@capybara-hq/kernel 0.1.0-rc.3 → 0.1.0-rc.5

package/dist/runtime/model-call.js

@@ -1,3 +1,4 @@
+import { createHash } from "node:crypto";
 import { isModelError, isRetryableKind, ModelError, sanitizeParams, } from "@capybara-hq/plugin-sdk";
 import { MEDIA_EXCLUDED_MAX_IDS } from "../conversations/events.js";
 import { ApiError } from "../errors.js";
@@ -6,16 +7,23 @@ import { hasInvocation, skillLoadOffered } from "../skills/catalog.js";
 import { buildModelRequest, latestSummary, mediaRefs, planMedia, } from "./context/build.js";
 import { acceptSummary, buildCompactionRequest, compactionBackoff, compactionPlan, shouldCompact, } from "./context/compaction.js";
 import { guardToolPairs } from "./context/pairing.js";
-import { turnLimitReminder } from "./context/reminders.js";
+import { TOOLS_UNSUPPORTED_REMINDER, turnLimitReminder } from "./context/reminders.js";
 import { calibrate, estimateTokens } from "./context/tokens.js";
+import { toolSetChange } from "./context/tool-changes.js";
 import { fireObserve, hookDeniedAudit, hookDeniedEvent, hookRequestOf, hookScope, recordHookDenial, } from "./hook-calls.js";
 import { MediaCache } from "./media-cache.js";
 import { ProviderListError } from "./providers.js";
-import { subagentBudgetUsedUp } from "./subagents/limits.js";
+import { budgetRunOf, subagentBudgetUsedUp } from "./subagents/limits.js";
 /** 单轮收尾的尾部提醒（模型可见，英文；spec 6 轮次额度收尾）。 */
 export const WRAP_UP_REMINDER = "Step limit reached. Do not call tools. Summarize what you found and what remains for the user.";
-/** 所用模型不支持工具（`capabilities.tools` 为 false）时的提醒（模型可见，英文）。 */
-export const TOOLS_UNSUPPORTED_REMINDER = "Tools are unavailable for this model. Answer without calling tools.";
+/**
+ * The tail reminder of a sub-agent's final call once its tree used its token budget (model-visible,
+ * English; sub-agents spec §8.3): the child hands back what it found instead of ending with nothing.
+ */
+export const BUDGET_WRAP_UP_REMINDER = "The sub-agents' token budget is used up, so this is your last reply. Do not call tools. " +
+    "Report what you found so far, with the sources and details that matter, and what remains " +
+    "unfinished.";
+export { TOOLS_UNSUPPORTED_REMINDER } from "./context/reminders.js";
 /** 不可用工具提醒的前缀（模型可见，英文）。 */
 export const TOOLS_UNAVAILABLE_PREFIX = "The following tools are unavailable in this run: ";
 /** 本 run 中此前有工具不可用、现在全部恢复时的提醒（模型可见，英文）。 */
@@ -70,6 +78,10 @@ const idem = {
     reminder: (runId, step) => `reminder:${runId}:${step}`,
     toolsUnavailable: (runId, step) => `tools_unavailable:${runId}:${step}`,
     toolsUnsupported: (runId, step) => `tools_unsupported:${runId}:${step}`,
+    toolsChanged: (runId, step, keys) => `tools_changed:${runId}:${step}:${createHash("sha256")
+        .update(JSON.stringify([...keys].sort()))
+        .digest("hex")
+        .slice(0, 16)}`,
     modelCall: modelCallIdemKey,
     hookDenied: (runId, step, attempt) => `hook_denied:${runId}:${step}:${attempt}`,
     /** Usage of a compaction call: `n` is the compaction attempt (consecutive failures + 1). */
@@ -88,18 +100,24 @@ class StepAborted extends Error {
  * 模型看到的一切先写进日志（`model_call` / `reminder` / `compaction_summary`），请求可由日志还原。
  */
 export function runModelStep(deps, input) {
-    return runStep(deps, input, false);
+    return runStep(deps, input, null);
 }
 /**
  * 单轮收尾：到达步数上限时调用一次。写入收尾 `reminder`（位于请求末尾）；照常带工具（历史里有工具调用时
  * 部分供应商要求请求声明工具，也保持缓存前缀稳定），但 `toolChoice: "none"`——模型只能回复文本。
  * 其余与 `runModelStep` 相同（幂等、降级、中止）。
+ *
+ * `reason: "budget"`: a sub-agent whose tree used its token budget makes this one call to hand back
+ * what it found (sub-agents spec §8.3); it is not stopped by the budget again (one call over).
  */
-export function runFinalWrapUp(deps, input) {
-    return runStep(deps, input, true);
+export function runFinalWrapUp(deps, input, reason = "max_steps") {
+    return runStep(deps, input, reason);
 }
-async function runStep(deps, input, wrapUp) {
-    const { conversationId, runId, step, signal } = input;
+async function runStep(deps, input, 
