dsh-qoder-connect 0.1.2 → 0.1.4

package/lib/index.js

@@ -1,4 +1,4 @@
-import { A as QODER_GLOBAL_AUTH_PATH, B as qoderCredentialIdentity, C as probeModel, D as FALLBACK_QODER_MODELS, E as qoderCatalogPath, F as QODER_AUTH_FILENAME, G as qoderPluginDataDir, H as QODER_DATA_DIR_ENV, I as QODER_PAT_ENV_CN, K as qoderStateDir, L as QODER_PAT_ENV_GLOBAL, M as QODER_GLOBAL_STATUS_PATH, N as QODER_PROBE_PATH, O as QoderCatalog, P as QODER_STATUS_PATH, R as QoderCredentialStore, S as PROBE_EFFORT_CANDIDATES, T as QoderCatalogStore, U as QODER_DATA_DIR_NAME, V as qoderOwnAuthPath, W as qoderMachineIdPath, _ as modelInfoOf, a as QODER_HOST_HEARTBEAT_FILENAME, b as createQoderTransport, c as processStartTimeMs, d as writeHostHeartbeat, f as QODER_CONNECT_VERSION, g as kindFromQoderFailure, h as classifyUpstreamError, i as variantFor, j as QODER_GLOBAL_PROBE_PATH, k as QODER_AUTH_PATH, l as qoderHostHeartbeatPath, m as QoderUpstreamClient, n as GLOBAL_VARIANT, o as clearHostHeartbeat, p as KIND_STATUS, r as QODER_VARIANTS, s as isHeartbeatProcessAlive, t as CHINA_VARIANT, u as readHostHeartbeat, v as normalizeCredits, w as randomSentinel, x as getMachineId, y as validateApiKey, z as patTail } from "./variants-C0iZD8AD.js";
+import { A as QODER_GLOBAL_AUTH_PATH, B as qoderCredentialIdentity, C as probeModel, D as FALLBACK_QODER_MODELS, E as qoderCatalogPath, F as QODER_AUTH_FILENAME, G as qoderPluginDataDir, H as QODER_DATA_DIR_ENV, I as QODER_PAT_ENV_CN, K as qoderStateDir, L as QODER_PAT_ENV_GLOBAL, M as QODER_GLOBAL_STATUS_PATH, N as QODER_PROBE_PATH, O as QoderCatalog, P as QODER_STATUS_PATH, R as QoderCredentialStore, S as PROBE_EFFORT_CANDIDATES, T as QoderCatalogStore, U as QODER_DATA_DIR_NAME, V as qoderOwnAuthPath, W as qoderMachineIdPath, _ as modelInfoOf, a as QODER_HOST_HEARTBEAT_FILENAME, b as createQoderTransport, c as processStartTimeMs, d as writeHostHeartbeat, f as QODER_CONNECT_VERSION, g as kindFromQoderFailure, h as classifyUpstreamError, i as variantFor, j as QODER_GLOBAL_PROBE_PATH, k as QODER_AUTH_PATH, l as qoderHostHeartbeatPath, m as QoderUpstreamClient, n as GLOBAL_VARIANT, o as clearHostHeartbeat, p as KIND_STATUS, r as QODER_VARIANTS, s as isHeartbeatProcessAlive, t as CHINA_VARIANT, u as readHostHeartbeat, v as normalizeCredits, w as randomSentinel, x as getMachineId, y as validateApiKey, z as patTail } from "./variants-CSeJqJpR.js";
 import z from "@deepseek-ai/schemastery";
 import { createHash, randomBytes, timingSafeEqual } from "node:crypto";
 import { dirname, join, resolve } from "node:path";
@@ -1290,16 +1290,42 @@ function emitJobTokenHint(at, text) {
 	const agent = lastRunningAgent;
 	if (agent?.session === void 0) return;
 	lastRefreshAt.set(String(agent.id), at);
+	appendHintRow(agent, "success", text);
+}
+/**
+* Print the failed-heal row into the conversation whose request triggered it.
+*
+* The mirror of {@link emitJobTokenHint}: the self-heal ran, exchanged a fresh
+* job token, retried, and the upstream still refused. Without this row that
+* outcome was silent — the user saw only the resulting authorization failure
+* and could not tell that recovery had already been attempted and lost.
