@capybara-hq/kernel 0.1.0-rc.2 → 0.1.0-rc.4

package/dist/db/kernel-schema.d.ts

@@ -14,6 +14,8 @@ export interface AccountsTable {
     is_enterprise_admin: Generated<boolean>;
     created_at: Ts;
     updated_at: Ts;
+    /** 0064: the org the account last chose as current; a new session starts in it. */
+    last_org_id: ColumnType<string | null, string | null | undefined, string | null>;
 }
 export interface IdentitiesTable {
     source: string;
@@ -149,6 +151,11 @@ export interface AgentsTable {
     org_role: Generated<"admin" | "member" | "guest">;
     current_version: number;
     needs_reassignment: Generated<boolean>;
+    /** 0067: the profile (display data, outside the versions). */
+    profile_avatar: ColumnType<Record<string, unknown> | null, string | null | undefined, string | null>;
+    profile_description: Generated<string>;
+    profile_quick_actions: ColumnType<unknown[], string | undefined, string>;
+    profile_updated_at: Ts;
     created_at: Ts;
     updated_at: Ts;
     deleted_at: TsNullable;
@@ -199,6 +206,10 @@ export interface ConversationsTable {
     queue_paused: ColumnType<boolean, boolean | undefined, boolean>;
     /** 0062: why the queue paused (`QueuePauseReason`); null when not paused or not recorded. */
     queue_pause_reason: ColumnType<string | null, string | null | undefined, string | null>;
+    /** 0065: when the owner archived it (hidden from the default lists); null = not archived. */
+    archived_at: ColumnType<Date | null, Date | string | null | undefined, Date | string | null>;
+    /** 0065: when the owner deleted it (soft delete: hidden everywhere for the owner); null = live. */
+    deleted_at: ColumnType<Date | null, Date | string | null | undefined, Date | string | null>;
 }
 /** 0060: an org's order of file-preview@1 previewers (UI refresh B Task 11). */
 export interface FilePreviewersTable {
@@ -208,6 +219,15 @@ export interface FilePreviewersTable {
     position: number;
     disabled: ColumnType<boolean, boolean | undefined, boolean>;
 }
+/** 0073: an org's default provider of an `org-default` extension point. */
+export interface OrgExtensionPointDefaultsTable {
+    org_id: string;
+    /** The extension point name without the major (`web-search`). */
+    point: string;
+    extension: string;
+    updated_at: Ts;
+    updated_by: string | null;
+}
 /** 0059: a queued message (not in the event log until sent). */
 export interface ConversationQueueTable {
     id: string;
@@ -269,7 +289,7 @@ export interface RunsTable {
     /** 0054: when this run's `subagents` wait expires (set while it waits on sub-agents). */
     wait_expires_at: TsNullable;
 }
-/** 0053: one row per sub-agent dispatch (an `agent` call of a parent run). */
+/** 0053: one row per sub-agent dispatch (an `Agent` call of a parent run). */
 export interface SubagentDispatchesTable {
     parent_run_id: string;
     parent_call_id: string;
@@ -400,7 +420,7 @@ export interface ApprovalsTable {
     /** 0056: the reason given with a denial (null without one). */
     deny_reason: Generated<string | null>;
 }
-/** 0010_tools.sql：ask_user 提问。 */
+/** 0010_tools.sql：AskUserQuestion 提问。 */
 export interface QuestionsTable {
     id: string;
     org_id: string;
@@ -547,9 +567,11 @@ export interface AttachmentsTable {
     expires_at: TsNullable;
     /** 0024：上传用途（技能包不能绑定到消息）。 */