+/** The final call's reason, or null for an ordinary step. */
+wrapUpReason) {
+    const wrapUp = wrapUpReason !== null;
+    const { conversationId, runId, step, signal, sink } = input;
     const existing = await findExisting(deps.eventLog, conversationId, runId, step);
     if (existing) {
         // resumed after a crash between writing usage and enqueueing it: enqueue again (job id dedupes)
@@ -108,6 +126,8 @@ async function runStep(deps, input, wrapUp) {
             if (u)
                 void deps.usageReport.enqueue(conversationId, u.seq);
         }
+        // a crash between the assistant_message and clearing its draft leaves one behind
+        sink({ kind: "step_end", runId, step });
         return existing;
     }
     const runtime = deps.config.get().runtime;
@@ -117,6 +137,11 @@ async function runStep(deps, input, wrapUp) {
         if (signal.aborted)
             throw new StepAborted();
         let events = await readAll(deps.eventLog, conversationId);
+        // an earlier execution of this run already sent this step (it streamed, then crashed or was
+        // handed over): this one starts at offset 0, so viewers drop that attempt's draft first
+        if (events.some((e) => e.type === "model_call" && e.runId === runId && e.payload.step === step)) {
+            sink({ kind: "step_reset", runId, step });
+        }
         // sticky (M-4): a model.request denial already logged for this step (any attempt, e.g. a crash
         // after recording it) is final: fail the same way, before compaction or any model call. Read from
         // the events just loaded, so it costs no query (M-2).
@@ -132,6 +157,8 @@ async function runStep(deps, input, wrapUp) {
         const resolved = await deps.tools.forAgent(input.orgId, {
             tools: input.agent.tools ?? [],
             subagents: input.agent.subagents,
+            child: input.parent !== undefined || input.parentRunId !== undefined || input.subagent === true,
+            forkChild: input.parent?.forSkill === true,
             skillLoad: skillLoadOffered(events.findLast((e) => e.type === "skills_catalog"), hasInvocation(events)),
         });
         const tools = {
@@ -147,10 +174,12 @@ async function runStep(deps, input, wrapUp) {
         const beforeCompaction = events.length;
         events = await maybeCompact(deps, input, events, runtime, tools);
         // a sub-agent checks its budget again after a compaction call, so the step's own request is
-        // not a second call over the cap (review M5; the loop checks before each step)
-        if (input.parentRunId !== undefined &&
+        // not a second call over the cap (review M5; the loop checks before each step); the budget
+        // wrap-up is the one call allowed over it
+        if (wrapUpReason !== "budget" &&
+            input.parentRunId !== undefined &&
             events.length !== beforeCompaction &&
-            (await subagentBudgetUsedUp(deps.db, deps.config, input.orgId, input.parentRunId))) {
+            (await subagentBudgetUsedUp(deps.db, deps.config, input.orgId, input.parent?.budgetRunId ?? (await budgetRunOf(deps.db, input.parentRunId))))) {
             return {
                 content: [],
                 stopReason: "end_turn",
@@ -160,6 +189,19 @@ async function runStep(deps, input, wrapUp) {
                 budget: true,
             };
         }
+        // the tool set changed since the previous request (an extension switched on, a sub-agent
+        // roster added): tell the model before this request (tool naming spec §9.1). Keyed by the new
+        // set, so a retried step finds it in the log (previousToolSet applies it) and writes nothing
+        const toolKeys = resolved.tools.map((t) => t.key);
+        const change = toolSetChange(events, toolKeys);
+        if (change !== null) {
+            events = await appendLocal(deps.eventLog, conversationId, events, {
+                type: "tools_changed",
+                runId,
+                idemKey: idem.toolsChanged(runId, step, toolKeys),
+                payload: { step, ...change },
+            });
