dsh-continual-evolve 0.15.11 → 0.15.12

package/README.md

@@ -7,7 +7,7 @@ English | [中文](README.zh.md)
 [![CI](https://github.com/ZK-Andy/dsh-continual-evolve/actions/workflows/ci.yml/badge.svg)](https://github.com/ZK-Andy/dsh-continual-evolve/actions/workflows/ci.yml)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
 [![Node](https://img.shields.io/badge/node-%5E22.19%20%7C%7C%20%3E%3D24-339933)](package.json)
-[![Tests](https://img.shields.io/badge/tests-337%20passing-brightgreen)]()
+[![Tests](https://img.shields.io/badge/tests-362%20passing-brightgreen)]()
 [![Coverage · statements](https://img.shields.io/badge/coverage_statements-96%25-brightgreen)]()
 [![Coverage · branches](https://img.shields.io/badge/coverage_branches-90%25-green)]()
 [![Coverage · functions](https://img.shields.io/badge/coverage_functions-100%25-brightgreen)]()
@@ -81,7 +81,7 @@ Profile patch example:
 
 ```bash
 pnpm install && pnpm build   # deps + clean build -> lib/
-pnpm test                    # vitest (337 tests)
+pnpm test                    # vitest (362 tests)
 pnpm test:coverage           # v8 coverage, CI-enforced thresholds
 pnpm lint                    # oxlint src test client
 pnpm check:pack              # published-artifact consistency (also runs on prepack)
@@ -95,6 +95,7 @@ Layout:
 │   ├── store.ts            # central SQLite store: DDL/WAL/FTS, reads, ledger, patrol, cursors
 │   ├── store-apply.ts      # the sole write gate: validate all, apply atomically, ledger
 │   ├── token-usage.ts      # single accounting entry: normalize provider usage, pure sums
+│   ├── llm-call.ts         # shared direct-call boundary: route, reasoning effort, failure text
 │   ├── review.ts           # autonomous quality loop: cheap gate → evolve, budget, scheduler
 │   ├── review-prompt.ts    # review prompts + verdict parsing (gate + memory/skill rubrics)
 │   ├── store-apply.ts      # the sole write gate: validate all, apply atomically, ledger
@@ -114,7 +115,7 @@ Layout:
 ├── client/
 │   └── client.js          # hand-written client bundle: read-only card in the plugin manager
 ├── locale/                # host package metadata: localized display name + description (zh/en)
-├── test/                  # vitest suite (25 files)
+├── test/                  # vitest suite (26 files)
 ├── lib/                   # build output (tsc)
 ├── docs/                  # design.md (design) · FAQ.md (real-world pitfalls)
 └── .agents/               # AI collaboration layer (AGENTS.md, skills, ADR notes)

package/README.zh.md

@@ -7,7 +7,7 @@
 [![CI](https://github.com/ZK-Andy/dsh-continual-evolve/actions/workflows/ci.yml/badge.svg)](https://github.com/ZK-Andy/dsh-continual-evolve/actions/workflows/ci.yml)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
 [![Node](https://img.shields.io/badge/node-%5E22.19%20%7C%7C%20%3E%3D24-339933)](package.json)
-[![Tests](https://img.shields.io/badge/tests-337%20passing-brightgreen)]()
+[![Tests](https://img.shields.io/badge/tests-362%20passing-brightgreen)]()
 [![Coverage · statements](https://img.shields.io/badge/coverage_statements-96%25-brightgreen)]()
 [![Coverage · branches](https://img.shields.io/badge/coverage_branches-90%25-green)]()
 [![Coverage · functions](https://img.shields.io/badge/coverage_functions-100%25-brightgreen)]()
@@ -81,7 +81,7 @@ profile patch 示例：
 
 ```bash
 pnpm install && pnpm build   # 依赖 + 干净构建 -> lib/
-pnpm test                    # vitest（337 例）
+pnpm test                    # vitest（362 例）
 pnpm test:coverage           # v8 覆盖率，CI 强制阈值
 pnpm lint                    # oxlint src test client
 pnpm check:pack              # 发布产物一致性（prepack 也会跑）
@@ -95,6 +95,7 @@ pnpm check:pack              # 发布产物一致性（prepack 也会跑）
 │   ├── store.ts            # 中央 SQLite 库：DDL/WAL/FTS、读面、账本、巡检、游标
 │   ├── store-apply.ts      # 唯一写门：全量校验、事务原子应用、账本
 │   ├── token-usage.ts      # 统一记账口：归一各家用量、纯累加
+│   ├── llm-call.ts         # 直属调用边界：路由、推理档位、失败原文
 │   ├── review.ts           # 内容质量自进化：便宜门 → 贵改写、预算、调度
 │   ├── review-prompt.ts    # 评审提示词 + verdict 解析（门 + 记忆/技能 rubric）
 │   ├── store-apply.ts      # 唯一写门：全量校验、事务原子应用、账本
@@ -114,7 +115,7 @@ pnpm check:pack              # 发布产物一致性（prepack 也会跑）
 ├── client/
 │   └── client.js          # 手写 client bundle：官方插件管理内只读卡片
 ├── locale/                # 宿主包元数据：显示名 + 描述双语（zh/en）
-├── test/                  # vitest 测试（25 个文件）
+├── test/                  # vitest 测试（26 个文件）
 ├── lib/                   # 构建产物（tsc）
 ├── docs/                  # design.md（设计）· FAQ.md（踩坑）
 └── .agents/               # AI 协作层（AGENTS.md、技能、ADR 笔记）