     purpose: ColumnType<AttachmentPurpose, AttachmentPurpose | undefined, AttachmentPurpose>;
+    /** 0068: the agent whose profile uses this avatar upload (set once, when it is saved). */
+    agent_id: ColumnType<string | null, string | null | undefined, string | null>;
 }
-/** 预签名上传的用途：对话附件 / 技能包 zip。 */
-export type AttachmentPurpose = "chat" | "skill-package";
+/** 预签名上传的用途：对话附件 / 技能包 zip / agent avatar image. */
+export type AttachmentPurpose = "chat" | "skill-package" | "agent-avatar";
 /** 技能版本的清单（`skill_versions.manifest`）。 */
 export interface SkillManifest {
     name: string;
@@ -630,6 +652,33 @@ export interface SkillChangeRequestsTable {
     decided_at: TsNullable;
     created_at: Ts;
 }
+/** 0070_background_tasks.sql: a Bash command running in the background in a conversation's sandbox. */
+export interface BackgroundTasksTable {
+    id: string;
+    org_id: string;
+    conversation_id: string;
+    sandbox_conversation_id: string;
+    run_id: string;
+    call_id: string;
+    trigger_account_id: string;
+    description: string;
+    command: string;
+    /** NULL while the launch is reserved but not started (0071). */
+    sid: string | null;
+    output_path: string;
+    status: ColumnType<"running" | "completed" | "failed" | "stopped" | "timed_out" | "lost", string | undefined, string>;
+    lost_reason: ColumnType<string | null, string | null | undefined, string | null>;
+    exit_code: ColumnType<number | null, number | null | undefined, number | null>;
+    started_at: Ts;
+    finished_at: TsNullable;
+    deadline_at: Date | string;
+    /** NULL while the launch is reserved but not started (0071). */
+    sandbox_created_at: Date | string | null;
+    kill_pending: ColumnType<boolean, boolean | undefined, boolean>;
+    read_offset: ColumnType<string, number | string | undefined, number | string>;
+    /** Set while a stop is under way (0072); the poller skips such a task. */
+    stopping_at: ColumnType<Date | null, Date | string | null | undefined, Date | string | null>;
+}
 /** 0023_sandboxes.sql：按对话的沙箱（见 sandbox/sandbox.service.ts）。 */
 export interface SandboxesTable {
     conversation_id: string;
@@ -788,6 +837,11 @@ export interface ExtensionPackagesTable {
     source_ref: string;
     uploaded_by: string | null;
     created_at: Ts;
+    /** 0066: the plugin.json `displayName`; null without one and for packages stored before 0066. */
+    display_name: ColumnType<{
+        zh: string;
+        en: string;
+    } | null, string | null | undefined, string | null>;
 }
 /** 0030, renamed in 0041: package bytes when there is no object storage. */
 export interface ExtensionBlobsTable {
@@ -867,6 +921,7 @@ export interface KernelDatabase {
     conversation_events: ConversationEventsTable;
     conversation_queue: ConversationQueueTable;
     file_previewers: FilePreviewersTable;
+    org_extension_point_defaults: OrgExtensionPointDefaultsTable;
     runs: RunsTable;
     subagent_dispatches: SubagentDispatchesTable;
     agent_templates: AgentTemplatesTable;
@@ -890,6 +945,7 @@ export interface KernelDatabase {
     directory_exclusions: DirectoryExclusionsTable;
     attachments: AttachmentsTable;
     sandboxes: SandboxesTable;
+    background_tasks: BackgroundTasksTable;
     skills: SkillsTable;
     skill_versions: SkillVersionsTable;
     skill_change_requests: SkillChangeRequestsTable;

