@goodandready/dsh-opencode-provider 0.5.4 → 0.5.6

package/CHANGELOG.md

@@ -1,5 +1,15 @@
 # Changelog
 
+## 0.5.6
+
+- **Persistent sticky routing**: the last successful account/model per conversation survives adapter and Desktop restarts. `~/.dsh/storages/dsh-opencode/sticky.json` stores hashed session ids and key fingerprints only (no keys or prompt text), keeps 500 entries for 14 days.
+- 88 tests.
+
+## 0.5.5
+
+- **Per-account request stats**: every upstream attempt now records its outcome, latency and TTFT in the local request log; the heatmap adds a per-account table with requests, errors and average latency/TTFT. No prompt or response content is stored.
+- 87 tests.
+
 ## 0.5.4
 
 - **Sidebar settings**: the quota-card switch moved out of Connection into its own **Sidebar** section in Settings → OpenCode and into the TUI; saving now refreshes the sidebar card immediately instead of waiting for the 60-second poll.

package/README.md

@@ -88,14 +88,14 @@ graph LR
 - **Per-account quotas**: each account row shows its own 5-hour / weekly / monthly usage, the active account drives the sidebar and quota panel, and the panel lists every account side by side.
 - **Per-account runway**: estimates time to exhaustion from the average burn rate of each window and marks the first window projected to run out before its reset; approximate because the API exposes only the current percentage and reset time.
 - **Model → account rules**: first matching glob pins a preferred account for that model; the other accounts remain fallbacks if it fails or is cooling.
-- **Request heatmap** in the quota panel: a Monday-first weekday × hour grid in local time, plus a 24-hour aggregate strip underneath. Selectors switch between all requests, any single model, and any single account. The adapter appends `{at, provider, model, purpose?, account, reason}` to `~/.dsh/storages/dsh-opencode/requests.jsonl`; no prompt, response or key is stored. Auxiliary `session-title`/`compaction` calls are excluded from the heatmap and exact duplicate records collapse. The log compacts at 2 MB to the newest 5000 lines. Collection starts with this build — the projection cache only has `lastPromptAt`, not per-request timestamps.
+- **Request heatmap** in the quota panel: a Monday-first weekday × hour grid in local time, plus a 24-hour aggregate strip underneath. Selectors switch between all requests, any single model, and any single account. A per-account table shows requests, errors and average latency/TTFT. The adapter appends `{at, provider, model, purpose?, account, reason, ok, latencyMs, ttftMs?, error?}` to `~/.dsh/storages/dsh-opencode/requests.jsonl`; no prompt, response or key is stored. Auxiliary `session-title`/`compaction` calls are excluded from the heatmap and exact duplicate records collapse. The log compacts at 2 MB to the newest 5000 lines. Collection starts with this build — the projection cache only has `lastPromptAt`, not per-request timestamps.
 - **Stable `x-opencode-session`** derived from the durable session id, so retries and every turn of a conversation reuse the same routing key.
 - **Live catalog** from `/v1/models` with a generated `models.dev` fallback.
 - **Per-model subscription limits** from the OpenCode Go pricing table: monthly dollar allowance plus estimated requests for the 5-hour, weekly and monthly windows, for both Go and Go Plus. Shown in the model picker and the settings page; estimates only — the gateway remains the authority.
 - **`/opencode` command** for status, usage, accounts, models, rules, and `test [model]`.
 - **TUI settings section** for the terminal profile (`/settings → OpenCode`) with the key, endpoints, allowlist, active account, fallback model, sticky routing and timeouts.
 - **Quota fallback**: choose a free model in Settings → OpenCode; a Go quota rejection retries once through Console/Zen.
-- **Sticky account**: the last successful account is preferred for the same conversation; explicit model rules and pinned accounts win.
+- **Sticky account**: the last successful account is preferred for the same conversation and survives restarts via a local `sticky.json` (hashed session ids and key fingerprints only); explicit model rules and pinned accounts win.
 - **Routing trace** in the quota panel and per-account heatmap filter.
 - **One-click plugin update** in Settings → OpenCode: current/latest version, an update badge, and an **Update now** button that installs the exact npm version through the DSH CLI and asks for a restart. The host endpoint is `GET/POST /dsh-opencode/update`.
 - **Plan-aware Go picker**: hides models outside the Go table unless turned off; descriptions carry `Go`, `Console` or `FREE`.
@@ -123,12 +123,12 @@ npm pack
 Install the **packed tarball** into a profile:
 
 ```bash
-dsh plugin --profile web add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.4.tgz
+dsh plugin --profile web add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.6.tgz
 # DSH Desktop:
-dsh plugin --profile desktop add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.4.tgz
+dsh plugin --profile desktop add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.6.tgz
 
 # Or install the published package:
-dsh plugin --profile web add @goodandready/dsh-opencode-provider@0.5.4
+dsh plugin --profile web add @goodandready/dsh-opencode-provider@0.5.6
 ```
 
