Render RewardKit reward design and outputs

AGENTS.md

@@ -985,6 +985,16 @@ trials without a persisted `_verifier` summary lazily discover and parse their
 `verifier/ctrf.json` artifact through the already-scoped trial files API; do
 not add an unscoped artifact lookup for this fallback.
 
+RewardKit output follows the same pattern: the drawer renders the reward
+breakdown from the embedded `_rewards` / `_reward_details` summaries in
+`trial.result` (written by `harbor_artifacts.py`), falling back to the scoped
+trial files API for `verifier/reward.json` / `verifier/reward-details.json`
+when the embedded copies are missing (imports) or truncated. The task page's
+Reward design section (`reward-design-card.tsx`) reconstructs the reward
+program client-side from the task's `tests/` TOMLs plus a best-effort static
+scan of Python criteria, enriched with criteria observed in a completed
+trial's breakdown — it executes nothing.
+
 On an experiment page, removing a task always calls the scoped
 `DELETE /experiments/{experiment_id}/tasks/{task_id}` proxy. It unlinks that
 experiment membership and its scoped trials without deleting the task, even

DOCS.md

@@ -782,3 +782,42 @@ dashboard shows those counts as a small passed/total line in the trial
 drawer's summary. Missing, malformed, or oversized CTRF reports are ignored
 and never change the settled `reward`; verifiers without a test report simply
 show no test line.
+
+## Reward Kit rewards (reward.json + reward-details.json)
+
+Tasks whose verifier uses [Harbor's RewardKit](https://www.harborframework.com/docs/rewardkit)
+(`harbor-rewardkit`) report **multiple named rewards** — one score per
+dimension (`tests/<dimension>/`) plus the aggregates defined in
+`tests/reward.toml` — in `/logs/verifier/reward.json`, and a per-criterion
+breakdown (scores, weights, judge reasoning, errors) in
+`/logs/verifier/reward-details.json`.
+
+Oddish captures both:
+
+- The trial's headline scalar (`trial.reward`) stays exactly as before: the
+  `"reward"` key of `reward.json`, or the sole value when only one is
+  reported. **Name your strict aggregate `reward`** in `reward.toml` so
+  pass@k, baseline gates, and the matrix keep working — other names are
+  treated as extra dimensions.
+- The full named-rewards map is persisted under the reserved
+  `trial.result._rewards` key.
+- A bounded summary of the breakdown (long judge reasoning truncated,
+  criterion lists capped) is persisted under `trial.result._reward_details`,
+  with the artifact path of the full document. The complete
+  `reward-details.json` stays with the trial artifacts.
+
+The dashboard renders these in the trial drawer as a reward breakdown —
+each dimension with its criteria, the judge's raw answer and reasoning, and
+the files the judge read — and renders `reward.json` /
+`reward-details.json` as reward trees in the Files and Artifacts tabs. On
+the task page, RewardKit tasks additionally get a **Reward design** section
+reconstructed from `tests/reward.toml`, the judge TOMLs, and a static scan
+of Python criteria files, including a what-if mode that simulates how
+criterion outcomes flow through each aggregation into the final rewards.
+
+Both keys follow the metrics.json rules: `_rewards` and `_reward_details`
+are Oddish-owned (task-authored keys of the same name are dropped), and a
+missing, malformed, or oversized details file is ignored — it never changes
+the settled `reward`. Imported trials rely on the artifact fallback: the
+drawer lazy-loads `verifier/reward.json` and `verifier/reward-details.json`
+through the scoped trial files API when the embedded copies are absent.

frontend/package.json

@@ -56,6 +56,7 @@
     "remark-breaks": "4.0.0",
     "remark-gfm": "4.0.1",
     "shiki": "4.0.2",
+    "smol-toml": "1.4.2",
     "swr": "2.4.1",
     "tailwind-merge": "3.5.0",
     "tailwindcss-animate": "1.0.7",

frontend/src/app/(app)/tasks/[task_id]/task-detail-client.tsx

@@ -70,6 +70,7 @@ import {
   Loader2,
   Star,
 } from "lucide-react";
+import { RewardDesignCard } from "@/components/reward-design-card";
 
 const TaskFilesPanel = dynamic(
   () =>
@@ -1374,6 +1375,14 @@ export function TaskDetailClient({
           error={judgeError}
         />
 