+*
+* Rate limited by the transport (once per unresolved outage), not here: this
+* module stays a pure printer so the dedupe policy lives with the state that
+* knows when an outage ends.
+*
+* @param at - When the failed heal was recorded.
+* @param text - The row's summary line.
+*/
+function emitJobTokenRefreshFailedHint(at, text) {
+	const agent = lastRunningAgent;
+	if (agent?.session === void 0) return;
+	appendHintRow(agent, "error", text);
+}
+/** Append one log-only command pair carrying `text` as its outcome. */
+function appendHintRow(agent, kind, text) {
+	const session = agent.session;
+	if (session === void 0) return;
 	const commandId = `cmd-qoder-hint-${Date.now().toString(36)}-${++hintSeq}`;
 	try {
-		agent.session.append("command/run", {
+		session.append("command/run", {
 			commandId,
 			name: JOB_TOKEN_HINT_NAME,
 			source: { kind: "user" }
 		});
-		agent.session.append("command/done", {
+		session.append("command/done", {
 			commandId,
-			kind: "success",
+			kind,
 			text
 		});
 	} catch {}
@@ -1475,6 +1501,17 @@ function fallbackFor(_variant) {
 function jobTokenHintText(at) {
 	return `jobToken 已自动刷新（${new Date(at).toLocaleTimeString("zh-CN", { hour12: false })}）— 旧令牌被上游拒绝，已自动重换并恢复`;
 }
+/**
+* The failed-heal row's summary line.
+*
+* Deliberately does not name a cause: the upstream rejection that survived a
+* fresh token is not necessarily an authorization problem at all, and the real
+* reason travels in the failure message itself (see the SSE envelope body).
+* Claiming "quota" or "revoked" here would be a guess presented as a finding.
+*/
+function jobTokenRefreshFailedHintText(at) {
+	return `jobToken 已重换但仍被上游拒绝（${new Date(at).toLocaleTimeString("zh-CN", { hour12: false })}）— 自愈未恢复，请查看上方错误详情`;
+}
 /** Build one variant's stores, transport, and probe state. */
 function createVariantRuntime(ctx, config, variant, current, identityOf) {
 	const store = new QoderCredentialStore({
@@ -1490,6 +1527,10 @@ function createVariantRuntime(ctx, config, variant, current, identityOf) {
 			jobTokenRefreshedAt = info.at;
 			ctx.logger.warn(`dsh-qoder-connect: ${variant.displayName} job token was auto-refreshed after an upstream rejection`);
 			emitJobTokenHint(info.at, jobTokenHintText(info.at));
+		},
+		onJobTokenRefreshFailed: (info) => {
+			ctx.logger.warn(`dsh-qoder-connect: ${variant.displayName} job token refresh did not recover the chat (upstream status ${info.status ?? "unknown"})`);
+			emitJobTokenRefreshFailedHint(info.at, jobTokenRefreshFailedHintText(info.at));
 		}
 	});
 	const client = new QoderUpstreamClient({

package/lib/{variants-C0iZD8AD.js → variants-CSeJqJpR.js}

@@ -2623,7 +2623,10 @@ function parseEnvelope(rawData) {
 	const envelope = value;
 	if (envelope.statusCodeValue !== void 0 && typeof envelope.statusCodeValue !== "number") throw malformed("Qoder SSE envelope has an invalid status code.");
 	if (envelope.statusCodeValue !== void 0 && (!Number.isInteger(envelope.statusCodeValue) || envelope.statusCodeValue < 100 || envelope.statusCodeValue > 599)) throw malformed("Qoder SSE envelope has an invalid status code.");
-	if (envelope.statusCodeValue !== void 0 && envelope.statusCodeValue !== 200) throw qoderHttpError(`Qoder service returned upstream error status ${envelope.statusCodeValue}.`, { status: envelope.statusCodeValue });
+	if (envelope.statusCodeValue !== void 0 && envelope.statusCodeValue !== 200) {
+		const status = envelope.statusCodeValue;
+		throw qoderHttpError(envelope.body ? `Qoder service returned upstream error status ${status}: ${envelope.body}` : `Qoder service returned upstream error status ${status}.`, { status });
+	}
 	if (envelope.body !== void 0 && typeof envelope.body !== "string") throw malformed("Qoder SSE envelope has an invalid model body.");
 	return envelope;
 }
@@ -3030,6 +3033,18 @@ async function* streamQoderChat(options, model, credentials, messages, dependenc
 		requestController.abort("request complete");
 	}
 }
+/**
+* How long a rotated-token notice may wait for an accepting chat.