 Restart dsh (or DSH Desktop), then pick `OpenCode` or `OpenCode Go` in the model picker.

package/README.ru.md

@@ -94,9 +94,10 @@ graph LR
 - **Онбординг аккаунтов**: **Import all from auth.json** переносит все ключи OpenCode в отдельные credential ref и добавляет аккаунты; **Check all accounts** обновляет квоты и health по каждому аккаунту.
 - **Правила «модель ➔ аккаунт»**: сопоставление по glob-паттернам закрепляет предпочтительный аккаунт за моделью; остальные аккаунты остаются в резерве.
 - **Квотный fallback**: выберите бесплатную модель в настройках; при ошибке квоты Go тот же запрос один раз повторяется через Console/Zen. Если fallback тоже не сработал, возвращается исходная ошибка квоты.
-- **Липкий аккаунт**: последний успешный аккаунт предпочитается в том же диалоге; явные правила модели, закреплённый аккаунт и квотный fallback имеют приоритет.
+- **Липкий аккаунт**: последний успешный аккаунт предпочитается в том же диалоге и сохраняется между перезапусками в локальном `sticky.json` (только хешированные session id и fingerprint ключей); явные правила модели, закреплённый аккаунт и квотный fallback имеют приоритет.
 - **Трассировка маршрутизации**: панель квот показывает последние решения (`rule`, `sticky`, `rotation`, `quota-fallback`) без промптов и ключей.
-- **Почасовой Heatmap запросов**: сетка «день недели × час» в локальном времени плюс суточная полоса агрегации. Фильтры переключают отображение между всеми запросами, конкретной моделью и аккаунтом. Лог `~/.dsh/storages/dsh-opencode/requests.jsonl` хранит только метаданные `{at, provider, model, purpose, account, reason}` (промпты и ответы не сохраняются).
+- **Статистика по аккаунтам**: в request log пишутся `ok`, `latencyMs`, `ttftMs` и `error`; heatmap показывает таблицу запросов, ошибок и средних latency/TTFT по каждому аккаунту.
+- **Почасовой Heatmap запросов**: сетка «день недели × час» в локальном времени плюс суточная полоса агрегации. Фильтры переключают отображение между всеми запросами, конкретной моделью и аккаунтом. Лог `~/.dsh/storages/dsh-opencode/requests.jsonl` хранит только метаданные `{at, provider, model, purpose, account, reason, ok, latencyMs, ttftMs, error}` (промпты и ответы не сохраняются).
 - **Стабильный заголовок `x-opencode-session`**: вычисляется из устойчивого ID сессии DSH, обеспечивая одинаковый ключ маршрутизации при повторах и диалоге.
 - **Живой каталог моделей**: автоматическое чтение `/v1/models` с резервным fallback из `models.dev`.
 - **Оценки лимитов подписки OpenCode Go**: долларовый лимит и расчетное число запросов для окон 5h, недели и месяца для планов Go и Go Plus.
@@ -121,12 +122,12 @@ npm pack
 Установка из **упакованного tarball**:
 
 ```bash
-dsh plugin --profile web add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.4.tgz
+dsh plugin --profile web add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.6.tgz
 # DSH Desktop:
-dsh plugin --profile desktop add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.4.tgz
+dsh plugin --profile desktop add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.6.tgz
 
 # Либо установка опубликованного пакета из npm:
-dsh plugin --profile web add @goodandready/dsh-opencode-provider@0.5.4
+dsh plugin --profile web add @goodandready/dsh-opencode-provider@0.5.6
 ```
 
 После установки перезапустите dsh (или DSH Desktop) и выберите **OpenCode** или **OpenCode Go** в списке моделей.

