@capybara-hq/kernel 0.1.0-rc.2 → 0.1.0-rc.20

package/dist/evals/sets.service.js

@@ -152,8 +152,8 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
             throw setNotFound();
         return EvalSetsService_1.toView(r);
     }
-    async create(orgId, actor, input, actorTokenId) {
-        const id = await this.createIn(orgId, actor, input, actorTokenId);
+    async create(orgId, actor, input, actorTokenId, auditDetail) {
+        const id = await this.createIn(orgId, actor, input, actorTokenId, auditDetail);
         return this.view(orgId, id);
     }
     async list(orgId) {
@@ -175,22 +175,22 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
             .execute();
         return { ...set, cases: cases.map(toCaseView) };
     }
-    async update(orgId, actor, setId, patch, actorTokenId) {
+    async update(orgId, actor, setId, patch, actorTokenId, auditDetail) {
         if (!UUID.test(setId))
             throw setNotFound();
-        await this.updateIn(orgId, actor, setId, patch, actorTokenId);
+        await this.updateIn(orgId, actor, setId, patch, actorTokenId, auditDetail);
         return this.view(orgId, setId);
     }
     /**
      * 删除评测集（Ruling S1-a）：集合行软删除，留作审计存根；用例（含快照）硬删，结果随之级联删除，不留复制来的对话内容；
      * 在执行的运行取消，运行行保留（无入口）。
      */
-    async remove(orgId, actor, setId, actorTokenId) {
+    async remove(orgId, actor, setId, actorTokenId, auditDetail) {
         if (!UUID.test(setId))
             throw setNotFound();
-        await this.removeIn(orgId, actor, setId, actorTokenId);
+        await this.removeIn(orgId, actor, setId, actorTokenId, auditDetail);
     }
-    async createIn(orgId, actor, input, actorTokenId) {
+    async createIn(orgId, actor, input, actorTokenId, auditDetail) {
         await this.orgLock.lockAndAuthorize(orgId, actor, "evals.manage");
         const agent = await this.db
             .selectFrom("agents")
@@ -219,11 +219,11 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
             orgId,
             action: "eval.set_created",
             target: row.id,
-            detail: { agentId: input.agentId, name: input.name },
+            detail: { agentId: input.agentId, name: input.name, ...auditDetail },
         });
         return row.id;
     }
-    async updateIn(orgId, actor, setId, patch, actorTokenId) {
+    async updateIn(orgId, actor, setId, patch, actorTokenId, auditDetail) {
         await this.orgLock.lockAndAuthorize(orgId, actor, "evals.manage");
         const r = await this.db
             .updateTable("eval_sets")
@@ -248,10 +248,11 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
             // 只记改了哪些字段（评分标准可能很长，不进审计）
             detail: {
                 fields: ["name", "description", "rubric"].filter((f) => patch[f] !== undefined),
+                ...auditDetail,
             },
         });
     }
-    async removeIn(orgId, actor, setId, actorTokenId) {
+    async removeIn(orgId, actor, setId, actorTokenId, auditDetail) {
         await this.orgLock.lockAndAuthorize(orgId, actor, "evals.manage");
         const r = await this.db
             .updateTable("eval_sets")
@@ -269,6 +270,7 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
             orgId,
             action: "eval.set_deleted",
             target: setId,
+            ...(auditDetail !== undefined && { detail: auditDetail }),
         });
     }
     async insertCaseIn(orgId, actor, setId, snapshot, extra, prepared) {
@@ -319,11 +321,12 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
                 conversationId: snapshot.conversationId,
                 runId: snapshot.runId,
                 ...(extra.sampled && { sampled: true }),
+                ...extra.auditDetail,
             },
         });
         return toCaseView(row);
     }
