: a CLI whose terminal figure is LOWER than its own per-turn sum has
  * reported the two inconsistently, and negative spend is not a thing to record.
+ *
+ * {@link HarnessCallMetric.spendOnly} is decided HERE, and it is not the same question as
+ * `standsForJob`. The row never occupies a turn, but whether it is a CALL depends on whether any
+ * turn was narrated beside it: with costed turns present this only corrects THEIR under-reporting
+ * (counting it would report one phantom call per dispatch), while with none it is the job's ONLY
+ * record and excluding it would report a step that spent tokens across zero calls. Only this
+ * function can answer that — the backend records a job's calls in BATCHES as the live drain
+ * delivers them, so a batch holding just this row cannot tell the two cases apart.
  */
 export declare function unaccountedUsageCall(parentCalls: readonly HarnessCallMetric[], usage: {
     inputTokens: number;

package/dist/usage-attribution.js

@@ -62,6 +62,14 @@ export function claudeUsage(raw) {
  *
  * Clamped at 0 per side: a CLI whose terminal figure is LOWER than its own per-turn sum has
  * reported the two inconsistently, and negative spend is not a thing to record.
+ *
+ * {@link HarnessCallMetric.spendOnly} is decided HERE, and it is not the same question as
+ * `standsForJob`. The row never occupies a turn, but whether it is a CALL depends on whether any
+ * turn was narrated beside it: with costed turns present this only corrects THEIR under-reporting
+ * (counting it would report one phantom call per dispatch), while with none it is the job's ONLY
+ * record and excluding it would report a step that spent tokens across zero calls. Only this
+ * function can answer that — the backend records a job's calls in BATCHES as the live drain
+ * delivers them, so a batch holding just this row cannot tell the two cases apart.
  */
 export function unaccountedUsageCall(parentCalls, usage) {
     if (!usage)
@@ -92,5 +100,6 @@ export function unaccountedUsageCall(parentCalls, usage) {
         outputTokens,
         finishReason: null,
         standsForJob: true,
+        spendOnly: parentCalls.length > 0,
     };
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cat-factory/executor-harness",
-  "version": "1.131.0",
+  "version": "1.132.3",
   "description": "Container payload: a thin TypeScript wrapper that runs the Pi coding agent against a cloned repo and opens a PR. Runs in the Cloudflare Container (and, in local native mode, as a host process); carries no secrets.",
   "repository": {
     "type": "git",
@@ -25,12 +25,12 @@
     "access": "public"
   },
   "devDependencies": {
-    "@cat-factory/kernel": "0.318.0",
-    "@cat-factory/server": "0.305.0",
-    "@cat-factory/spend": "0.16.12",
+    "@cat-factory/kernel": "0.320.0",
+    "@cat-factory/server": "0.306.4",
+    "@cat-factory/spend": "0.16.16",
     "@hono/node-server": "^2.1.1",
     "@types/node": "^26.2.0",
-    "hono": "^4.13.3",
+    "hono": "^4.13.4",
     "typescript": "7.0.2",
     "vitest": "^4.1.11"
   },

package/src/pi.ts

@@ -652,6 +652,21 @@ export interface HarnessCallMetric {
    * turn. `CliInlineLanguageModel`'s step-level row is the same idea on the inline path.
    */
   standsForJob?: boolean
+  /**
+   * This row carries only TOKENS and stands for no model call of its own, so the backend counts it
+   * in every token sum and in no `calls` figure.
+   *
+   * Separate from {@link standsForJob}, which is only about the missing TURN ordinal. The
+   * shortfall row occupies no turn either way, but when the CLI narrated NO turns at all it is the
+   * job's only record and IS the call — nothing else recorded it, and a count excluding it would
+   * report a step that burned tokens across zero calls. The producer is the only layer that can
+   * tell the two apart: the backend sees one BATCH of a job's calls at a time (the live drain
+   * splits them across polls), so it cannot ask whether any turn was narrated.
+   *
+   * Absent on every real turn. `CliInlineLanguageModel.fileUnaccounted` makes the same split on
+   * the inline path, off its own `reported > 0`.
+   */
+  spendOnly?: boolean
 }
 
 /**

package/src/usage-attribution.ts

@@ -67,6 +67,14 @@ export function claudeUsage(
  *
  * Clamped at 0 per side: a CLI whose terminal figure is LOWER than its own per-turn sum has
  * reported the two inconsistently, and negative spend is not a thing to record.
+ *
+ * {@link HarnessCallMetric.spendOnly} is decided HERE, and it is not the same question as
+ * `standsForJob`. The row never occupies a turn, but whether it is a CALL depends on whether any
+ * turn was narrated beside it: with costed turns present this only corrects THEIR under-reporting
+ * (counting it would report one phantom call per dispatch), while with none it is the job's ONLY
+ * record and excluding it would report a step that spent tokens across zero calls. Only this
+ * function can answer that — the backend records a job's calls in BATCHES as the live drain
+ * delivers them, so a batch holding just this row cannot tell the two cases apart.
  */
 export function unaccountedUsageCall(
   parentCalls: readonly HarnessCallMetric[],
@@ -98,5 +106,6 @@ export function unaccountedUsageCall(
     outputTokens,
     finishReason: null,
     standsForJob: true,
+    spendOnly: parentCalls.length > 0,
   }
 }