package/README.zh.md

@@ -94,9 +94,10 @@ graph LR
 - **每账户预计可用时长**：按各窗口的平均消耗速度估算耗尽时间，并标记在重置前最先耗尽的窗口；API 仅提供当前百分比与重置时间，因此为近似值。
 - **模型 ➔ 账户路由规则**：支持 Glob 通配符（如 `*free*`、`deepseek-*`）将模型优先绑定至特定账户，其他账户保留为备用。
 - **配额回退**：在设置中选择免费模型；Go 返回配额错误时，同一请求通过 Console/Zen 重试一次。若回退也失败，返回原始配额错误。
-- **会话粘性账户**：同一会话优先使用上次成功的账户；显式模型规则、固定账户和配额回退优先。
+- **会话粘性账户**：同一会话优先使用上次成功的账户，并通过本地 `sticky.json` 跨重启保留（仅哈希后的 session id 和密钥指纹）；显式模型规则、固定账户和配额回退优先。
 - **路由追踪**：配额面板显示最近的账户/模型决策（`rule`、`sticky`、`rotation`、`quota-fallback`），不记录提示词或密钥。
-- **本地时间请求热力图**：星期×小时交互式热力图，排除系统心跳与重复记录，可按模型或账户过滤，仅保存 `{at, provider, model, purpose, account, reason}` 元数据，不保存任何提示词与回复。
+- **按账户请求统计**：请求日志记录 `ok`、`latencyMs`、`ttftMs` 和 `error`；热力图表格显示每个账户的请求数、错误数及平均延迟/TTFT。
+- **本地时间请求热力图**：星期×小时交互式热力图，排除系统心跳与重复记录，可按模型或账户过滤，仅保存 `{at, provider, model, purpose, account, reason, ok, latencyMs, ttftMs, error}` 元数据，不保存任何提示词与回复。
 - **稳定的 `x-opencode-session` 标头**：根据 DSH 会话 ID 派生，确保重试与连续对话复用相同的网关路由键。
 - **实时模型目录**：自动从 `/v1/models` 同步，并以 `models.dev` 生成文件作为回退。
 - **Go 订阅额度估算**：每月美元额度以及 5 小时、每周和每月的估算请求次数。
@@ -118,12 +119,12 @@ npm pack
 通过**打包文件 (.tgz)**安装：
 
 ```bash
-dsh plugin --profile web add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.4.tgz
+dsh plugin --profile web add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.6.tgz
 # DSH 桌面端：
-dsh plugin --profile desktop add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.4.tgz
+dsh plugin --profile desktop add /absolute/path/to/goodandready-dsh-opencode-provider-0.5.6.tgz
 
 # 或从 npm 安装已发布版本：
-dsh plugin --profile web add @goodandready/dsh-opencode-provider@0.5.4
+dsh plugin --profile web add @goodandready/dsh-opencode-provider@0.5.6
 ```
 
 安装完成后重启 dsh 或 DSH 桌面端，在模型选择器中选择 **OpenCode** 或 **OpenCode Go** 即可。

package/lib/client.js

@@ -104,6 +104,8 @@ window.__ModuleLoader__.load({
 				avgTtft: "Avg TTFT",
 				avgLatency: "Avg latency",
 				errors: "Errors",
+				latency: "Latency",
+				ttft: "TTFT",
 				requestHeatmap: "Request heatmap",
 				heatmapPeak: "Peak",
 				heatmapEmpty: "No requests recorded yet. Send a message or run Test connection.",
@@ -286,6 +288,8 @@ window.__ModuleLoader__.load({
 				avgTtft: "平均 TTFT",
 				avgLatency: "平均延迟",
 				errors: "错误",
+				latency: "延迟",
+				ttft: "首字延迟",
 				requestHeatmap: "请求热力图",
 				heatmapPeak: "峰值",
 				heatmapEmpty: "尚无请求记录。发送消息或运行测试连接。",
@@ -482,6 +486,11 @@ window.__ModuleLoader__.load({
 		function pad(hour) {
 			return String(hour).padStart(2, "0");
 		}
+		function formatMs(value) {
+			if (value === null || !Number.isFinite(value)) return "—";
+			if (value >= 1e3) return `${(value / 1e3).toFixed(1)}s`;
+			return `${Math.round(value)}ms`;
+		}
 		function RequestHeatmapView(props) {
 			const { heatmap, t } = props;
 			const [selected, setSelected] = (0, react.useState)("all");
@@ -572,6 +581,26 @@ window.__ModuleLoader__.load({
 							})
 						]
 					}),
+					accounts.length === 0 ? null : /* @__PURE__ */ (0, react_jsx_runtime.jsx)("div", {
+						className: "oc-table-scroll",
+						style: { marginTop: 8 },
+						children: /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("table", {
+							className: "oc-table",
+							children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("thead", { children: /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("tr", { children: [
+								/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t("accountLabel") }),
+								/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t("requests") }),
+								/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t("errors") }),
+								/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t("latency") }),
+								/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t("ttft") })
+							] }) }), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("tbody", { children: accounts.map((entry) => /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("tr", { children: [
+								/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", { children: entry.account }),
+								/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", { children: entry.total }),
+								/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", { children: entry.errors }),
+								/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", { children: formatMs(entry.avgLatencyMs) }),
+								/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", { children: formatMs(entry.avgTtftMs) })
+							] }, entry.account)) })]
+						})
+					}),
 					/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
 						className: "oc-heat-scroll",
 						children: [