-    async removeCaseIn(orgId, actor, setId, caseId, actorTokenId) {
+    async removeCaseIn(orgId, actor, setId, caseId, actorTokenId, auditDetail) {
         await this.orgLock.lockAndAuthorize(orgId, actor, "evals.manage");
         const set = await this.db
             .selectFrom("eval_sets")
@@ -354,6 +357,7 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
                 caseId,
                 conversationId: r.source_conversation_id,
                 runId: r.source_run_id,
+                ...auditDetail,
             },
         });
     }
@@ -402,7 +406,7 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
      * `eval.agent_mismatch`；已在集合里 → `eval.case_exists`；集合满了 → `eval.too_many_cases`；
      * 运行未终态 / 快照超大 → `eval.run_not_finished` / `eval.case_too_large`（`takeSnapshot`）。
      */
-    async addCase(orgId, actor, setId, input, actorTokenId) {
+    async addCase(orgId, actor, setId, input, actorTokenId, auditDetail) {
         const set = await this.liveSet(orgId, setId);
         const conv = await this.db
             .selectFrom("conversations")
@@ -434,6 +438,7 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
         const r = await this.insertCase(orgId, actor, setId, snapshot, {
             sampled: false,
             ...(actorTokenId !== undefined && { actorTokenId }),
+            ...(auditDetail !== undefined && { auditDetail }),
         });
         if (r === "exists")
             throw caseExists();
@@ -446,7 +451,7 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
      * `min(count, 剩余容量)` 个，逐个取快照写入。快照超大（或其间被删除 / 写入时已满 / 并发加入）的计入 `skipped`。
      * 集合已满 → `eval.too_many_cases`。逐个写入，内存里一次只有一份快照。
      */
-    async sample(orgId, actor, setId, input, actorTokenId) {
+    async sample(orgId, actor, setId, input, actorTokenId, auditDetail) {
         const set = await this.liveSet(orgId, setId);
         const { maxCasesPerSet, maxCaseBytes } = this.config.get().evals;
         const room = maxCasesPerSet - (await this.caseCount(this.db, setId));
@@ -497,6 +502,7 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
             const r = await this.insertCase(orgId, actor, setId, snapshot, {
                 sampled: true,
                 ...(actorTokenId !== undefined && { actorTokenId }),
+                ...(auditDetail !== undefined && { auditDetail }),
             });
             if (typeof r === "string")
                 skipped++;
@@ -505,18 +511,18 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
         }
         return { added, skipped };
     }
-    async removeCase(orgId, actor, setId, caseId, actorTokenId) {
+    async removeCase(orgId, actor, setId, caseId, actorTokenId, auditDetail) {
         if (!UUID.test(setId))
             throw setNotFound();
         if (!UUID.test(caseId))
             throw caseNotFound();
-        await this.removeCaseIn(orgId, actor, setId, caseId, actorTokenId);
+        await this.removeCaseIn(orgId, actor, setId, caseId, actorTokenId, auditDetail);
     }
     /**
      * 用例详情：由快照还原轨迹（原对话删除后照常可读）。快照即对话内容，每次查看写审计 `trajectory.viewed`
      * （target = 原对话 id，detail 带 `caseId` / `setId`）。
      */
-    async getCase(orgId, actor, setId, caseId, actorTokenId) {
+    async getCase(orgId, actor, setId, caseId, actorTokenId, auditDetail) {
         await this.liveSet(orgId, setId);
         if (!UUID.test(caseId))
             throw caseNotFound();
@@ -537,7 +543,7 @@ let EvalSetsService = EvalSetsService_1 = class EvalSetsService {
             orgId,
             action: "trajectory.viewed",
             target: row.source_conversation_id,
-            detail: { agentId: snapshot.agentId, caseId, setId },
+            detail: { agentId: snapshot.agentId, caseId, setId, ...auditDetail },
         });
         return {
             case: toCaseView(row),