package/dist/db/pool.d.ts

@@ -18,6 +18,11 @@ export declare function pgConnectOptions(cfg: {
  */
 export declare function createDb(url: string, poolMax: number, onError?: (err: Error) => void, opts?: {
     connectTimeoutMs?: number | undefined;
+    /**
+     * How long an idle client stays in the pool (pg's `idleTimeoutMillis`; unset = pg's 10 s).
+     * Every new client costs a TCP + auth handshake, several round trips to a distant database.
+     */
+    idleTimeoutMs?: number | undefined;
     /**
      * Error of a client while it is checked out: pg-pool drops its own listener on checkout, so
      * without this a backend killed between two statements is an uncaught exception.

package/dist/db/pool.js

@@ -22,6 +22,9 @@ export function createDb(url, poolMax, onError = (err) => console.error("[db] id
         connectionString: url,
         max: poolMax,
         ...pgConnectOptions(opts),
+        ...(opts.idleTimeoutMs !== undefined && { idleTimeoutMillis: opts.idleTimeoutMs }),
+        // a client kept idle longer must notice a dead peer (a dropped VPN, a NAT timeout) by itself
+        keepAlive: true,
     });
     pool.on("error", onError);
     pool.on("connect", (client) => client.on("error", opts.onClientError ?? onError));

package/dist/errors.d.ts

@@ -110,7 +110,11 @@ export declare const ERROR_CODES: {
     readonly "setup.token_invalid": 403;
     readonly "agent.not_found": 404;
     readonly "agent.version_conflict": 409;
+    /** The profile was changed since the client read it (`expectedUpdatedAt` is stale). */
+    readonly "agent.profile_conflict": 409;
     readonly "agent.model_unavailable": 400;
+    /** A save gave `instructionFiles` and an `instructions` that is not their join. */
+    readonly "agent.instructions_mismatch": 400;
     /**
      * `skills.pinned` of an agent definition is refused (details: `{ skill, reason, totalBytes? }`;
      * reasons `not_selected`, `not_found`, `duplicate`, `too_many`, `too_large`, `manual_only`).
@@ -130,6 +134,10 @@ export declare const ERROR_CODES: {
     /** A sub-agent's conversation takes no messages from people (sub-agents spec §6.1). */
     readonly "conversation.read_only": 409;
     readonly "run.not_active": 409;
+    /** A background task that is not in this conversation (or its sub-agents), or does not exist. */
+    readonly "background_task.not_found": 404;
+    /** The task's output cannot be read now (its sandbox is not running and no finish recorded it). */
+    readonly "background_task.output_unavailable": 409;
     readonly "connector.not_found": 404;
     readonly "connector.slug_taken": 409;
     readonly "connector.connect_failed": 502;

package/dist/errors.js

@@ -110,7 +110,11 @@ export const ERROR_CODES = {
     "setup.token_invalid": 403,
     "agent.not_found": 404,
     "agent.version_conflict": 409,
+    /** The profile was changed since the client read it (`expectedUpdatedAt` is stale). */
+    "agent.profile_conflict": 409,
     "agent.model_unavailable": 400,
+    /** A save gave `instructionFiles` and an `instructions` that is not their join. */
+    "agent.instructions_mismatch": 400,
     /**
      * `skills.pinned` of an agent definition is refused (details: `{ skill, reason, totalBytes? }`;
      * reasons `not_selected`, `not_found`, `duplicate`, `too_many`, `too_large`, `manual_only`).
@@ -130,6 +134,10 @@ export const ERROR_CODES = {
     /** A sub-agent's conversation takes no messages from people (sub-agents spec §6.1). */
     "conversation.read_only": 409,
     "run.not_active": 409,
+    /** A background task that is not in this conversation (or its sub-agents), or does not exist. */
+    "background_task.not_found": 404,
+    /** The task's output cannot be read now (its sandbox is not running and no finish recorded it). */
+    "background_task.output_unavailable": 409,
     "connector.not_found": 404,
     "connector.slug_taken": 409,
     "connector.connect_failed": 502,