package/lib/index.d.ts

@@ -116,6 +116,44 @@ declare function createKeyPool(ctx: Context, config: () => ResolvedConfig): KeyP
 /** Backwards-compatible single-key facade over the pool. */
 declare function createKeyResolver(ctx: Context, config: () => ResolvedConfig): KeyResolver;
 //#endregion
+//#region src/request-log.d.ts
+interface RequestOutcome {
+  ok: boolean;
+  latencyMs: number;
+  /** Request start; the log uses it as the event timestamp. */
+  at?: number;
+  ttftMs?: number;
+  error?: string;
+}
+interface ModelRequestHeatmap {
+  provider: string;
+  model: string;
+  total: number;
+  hours: number[];
+  grid: number[][];
+}
+interface AccountRequestHeatmap {
+  account: string;
+  total: number;
+  errors: number;
+  avgLatencyMs: number | null;
+  avgTtftMs: number | null;
+  hours: number[];
+  grid: number[][];
+}
+interface RequestHeatmap {
+  since: number;
+  total: number;
+  /** Requests per hour of day, local time. */
+  hours: number[];
+  /** Monday-first weekday × hour grid, local time. */
+  grid: number[][];
+  firstAt: number | null;
+  lastAt: number | null;
+  models: ModelRequestHeatmap[];
+  accounts: AccountRequestHeatmap[];
+}
+//#endregion
 //#region src/adapter.d.ts
 type RoutingReason = 'rule' | 'sticky' | 'pinned' | 'default' | 'rotation' | 'quota-fallback';
 interface RoutingTraceEntry {
@@ -135,9 +173,11 @@ interface OpenCodeAdapterDeps {
   readonly resolveAttachments?: () => AttachmentStore | undefined;
   readonly fetchImpl?: typeof fetch;
   /** Test seam; defaults to the local append-only request log. */
-  readonly recordRequest?: (provider: string, model: string, purpose?: string, account?: string, reason?: string) => void;
+  readonly recordRequest?: (provider: string, model: string, purpose?: string, account?: string, reason?: string, outcome?: RequestOutcome) => void;
   /** Test seam; defaults to ~/.dsh/storages/dsh-opencode/cooldowns.json. */
   readonly cooldownPath?: string;
+  /** Test seam; defaults to ~/.dsh/storages/dsh-opencode/sticky.json. */
+  readonly stickyPath?: string;
   /** Receives a one-time warning when the provider tool limit drops tools. */
   readonly warn?: (message: string) => void;
 }
@@ -145,7 +185,9 @@ declare class OpenCodeAdapter extends LlmAdapter {
   private readonly deps;
   private readonly fetchImpl;
   private readonly cooldownPath;
+  private readonly stickyPath;
   private cooldownWrites;
+  private stickyWrites;
   /** key fingerprint → cooldown deadline and the rejection that caused it. */
   private readonly cooldowns;
   private readonly sticky;
@@ -163,6 +205,11 @@ declare class OpenCodeAdapter extends LlmAdapter {
   stream(options: GenerateOptions): AsyncIterable<StreamChunk>;
   private run;
   private rememberSticky;
+  /** Await pending sticky persistence writes (tests and shutdown). */
+  flushSticky(): Promise<void>;
+  private loadSticky;
+  private pruneSticky;
+  private persistSticky;
   private pushTrace;
   private parse;
   private send;
@@ -307,33 +354,6 @@ interface AnalyticsReport {
   sessions: SessionUsageRow[];
 }
 //#endregion
-//#region src/request-log.d.ts
-interface ModelRequestHeatmap {
-  provider: string;
-  model: string;
-  total: number;
-  hours: number[];
-  grid: number[][];
-}
-interface AccountRequestHeatmap {
-  account: string;
-  total: number;
-  hours: number[];
-  grid: number[][];
-}
-interface RequestHeatmap {
-  since: number;
-  total: number;
-  /** Requests per hour of day, local time. */
-  hours: number[];
-  /** Monday-first weekday × hour grid, local time. */
-  grid: number[][];
-  firstAt: number | null;
-  lastAt: number | null;
-  models: ModelRequestHeatmap[];
-  accounts: AccountRequestHeatmap[];
-}
-//#endregion
 //#region src/routes.d.ts
 interface RouteDeps {
   readonly config: () => ResolvedConfig;