package/dist/evals/snapshot.d.ts

@@ -59,8 +59,9 @@ export interface CaseSnapshot {
  */
 export declare function takeSnapshot(db: KernelSchemaDb, orgId: string, conversationId: string, runId: string, maxBytes: number): Promise<CaseSnapshot>;
 /**
- * 读出库里存的快照（`eval_cases.snapshot`）：把各版本定义里改名前（D17）的 `{ kind: "app" }` 工具引用规范化为
- * connector 形（C27）。整轮回放会把定义交给实时注册表，不规范化的话连接器工具全部失效。其余字段原样。
+ * Reads a stored snapshot (`eval_cases.snapshot`): each version's definition goes through
+ * `normalizeStoredDefinition` (rc tool keys, and an enabled tool list read as nothing turned off).
+ * Other fields as they are.
  */
 export declare function readSnapshot(raw: unknown): CaseSnapshot;
 /** 快照事件还原为日志事件（`createdAt` 转回 Date）。 */

package/dist/evals/snapshot.js

@@ -97,7 +97,7 @@ export async function takeSnapshot(db, orgId, conversationId, runId, maxBytes) {
         .orderBy("version")
         .execute();
     for (const r of defRows) {
-        definitions[String(r.version)] = normalizeStoredDefinition(r.definition);
+        definitions[String(r.version)] = snapshotDefinition(r.definition);
     }
     const toolSpecs = {};
     if (hashes.size > 0) {
@@ -151,14 +151,27 @@ export async function takeSnapshot(db, orgId, conversationId, runId, maxBytes) {
     return snapshot;
 }
 /**
- * 读出库里存的快照（`eval_cases.snapshot`）：把各版本定义里改名前（D17）的 `{ kind: "app" }` 工具引用规范化为
- * connector 形（C27）。整轮回放会把定义交给实时注册表，不规范化的话连接器工具全部失效。其余字段原样。
+ * A definition as a snapshot holds it: normalized (`normalizeStoredDefinition`), and one from
+ * before tools were on by default keeps its enabled list next to it: rc tool keys are read by its
+ * extension refs (`legacyScopeOf`); the runtime ignores it (tools default-on spec §2).
+ */
+function snapshotDefinition(d) {
+    const enabled = d?.tools;
+    return {
+        ...normalizeStoredDefinition(d),
+        ...(Array.isArray(enabled) && { tools: enabled }),
+    };
+}
+/**
+ * Reads a stored snapshot (`eval_cases.snapshot`): each version's definition goes through
+ * `normalizeStoredDefinition` (rc tool keys, and an enabled tool list read as nothing turned off).
+ * Other fields as they are.
  */
 export function readSnapshot(raw) {
     const s = raw;
     const definitions = {};
     for (const [v, d] of Object.entries(s.definitions ?? {})) {
-        definitions[v] = normalizeStoredDefinition(d);
+        definitions[v] = snapshotDefinition(d);
     }
     return {
         ...s,

package/dist/evals/step-replay.d.ts

@@ -4,7 +4,7 @@ import { type Selection } from "../conversations/selection.js";
 import type { KernelConfigService } from "../infra/config/kernel-config.service.js";
 import { type MediaSourceDeps } from "../runtime/model-call.js";
 import type { ProviderRegistry } from "../runtime/providers.js";
-import { type DriftReason } from "./reconstruct.js";
+import { type DriftReason, type SubagentTemplateRef } from "./reconstruct.js";
 import type { CaseSnapshot } from "./snapshot.js";
 /**
  * 单步回放（D2 / C5）：由用例快照还原某一步的模型请求，换成候选（agent 版本 / 模型 / 参数 / 指令）重发，
@@ -33,10 +33,7 @@ export interface ResolvedCandidate {
      * A plugin template's child case: the template's current version (set by the worker), compared
      * with the logged one for `subagent_changed` when the candidate changed.
      */
-    subagentTemplate?: {
-        id: string;
-        version: number;
-    };
+    subagentTemplate?: SubagentTemplateRef;
     /**
      * A child case under another agent version: that version's own definition (with the candidate's
      * instructions), compared with the logged sub-agent for `subagent_changed`. Never rendered.
@@ -95,7 +92,8 @@ export declare class ReplayAttemptError extends Error {
  * 否则带候选的定义、版本与参数——参数按候选模型的声明宽松处理，与运行时一致），用 `consumeModelStream` 收完。
  * 请求发出之前的错误原样抛出（`agent.model_unavailable`、取附件暂时失败为 `MediaUnavailableError`）；请求发出之后的
  * 错误包成 {@link ReplayAttemptError}（带估算用量，调用方计入预算后按 `cause` 分类重试）。
- * 模型按部署范围解析（与运行时一致；组织级模型准入若在第二段落地，须同样覆盖评测候选）。
+ * The candidate model resolves in the snapshot's org like the runtime (its effective provider
+ * config and disabled lists: a model disabled there is `agent.model_unavailable`, spec B §5.3).
  * 不另设单次请求的墙钟时限：与运行时一样只有流空闲看门狗（`streamIdleTimeoutMs` / `runtime.streamIdleMs`）。
  */
 export declare function replayStep(d: StepReplayDeps, s: CaseSnapshot, call: {