package/lib/extraction.d.ts

@@ -1,14 +1,18 @@
 import { type SkipReason, type SurfaceReader } from "./extraction-surface.js";
 import type { LedgerEntry, MemoryStore } from "./store.js";
-/** The minimal LLM face extraction calls (duck-typed host service). */
-export interface LlmStream {
-    stream(options: Record<string, unknown>): AsyncIterable<unknown>;
-}
+import { type LlmRuntime, type RouteResolver } from "./llm-call.js";
+export type { LlmStream, LlmRuntime, RouteResolver } from "./llm-call.js";
 /** Everything a run needs: the store, the session reader, and the LLM. */
 export interface ExtractionDeps {
     store: MemoryStore;
     surface: SurfaceReader | undefined;
-    llm: LlmStream | undefined;
+    llm: LlmRuntime | undefined;
+    /**
+     * Route resolution for a direct call: the agent's explicit selection first,
+     * the host default model second. Absent keeps the legacy agent-options-only
+     * behavior (tests and older hosts).
+     */
+    resolveRoute?: RouteResolver | undefined;
 }
 /** One extraction target: whose session, in which workspace, on which route. */
 export interface ExtractionTarget {

package/lib/extraction.js

@@ -1,6 +1,7 @@
 import { boundaryCursorOf, evaluateSlice, eventsAfterCursor, isInternalAgent, seqOfCursor, serializeIncrement } from "./extraction-surface.js";
 import { CHEAP_INCREMENT_MAX, EXTRACTION_CHEAP_SYSTEM_PROMPT, EXTRACTION_SYSTEM_PROMPT, cheapPrompt, extractionPrompt, manifestLineOf, parseCheapAnswer, parseExtractionAnswer } from "./extraction-prompt.js";
 import { cwdOf } from "./memory-section.js";
+import { resolveCallRoute, streamOnce } from "./llm-call.js";
 /** Head length of a failed run's raw answer kept in the ledger `files` column. */
 const FAILURE_RAW_MAX = 2048;
 /** Head length of an unexpected error's message/stack trace kept in the ledger `files` column. */
@@ -11,6 +12,8 @@ const ERROR_REASON_MAX = 300;
 const GATE_ERROR_MAX = 200;
 /** Output budget for one cheap-gate call (see review.ts's `GATE_MAX_TOKENS`). */
 const GATE_MAX_TOKENS = 1024;
+/** Output budget for the full proposal call: a decision plus replacement bodies. */
+const EXTRACTION_MAX_TOKENS = 4096;
 /** Split an unexpected throw into a ledger-sized reason head and file body. */
 function errorText(error) {
     const message = error instanceof Error && error.message.length > 0 ? error.message : String(error);
@@ -84,6 +87,9 @@ async function runExtractionInner(deps, target, opts) {
         store.logSkipped({ ...ledgerBase, skipReason: "no-route", status: "skipped" });
         return { status: "skipped", skipReason: "no-route" };
     }
+    // Narrowed once, here: the shared call boundary takes a concrete route.
+    const llm = deps.llm;
+    const route = { provider: target.provider, model: target.model };
     const cursor = store.cursor(target.workspaceId, target.sessionId);
     let all;
     try {
@@ -129,7 +135,7 @@ async function runExtractionInner(deps, target, opts) {
     {
         const gateStarted = Date.now();
         try {
-            const gate = await streamOnce(deps.llm, target, cheapPrompt({ incrementHead: increment.slice(0, CHEAP_INCREMENT_MAX), userSeqs: slice.userSeqs }), EXTRACTION_CHEAP_SYSTEM_PROMPT, GATE_MAX_TOKENS);
+            const gate = await streamOnce(llm, route, target.sessionId, EXTRACTION_CHEAP_SYSTEM_PROMPT, cheapPrompt({ incrementHead: increment.slice(0, CHEAP_INCREMENT_MAX), userSeqs: slice.userSeqs }), GATE_MAX_TOKENS);
             cheapUsage = gate.usage;
             const verdict = parseCheapAnswer(gate.text);
             // A gate cut at its budget that still answered is a usable verdict;
@@ -170,7 +176,7 @@ async function runExtractionInner(deps, target, opts) {
     let text;
     let usage;
     try {
-        const answer = await streamOnce(deps.llm, target, prompt);
+        const answer = await streamOnce(llm, route, target.sessionId, EXTRACTION_SYSTEM_PROMPT, prompt, EXTRACTION_MAX_TOKENS);
         // The full call carries the proposal payload: a cut stream is a failure.
         if (answer.finish !== undefined && answer.finish !== "stop") {
             throw new Error(`stream finished with ${answer.finish}`);
@@ -181,7 +187,7 @@ async function runExtractionInner(deps, target, opts) {
     catch (error) {
         const { message, detail } = errorText(error);
         try {
-            store.logSkipped({ ...ledgerBase, skipReason: "llm-failed", status: "failed", durationMs: Date.now() - started, cheapUsage, cheapError, files: detail });
+            store.logSkipped({ ...ledgerBase, skipReason: "llm-failed", status: "failed", model: target.model, durationMs: Date.now() - started, cheapUsage, cheapError, files: detail });
         }
         catch (ledgerError) {
             throw new Error(`llm call failed (${message}); failure ledger write also failed (${errorText(ledgerError).message})`);
@@ -253,41 +259,6 @@ function candidateRecords(store, workspaceId, userText) {
 function formatCandidate(record) {
     return [`# 候选 id=${record.id}（type=${record.type}）`, `title=${record.title}`, `钩子=${record.description}`, "全文：", record.body].join("\n");
 }
-/** One streaming text call with the host session identity forwarded. */
-async function streamOnce(llm, target, prompt, system = EXTRACTION_SYSTEM_PROMPT, maxTokens = 4096) {
-    let text = "";
-    let usage;
-    let finish;
-    const options = {
-        provider: target.provider,
-        model: target.model,
-        messages: [{ role: "user", content: [{ type: "text", text: prompt }] }],
-        system,
-        maxTokens,
-        // MissingSessionID pitfall: direct calls must carry the host session.
-        sessionId: target.sessionId,
-    };
-    for await (const chunk of llm.stream(options)) {
-        const piece = chunk;
-        if (piece?.type === "text-delta" && typeof piece.text === "string") {
-            text += piece.text;
-        }
-        else if (piece?.type === "usage") {
-            usage = piece.usage;
-        }
-        else if (piece?.type === "finish") {
-            finish = piece.reason;
-        }
-    }
-    const finishKind = finish?.kind;
-    if (text.trim().length === 0) {
-        // Nothing usable arrived: the finish reason is the only diagnosis.
-        throw new Error(finishKind !== undefined && finishKind !== "stop" ? `stream finished with ${finishKind}` : "empty model output");
-    }
-    // The caller owns the truncation verdict: a gate answer often completes
-    // before its budget does, a proposal body never does.
-    return { text, usage, finish: finishKind };
-}
 /**
  * Wire the extraction triggers onto the host context. Returns a disposer
  * that removes the listeners. The scheduler is the only background
@@ -310,12 +281,14 @@ export function createExtractionScheduler(host, deps) {
         if (typeof id !== "string" || id.length === 0) {
             return null;
         }
-        const route = agent?.options;
+        // Absent resolver (older host, tests) still reads the agent's explicit
+        // selection: the fallback is the legacy behavior, never "no route".
+        const route = deps.resolveRoute?.(agent) ?? resolveCallRoute(agent, undefined);
         return {
             agent,
             sessionId: id,
-            provider: typeof route?.provider === "string" ? route.provider : undefined,
-            model: typeof route?.model === "string" ? route.model : undefined,
+            provider: route?.provider,
+            model: route?.model,
         };
     };
     const drain = (sessionId, trigger) => {

