@oneharness/sdk 0.10.2 → 0.11.0

package/dist/generated/history-stream-envelope.d.ts

@@ -14,7 +14,7 @@ export type HistoryStreamEnvelope = {
         [k: string]: unknown;
     } | ({
         error: string;
-        schema_version?: "1.3" | "1.4" | "1.5" | "1.6" | undefined;
+        schema_version?: "1.3" | "1.4" | "1.5" | "1.6" | "1.7" | undefined;
         [k: string]: unknown;
     } & ({
         status?: "nonzero" | "timeout" | "cancelled" | "spawn-error" | "skipped" | undefined;
@@ -23,22 +23,30 @@ export type HistoryStreamEnvelope = {
         failure_kind: "tool_deferred";
         [k: string]: unknown;
     }))) & ({
+        work?: null | undefined;
+        [k: string]: unknown;
+    } | {
+        schema_version?: "1.7" | undefined;
+        status?: "nonzero" | "timeout" | "cancelled" | undefined;
+        work: "done" | "none";
+        [k: string]: unknown;
+    }) & ({
         status?: "ok" | "nonzero" | "timeout" | "spawn-error" | "skipped" | "planned" | undefined;
         [k: string]: unknown;
     } | {
-        schema_version?: "1.4" | "1.5" | "1.6" | undefined;
+        schema_version?: "1.4" | "1.5" | "1.6" | "1.7" | undefined;
         [k: string]: unknown;
     }) & (({
         failure_kind?: "auth" | "rate_limit" | "model_not_found" | "quota" | "tool_deferred" | "untrusted_directory" | "input_too_large" | null;
         [k: string]: unknown;
     } | {
-        schema_version?: "1.5" | "1.6" | undefined;
+        schema_version?: "1.5" | "1.6" | "1.7" | undefined;
         [k: string]: unknown;
     }) & ({
         failure_kind?: "auth" | "rate_limit" | "model_not_found" | "quota" | "session_not_found" | "tool_deferred" | null;
         [k: string]: unknown;
     } | {
-        schema_version?: "1.6" | undefined;
+        schema_version?: "1.6" | "1.7" | undefined;
         [k: string]: unknown;
     })) & HistoryRecord;
     type: "record";
@@ -143,7 +151,7 @@ export type HistoryRecord = {
      * run's single prompt).
      */
     prompt: string;
-    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6";
+    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6" | "1.7";
     /**
      * The oneharness session id this run belongs to (the history file's stem).
      */
@@ -170,6 +178,22 @@ export type HistoryRecord = {
     tool_ms: number;
     usage: Usage;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number;
@@ -296,6 +320,22 @@ export type HistoryRecord = {
     tool_ms: number;
     usage: Usage1;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number | null;
@@ -373,7 +413,7 @@ export type HistoryRecord = {
      * run's single prompt).
      */
     prompt: string;
-    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6";
+    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6" | "1.7";
     /**
      * The oneharness session id this run belongs to (the history file's stem).
      */
@@ -400,6 +440,22 @@ export type HistoryRecord = {
     tool_ms?: never | undefined;
     usage: Usage2;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number;
@@ -465,7 +521,7 @@ export type HistoryRecord = {
      * run's single prompt).
      */
     prompt: string;
-    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6";
+    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6" | "1.7";
     /**
      * The oneharness session id this run belongs to (the history file's stem).
      */
@@ -492,6 +548,22 @@ export type HistoryRecord = {
     tool_ms?: never | undefined;
     usage: Usage3;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number | null;
@@ -557,7 +629,7 @@ export type HistoryRecord = {
      * run's single prompt).
      */
     prompt: string;
-    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6";
+    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6" | "1.7";
     /**
      * The oneharness session id this run belongs to (the history file's stem).
      */
@@ -584,6 +656,22 @@ export type HistoryRecord = {
     tool_ms?: never | undefined;
     usage: Usage4;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number;
@@ -649,7 +737,7 @@ export type HistoryRecord = {
      * run's single prompt).
      */
     prompt: string;
-    schema_version: "1.3" | "1.4" | "1.5" | "1.6";
+    schema_version: "1.3" | "1.4" | "1.5" | "1.6" | "1.7";
     /**
      * The oneharness session id this run belongs to (the history file's stem).