package/dist/evals/step-replay.js

@@ -2,7 +2,7 @@ import { sanitizeParams, } from "@capybara-hq/plugin-sdk";
 import { effectiveDefinition, NO_SELECTION, sameModel, } from "../conversations/selection.js";
 import { estimateTokens } from "../runtime/context/tokens.js";
 import { consumeModelStream, MediaUnavailableError, } from "../runtime/model-call.js";
-import { loadSnapshotMedia, reconstructRequest } from "./reconstruct.js";
+import { loadSnapshotMedia, reconstructRequest, } from "./reconstruct.js";
 /** 键排序后的 JSON（比较参数是否相同）。 */
 const canonical = (v) => JSON.stringify(Object.keys(v)
     .sort()
@@ -54,10 +54,9 @@ subagent) {
                 ? { ...(from.modelParams ?? {}) }
                 : {});
     const definition = {
-        ...from,
+        ...withInstructions(from, c.instructions),
         model,
         modelParams: params,
-        ...(c.instructions !== undefined && { instructions: c.instructions }),
     };
     const changed = agentVersion !== base.agentVersion ||
         !sameModel(model, base.definition.model) ||
@@ -73,13 +72,20 @@ subagent) {
         changed,
         overridesSwitch,
         ...(otherVersion && {
-            subagentVersion: {
-                ...versionDef,
-                ...(c.instructions !== undefined && { instructions: c.instructions }),
-            },
+            subagentVersion: withInstructions(versionDef, c.instructions),
         }),
     };
 }
+/**
+ * A candidate's `instructions` override: the prompt files it was joined from no longer describe
+ * it, so they are dropped (never model-visible either way).
+ */
+function withInstructions(def, instructions) {
+    if (instructions === undefined)
+        return def;
+    const { instructionFiles: _files, ...rest } = def;
+    return { ...rest, instructions };
+}
 /**
  * 候选请求已经发出、但没有正常收完（出错、超时、中止）：`tokens` 是这次尝试应计入预算的估算用量，
  * `cause` 是原始错误（调用方按它分类重试）。请求发出之前的失败（解析模型、取附件）原样抛出，不计费。
@@ -120,12 +126,13 @@ function baselineOf(s, call) {
  * 否则带候选的定义、版本与参数——参数按候选模型的声明宽松处理，与运行时一致），用 `consumeModelStream` 收完。
  * 请求发出之前的错误原样抛出（`agent.model_unavailable`、取附件暂时失败为 `MediaUnavailableError`）；请求发出之后的
  * 错误包成 {@link ReplayAttemptError}（带估算用量，调用方计入预算后按 `cause` 分类重试）。
- * 模型按部署范围解析（与运行时一致；组织级模型准入若在第二段落地，须同样覆盖评测候选）。
+ * The candidate model resolves in the snapshot's org like the runtime (its effective provider
+ * config and disabled lists: a model disabled there is `agent.model_unavailable`, spec B §5.3).
  * 不另设单次请求的墙钟时限：与运行时一样只有流空闲看门狗（`streamIdleTimeoutMs` / `runtime.streamIdleMs`）。
  */
 export async function replayStep(d, s, call, candidate, signal) {
     const baseline = baselineOf(s, call);
-    const { provider, info } = await d.registry.resolve(candidate.model);
+    const { provider, info } = await d.registry.resolve(s.orgId, candidate.model);
     let media;
     if (d.media) {
         try {