+        }
         const last = lastToolsReminder(events, runId);
         // 有不可用工具 → 列出；此前列过、现在全部恢复 → 说明已恢复，避免过时的「不可用」成为最后一句
         // （收尾调用不得调用工具，可用性变化不写）。
@@ -181,7 +223,7 @@ async function runStep(deps, input, wrapUp) {
                 });
             }
         }
-        const { outcome, ref } = await ladder(deps, input, events, runtime, tools, wrapUp, streamed);
+        const { outcome, ref } = await ladder(deps, input, events, runtime, tools, wrapUpReason, streamed);
         // assistant_message 与 usage 同一次 append = 同一事务。
         const appended = await appendCounted(deps, input, [
             {
@@ -201,6 +243,7 @@ async function runStep(deps, input, wrapUp) {
         const usageEvent = appended.find((e) => e.type === "usage");
         if (deps.usageReport && usageEvent)
             void deps.usageReport.enqueue(conversationId, usageEvent.seq);
+        sink({ kind: "step_end", runId, step });
         return {
             content: outcome.content,
             stopReason: outcome.stopReason,
@@ -222,21 +265,24 @@ async function runStep(deps, input, wrapUp) {
                 payload: { step, content, stopReason: "aborted" },
             },
         ]);
+        sink({ kind: "step_end", runId, step });
         return { content, stopReason: "aborted", reused: false, aborted: true, usage: null };
     }
 }
 /**
- * Appends a model call's events (its reply or compaction outcome with its `usage`). For a
+ * Appends a model call's events (its reply or compaction outcome with its `usage`; also the usage
+ * of a `WebFetch` extraction call, `tools/builtins/web/extract.ts`). For a
  * sub-agent's run (`input.parentRunId`) in one transaction that also adds the tokens of the usage
- * events it newly wrote to the parent run's `subagent_tokens_used` (sub-agents spec §8.3): a usage
+ * events it newly wrote to the budget run's `subagent_tokens_used` (sub-agents spec §8.3): a usage
  * event already in the log (an idempotent hit after a crash) adds nothing, so each call counts
  * exactly once.
  */
-async function appendCounted(deps, input, events) {
+export async function appendCounted(deps, input, events) {
     const { conversationId, parentRunId } = input;
     if (parentRunId === undefined)
         return deps.eventLog.append(conversationId, events);
-    const out = await deps.stepLog.appendCounted(conversationId, parentRunId, events);
+    const budgetRunId = input.budgetRunId ?? input.parent?.budgetRunId;
+    const out = await deps.stepLog.appendCounted(conversationId, { parentRunId, ...(budgetRunId !== undefined && { budgetRunId }) }, events);
     if (out.toSeq !== null)
         await deps.eventLog.notify(conversationId, out.toSeq);
     return out.written;
@@ -657,11 +703,14 @@ function mergeLocal(events, written) {
  * 每次实际尝试写自己的 `model_call`（attempt 从 1 连续递增，跳过的不计）；重试前经 sink 发 `step_reset`。
  * 不可重试的失败或梯子用尽 → `RunError`。
  */
-async function ladder(deps, input, events, runtime, tools, wrapUp, streamed) {
+async function ladder(deps, input, events, runtime, tools, wrapUpReason, streamed) {
     const { conversationId, runId, step, signal, sink, agent } = input;
-    const turnReminder = wrapUp
-        ? WRAP_UP_REMINDER
-        : turnLimitReminder(step, input.maxSteps ?? runtime.maxSteps);
+    const wrapUp = wrapUpReason !== null;
+    const turnReminder = wrapUpReason === "budget"
+        ? BUDGET_WRAP_UP_REMINDER
+        : wrapUp
+            ? WRAP_UP_REMINDER
+            : turnLimitReminder(step, input.maxSteps ?? runtime.maxSteps);
     let evs = events;
     const rungs = [
         { ref: agent.model, fallback: false },
@@ -784,6 +833,8 @@ async function ladder(deps, input, events, runtime, tools, wrapUp, streamed) {
                     sink({ kind: "thinking", runId, step, text: ev.text });
                 }
             });
+            // the draft's last write, before the assistant_message is logged
+            sink({ kind: "stream_end", runId, step });
             if (scope && deps.hooks) {
                 fireObserve(deps.hooks, "model.response", {
                     ...scope,