+        {/* RewardKit tasks get their reward program rendered as an
+            explorable design; other tasks render nothing here. */}
+        <RewardDesignCard
+          taskId={task.id}
+          taskVersion={selectedVersion?.version}
+          trials={trialsForVersion}
+        />
+
         <div className="space-y-3">
           <div className="flex items-baseline justify-between">
             <h2 className="font-mono text-[12px] font-semibold tracking-[0.06em] text-[color:var(--paper-ink-2)] uppercase">

frontend/src/components/renderers/file-renderer.tsx

@@ -54,6 +54,11 @@ const JsonRenderer = dynamic(
   { ssr: false, loading: () => <LoadingStub label="Rendering JSON..." /> }
 );
 
+const RewardJsonRenderer = dynamic(
+  () => import("./reward-json-renderer").then((m) => m.RewardJsonRenderer),
+  { ssr: false, loading: () => <LoadingStub label="Rendering rewards..." /> }
+);
+
 function LoadingStub({ label }: { label: string }) {
   return (
     <div className="text-muted-foreground flex h-full items-center justify-center gap-2 p-8">
@@ -75,6 +80,7 @@ type FileRendererKind =
   | "json"
   | "config-json"
   | "result-json"
+  | "reward-json"
   | "diff"
   | "csv"
   | "log"
@@ -155,6 +161,16 @@ function getFileRendererKind(fileName: string): FileRendererKind {
   if (lower.endsWith("/result.json") || lower === "result.json") {
     return "result-json";
   }
+  // RewardKit verifier output (also copied into task artifacts by some
+  // tasks): render as a reward tree rather than raw JSON.
+  if (
+    lower.endsWith("/reward.json") ||
+    lower === "reward.json" ||
+    lower.endsWith("/reward-details.json") ||
+    lower === "reward-details.json"
+  ) {
+    return "reward-json";
+  }
   if (ext === "json") return "json";
   if (ext === "diff" || ext === "patch") return "diff";
 
@@ -269,6 +285,8 @@ export function FileRenderer({
       return <ConfigJsonRenderer content={content ?? ""} />;
     case "result-json":
       return <ResultJsonRenderer content={content ?? ""} />;
+    case "reward-json":
+      return <RewardJsonRenderer content={content ?? ""} />;
     case "diff":
       return <DiffRenderer content={content ?? ""} fileName={fileName} />;
     case "csv": {

frontend/src/components/renderers/reward-json-renderer.tsx

@@ -0,0 +1,40 @@
+"use client";
+
+import { JsonRenderer } from "./json-renderer";
+import { RewardBreakdownView } from "@/components/reward-breakdown-view";
+import {
+  parseRewardDetailsDocument,
+  parseRewardsMap,
+  type RewardBreakdown,
+  type RewardsMap,
+} from "@/lib/reward-kit";
+
+/**
+ * Renders RewardKit verifier output files as a reward tree instead of raw
+ * JSON: `reward.json` (flat named-score map) and `reward-details.json`
+ * (per-criterion breakdown with judge reasoning). Anything that doesn't
+ * parse as either falls back to the plain JSON view.
+ */
+export function RewardJsonRenderer({ content }: { content: string }) {
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(content);
+  } catch {
+    return <JsonRenderer content={content} />;
+  }
+
+  const breakdown: RewardBreakdown | null = parseRewardDetailsDocument(parsed);
+  const rewards: RewardsMap | null = breakdown ? null : parseRewardsMap(parsed);
+
+  if (!breakdown && !rewards) {
+    return <JsonRenderer content={content} />;
+  }
+
+  // A bare reward.json map has no dimension entries to classify against, so
+  // every named score renders in the chip row.
+  return (
+    <div className="p-4">
+      <RewardBreakdownView breakdown={breakdown} rewards={rewards} />
+    </div>
+  );
+}