package/dist/evals/trajectory.d.ts

@@ -54,17 +54,19 @@ export interface TrajectoryToolCall {
         byAccountId: string | null;
         decidedAt: string | null;
     } | null;
-    /** 最近一次输入请求（`ask_user`、插件 `input_required`、MCP elicitation）。 */
+    /** 最近一次输入请求（`AskUserQuestion`、插件 `input_required`、MCP elicitation）。 */
     question: {
         questionId: string;
         askedAt: string;
-        outcome: "answered" | "expired" | "cancelled" | null;
+        outcome: "answered" | "skipped" | "expired" | "cancelled" | null;
         byAccountId: string | null;
+        /** A skipped question: the seq of the user message that skipped it. */
+        skippedBySeq?: number;
     } | null;
     /** A `tool.preExecute` policy hook denied the call (it never started). */
     hookDenied?: TrajectoryHookDenial;
     /**
-     * An `agent` call that dispatched (or resumed) a sub-agent (phase C, spec §15.4): the child
+     * An `Agent` call that dispatched (or resumed) a sub-agent (phase C, spec §15.4): the child
      * conversation (`subagentId`, whose trajectory the viewer links to), the run it started and that
      * run's end (`running` until its `subagent_finished`).
      */

package/dist/evals/trajectory.js

@@ -130,7 +130,7 @@ function buildRun({ started, events }, prices, forks) {
                 assistants.set(e.payload.step, [...(assistants.get(e.payload.step) ?? []), e]);
                 break;
             case "usage":
-                // a compaction or title call's usage is not a step call's: counted in the totals only (I-1)
+                // a compaction, title or web_fetch call's usage is not a step call's: counted in the totals only (I-1)
                 if (e.payload.purpose !== undefined) {
                     compactions.push(e);
                     break;
@@ -281,6 +281,9 @@ function buildRun({ started, events }, prices, forks) {
                         askedAt: iso(q.createdAt),
                         outcome: ans?.payload.outcome ?? null,
                         byAccountId: ans?.payload.byAccountId ?? null,
+                        ...(ans?.payload.skippedBySeq !== undefined && {
+                            skippedBySeq: ans.payload.skippedBySeq,
+                        }),
                     },
             };
         });

package/dist/evals/trajectory.service.d.ts

@@ -58,7 +58,7 @@ export type RunStatus = (typeof RUN_STATUSES)[number];
 /**
  * 轨迹查看器（D6、C1）：只给组织管理员（admin / owner）与 agent 负责人。负责人与角色每次实时读（R5）；
  * 无权查看与不存在一律 `trajectory.not_found`，不泄露对话是否存在。每次查看详情写审计 `trajectory.viewed`。
- * 企业管理员不因身份而能看（不是组织成员时 OrgGuard 已 404）。
+ * 超级管理员不因身份而能看（不是组织成员时 OrgGuard 已 404）。
  *
  * 列表按 `conversations.kind` 过滤，不传时只列 `chat`（评测回放的隐藏对话用 `kind: "eval"` 列出）。
  * 评测对话（`kind = 'eval'`）只给组织管理员（Ruling S1-c）：里面是管理员的候选指令，非管理员的 agent 负责人
@@ -98,8 +98,13 @@ export declare class TrajectoryService {
     get(orgId: string, accountId: string, conversationId: string, actorTokenId?: string, window?: RunWindow): Promise<TrajectoryDetail>;
     /** The parent call that first dispatched a sub-agent conversation. */
     private firstDispatch;
-    /** The agent of a sub-agent conversation's parent conversation, or null. */
-    private parentAgentId;
+    /**
+     * The agent of the root conversation of a sub-agent conversation's tree, from its parent `parentId`
+     * up (the same walk as `conversations/access.ts`: at most 16 levels, every conversation between
+     * a sub-agent conversation of the org, ending at one without a parent), or null when the chain
+     * leaves the org, passes another kind or does not end.
+     */
+    private rootAgentId;
     /**
      * 窗口内已开始（`started_at` 非空，与轨迹只列有 `run_started` 的运行一致）的运行，按 (created_at, id) 倒序取
      * `limit` 次；`before` 是上一页的 `next`（运行 id）。`next`：还有更早的运行时为本页最早一次的 id。