package/dist/evals/recorded-tools.d.ts

@@ -1,4 +1,5 @@
 import type { EventLog } from "../conversations/event-log.js";
+import { type LegacyScope } from "../tools/legacy-keys.js";
 import { type ToolPipeline, type ToolStepContext } from "../tools/pipeline.js";
 import type { CaseSnapshot } from "./snapshot.js";
 /**
@@ -15,12 +16,34 @@ export interface RecordedCall {
     inputKey: string;
     content: string;
     isError: boolean;
-    /** 原结果里的 `tool_search` 已加载工具（照抄：之后请求的已加载集合由日志还原）。 */
+    /** 原结果里的 `ToolSearch` 已加载工具（照抄：之后请求的已加载集合由日志还原）。 */
     loadedTools?: string[];
     contentTypes?: string[];
     /** 原 `tool_call_started` 的定义指纹（没有开始事件的调用为空串）。 */
     specHash: string;
+    /** The recorded call used release-candidate names; `toolKey` and `inputKey` are already the current ones. */
+    legacy?: true;
 }
+/**
+ * The connectors and extensions an rc `<x>__<tool>` key can name, from positive evidence in the
+ * snapshot only (a head with no evidence, or evidence of both kinds, stays unmapped, so its
+ * recorded calls are unmatched, which is a safe miss):
+ * - a connector: the tool definition the key was recorded with (`toolSpecs`, found by the specHash of
+ *   its `tool_call_started` or its `model_call` tools / catalog entry) has `connectorData`;
+ * - an extension: that definition exists without `connectorData` and its name is the key's tool
+ *   part, or an agent definition names exactly that tool as a `{ kind: "plugin", plugin: x, name }` ref.
+ */
+export declare function legacyScopeOf(s: CaseSnapshot): LegacyScope;
+/**
+ * A recorded call in the current names: an rc builtin or `<x>__<tool>` key becomes its current
+ * key with the input's parameters renamed. Only the recorded side is ever normalized; a call
+ * already in current names (or one that cannot be mapped) is returned as is.
+ */
+export declare function currentCall(name: string, input: unknown, scope: LegacyScope): {
+    name: string;
+    input: unknown;
+    legacy: boolean;
+};
 /** 规范化 JSON：对象键递归按码点排序，数组保持顺序；`undefined` 按 JSON 规则省略。 */
 export declare function canonicalJson(v: unknown): string;
 /** The matching key of a call: the canonical JSON of its input without the ignored fields. */
@@ -43,7 +66,10 @@ export declare class RecordedToolPipeline implements Pick<ToolPipeline, "runStep
     private readonly d;
     private readonly used;
     private misses;
+    private legacyHits;
     constructor(d: RecordedToolPipelineDeps);
+    /** Calls so far matched with a recording made under release-candidate tool names. */
+    legacyMatches(): number;
     /** 至今没有匹配到记录的调用数。 */
     unmatched(): number;
     runStep(ctx: ToolStepContext): Promise<{