+*
+* The heal's own retry can lose a race with a transient fault and be rescued
+* by a later attempt, so the notice is deferred until some chat is accepted.
+* That deferral has to be bounded: an unbounded one let a rotation from
+* minutes earlier surface as a fresh-looking row whose timestamp named a
+* moment the user could not connect to the row appearing now. Five minutes
+* comfortably covers the host's own retry cycle (its backoff caps at 10s)
+* while keeping the row adjacent to the event it describes.
+*/
+const jobTokenNoticeMaxAgeMs = 3e5;
 function aborted(message) {
 	return new QoderLlmError(message, "ABORTED");
 }
@@ -3048,6 +3063,7 @@ var DefaultQoderTransport = class {
 	imageUploader;
 	preserveThinking;
 	onJobTokenRefreshed;
+	onJobTokenRefreshFailed;
 	/**
 	* Set when the self-heal exchanged a fresh job token, cleared when a chat
 	* afterwards succeeds. The notice reports "the token was rotated because the
@@ -3056,6 +3072,16 @@ var DefaultQoderTransport = class {
 	* transient upstream timeout and be rescued by a later attempt.
 	*/
 	pendingRefreshAt;
+	/**
+	* Whether the current unresolved heal failure has already been reported.
+	*
+	* One upstream rejection can outlive many host-level retries (an observed
+	* storm ran 59 of them across 75 steps), and the transport re-heals inside
+	* every one of them. Without this latch the failure notice would print once
+	* per retry; with it, the user is told once per outage. Cleared as soon as
+	* any chat is accepted, so the next outage announces itself again.
+	*/
+	refreshFailureAnnounced = false;
 	constructor(options) {
 		this.region = options.region;
 		this.resolvePat = options.resolvePat;
@@ -3067,6 +3093,7 @@ var DefaultQoderTransport = class {
 		this.attachments = options.attachments;
 		this.preserveThinking = options.preserveThinking;
 		this.onJobTokenRefreshed = options.onJobTokenRefreshed;
+		this.onJobTokenRefreshFailed = options.onJobTokenRefreshFailed;
 		this.auth = new QoderAuthService({
 			fetch: this.fetchImpl,
 			logger: this.logger,
@@ -3107,12 +3134,35 @@ var DefaultQoderTransport = class {
 		if (this.pendingRefreshAt === void 0) return;
 		const at = this.pendingRefreshAt;
 		this.pendingRefreshAt = void 0;
+		if (Date.now() - at > jobTokenNoticeMaxAgeMs) return;
 		this.logger?.warn?.("[Qoder Stream] Job token was auto-refreshed after an upstream rejection; the chat has recovered");
 		this.onJobTokenRefreshed?.({
 			region: this.region,
 			at
 		});
 	}
+	/**
+	* Drop a pending rotation notice whose heal did not rescue anything.
+	*
+	* The heal's own retry was rejected too, so "the rotation fixed it" is not
+	* what happened. Leaving the notice pending made a LATER, unrelated chat
+	* acceptance flush it — printing a success row that contradicts the failure
+	* row already shown, stamped with the old rotation time.
+	*/
+	discardPendingRefreshNotice() {
+		this.pendingRefreshAt = void 0;
+	}
+	/**
+	* Note that the upstream accepted a chat, ending any unresolved heal failure.
+	*
+	* `streamChat` calls this the moment its first chunk arrives — the only point
+	* at which "the credential was accepted" is actually known. A rejected chat
+	* throws before that, so this never fires on a failing attempt.