package/dist/runtime/providers.d.ts

@@ -67,6 +67,12 @@ export declare class ProviderRegistry {
     listTimeoutMs?: number);
     /** `listModels()` 加时限（到时按失败处理，供应商的调用留在后台、结果丢弃）。 */
     private listModelsWithTimeout;
+    /**
+     * Whether at least one model-provider extension is loaded on this replica (enabled, configured,
+     * set up). Read from the extension host only: never calls a vendor, so a provider whose vendor
+     * is down still counts (org readiness, web-order spec §3.8).
+     */
+    hasProvider(): boolean;
     /** 每个插件取一个实现（`@2` 优先），按插件首次出现的顺序；不支持的主版本记 warn 后忽略。 */
     private providers;
     private models;

package/dist/runtime/providers.js

@@ -132,6 +132,14 @@ let ProviderRegistry = class ProviderRegistry {
             clearTimeout(timer);
         }
     }
+    /**
+     * Whether at least one model-provider extension is loaded on this replica (enabled, configured,
+     * set up). Read from the extension host only: never calls a vendor, so a provider whose vendor
+     * is down still counts (org readiness, web-order spec §3.8).
+     */
+    hasProvider() {
+        return this.providers().length > 0;
+    }
     /** 每个插件取一个实现（`@2` 优先），按插件首次出现的顺序；不支持的主版本记 warn 后忽略。 */
     providers() {
         const chosen = new Map();

package/dist/runtime/run-control.d.ts

@@ -13,9 +13,11 @@ export declare class RunControl {
     private readonly abort;
     private readonly events;
     private readonly log;
-    constructor(queue: Pick<RunQueue, "ensureRun" | "abortIfOrphaned">, waits: Pick<Waits, "abortWaiting" | "childFinished" | "stopChildren">, abort: Pick<RunAbort, "request">, events: KernelEvents, log: Pick<KernelLogger, "error">);
+    constructor(queue: Pick<RunQueue, "ensureRun" | "abortIfOrphaned" | "wakeHeldTree">, waits: Pick<Waits, "abortWaiting" | "childFinished" | "stopChildren" | "stopTree">, abort: Pick<RunAbort, "request">, events: KernelEvents, log: Pick<KernelLogger, "error">);
     /** 有未消费的消息且没有进行中的运行时投递一轮运行；返回 runId 或 null。 */
     ensureRun(conversationId: string): Promise<string | null>;
+    /** After a root's queue was unpaused: the wakes it held for its tree ({@link RunQueue.wakeHeldTree}). */
+    wakeHeldTree(rootId: string): Promise<void>;
     /**
      * 运行是孤儿（会话锁空闲、任务已不存在，没有 worker 会收到中止广播）时直接把它写成 aborted，返回 true；
      * 否则返回 false。A stopped sub-agent then reports to its parent at once, and a stopped parent
@@ -28,6 +30,14 @@ export declare class RunControl {
      * changed: `run.wait-stopped` lets the channel outbound update them.
      */
     abortWaiting(conversationId: string, runId: string, byAccountId: string | null): Promise<boolean>;
+    /**
+     * Stops a sub-agent with its descendants, or with `includeSelf: false` only the sub-agent
+     * conversations below `conversationId` (agent map spec §6; {@link Waits.stopTree}). Returns
+     * the number of runs asked to stop.
+     */
+    stopTree(conversationId: string, byAccountId: string | null, opts?: {
+        includeSelf?: boolean;
+    }): Promise<number>;
     /** 请求中止运行（广播到所有副本）。 */
     requestAbort(runId: string, byAccountId: string | null): Promise<void>;
 }

package/dist/runtime/run-control.js

@@ -34,6 +34,10 @@ let RunControl = class RunControl {
     ensureRun(conversationId) {
         return this.queue.ensureRun(conversationId);
     }
+    /** After a root's queue was unpaused: the wakes it held for its tree ({@link RunQueue.wakeHeldTree}). */
+    wakeHeldTree(rootId) {
+        return this.queue.wakeHeldTree(rootId);
+    }
     /**
      * 运行是孤儿（会话锁空闲、任务已不存在，没有 worker 会收到中止广播）时直接把它写成 aborted，返回 true；
      * 否则返回 false。A stopped sub-agent then reports to its parent at once, and a stopped parent
@@ -61,6 +65,14 @@ let RunControl = class RunControl {
         await this.events.emit("run.wait-stopped", { conversationId, runId });
         return r;
     }
+    /**
+     * Stops a sub-agent with its descendants, or with `includeSelf: false` only the sub-agent
+     * conversations below `conversationId` (agent map spec §6; {@link Waits.stopTree}). Returns
+     * the number of runs asked to stop.
+     */
+    stopTree(conversationId, byAccountId, opts = {}) {
+        return this.waits.stopTree(conversationId, byAccountId, opts);
+    }
     /** 请求中止运行（广播到所有副本）。 */
     requestAbort(runId, byAccountId) {
         return this.abort.request(runId, byAccountId);