package/lib/index.js

@@ -31,6 +31,7 @@
 import z from "@deepseek-ai/schemastery";
 import { mountCardRoutes } from "./card-routes.js";
 import { createExtractionScheduler } from "./extraction.js";
+import { resolveCallRoute } from "./llm-call.js";
 import { createReviewScheduler, REVIEW_DEFAULT_BATCH_SIZE, REVIEW_DEFAULT_DAILY_BUDGET, REVIEW_DEFAULT_DRY_RUN, REVIEW_DEFAULT_STALE_DAYS, } from "./review.js";
 import { createFrozenMemorySection, cwdOf, DEFAULT_MAX_CHARS, DEFAULT_MEMORY_SECTION_ORDER, MEMORY_SECTION_NAME, memorySectionText, } from "./memory-section.js";
 import { registerMemoryTools } from "./memory-tools.js";
@@ -87,6 +88,34 @@ export const Config = z.object({
         enabled: z.boolean(),
     }),
 });
+/**
+ * Build the route resolver a scheduler uses for its direct calls: the agent's
+ * explicit selection first, the host's default-model service second.
+ *
+ * The default-model service is read through `ctx.get` on every call and never
+ * joins an inject list — a host without it (or an older one) must keep working,
+ * and its absence only means a session with no explicit selection has no route.
+ */
+function routeResolverOf(scoped, get) {
+    if (typeof get !== "function") {
+        return undefined;
+    }
+    const read = get.bind(scoped);
+    return (agent) => resolveCallRoute(agent, read);
+}
+/**
+ * Keep one scheduler's disposer on the owning fiber.
+ *
+ * Without this a remount (HMR, a live config edit) leaves the previous
+ * listeners attached: two schedulers then drain the same boundary, and the
+ * older one holds an LLM service the remount already replaced.
+ */
+function registerSchedulerDisposer(scoped, dispose) {
+    const effect = scoped.effect;
+    if (typeof effect === "function") {
+        effect.call(scoped, () => dispose);
+    }
+}
 /** Register the memory tools through a nested inject (see memory-tools.ts). */
 function mountToolsWhenAvailable(host, store, state) {
     const nested = host;
@@ -134,7 +163,11 @@ function mountExtractionWhenAvailable(host, store, config) {
             log.warn("sessionQuery service present but empty — extraction not armed (injection and tools unaffected)");
             return;
         }
-        createExtractionScheduler(events, { store, surface, llm });
+        const resolveRoute = routeResolverOf(scoped, context.get);
+        registerSchedulerDisposer(scoped, createExtractionScheduler(events, {
+            store, surface, llm,
+            ...(resolveRoute === undefined ? {} : { resolveRoute }),
+        }));
     });
 }
 /**
@@ -158,7 +191,11 @@ function mountReviewWhenAvailable(host, store, config) {
     nested.inject(["llm"], (scoped) => {
         const context = scoped;
         const llm = (context.llm ?? context.get?.("llm"));
-        createReviewScheduler(events, { store, llm }, {
+        const resolveRoute = routeResolverOf(scoped, context.get);
+        registerSchedulerDisposer(scoped, createReviewScheduler(events, {
+            store, llm,
+            ...(resolveRoute === undefined ? {} : { resolveRoute }),
+        }, {
             enabled: true,
             read: () => ({
                 dryRun: config.review?.dryRun?.get() ?? REVIEW_DEFAULT_DRY_RUN,
@@ -167,7 +204,7 @@ function mountReviewWhenAvailable(host, store, config) {
                 staleDays: config.review?.staleDays?.get() ?? REVIEW_DEFAULT_STALE_DAYS,
             }),
             minIntervalMs: 5 * 60 * 1000,
-        });
+        }));
     });
 }
 /**