+	*/
+	onChatAccepted() {
+		this.refreshFailureAnnounced = false;
+		this.flushPendingRefreshNotice();
+	}
 	async discoverModels(signal) {
 		const pat = await this.requirePat(signal);
 		const key = opaqueCredentialKey(pat);
@@ -3178,7 +3228,7 @@ var DefaultQoderTransport = class {
 		const chatCredentials = await this.auth.getCredentials(pat, options.signal);
 		try {
 			yield* this.streamChat(options, model, chatCredentials, messages, () => {
-				this.flushPendingRefreshNotice();
+				this.onChatAccepted();
 			});
 			return;
 		} catch (error) {
@@ -3191,7 +3241,7 @@ var DefaultQoderTransport = class {
 			this.pendingRefreshAt = Date.now();
 			try {
 				yield* this.streamChat(options, model, refreshed, messages, () => {
-					this.flushPendingRefreshNotice();
+					this.onChatAccepted();
 				});
 				return;
 			} catch (error) {
@@ -3212,6 +3262,16 @@ var DefaultQoderTransport = class {
 				}
 			}
 		}
+		this.discardPendingRefreshNotice();
+		if (!this.refreshFailureAnnounced) {
+			this.refreshFailureAnnounced = true;
+			this.logger?.warn?.("[Qoder Stream] Job token refresh did not recover the chat; the upstream still rejects it", { status: lastRejection instanceof LlmError ? lastRejection.failure.status : void 0 });
+			this.onJobTokenRefreshFailed?.({
+				region: this.region,
+				at: Date.now(),
+				...lastRejection instanceof LlmError && lastRejection.failure.status !== void 0 ? { status: lastRejection.failure.status } : {}
+			});
+		}
 		throw lastRejection;
 	}
 };
@@ -3618,14 +3678,23 @@ var ChunkEncoder = class {
 				}] })];
 			}
 			case "usage": {
-				const prompt = chunk.usage.inputTokens + (chunk.usage.cacheReadTokens ?? 0) + (chunk.usage.cacheWriteTokens ?? 0);
+				const cacheRead = chunk.usage.cacheReadTokens;
+				const cacheWrite = chunk.usage.cacheWriteTokens;
+				const prompt = chunk.usage.inputTokens + (cacheRead ?? 0) + (cacheWrite ?? 0);
 				const completion = chunk.usage.outputTokens;
+				const reasoning = chunk.usage.reasoningTokens;
+				const details = cacheRead === void 0 && cacheWrite === void 0 ? void 0 : {
+					...cacheRead === void 0 ? {} : { cached_tokens: cacheRead },
+					...cacheWrite === void 0 ? {} : { cache_write_tokens: cacheWrite }
+				};
 				return [this.frame({
 					choices: [],
 					usage: {
 						prompt_tokens: prompt,
 						completion_tokens: completion,
-						total_tokens: chunk.usage.totalTokens ?? prompt + completion
+						total_tokens: chunk.usage.totalTokens ?? prompt + completion,
+						...details === void 0 ? {} : { prompt_tokens_details: details },
+						...reasoning === void 0 ? {} : { completion_tokens_details: { reasoning_tokens: reasoning } }
 					}
 				})];
 			}
@@ -4013,7 +4082,7 @@ const DEFAULT_MAX_TOKENS = 32768;
 const DEFAULT_CONTEXT_WINDOW = 18e4;
 //#endregion
 //#region src/version.ts
-const QODER_CONNECT_VERSION = "0.1.2";
+const QODER_CONNECT_VERSION = "0.1.4";
 //#endregion
 //#region src/host-heartbeat.ts
 /**

package/package.json

@@ -2,7 +2,7 @@
   "name": "dsh-qoder-connect",
   "displayName": "DSH Qoder Connect",
   "description": "将 Qoder（国内版 / 国际版）的模型接入 DeepSeek Harness — bring Qoder (CN & international) models into DeepSeek Harness.",
-  "version": "0.1.2",
+  "version": "0.1.4",
   "author": "corrinehu",
   "keywords": [
     "dsh-plugin",
@@ -23,7 +23,7 @@
   "scripts": {
     "prepack": "npm run build",
     "build": "tsdown",
-    "test": "vitest run",
+    "test": "node -r ./tests/sandbox-shim.cjs ./node_modules/vitest/vitest.mjs run",
     "test:qoder": "node --test --test-isolation=none --experimental-strip-types tests/qoder/*.test.ts",
     "typecheck": "tsc -p tsconfig.json && tsc -p tsconfig.client.json",
     "check": "pnpm run typecheck && pnpm run test && pnpm run test:qoder && pnpm run build"