package/dist/evals/recorded-tools.js

@@ -1,4 +1,6 @@
 import { assistantIdemKey } from "../runtime/model-call.js";
+import { MATCH_IGNORED_FIELDS } from "../tools/builtin-names.js";
+import { isLegacyPrefixedKey, legacyToolInput, legacyToolKey, } from "../tools/legacy-keys.js";
 import { RUN_STOPPED_TEXT } from "../tools/pipeline.js";
 /**
  * 整轮回放的工具管线（C6 / Review Focus 3）：只回放原运行记录的工具结果，**从不执行任何工具**——
@@ -7,6 +9,77 @@ import { RUN_STOPPED_TEXT } from "../tools/pipeline.js";
  */
 /** 没有匹配记录时模型看到的工具结果（模型可见，英文）。 */
 export const REPLAY_NO_RECORDING = "[replay] No recorded result matches this tool call (same tool and arguments); the tool was not executed during replay.";
+/**
+ * The connectors and extensions an rc `<x>__<tool>` key can name, from positive evidence in the
+ * snapshot only (a head with no evidence, or evidence of both kinds, stays unmapped, so its
+ * recorded calls are unmatched, which is a safe miss):
+ * - a connector: the tool definition the key was recorded with (`toolSpecs`, found by the specHash of
+ *   its `tool_call_started` or its `model_call` tools / catalog entry) has `connectorData`;
+ * - an extension: that definition exists without `connectorData` and its name is the key's tool
+ *   part, or an agent definition names exactly that tool as a `{ kind: "plugin", plugin: x, name }` ref.
+ */
+export function legacyScopeOf(s) {
+    const hashes = new Map();
+    const note = (key, hash) => {
+        if (!isLegacyPrefixedKey(key))
+            return;
+        const set = hashes.get(key) ?? new Set();
+        if (hash !== undefined && hash !== "")
+            set.add(hash);
+        hashes.set(key, set);
+    };
+    for (const e of s.events) {
+        if (e.type === "tool_call_started") {
+            const p = e.payload;
+            note(p.toolKey, p.specHash);
+        }
+        else if (e.type === "model_call") {
+            const r = e.payload.request;
+            for (const t of [...(r?.tools ?? []), ...(r?.catalog ?? [])])
+                note(t.key, t.specHash);
+        }
+    }
+    const extensionNames = new Set();
+    const connectorSlugs = new Set();
+    const planted = new Set();
+    for (const def of Object.values(s.definitions)) {
+        for (const t of def.tools)
+            if (t.kind === "plugin")
+                planted.add(`${t.plugin}__${t.name}`);
+    }
+    for (const [key, set] of hashes) {
+        const i = key.indexOf("__");
+        const head = key.slice(0, i);
+        const tool = key.slice(i + 2);
+        for (const h of set) {
+            const spec = s.toolSpecs[h];
+            if (spec === undefined)
+                continue;
+            if (spec.connectorData !== undefined) {
+                connectorSlugs.add(head);
+            }
+            else if (spec.name === tool) {
+                extensionNames.add(head);
+            }
+        }
+    }
+    for (const key of planted) {
+        if (hashes.has(key))
+            extensionNames.add(key.slice(0, key.indexOf("__")));
+    }
+    return { connectorSlugs, extensionNames };
+}
+/**
+ * A recorded call in the current names: an rc builtin or `<x>__<tool>` key becomes its current
+ * key with the input's parameters renamed. Only the recorded side is ever normalized; a call
+ * already in current names (or one that cannot be mapped) is returned as is.
+ */
+export function currentCall(name, input, scope) {
+    const key = legacyToolKey(name, scope);
+    if (key === null)
+        return { name, input, legacy: false };
+    return { name: key, input: legacyToolInput(name, input), legacy: true };
+}
 /** 规范化 JSON：对象键递归按码点排序，数组保持顺序；`undefined` 按 JSON 规则省略。 */
 export function canonicalJson(v) {
     const norm = (x) => {
@@ -23,14 +96,6 @@ export function canonicalJson(v) {
     };
     return JSON.stringify(norm(v) ?? null);
 }
-/**
- * Input fields left out when matching, per tool: display-only arguments the model writes that do not
- * change what the tool does. `sandbox_bash`'s `description` became required in UI refresh B, so a
- * recording without it and a replayed call with one (or a different wording) still match.
- */
-const MATCH_IGNORED_FIELDS = {
-    sandbox_bash: ["description"],
-};
 /** The matching key of a call: the canonical JSON of its input without the ignored fields. */
 export function matchKey(toolKey, input) {
     const ignored = MATCH_IGNORED_FIELDS[toolKey];
@@ -64,6 +129,7 @@ export function recordedCalls(s) {
                 started.set(p.callId, p.specHash);
         }
     }
+    const scope = legacyScopeOf(s);
     const out = [];
     for (const e of own) {
         if (e.type !== "assistant_message")
@@ -76,9 +142,11 @@ export function recordedCalls(s) {
             const r = results.get(call.id);
             if (!r)
                 continue;
+            const cur = currentCall(call.name, call.input, scope);
             out.push({
-                toolKey: call.name,
-                inputKey: matchKey(call.name, call.input),
+                toolKey: cur.name,
+                inputKey: matchKey(cur.name, cur.input),
+                ...(cur.legacy && { legacy: true }),
                 content: r.content,
                 isError: r.isError,
                 ...(r.loadedTools !== undefined && { loadedTools: r.loadedTools }),
@@ -98,9 +166,14 @@ export class RecordedToolPipeline {
     d;
     used = new Set();
     misses = 0;
+    legacyHits = 0;
     constructor(d) {
         this.d = d;
     }
+    /** Calls so far matched with a recording made under release-candidate tool names. */
+    legacyMatches() {
+        return this.legacyHits;
+    }
     /** 至今没有匹配到记录的调用数。 */
     unmatched() {
         return this.misses;
@@ -162,6 +235,8 @@ export class RecordedToolPipeline {
         if (i < 0)
             return undefined;
         this.used.add(i);
+        if (this.d.recordings[i]?.legacy === true)
+            this.legacyHits++;
         return this.d.recordings[i];
     }
 }