package/lib/llm-call.d.ts

@@ -0,0 +1,116 @@
+/**
+ * The shared boundary for the plugin's own LLM calls: extraction's cheap
+ * pre-gate and full proposal call, and review's gate and evolve passes all
+ * stream through here.
+ *
+ * Three disciplines live in one place because all three were learned the hard
+ * way, and none of them is visible at a call site:
+ *
+ * - **The route.** `agent.options` carries an *explicit* selection only. A
+ *   session running on the host's default model carries nothing there, so the
+ *   route falls back to the host's `agentDefaultModel` service — the same
+ *   value the host Agent loop itself runs on. A caller that still has no
+ *   route refuses to send instead of streaming into a provider rejection.
+ * - **The reasoning effort.** An exact provider/model advertises the effort
+ *   ids it accepts, and a request naming one it does not accept is rejected
+ *   by capability validation before any provider I/O. The lowest *enabled*
+ *   effort the model advertises is used; a model with no reasoning metadata
+ *   gets no caller-supplied effort at all. This is the pre-teardown
+ *   `llm-text.ts` discipline (see the 2026-09-24 capability note), restored
+ *   after the direct calls that replaced that boundary shipped without it.
+ * - **The failure text.** A terminal `error`/`aborted` finish carries the
+ *   host's own failure object. Keeping only its `kind` produced a ledger full
+ *   of "stream finished with error" that named no cause for a whole day.
+ */
+/** The minimal LLM face a direct call needs (duck-typed host service). */
+export interface LlmStream {
+    stream(options: Record<string, unknown>): AsyncIterable<unknown>;
+}
+/** Exact-model capability metadata, narrowed to the field this plugin reads. */
+export interface ModelInfo {
+    reasoning?: {
+        efforts?: unknown;
+    } | undefined;
+}
+/** The LLM service plus the optional exact-model capability query. */
+export interface LlmRuntime extends LlmStream {
+    /** Resolve one exact provider/model's capabilities; absent on older hosts and test fakes. */
+    resolveModelInfo?(provider: string, model: string): Promise<ModelInfo | undefined> | ModelInfo | undefined;
+}
+/** One concrete provider/model route for a direct call. */
+export interface ModelSelection {
+    provider: string;
+    model: string;
+    reasoningEffort?: string | undefined;
+}
+/** The host's default-model service, reached as `ctx.get("agentDefaultModel")`. */
+export interface DefaultModelService {
+    currentSelection(): unknown;
+}
+/** Read one sibling service off the host context (`ctx.get`). */
+export type ServiceGetter = (name: string) => unknown;
+/** Resolve one direct call's route from the agent plus the host default model. */
+export type RouteResolver = (agent: unknown) => ModelSelection | undefined;
+/** Effort ids that mean "reasoning off"; a model may advertise only these. */
+export declare const CLOSED_REASONING_EFFORTS: ReadonlySet<string>;
+/**
+ * Select the lowest reasoning effort an exact model advertises.
+ *
+ * An enabled effort wins over a declared closing effort: the adapter's first
+ * enabled effort is the lowest enabled level it exposes, and a closing effort
+ * is only a fallback when the model publishes no enabled level at all. No
+ * capability list means no caller-supplied effort.
+ *
+ * @param modelInfo - Exact resolved-model metadata, when the provider exposes it.
+ * @returns The provider-owned effort id to send, or `undefined` to omit the field.
+ * @throws When the host reports reasoning metadata in a shape this plugin cannot trust.
+ */
+export declare function selectLowestReasoningEffort(modelInfo: ModelInfo | undefined): string | undefined;
+/**
+ * The host's default model selection, when the host exposes that service.
+ *
+ * The service is read through `ctx.get` and never joins an inject list: a host
+ * without it (or an older one) must keep working, and a failed read is simply
+ * "no default", never a mount failure.
+ *
+ * @param get - The host context's sibling-service reader, when available.
+ * @returns The default provider/model (plus effort), or `undefined` when unknown.
+ */
+export declare function defaultRouteOf(get: ServiceGetter | undefined): ModelSelection | undefined;
+/**
+ * Resolve the route one direct call runs on: the agent's explicit selection
+ * first, the host default second.
+ *
+ * @param agent - The host Agent the call belongs to (its `options` may be empty).
+ * @param get - The host context's sibling-service reader, when available.
+ * @returns The route to call with, or `undefined` when there is none to use.
+ */
+export declare function resolveCallRoute(agent: unknown, get: ServiceGetter | undefined): ModelSelection | undefined;
+/** One completed direct call: its text, usage, terminal finish kind, and chosen effort. */
+export interface StreamOnceResult {
+    text: string;
+    usage: unknown;
+    /** The host's finish kind (`stop`, `max-tokens`, …); `undefined` when the stream omitted it. */
+    finish: string | undefined;
+    /** The capability-derived effort actually sent, for diagnostics. */
+    reasoningEffort: string | undefined;
+}
+/**
+ * Stream one single-turn text call on one route.
+ *
+ * A terminal `error` or `aborted` finish always throws, carrying the host's own
+ * failure text — partial text never rescues a failed call. A `max-tokens` cut
+ * is *not* thrown here: a gate answer often completes before its budget does,
+ * so the caller owns that verdict and reads `finish`.
+ *
+ * @param llm - The LLM service; `undefined` means the host exposed no route at all.
+ * @param route - The exact provider/model to call.
+ * @param sessionId - The host Agent identity providers use for request routing.
+ * @param system - The system prompt for this call.
+ * @param prompt - The user prompt for this call.
+ * @param maxTokens - Output budget for this call.
+ * @returns The assembled text, usage, finish kind, and chosen effort.
+ * @throws On a missing LLM service, capability lookup failure, terminal provider error or abort, and empty output.
+ */
+export declare function streamOnce(llm: LlmRuntime | undefined, route: ModelSelection, sessionId: string | undefined, system: string, prompt: string, maxTokens: number): Promise<StreamOnceResult>;
+//# sourceMappingURL=llm-call.d.ts.map