      */
@@ -676,6 +764,22 @@ export type HistoryRecord = {
     tool_ms?: never | undefined;
     usage: Usage5;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number | null;
@@ -771,6 +875,22 @@ export type HistoryRecord = {
     tool_ms?: never | undefined;
     usage: Usage6;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number | null;
@@ -878,6 +998,22 @@ export type HistoryRecord = {
     tool_ms?: never | undefined;
     usage: Usage7;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number | null;
@@ -983,6 +1119,22 @@ export type HistoryRecord = {
     tool_ms?: number | null | undefined;
     usage: Usage8;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 };
 export type ToolCallStatus = "completed" | "failed" | "timeout" | "interrupted";
@@ -1006,12 +1158,32 @@ export type FailureKind = "auth" | "rate_limit" | "model_not_found" | "quota" |
  * The outcome of attempting to run one harness.
  */
 export type Status = "ok" | "nonzero" | "timeout" | "cancelled" | "spawn-error" | "skipped" | "planned";
+/**
+ * Whether a candidate's normalized result carries **evidence it did the task's
+ * work** — the first thing [`startup_failure_reason`] consults.
+ *
+ * Two independent witnesses, either of which is decisive:
+ *
+ * - **Tool events.** A recorded tool call is the harness acting on the task.
+ * - **Usage accounting.** [`Usage::reports_billed_work`][billed] — the same definition
+ *   `signals::record_reports_work` classifies a raw harness record with, so the
+ *   two readings of "billed" are one contract with one implementation.
+ *
+ * It is also a **published reading**, not only an internal one: a run that
+ * failed with nothing to classify carries it as [`RunResult::work`] and in its
+ * history record, because there the question "did this candidate do anything?"
+ * is the only one left. One type for both, so the value a reader sees is the
+ * same value the fall-through verdict consulted.
+ *
+ * [billed]: crate::domain::signals::Usage::reports_billed_work
+ */
+export type RunWork = "done" | "none";
 export type HistoryEventLine = {
     event: ActionEvent;
     harness: string;
     harness_id?: string | null | undefined;
     run_id: string;
-    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6";
+    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6" | "1.7";
     variant?: string | null | undefined;
     [k: string]: unknown;
 } | {

package/dist/generated/history.d.ts

@@ -8,7 +8,7 @@ export type HistoryRecord = ({
     [k: string]: unknown;
 } | ({
     error: string;
-    schema_version?: "1.3" | "1.4" | "1.5" | "1.6" | undefined;
+    schema_version?: "1.3" | "1.4" | "1.5" | "1.6" | "1.7" | undefined;
     [k: string]: unknown;
 } & ({
     status?: "nonzero" | "timeout" | "cancelled" | "spawn-error" | "skipped" | undefined;
@@ -17,22 +17,30 @@ export type HistoryRecord = ({
     failure_kind: "tool_deferred";
     [k: string]: unknown;
 }))) & ({
+    work?: null | undefined;
+    [k: string]: unknown;
+} | {
+    schema_version?: "1.7" | undefined;
+    status?: "nonzero" | "timeout" | "cancelled" | undefined;
+    work: "done" | "none";
+    [k: string]: unknown;
+}) & ({
     status?: "ok" | "nonzero" | "timeout" | "spawn-error" | "skipped" | "planned" | undefined;
     [k: string]: unknown;
 } | {
-    schema_version?: "1.4" | "1.5" | "1.6" | undefined;
+    schema_version?: "1.4" | "1.5" | "1.6" | "1.7" | undefined;
     [k: string]: unknown;
 }) & (({
     failure_kind?: "auth" | "rate_limit" | "model_not_found" | "quota" | "tool_deferred" | "untrusted_directory" | "input_too_large" | null;
     [k: string]: unknown;
 } | {
-    schema_version?: "1.5" | "1.6" | undefined;
+    schema_version?: "1.5" | "1.6" | "1.7" | undefined;
     [k: string]: unknown;
 }) & ({
     failure_kind?: "auth" | "rate_limit" | "model_not_found" | "quota" | "session_not_found" | "tool_deferred" | null;
     [k: string]: unknown;
 } | {
-    schema_version?: "1.6" | undefined;
+    schema_version?: "1.6" | "1.7" | undefined;
     [k: string]: unknown;
 })) & ({
     duration_ms: number;
@@ -129,7 +137,7 @@ export type HistoryRecord = ({
      * run's single prompt).
      */
     prompt: string;
-    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6";
+    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6" | "1.7";
     /**
      * The oneharness session id this run belongs to (the history file's stem).
      */
@@ -156,6 +164,22 @@ export type HistoryRecord = ({
     tool_ms: number;
     usage: Usage;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number;
@@ -282,6 +306,22 @@ export type HistoryRecord = ({
     tool_ms: number;
     usage: Usage1;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number | null;
@@ -359,7 +399,7 @@ export type HistoryRecord = ({
      * run's single prompt).