package/dist/runtime/run-drafts.d.ts

@@ -0,0 +1,85 @@
+import type { ExtensionContext, Logger } from "@capybara-hq/plugin-sdk";
+import type { Redis } from "ioredis";
+import { EventLog } from "../conversations/event-log.js";
+import { KernelDb } from "../infra/database/kernel-db.js";
+import { type StreamSink } from "./model-call.js";
+/** How long a draft lives without a write (a step that keeps streaming keeps refreshing it). */
+export declare const DRAFT_TTL_SECONDS = 600;
+/** At most one write per step per window, plus the one at the end of the model stream. */
+export declare const DRAFT_FLUSH_MS = 200;
+/** One step's draft: `{ text, thinking }` as JSON (spec 2026-11-07 §3.2). */
+export declare const draftKey: (runId: string, step: number) => string;
+/** What one step of a run has streamed so far. */
+export interface RunDraft {
+    step: number;
+    text: string;
+    thinking: string;
+}
+/** A draft of a conversation's active run (the `event: draft` frame's data). */
+export interface LiveDraft extends RunDraft {
+    runId: string;
+}
+/**
+ * Live drafts (spec 2026-11-07-run-input-and-live-drafts §3.2): what each step of a running run
+ * has streamed so far, kept in Redis so a viewer that (re)connects on any replica gets the text
+ * already shown, then the deltas from there on.
+ *
+ * The writer is the process executing the run: it holds the step's draft in memory (only for the
+ * runs it executes; {@link release} drops them when the execution ends) and writes it at most once
+ * per {@link DRAFT_FLUSH_MS}: on the first delta past the window, or, when deltas stop inside it,
+ * by one trailing one-shot timer per held step (process-local, unref'd, cancelled by a write,
+ * {@link clear} and {@link release}), plus once when the model stream ends ({@link flush}). Writes are issued before the delta is published, on the same
+ * connection as the bus publish, so a viewer that sees a delta past its snapshot and asks again
+ * gets a draft at least that far whenever that delta was written. Every write refreshes the
+ * {@link DRAFT_TTL_SECONDS} TTL; {@link clear} deletes the step's draft once its
+ * `assistant_message` is logged or the step is reset. Redis failures are logged, never thrown to
+ * the run: a draft is a convenience, the log stays the truth.
+ */
+export declare class RunDrafts {
+    private readonly redis;
+    private readonly kdb;
+    private readonly eventLog;
+    private readonly log;
+    private readonly held;
+    constructor(redis: Redis, kdb: KernelDb, eventLog: Pick<EventLog, "findByIdem">, log: Pick<Logger, "warn">);
+    private get db();
+    /** Adds a delta to the step's draft; writes it when the last write is a window old. */
+    append(runId: string, step: number, channel: "text" | "thinking", text: string): void;
+    /** The end of the step's model stream: writes what the window still holds. */
+    flush(runId: string, step: number): void;
+    /** Deletes the step's draft (its `assistant_message` is logged, or the step starts over). */
+    clear(runId: string, step: number): void;
+    /**
+     * Drops what this process holds for the run (its execution here ended: finished, suspended,
+     * handed over). Redis keeps what was written: another worker may continue the run, and a step
+     * that never ended is not shown once the run is no longer running (TTL removes it).
+     */
+    release(runId: string): void;
+    /** The run's steps with a draft, by step. */
+    read(runId: string): Promise<RunDraft[]>;
+    /**
+     * The drafts a viewer of the conversation is shown: steps of its running run (or only `runId`)
+     * whose `assistant_message` is not logged. Redis is read before the log, so a step finalized
+     * meanwhile is either skipped here or its `assistant_message` reaches the viewer after the draft.
+     */
+    live(conversationId: string, runId?: string): Promise<LiveDraft[]>;
+    /** Whether the run belongs to the conversation (the draft route's 404). */
+    runOf(conversationId: string, runId: string): Promise<boolean>;
+    private drop;
+    private write;
+}
+/** What {@link streamSink} uses. */
+export interface StreamSinkDeps {
+    bus: Pick<ExtensionContext["bus"], "publish">;
+    drafts: Pick<RunDrafts, "append" | "flush" | "clear">;
+    log: Pick<Logger, "warn">;
+}
+/**
+ * The run's stream sink (one per execution of a run): publishes deltas on the conversation's topic
+ * (`conv.<id>`, shared with `{ kind: "events" }`, so a subscriber on any replica gets them) with
+ * `offset`, the characters of that channel (text or thinking) this attempt of the step streamed
+ * before the delta, and keeps the step's live draft. `step_reset` clears the draft and both
+ * offsets; `stream_end` / `step_end` only flush / clear the draft and are not published. A failed
+ * publish is only logged: deltas are drafts, the `assistant_message` in the log is the truth.
+ */
+export declare function streamSink(conversationId: string, deps: StreamSinkDeps): StreamSink;