package/lib/llm-call.js

@@ -0,0 +1,206 @@
+/**
+ * The shared boundary for the plugin's own LLM calls: extraction's cheap
+ * pre-gate and full proposal call, and review's gate and evolve passes all
+ * stream through here.
+ *
+ * Three disciplines live in one place because all three were learned the hard
+ * way, and none of them is visible at a call site:
+ *
+ * - **The route.** `agent.options` carries an *explicit* selection only. A
+ *   session running on the host's default model carries nothing there, so the
+ *   route falls back to the host's `agentDefaultModel` service — the same
+ *   value the host Agent loop itself runs on. A caller that still has no
+ *   route refuses to send instead of streaming into a provider rejection.
+ * - **The reasoning effort.** An exact provider/model advertises the effort
+ *   ids it accepts, and a request naming one it does not accept is rejected
+ *   by capability validation before any provider I/O. The lowest *enabled*
+ *   effort the model advertises is used; a model with no reasoning metadata
+ *   gets no caller-supplied effort at all. This is the pre-teardown
+ *   `llm-text.ts` discipline (see the 2026-09-24 capability note), restored
+ *   after the direct calls that replaced that boundary shipped without it.
+ * - **The failure text.** A terminal `error`/`aborted` finish carries the
+ *   host's own failure object. Keeping only its `kind` produced a ledger full
+ *   of "stream finished with error" that named no cause for a whole day.
+ */
+/** Effort ids that mean "reasoning off"; a model may advertise only these. */
+export const CLOSED_REASONING_EFFORTS = new Set(["disabled", "off", "none"]);
+/**
+ * Select the lowest reasoning effort an exact model advertises.
+ *
+ * An enabled effort wins over a declared closing effort: the adapter's first
+ * enabled effort is the lowest enabled level it exposes, and a closing effort
+ * is only a fallback when the model publishes no enabled level at all. No
+ * capability list means no caller-supplied effort.
+ *
+ * @param modelInfo - Exact resolved-model metadata, when the provider exposes it.
+ * @returns The provider-owned effort id to send, or `undefined` to omit the field.
+ * @throws When the host reports reasoning metadata in a shape this plugin cannot trust.
+ */
+export function selectLowestReasoningEffort(modelInfo) {
+    const efforts = modelInfo?.reasoning?.efforts;
+    if (efforts === undefined) {
+        return undefined;
+    }
+    if (!Array.isArray(efforts)) {
+        throw new Error("continual-evolve: invalid model reasoning metadata");
+    }
+    const ids = [];
+    for (const effort of efforts) {
+        const id = effort?.id;
+        if (typeof id !== "string") {
+            throw new Error("continual-evolve: invalid model reasoning metadata");
+        }
+        ids.push(id);
+    }
+    return ids.find((id) => !CLOSED_REASONING_EFFORTS.has(id)) ?? ids[0];
+}
+/**
+ * The host's default model selection, when the host exposes that service.
+ *
+ * The service is read through `ctx.get` and never joins an inject list: a host
+ * without it (or an older one) must keep working, and a failed read is simply
+ * "no default", never a mount failure.
+ *
+ * @param get - The host context's sibling-service reader, when available.
+ * @returns The default provider/model (plus effort), or `undefined` when unknown.
+ */
+export function defaultRouteOf(get) {
+    if (get === undefined) {
+        return undefined;
+    }
+    let service;
+    try {
+        service = get("agentDefaultModel");
+    }
+    catch {
+        // An absent service is not an error here: the caller falls back to refusing.
+        return undefined;
+    }
+    const currentSelection = service?.currentSelection;
+    if (typeof currentSelection !== "function") {
+        return undefined;
+    }
+    try {
+        return selectionOf(currentSelection.call(service));
+    }
+    catch {
+        // A throwing service is "no default" — the refusal path stays reachable.
+        return undefined;
+    }
+}
+/**
+ * Resolve the route one direct call runs on: the agent's explicit selection
+ * first, the host default second.
+ *
+ * @param agent - The host Agent the call belongs to (its `options` may be empty).
+ * @param get - The host context's sibling-service reader, when available.
+ * @returns The route to call with, or `undefined` when there is none to use.
+ */
+export function resolveCallRoute(agent, get) {
+    const options = agent?.options;