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

@@ -19,7 +19,10 @@ import type { ResolvedCandidate } from "./step-replay.js";
  *
  * - 种子：原运行 `run_started` 之前的事件按原 seq 插入（`last_seq` 停在其前一位，新的 `run_started` 因此与原来同 seq），
  *   原运行开始时的 `skills_catalog` 照抄；然后写 `runs` 行与 `run_started`，`agent` 用候选定义。
- * - 插话：原运行中途吸收的 user_message，在「原吸收步骤 ≤ 当前步骤」时追加进评测对话并交给 loop。
+ * - 插话：原运行中途吸收的可吸收事件（user_message, and finished background tasks: background Bash
+ *   spec §4.4），在「原吸收步骤 ≤ 当前步骤」时追加进评测对话并交给 loop（a recorded finish is
+ *   appended as recorded: no task starts, polls or stops）。
+ * - A wake run (its boundary a finished background task, not a message) is seeded from that event.
  * - 计费（C7，与单步回放同一套）：每次模型请求（含压缩、重试、兜底）发出前查运行仍在执行、仍持有认领；
  *   收完按真实用量、没收完按估算 `consumeTokens`（越限即 `capped`，下一次请求前停下）。
  * - 附件按快照自己的组织与原对话取（C3）。
@@ -37,6 +40,11 @@ export interface RunReplayResult {
     usage: ModelUsage;
     steps: number;
     unmatched: number;
+    /**
+     * Set when a call was matched with a recording made under the release-candidate tool names
+     * (`sandbox_edit` for `Edit`, `docs__search` for `mcp__docs__search`, old parameter names).
+     */
+    legacyKeyMatch?: true;
 }
 /** 原运行（基线）的同形摘要。 */
 export type RunBaseline = Omit<RunReplayResult, "conversationId" | "outcome" | "unmatched">;
@@ -96,7 +104,7 @@ export interface ReplaySeed {
     /** The conversation's `consumed_seq` when the original run started. */
     consumed: number;
     seeds: SnapshotEvent[];
-    /** The run's trigger seq (the last user message before it). */
+    /** The run's trigger seq (its boundary: the last unconsumed message, or the waking finish). */
     trigger: number;
     agentVersion: number;
     model: ModelRef;