package/dist/runtime/run-drafts.js

@@ -0,0 +1,230 @@
+var __decorate = (this && this.__decorate) || function (decorators, target, key, desc) {
+    var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
+    if (typeof Reflect === "object" && typeof Reflect.decorate === "function") r = Reflect.decorate(decorators, target, key, desc);
+    else for (var i = decorators.length - 1; i >= 0; i--) if (d = decorators[i]) r = (c < 3 ? d(r) : c > 3 ? d(target, key, r) : d(target, key)) || r;
+    return c > 3 && r && Object.defineProperty(target, key, r), r;
+};
+var __param = (this && this.__param) || function (paramIndex, decorator) {
+    return function (target, key) { decorator(target, key, paramIndex); }
+};
+import { Inject, Injectable } from "@nestjs/common";
+import { conversationTopic, EventLog, } from "../conversations/event-log.js";
+import { KernelDb } from "../infra/database/kernel-db.js";
+import { KERNEL_LOGGER, REDIS } from "../infra/tokens.js";
+import { assistantIdemKey } from "./model-call.js";
+/** How long a draft lives without a write (a step that keeps streaming keeps refreshing it). */
+export const DRAFT_TTL_SECONDS = 600;
+/** At most one write per step per window, plus the one at the end of the model stream. */
+export const DRAFT_FLUSH_MS = 200;
+/** One step's draft: `{ text, thinking }` as JSON (spec 2026-11-07 §3.2). */
+export const draftKey = (runId, step) => `capk:run-draft:${runId}:${step}`;
+/** The steps of a run that have a draft (a set, so a reader finds them without a key scan). */
+const stepsKey = (runId) => `capk:run-draft-steps:${runId}`;
+const parse = (step, v) => {
+    if (v === null)
+        return null;
+    try {
+        const d = JSON.parse(v);
+        return {
+            step,
+            text: typeof d.text === "string" ? d.text : "",
+            thinking: typeof d.thinking === "string" ? d.thinking : "",
+        };
+    }
+    catch {
+        return null;
+    }
+};
+/**
+ * Live drafts (spec 2026-11-07-run-input-and-live-drafts §3.2): what each step of a running run
+ * has streamed so far, kept in Redis so a viewer that (re)connects on any replica gets the text
+ * already shown, then the deltas from there on.
+ *
+ * The writer is the process executing the run: it holds the step's draft in memory (only for the
+ * runs it executes; {@link release} drops them when the execution ends) and writes it at most once
+ * per {@link DRAFT_FLUSH_MS}: on the first delta past the window, or, when deltas stop inside it,
+ * by one trailing one-shot timer per held step (process-local, unref'd, cancelled by a write,
+ * {@link clear} and {@link release}), plus once when the model stream ends ({@link flush}). Writes are issued before the delta is published, on the same
+ * connection as the bus publish, so a viewer that sees a delta past its snapshot and asks again
+ * gets a draft at least that far whenever that delta was written. Every write refreshes the
+ * {@link DRAFT_TTL_SECONDS} TTL; {@link clear} deletes the step's draft once its
+ * `assistant_message` is logged or the step is reset. Redis failures are logged, never thrown to
+ * the run: a draft is a convenience, the log stays the truth.
+ */
+let RunDrafts = class RunDrafts {
+    redis;
+    kdb;
+    eventLog;
+    log;
+    held = new Map();
+    constructor(redis, kdb, eventLog, log) {
+        this.redis = redis;
+        this.kdb = kdb;
+        this.eventLog = eventLog;
+        this.log = log;
+    }
+    get db() {
+        return this.kdb.db;
+    }
+    /** Adds a delta to the step's draft; writes it when the last write is a window old. */
+    append(runId, step, channel, text) {
+        let h = this.held.get(runId);
+        if (h?.step !== step) {
+            if (h)
+                this.drop(runId, h);
+            h = {
+                step,
+                text: "",
+                thinking: "",
+                writtenAt: Number.NEGATIVE_INFINITY,
+                dirty: false,
+                timer: undefined,
+            };
+            this.held.set(runId, h);
+        }
+        h[channel] += text;
+        h.dirty = true;
+        const elapsed = Date.now() - h.writtenAt;
+        if (elapsed >= DRAFT_FLUSH_MS)
+            this.write(runId, h);
+        else if (h.timer === undefined) {
+            // held: written when the window ends even if no delta follows (a tool input, a pause)
+            h.timer = setTimeout(() => {
+                h.timer = undefined;
+                this.flush(runId, step);
+            }, DRAFT_FLUSH_MS - elapsed).unref();
+        }
+    }
+    /** The end of the step's model stream: writes what the window still holds. */
+    flush(runId, step) {
+        const h = this.held.get(runId);
+        if (h?.step === step && h.dirty)
+            this.write(runId, h);
+    }
+    /** Deletes the step's draft (its `assistant_message` is logged, or the step starts over). */
+    clear(runId, step) {
+        const h = this.held.get(runId);
+        if (h?.step === step)
+            this.drop(runId, h);
+        this.redis
+            .multi()
+            .del(draftKey(runId, step))
+            .srem(stepsKey(runId), String(step))