+    return selectionOf(options) ?? defaultRouteOf(get);
+}
+/** Narrow one candidate selection object; empty strings count as absent. */
+function selectionOf(value) {
+    if (typeof value !== "object" || value === null) {
+        return undefined;
+    }
+    const { provider, model, reasoningEffort } = value;
+    if (typeof provider !== "string" || provider.length === 0) {
+        return undefined;
+    }
+    if (typeof model !== "string" || model.length === 0) {
+        return undefined;
+    }
+    return typeof reasoningEffort === "string" && reasoningEffort.length > 0
+        ? { provider, model, reasoningEffort }
+        : { provider, model };
+}
+/**
+ * Stream one single-turn text call on one route.
+ *
+ * A terminal `error` or `aborted` finish always throws, carrying the host's own
+ * failure text — partial text never rescues a failed call. A `max-tokens` cut
+ * is *not* thrown here: a gate answer often completes before its budget does,
+ * so the caller owns that verdict and reads `finish`.
+ *
+ * @param llm - The LLM service; `undefined` means the host exposed no route at all.
+ * @param route - The exact provider/model to call.
+ * @param sessionId - The host Agent identity providers use for request routing.
+ * @param system - The system prompt for this call.
+ * @param prompt - The user prompt for this call.
+ * @param maxTokens - Output budget for this call.
+ * @returns The assembled text, usage, finish kind, and chosen effort.
+ * @throws On a missing LLM service, capability lookup failure, terminal provider error or abort, and empty output.
+ */
+export async function streamOnce(llm, route, sessionId, system, prompt, maxTokens) {
+    if (llm === undefined) {
+        throw new Error("no llm route");
+    }
+    const reasoningEffort = await resolveReasoningEffort(llm, route);
+    let text = "";
+    let usage;
+    let finish;
+    for await (const chunk of llm.stream({
+        provider: route.provider,
+        model: route.model,
+        messages: [{ role: "user", content: [{ type: "text", text: prompt }] }],
+        system,
+        maxTokens,
+        sessionId,
+        ...(reasoningEffort === undefined ? {} : { reasoningEffort }),
+    })) {
+        const piece = chunk;
+        if (piece?.type === "text-delta" && typeof piece.text === "string") {
+            text += piece.text;
+        }
+        else if (piece?.type === "usage") {
+            usage = piece.usage;
+        }
+        else if (piece?.type === "finish") {
+            finish = piece.reason;
+        }
+    }
+    const finishKind = finish?.kind;
+    if (finishKind === "error" || finishKind === "aborted") {
+        throw new Error(failureTextOf(finishKind, finish?.failure));
+    }
+    if (text.trim().length === 0) {
+        // Nothing usable arrived: the finish reason is the only diagnosis.
+        throw new Error(finishKind !== undefined && finishKind !== "stop" ? `stream finished with ${finishKind}` : "empty model output");
+    }
+    return { text, usage, finish: finishKind, reasoningEffort };
+}
+/**
+ * Resolve the effort to send for one exact route.
+ *
+ * The host's capability query is the only source: an explicit selection's
+ * effort belongs to the *main* conversation and must not silently raise an
+ * auxiliary call's cost. No query (older host, test fake) means no effort.
+ */
+async function resolveReasoningEffort(llm, route) {
+    if (typeof llm.resolveModelInfo !== "function") {
+        return undefined;
+    }
+    const modelInfo = await llm.resolveModelInfo(route.provider, route.model);
+    return selectLowestReasoningEffort(modelInfo);
+}
+/** Render a terminal failure for the ledger: host code plus host message. */
+function failureTextOf(kind, failure) {
+    const detail = failureDetailOf(failure);
+    return detail === undefined ? `stream finished with ${kind}` : `LLM call ${kind}: ${detail}`;
+}
+/** Flatten the host failure object to `code: message` when either is present. */
+function failureDetailOf(failure) {
+    if (typeof failure !== "object" || failure === null) {
+        return undefined;
+    }
+    const { code, message } = failure;
+    const head = typeof message === "string" && message.length > 0 ? message : undefined;
+    const label = typeof code === "string" && code.length > 0 ? code : undefined;
+    if (head !== undefined && label !== undefined) {
+        return `${label}: ${head}`;
+    }
+    return head ?? label;
+}
+//# sourceMappingURL=llm-call.js.map