package/dist/evals/trajectory.service.js

@@ -34,7 +34,7 @@ const notFound = () => new ApiError("trajectory.not_found", "Trajectory not foun
 /**
  * 轨迹查看器（D6、C1）：只给组织管理员（admin / owner）与 agent 负责人。负责人与角色每次实时读（R5）；
  * 无权查看与不存在一律 `trajectory.not_found`，不泄露对话是否存在。每次查看详情写审计 `trajectory.viewed`。
- * 企业管理员不因身份而能看（不是组织成员时 OrgGuard 已 404）。
+ * 超级管理员不因身份而能看（不是组织成员时 OrgGuard 已 404）。
  *
  * 列表按 `conversations.kind` 过滤，不传时只列 `chat`（评测回放的隐藏对话用 `kind: "eval"` 列出）。
  * 评测对话（`kind = 'eval'`）只给组织管理员（Ruling S1-c）：里面是管理员的候选指令，非管理员的 agent 负责人
@@ -180,11 +180,13 @@ let TrajectoryService = TrajectoryService_1 = class TrajectoryService {
             .where("id", "=", conversationId)
             .where("org_id", "=", orgId)
             .executeTakeFirst();
-        // A sub-agent's trajectory is part of its parent conversation's: authorized by the parent
-        // conversation's agent, never by the dispatched agent (an org agent's owner must not read a
-        // child drawn from someone else's conversation; batch 9 review I1)
+        // A sub-agent's trajectory is part of its tree's: authorized by the ROOT conversation's agent
+        // (the rule `parentReachable` applies to the conversation itself), never by the dispatched
+        // agent nor by the direct parent's, which below an org agent's child is that org agent (an org
+        // agent's owner must not read work drawn from someone else's conversation; batch 9 review I1,
+        // agent map final review: depth 2 and below)
         const authAgentId = conv?.kind === "subagent"
-            ? await this.parentAgentId(orgId, conv.parent_conversation_id)
+            ? await this.rootAgentId(orgId, conv.parent_conversation_id)
             : conv?.agent_id;
         if (!conv || authAgentId === null || authAgentId === undefined)
             throw notFound();
@@ -276,17 +278,27 @@ let TrajectoryService = TrajectoryService_1 = class TrajectoryService {
                 callId: row.parent_call_id,
             };
     }
-    /** The agent of a sub-agent conversation's parent conversation, or null. */
-    async parentAgentId(orgId, parentId) {
+    /**
+     * The agent of the root conversation of a sub-agent conversation's tree, from its parent `parentId`
+     * up (the same walk as `conversations/access.ts`: at most 16 levels, every conversation between
+     * a sub-agent conversation of the org, ending at one without a parent), or null when the chain
+     * leaves the org, passes another kind or does not end.
+     */
+    async rootAgentId(orgId, parentId) {
         if (parentId === null)
             return null;
-        const p = await this.db
-            .selectFrom("conversations")
-            .select("agent_id")
-            .where("id", "=", parentId)
-            .where("org_id", "=", orgId)
-            .executeTakeFirst();
-        return p?.agent_id ?? null;
+        const row = await sql `
+      WITH RECURSIVE up(id, parent, kind, agent_id, n) AS (
+        SELECT c.id, c.parent_conversation_id, c.kind, c.agent_id, 0
+        FROM capybara.conversations c
+        WHERE c.id = ${parentId} AND c.org_id = ${orgId}
+        UNION ALL
+        SELECT c.id, c.parent_conversation_id, c.kind, c.agent_id, up.n + 1
+        FROM capybara.conversations c JOIN up ON c.id = up.parent
+        WHERE up.kind = 'subagent' AND up.n < 16 AND c.org_id = ${orgId}
+      )
+      SELECT agent_id FROM up WHERE parent IS NULL LIMIT 1`.execute(this.db.db);
+        return row.rows[0]?.agent_id ?? null;
     }
     /**
      * 窗口内已开始（`started_at` 非空，与轨迹只列有 `run_started` 的运行一致）的运行，按 (created_at, id) 倒序取