+            .exec()
+            .catch((err) => this.log.warn("run draft clear failed", { runId, step, err }));
+    }
+    /**
+     * Drops what this process holds for the run (its execution here ended: finished, suspended,
+     * handed over). Redis keeps what was written: another worker may continue the run, and a step
+     * that never ended is not shown once the run is no longer running (TTL removes it).
+     */
+    release(runId) {
+        const h = this.held.get(runId);
+        if (h)
+            this.drop(runId, h);
+    }
+    /** The run's steps with a draft, by step. */
+    async read(runId) {
+        const steps = (await this.redis.smembers(stepsKey(runId)))
+            .map(Number)
+            .filter((n) => Number.isSafeInteger(n))
+            .sort((a, b) => a - b);
+        if (steps.length === 0)
+            return [];
+        const values = await this.redis.mget(steps.map((s) => draftKey(runId, s)));
+        return steps.flatMap((s, i) => parse(s, values[i] ?? null) ?? []);
+    }
+    /**
+     * The drafts a viewer of the conversation is shown: steps of its running run (or only `runId`)
+     * whose `assistant_message` is not logged. Redis is read before the log, so a step finalized
+     * meanwhile is either skipped here or its `assistant_message` reaches the viewer after the draft.
+     */
+    async live(conversationId, runId) {
+        let q = this.db
+            .selectFrom("runs")
+            .select("id")
+            .where("conversation_id", "=", conversationId)
+            .where("status", "=", "running");
+        if (runId !== undefined)
+            q = q.where("id", "=", runId);
+        const out = [];
+        for (const run of await q.execute()) {
+            for (const d of await this.read(run.id)) {
+                const final = await this.eventLog.findByIdem(conversationId, assistantIdemKey(run.id, d.step));
+                if (final === undefined)
+                    out.push({ runId: run.id, ...d });
+            }
+        }
+        return out;
+    }
+    /** Whether the run belongs to the conversation (the draft route's 404). */
+    async runOf(conversationId, runId) {
+        const row = await this.db
+            .selectFrom("runs")
+            .select("id")
+            .where("id", "=", runId)
+            .where("conversation_id", "=", conversationId)
+            .executeTakeFirst();
+        return row !== undefined;
+    }
+    drop(runId, h) {
+        clearTimeout(h.timer);
+        h.timer = undefined;
+        this.held.delete(runId);
+    }
+    write(runId, h) {
+        clearTimeout(h.timer);
+        h.timer = undefined;
+        h.writtenAt = Date.now();
+        h.dirty = false;
+        const { step } = h;
+        this.redis
+            .multi()
+            .set(draftKey(runId, step), JSON.stringify({ text: h.text, thinking: h.thinking }), "EX", DRAFT_TTL_SECONDS)
+            .sadd(stepsKey(runId), String(step))
+            .expire(stepsKey(runId), DRAFT_TTL_SECONDS)
+            .exec()
+            .catch((err) => this.log.warn("run draft write failed", { runId, step, err }));
+    }
+};
+RunDrafts = __decorate([
+    Injectable(),
+    __param(0, Inject(REDIS)),
+    __param(1, Inject(KernelDb)),
+    __param(2, Inject(EventLog)),
+    __param(3, Inject(KERNEL_LOGGER))
+], RunDrafts);
+export { RunDrafts };
+/**
+ * The run's stream sink (one per execution of a run): publishes deltas on the conversation's topic
+ * (`conv.<id>`, shared with `{ kind: "events" }`, so a subscriber on any replica gets them) with
+ * `offset`, the characters of that channel (text or thinking) this attempt of the step streamed
+ * before the delta, and keeps the step's live draft. `step_reset` clears the draft and both
+ * offsets; `stream_end` / `step_end` only flush / clear the draft and are not published. A failed
+ * publish is only logged: deltas are drafts, the `assistant_message` in the log is the truth.
+ */
+export function streamSink(conversationId, deps) {
+    let at = { runId: "", step: -1, text: 0, thinking: 0 };
+    const publish = (msg) => deps.bus
+        .publish(conversationTopic(conversationId), msg)
+        .catch((err) => deps.log.warn("stream delta publish failed", { conversationId, err }));
+    return (d) => {
+        if (d.kind === "stream_end")
+            return deps.drafts.flush(d.runId, d.step);
+        if (d.kind === "step_end")
+            return deps.drafts.clear(d.runId, d.step);
+        if (at.runId !== d.runId || at.step !== d.step || d.kind === "step_reset") {
+            at = { runId: d.runId, step: d.step, text: 0, thinking: 0 };
+        }
+        if (d.kind === "step_reset") {
+            deps.drafts.clear(d.runId, d.step);
+            void publish({ kind: "delta", runId: d.runId, step: d.step, delta: "step_reset" });
+            return;
+        }
+        const text = d.text ?? "";
+        const offset = at[d.kind];
+        at[d.kind] += text.length;
+        deps.drafts.append(d.runId, d.step, d.kind, text);
+        void publish({ kind: "delta", runId: d.runId, step: d.step, delta: d.kind, text, offset });
+    };
+}