package/lib/review.d.ts

@@ -1,11 +1,16 @@
 import type { MemoryStore } from "./store.js";
-/** The minimal LLM face review calls (same duck-typed host service as extraction). */
-export interface ReviewLlm {
-    stream(options: Record<string, unknown>): AsyncIterable<unknown>;
-}
+import { type LlmRuntime, type RouteResolver } from "./llm-call.js";
+/** The LLM face review calls, aliased so each scheduler can name its own. */
+export type ReviewLlm = LlmRuntime;
 export interface ReviewDeps {
     store: MemoryStore;
-    llm: ReviewLlm | undefined;
+    llm: LlmRuntime | undefined;
+    /**
+     * Route resolution for a direct call: the agent's explicit selection first,
+     * the host default model second. Absent keeps the legacy agent-options-only
+     * behavior (tests and older hosts).
+     */
+    resolveRoute?: RouteResolver | undefined;
 }
 export interface ReviewCandidate {
     workspaceId: string;
@@ -55,6 +60,8 @@ export declare const REWRITE_MAGNITUDE_CAP = 0.3;
  * route failure, so the per-item failure rate read as a network problem.
  */
 export declare const GATE_MAX_TOKENS = 1024;
+/** Output budget for the evolve pass: a verdict plus full replacement text. */
+export declare const EVOLVE_MAX_TOKENS = 4096;
 /** Defaults for the review settings the host may override — and edit live. */
 export declare const REVIEW_DEFAULT_DRY_RUN = true;
 export declare const REVIEW_DEFAULT_DAILY_BUDGET = 100;