      */
     prompt: string;
-    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6";
+    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6" | "1.7";
     /**
      * The oneharness session id this run belongs to (the history file's stem).
      */
@@ -386,6 +426,22 @@ export type HistoryRecord = ({
     tool_ms?: never | undefined;
     usage: Usage2;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number;
@@ -451,7 +507,7 @@ export type HistoryRecord = ({
      * run's single prompt).
      */
     prompt: string;
-    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6";
+    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6" | "1.7";
     /**
      * The oneharness session id this run belongs to (the history file's stem).
      */
@@ -478,6 +534,22 @@ export type HistoryRecord = ({
     tool_ms?: never | undefined;
     usage: Usage3;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number | null;
@@ -543,7 +615,7 @@ export type HistoryRecord = ({
      * run's single prompt).
      */
     prompt: string;
-    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6";
+    schema_version: "1.2" | "1.3" | "1.4" | "1.5" | "1.6" | "1.7";
     /**
      * The oneharness session id this run belongs to (the history file's stem).
      */
@@ -570,6 +642,22 @@ export type HistoryRecord = ({
     tool_ms?: never | undefined;
     usage: Usage4;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number;
@@ -635,7 +723,7 @@ export type HistoryRecord = ({
      * run's single prompt).
      */
     prompt: string;
-    schema_version: "1.3" | "1.4" | "1.5" | "1.6";
+    schema_version: "1.3" | "1.4" | "1.5" | "1.6" | "1.7";
     /**
      * The oneharness session id this run belongs to (the history file's stem).
      */
@@ -662,6 +750,22 @@ export type HistoryRecord = ({
     tool_ms?: never | undefined;
     usage: Usage5;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number | null;
@@ -757,6 +861,22 @@ export type HistoryRecord = ({
     tool_ms?: never | undefined;
     usage: Usage6;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number | null;
@@ -864,6 +984,22 @@ export type HistoryRecord = ({
     tool_ms?: never | undefined;
     usage: Usage7;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 } | {
     duration_ms: number | null;
@@ -969,6 +1105,22 @@ export type HistoryRecord = ({
     tool_ms?: number | null | undefined;
     usage: Usage8;
     variant?: string | null | undefined;
+    /**
+     * What a run that failed with **nothing to classify** has to show for
+     * itself: [`RunWork::Done`] where the harness recorded a tool call or billed
+     * usage, [`RunWork::None`] where nothing says it got that far. Copied from
+     * [`RunResult::work`] — one value with one source, like every signal here —
+     * and omitted on the wire on every other run, since a success raises the
+     * question and a classified failure has already answered it.
+     *
+     * This is the field that tells apart, for a reader of the record, the two
+     * things an unexplained failure can be: a candidate that never got started,
+     * and one that ran the task and lost. Without it both are a `nonzero` line with a
+     * null `failure_kind` and empty accounting, and a chain that stopped at the
+     * first reads exactly like a chain that stopped at the second. Gated to
+     * [`FIRST_WORK_EVIDENCE_SCHEMA_VERSION`].
+     */
+    work?: RunWork | null | undefined;
     [k: string]: unknown;
 });
 export type ToolCallStatus = "completed" | "failed" | "timeout" | "interrupted";
@@ -992,6 +1144,26 @@ export type FailureKind = "auth" | "rate_limit" | "model_not_found" | "quota" |
  * The outcome of attempting to run one harness.
  */
 export type Status = "ok" | "nonzero" | "timeout" | "cancelled" | "spawn-error" | "skipped" | "planned";
+/**
+ * Whether a candidate's normalized result carries **evidence it did the task's
+ * work** — the first thing [`startup_failure_reason`] consults.
+ *
+ * Two independent witnesses, either of which is decisive:
+ *
+ * - **Tool events.** A recorded tool call is the harness acting on the task.
+ * - **Usage accounting.** [`Usage::reports_billed_work`][billed] — the same definition
+ *   `signals::record_reports_work` classifies a raw harness record with, so the
+ *   two readings of "billed" are one contract with one implementation.
+ *
+ * It is also a **published reading**, not only an internal one: a run that
+ * failed with nothing to classify carries it as [`RunResult::work`] and in its
+ * history record, because there the question "did this candidate do anything?"
+ * is the only one left. One type for both, so the value a reader sees is the
+ * same value the fall-through verdict consulted.
+ *
+ * [billed]: crate::domain::signals::Usage::reports_billed_work
+ */
+export type RunWork = "done" | "none";
 /**
  * One normalized action a harness took, harness-agnostic so a single consumer
  * assertion works across harnesses. Every field is always serialized (null when