package/dist/runtime/run-jobs.d.ts

@@ -34,12 +34,10 @@ export declare class RunJobs {
     queue(name?: string): Queue<RunJobData>;
     /** A run job by id on the queue `name`, or undefined. */
     getJob(jobId: string, name?: string): Promise<Job<RunJobData> | undefined>;
-    /** Adds a conversation run job (`ensureRun`'s own enqueue). */
-    add(data: RunJobData, jobId: string): Promise<void>;
     /**
-     * Enqueues a run job under the caller's job id (resumes, dispatches, repairs). The queue follows
-     * the conversation's kind (a sub-agent's runs go to {@link SUBAGENT_RUN_QUEUE_NAME}); pass
-     * `kind` when the caller knows it, else it is read.
+     * Enqueues a run job under the caller's job id (`ensureRun`, resumes, dispatches, repairs). The
+     * queue follows the conversation's kind (a sub-agent's runs go to
+     * {@link SUBAGENT_RUN_QUEUE_NAME}); pass `kind` when the caller knows it, else it is read.
      */
     enqueue(data: RunJobData, jobId: string, kind?: ConversationKind): Promise<void>;
 }

package/dist/runtime/run-jobs.js

@@ -54,14 +54,10 @@ let RunJobs = class RunJobs {
     getJob(jobId, name = RUN_QUEUE_NAME) {
         return this.queue(name).getJob(jobId);
     }
-    /** Adds a conversation run job (`ensureRun`'s own enqueue). */
-    async add(data, jobId) {
-        await this.queue().add("run", data, { jobId, ...RUN_JOB_OPTIONS });
-    }
     /**
-     * Enqueues a run job under the caller's job id (resumes, dispatches, repairs). The queue follows
-     * the conversation's kind (a sub-agent's runs go to {@link SUBAGENT_RUN_QUEUE_NAME}); pass
-     * `kind` when the caller knows it, else it is read.
+     * Enqueues a run job under the caller's job id (`ensureRun`, resumes, dispatches, repairs). The
+     * queue follows the conversation's kind (a sub-agent's runs go to
+     * {@link SUBAGENT_RUN_QUEUE_NAME}); pass `kind` when the caller knows it, else it is read.
      */
     async enqueue(data, jobId, kind) {
         const k = kind ??