package/lib/review.js

@@ -14,6 +14,7 @@
 import { normalizeUsage, zeroUsage } from "./token-usage.js";
 import { evolvePrompt, gatePrompt, parseEvolveAnswer, parseGateAnswer, REVIEW_CHEAP_SYSTEM_PROMPT, REVIEW_MEMORY_SYSTEM_PROMPT, REVIEW_SKILL_SYSTEM_PROMPT, } from "./review-prompt.js";
 import { cwdOf } from "./memory-section.js";
+import { resolveCallRoute, streamOnce } from "./llm-call.js";
 /** A rewrite lands only when the body-length ratio stays within this bound. */
 export const REWRITE_MAGNITUDE_CAP = 0.3;
 /**
@@ -23,6 +24,8 @@ export const REWRITE_MAGNITUDE_CAP = 0.3;
  * route failure, so the per-item failure rate read as a network problem.
  */
 export const GATE_MAX_TOKENS = 1024;
+/** Output budget for the evolve pass: a verdict plus full replacement text. */
+export const EVOLVE_MAX_TOKENS = 4096;
 /** Defaults for the review settings the host may override — and edit live. */
 export const REVIEW_DEFAULT_DRY_RUN = true;
 export const REVIEW_DEFAULT_DAILY_BUDGET = 100;
@@ -153,7 +156,10 @@ export async function runReviewBatch(deps, candidates, opts) {
     const runId = opts.runId ?? `review:${opts.day}:${Date.now()}`;
     let newItems = 0;
     let requeued = 0;
-    const noRoute = deps.llm === undefined || opts.provider === undefined || opts.model === undefined;
+    const route = opts.provider === undefined || opts.model === undefined
+        ? undefined
+        : { provider: opts.provider, model: opts.model };
+    const noRoute = deps.llm === undefined || route === undefined;
     let position = 0;
     for (const candidate of batch) {
         const remaining = deps.store.consumeReviewBudget(opts.day, opts.budget);
@@ -177,7 +183,7 @@ export async function runReviewBatch(deps, candidates, opts) {
         else {
             requeued += 1;
         }
-        if (noRoute) {
+        if (noRoute || route === undefined) {
             deps.store.logReview({
                 runId, workspaceId: candidate.workspaceId, itemKind: candidate.kind, itemId: candidate.id,
                 verdict: "skip", reason: "no-route", applied: 0,
@@ -186,7 +192,7 @@ export async function runReviewBatch(deps, candidates, opts) {
             position += 1;
             continue;
         }
-        await reviewOne(deps, runId, candidate, opts, summary, skipReasons);
+        await reviewOne(deps, runId, candidate, opts, route, summary, skipReasons);
         position += 1;
     }
     summary.backlog = Math.max(0, candidates.length - summary.reviewed - summary.dropped);
@@ -198,7 +204,7 @@ export async function runReviewBatch(deps, candidates, opts) {
     });
     return summary;
 }
-async function reviewOne(deps, runId, candidate, opts, summary, skipReasons) {
+async function reviewOne(deps, runId, candidate, opts, route, summary, skipReasons) {
     const base = {
         runId, workspaceId: candidate.workspaceId, itemKind: candidate.kind, itemId: candidate.id,
     };
@@ -206,7 +212,7 @@ async function reviewOne(deps, runId, candidate, opts, summary, skipReasons) {
     let gateUsage = zeroUsage();
     let gateRaw;
     try {
-        const answer = await streamOnce(deps.llm, opts, REVIEW_CHEAP_SYSTEM_PROMPT, gatePrompt(candidate), GATE_MAX_TOKENS);
+        const answer = await streamOnce(deps.llm, route, opts.sessionId, REVIEW_CHEAP_SYSTEM_PROMPT, gatePrompt(candidate), GATE_MAX_TOKENS);
         gateRaw = answer.usage;
         const gate = parseGateAnswer(answer.text);
         if (gate === null) {
@@ -246,7 +252,7 @@ async function reviewOne(deps, runId, candidate, opts, summary, skipReasons) {
     let evolveRaw;
     try {
         const system = candidate.kind === "memory" ? REVIEW_MEMORY_SYSTEM_PROMPT : REVIEW_SKILL_SYSTEM_PROMPT;
-        const answer = await streamOnce(deps.llm, opts, system, evolvePrompt({ id: candidate.id, title: candidate.title, description: candidate.description, body: candidate.body, extra, scope: candidate.scope }, neighbors), 4096);
+        const answer = await streamOnce(deps.llm, route, opts.sessionId, system, evolvePrompt({ id: candidate.id, title: candidate.title, description: candidate.description, body: candidate.body, extra, scope: candidate.scope }, neighbors), EVOLVE_MAX_TOKENS);
         evolveRaw = answer.usage;
         summary.evolveToken += totalOf(evolveRaw);
         // The evolve pass carries full replacement text: a cut stream may hold
@@ -379,42 +385,6 @@ export function failureHead(error, max = 120) {
 function normalizedOf(raw) {
     return normalizeUsage(raw) ?? zeroUsage();
 }
-/** One streaming text call with the host session identity forwarded. */
-async function streamOnce(llm, opts, system, prompt, maxTokens) {
-    if (llm === undefined) {
-        throw new Error("no llm route");
-    }
-    let text = "";
-    let usage;
-    let finish;
-    for await (const chunk of llm.stream({
-        provider: opts.provider,
-        model: opts.model,
-        messages: [{ role: "user", content: [{ type: "text", text: prompt }] }],
-        system,
-        maxTokens,
-        sessionId: opts.sessionId,
-    })) {
-        const piece = chunk;
-        if (piece?.type === "text-delta" && typeof piece.text === "string") {
-            text += piece.text;
-        }
-        else if (piece?.type === "usage") {
-            usage = piece.usage;
-        }
-        else if (piece?.type === "finish") {
-            finish = piece.reason;
-        }
-    }
-    const finishKind = finish?.kind;
-    if (text.trim().length === 0) {
-        // Nothing usable arrived: the finish reason is the only diagnosis.
-        throw new Error(finishKind !== undefined && finishKind !== "stop" ? `stream finished with ${finishKind}` : "empty model output");
-    }
-    // The caller owns the truncation verdict: a gate answer often completes
-    // before its budget does, a replacement body never does.
-    return { text, usage, finish: finishKind };
-}
 /** Current UTC day `YYYY-MM-DD` for the hard-cap account. */
 export function todayUtc() {
     return new Date().toISOString().slice(0, 10);
@@ -484,20 +454,21 @@ export function createReviewScheduler(host, deps, config) {
         if (typeof id !== "string" || id.length === 0) {
             return;
         }
-        const provider = typeof agent?.options?.provider === "string" ? agent.options.provider : undefined;
-        const model = typeof agent?.options?.model === "string" ? agent.options.model : undefined;
-        lastRoute = { provider, model, sessionId: id };
-        drain(cwdOf(payload.agent));
-    }));
-    listeners.push(host.on("session/event", (session, event) => {
-        if (event?.type !== "compaction/start" || typeof session?.id !== "string") {
+        // A drain needs a live route. A session running on the host default
+        // model carries no explicit selection, so the default-model service
+        // fills it in; with neither, the boundary is only remembered and the
+        // next turn boundary reviews it. This is what keeps a review call
+        // from being sent with an empty provider or a dead session identity
+        // (both of which the provider answers with a bare stream error).
+        // Absent resolver (older host, tests) still reads the agent's explicit
+        // selection: the fallback is the legacy behavior, never "no route".
+        const route = deps.resolveRoute?.(agent) ?? resolveCallRoute(agent, undefined);
+        if (route === undefined) {
+            pendingWorkspace = cwdOf(agent);
             return;
         }
-        drain(pendingWorkspace);
-    }));
-    listeners.push(host.on("agent/disposed", () => {
-        drain(pendingWorkspace);
-        pendingWorkspace = undefined;
+        lastRoute = { provider: route.provider, model: route.model, sessionId: id };
+        drain(cwdOf(agent));
     }));
     const armed = config.read();
     host.logger("continual-evolve").info(`review scheduler armed (daily cap ${armed.dailyBudget}, dry-run ${armed.dryRun ? "on" : "off"})`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-continual-evolve",
-  "version": "0.15.11",
+  "version": "0.15.12",
   "description": "ZCode-style workspace memory for DeepSeek Harness: one central SQLite store partitioned by workspace, index injection, and sole-writer gated writes.",
   "license": "MIT",
   "repository": {