package/dist/evals/run-replay.js

@@ -13,6 +13,7 @@ import { Inject, Injectable } from "@nestjs/common";
 import { sql } from "kysely";
 import { normalizeStoredDefinition } from "../agents/definition.js";
 import { EventLog } from "../conversations/event-log.js";
+import { isAbsorbable, } from "../conversations/events.js";
 import { ApiError } from "../errors.js";
 import { KernelDb } from "../infra/database/kernel-db.js";
 import { TransactionHost } from "../infra/database/transaction-host.js";
@@ -22,7 +23,7 @@ import { RunError } from "../runtime/model-call.js";
 import { recordedSkillForks } from "../runtime/subagents/skill-fork.js";
 import { skillsCatalogIdemKey } from "../skills/catalog.js";
 import { ACTIVE_EVAL_STATUSES, addUsage, consumeTokens, TOO_MANY_ERRORS, usageTokens, } from "./budget.js";
-import { RecordedToolPipeline, recordedCalls } from "./recorded-tools.js";
+import { currentCall, legacyScopeOf, RecordedToolPipeline, recordedCalls, } from "./recorded-tools.js";
 /** 写种子时发现已失去认领或运行已不在执行：本次执行停下，不留评测对话。 */
 export class ReplayClaimLost extends Error {
     constructor() {
@@ -177,7 +178,7 @@ const ZERO = {
     cacheWriteTokens: 0,
 };
 /** 一次运行的工具调用序列、最后的回复、停止原因、用量与步数（基线与候选同一算法）。 */
-function summarize(events, runId) {
+function summarize(events, runId, scope) {
     const toolCalls = [];
     let usage = ZERO;
     let last = null;
@@ -189,8 +190,14 @@ function summarize(events, runId) {
             const p = e.payload;
             steps.add(p.step);
             for (const b of p.content) {
-                if (b.type === "tool_call")
+                if (b.type !== "tool_call")
+                    continue;
+                if (scope === undefined)
                     toolCalls.push({ name: b.name, input: b.input });
+                else {
+                    const { name, input } = currentCall(b.name, b.input, scope);
+                    toolCalls.push({ name, input });
+                }
             }
             last = { content: p.content, stopReason: p.stopReason };
         }
@@ -214,7 +221,9 @@ function summarize(events, runId) {
 }
 /** 原运行（快照）的摘要，作为整轮回放的基线。 */
 export function runBaseline(s) {
-    return summarize(s.events, s.runId);
+    // an rc recording is read in the current tool names, so the diff against a candidate only
+    // shows real differences
+    return summarize(s.events, s.runId, legacyScopeOf(s));
 }
 /**
  * 删掉一个评测对话（崩溃重跑前清掉上一次的残留）：先删它各次运行的 checkpoint，再删对话（事件、运行随之级联删除）。
@@ -252,6 +261,28 @@ function absorbedByStep(s) {
     }
     return out;
 }
+/**
+ * The replayed run's boundary (`trigger_seq`): a wake run's recorded one (`run_started.triggerSeq`,
+ * background Bash spec §3); otherwise the newest message after `consumed`, else the newest finish
+ * after it that could have woken the run (not reported, no `stoppedBy`), else the last
+ * message (old logs, as before).
+ */
+function replayTrigger(started, seeds, consumed) {
+    const recorded = started;
+    if (recorded.trigger === "background_task" && recorded.triggerSeq !== undefined) {
+        return recorded.triggerSeq;
+    }
+    const lastMessage = seeds.findLast((e) => e.type === "user_message")?.seq;
+    if (lastMessage !== undefined && lastMessage > consumed)
+        return lastMessage;
+    const wake = seeds.findLast((e) => {
+        if (e.type !== "background_task_finished" || e.seq <= consumed)
+            return false;
+        const p = e.payload;
+        return !p.reported && p.stoppedBy === undefined;
+    })?.seq;
+    return wake ?? lastMessage ?? 0;
+}
 const readAll = async (eventLog, conversationId) => {
     const out = [];
     let after = 0;
@@ -286,7 +317,7 @@ export async function replayRun(d, s, candidate, evalRunId, signal) {
             consumed = Math.max(consumed, e.payload.consumedUpToSeq);
         }
     }
-    const trigger = seeds.findLast((e) => e.type === "user_message")?.seq ?? 0;