package/dist/evals/worker.d.ts

@@ -13,6 +13,7 @@ import { RunFinish } from "../runtime/run-finish.js";
 import { AgentLoops, RunSteps } from "../runtime/run-steps.js";
 import { EvalBudget } from "./budget.js";
 import { type EvalJobData } from "./jobs.js";
+import { type SubagentTemplateRef } from "./reconstruct.js";
 import { type EvalReplayHooks, EvalReplaySeeder } from "./run-replay.js";
 import { type CaseSnapshot } from "./snapshot.js";
 /**
@@ -110,9 +111,8 @@ export declare class EvalWorker {
 }
 /**
  * A plugin template's child case: the template's current version, which a new dispatch would use
- * (a template removed since gets version 0, so it always differs). Undefined for other cases.
+ * (a template removed since gets version 0, so it always differs), with the platform template
+ * whose content that is (absent: the org's own; a link or pin change under a copy's id changes
+ * it). Undefined for other cases.
  */
-export declare function currentTemplate(db: KernelSchemaDb, s: CaseSnapshot): Promise<{
-    id: string;
-    version: number;
-} | undefined>;
+export declare function currentTemplate(db: KernelSchemaDb, s: CaseSnapshot): Promise<SubagentTemplateRef | undefined>;

package/dist/evals/worker.js

@@ -34,7 +34,7 @@ import { replyParts, structuralDiff } from "./diff.js";
 import { evalJobId } from "./jobs.js";
 import { JUDGE_PROVIDER_UNAVAILABLE, judge } from "./judge.js";
 import { userRequestOf } from "./plan.js";
-import { subagentCaseChanged, subagentStartedOf } from "./reconstruct.js";
+import { subagentCaseChanged, subagentStartedOf, } from "./reconstruct.js";
 import { deleteEvalConversation, EvalReplaySeeder, ReplayClaimLost, replayRun, runBaseline, } from "./run-replay.js";
 import { caseBase, versionDefinition } from "./runs.service.js";
 import { readSnapshot, replayableCalls, toEvents } from "./snapshot.js";
@@ -878,12 +878,18 @@ EvalWorker = __decorate([
 export { EvalWorker };
 /**
  * A plugin template's child case: the template's current version, which a new dispatch would use
- * (a template removed since gets version 0, so it always differs). Undefined for other cases.
+ * (a template removed since gets version 0, so it always differs), with the platform template
+ * whose content that is (absent: the org's own; a link or pin change under a copy's id changes
+ * it). Undefined for other cases.
  */
 export async function currentTemplate(db, s) {
     const t = subagentStartedOf(toEvents(s))?.template;
     if (t?.source !== "plugin" || t.id === undefined)
         return undefined;
     const row = (await loadTemplates(db, s.orgId, [t.id])).get(t.id);
-    return { id: t.id, version: row?.version ?? 0 };
+    return {
+        id: t.id,
+        version: row?.version ?? 0,
+        ...(row?.platformTemplateId !== undefined && { platformTemplateId: row.platformTemplateId }),
+    };
 }