+    const trigger = replayTrigger(start.payload, seeds, consumed);
     const catalog = s.events.find((e) => e.runId === s.runId && e.type === "skills_catalog" && e.seq > R);
     // a sub-agent's case (spec §15.3) runs the effective definition frozen in its subagent_started
     // (the request renders that event's preamble and instructions; its tools come from it too), with
@@ -419,7 +450,7 @@ export async function replayRun(d, s, candidate, evalRunId, signal) {
             return "stop";
         },
     });
-    // 插话：原运行开始之后到达、且被原运行某一步吸收的 user_message
+    // 插话：原运行开始之后到达、且被原运行某一步吸收的可吸收事件（user_message / background_task_finished）
     const absorbed = absorbedByStep(s);
     const stepsInOrder = [...absorbed.keys()].sort((a, b) => a - b);
     /** 第 k 步（含）之前原运行吸收到的最大位置；原运行没有模型调用时不设限。 */
@@ -434,7 +465,7 @@ export async function replayRun(d, s, candidate, evalRunId, signal) {
         return cap;
     };
     const interjections = s.events
-        .filter((e) => e.type === "user_message" && e.seq > R)
+        .filter((e) => (e.type === "user_message" || e.type === "background_task_finished") && e.seq > R)
         .flatMap((e) => {
         const step = stepsInOrder.find((k) => absorbed.get(k) >= e.seq);
         return step === undefined ? [] : [{ event: e, step }];
@@ -442,7 +473,7 @@ export async function replayRun(d, s, candidate, evalRunId, signal) {
     /** 已追加的插话：原 seq → 评测对话里的 seq。 */
     const fed = new Map();
     let prepared = 0;
-    const pendingUserMessages = async (afterSeq) => {
+    const pendingAbsorbable = async (afterSeq) => {
         prepared++;
         // loop 的步骤从 1 起（与运行时相同），每步的 prepare 调用一次：第 n 次调用即第 n 步
         const k = prepared;
@@ -451,7 +482,7 @@ export async function replayRun(d, s, candidate, evalRunId, signal) {
                 continue;
             const [w] = await eventLog.append(conversationId, [
                 {
-                    type: "user_message",
+                    type: x.event.type,
                     idemKey: `replay_interjection:${runId}:${x.event.seq}`,
                     payload: x.event.payload,
                 },
@@ -463,7 +494,7 @@ export async function replayRun(d, s, candidate, evalRunId, signal) {
         const fedSeqs = new Set(fed.values());
         const out = [];
         for (const e of await readAll(eventLog, conversationId)) {
-            if (e.type !== "user_message" || e.seq <= afterSeq)
+            if (!isAbsorbable(e) || e.seq <= afterSeq)
                 continue;
             if (fedSeqs.has(e.seq) || (e.seq < R && e.seq <= cap))
                 out.push(e);
@@ -494,7 +525,7 @@ export async function replayRun(d, s, candidate, evalRunId, signal) {
             signal: loopSignal,
             shutdown: new AbortController().signal,
             sink: () => { },
-            pendingUserMessages,
+            pendingAbsorbable,
             limits: {
                 maxSteps: subagent === undefined ? runtime.maxSteps : Math.min(subagent.maxSteps, runtime.maxSteps),
                 deadline,
@@ -558,5 +589,11 @@ export async function replayRun(d, s, candidate, evalRunId, signal) {
     });
     await Promise.allSettled([...charges]);
     const summary = summarize(await readAll(eventLog, conversationId), runId);
-    return { conversationId, outcome: outcome.kind, ...summary, unmatched: pipeline.unmatched() };
+    return {
+        conversationId,
+        outcome: outcome.kind,
+        ...summary,
+        unmatched: pipeline.unmatched(),
+        ...(pipeline.legacyMatches() > 0 && { legacyKeyMatch: true }),
+    };
 }

package/dist/evals/step-replay.js

@@ -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。

package/dist/evals/trajectory.d.ts

@@ -54,7 +54,7 @@ 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;
@@ -64,7 +64,7 @@ export interface TrajectoryToolCall {
     /** 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;