package/lib/index.js

@@ -17,7 +17,7 @@ const DEFAULT_GO_API_BASE = "https://opencode.ai/inference/go";
 const DEFAULT_GO_ZEN_BASE = "https://opencode.ai/zen/go/v1";
 const DEFAULT_USAGE_URL = "https://opencode.ai/zen/go/v1/usage";
 const DEFAULT_AUTH_ENTRY = "opencode-go";
-const DEFAULT_USER_AGENT = "dsh-opencode-provider/0.5.4";
+const DEFAULT_USER_AGENT = "dsh-opencode-provider/0.5.6";
 const DEFAULT_CONTEXT_WINDOW = 262144;
 const DEFAULT_MAX_TOKENS = 32768;
 const DEFAULT_GENERATE_MAX_TOKENS = 32768;
@@ -1526,14 +1526,20 @@ function requestLogPath() {
 let queue = Promise.resolve();
 let readyDirectory = "";
 let appends = 0;
-function recordRequest(provider, model, at = Date.now(), purpose, account, reason) {
+function recordRequest(provider, model, at = Date.now(), purpose, account, reason, outcome) {
 	const line = `${JSON.stringify({
 		at,
 		provider,
 		model,
 		...purpose === void 0 ? {} : { purpose },
 		...account === void 0 ? {} : { account },
-		...reason === void 0 ? {} : { reason }
+		...reason === void 0 ? {} : { reason },
+		...outcome === void 0 ? {} : {
+			ok: outcome.ok,
+			latencyMs: outcome.latencyMs,
+			...outcome.ttftMs === void 0 ? {} : { ttftMs: outcome.ttftMs },
+			...outcome.error === void 0 ? {} : { error: outcome.error }
+		}
 	})}\n`;
 	queue = queue.then(async () => {
 		try {
@@ -1565,6 +1571,7 @@ async function readRequestHeatmap(days, now = Date.now()) {
 	const grid = Array.from({ length: 7 }, () => Array.from({ length: 24 }, () => 0));
 	const models = /* @__PURE__ */ new Map();
 	const accounts = /* @__PURE__ */ new Map();
+	const accountMetrics = /* @__PURE__ */ new Map();
 	const seen = /* @__PURE__ */ new Set();
 	let total = 0;
 	let firstAt = null;
@@ -1632,12 +1639,37 @@ async function readRequestHeatmap(days, now = Date.now()) {
 			accountEntry = {
 				account,
 				total: 0,
+				errors: 0,
+				avgLatencyMs: null,
+				avgTtftMs: null,
 				hours: Array.from({ length: 24 }, () => 0),
 				grid: Array.from({ length: 7 }, () => Array.from({ length: 24 }, () => 0))
 			};
 			accounts.set(account, accountEntry);
 		}
 		accountEntry.total += 1;
+		let metrics = accountMetrics.get(account);
+		if (metrics === void 0) {
+			metrics = {
+				errors: 0,
+				latencySum: 0,
+				latencyCount: 0,
+				ttftSum: 0,
+				ttftCount: 0
+			};
+			accountMetrics.set(account, metrics);
+		}
+		if (parsed.ok === false) metrics.errors += 1;
+		const latency = parsed.latencyMs;
+		if (typeof latency === "number" && Number.isFinite(latency) && latency >= 0) {
+			metrics.latencySum += latency;
+			metrics.latencyCount += 1;
+		}
+		const ttft = parsed.ttftMs;
+		if (typeof ttft === "number" && Number.isFinite(ttft) && ttft >= 0) {
+			metrics.ttftSum += ttft;
+			metrics.ttftCount += 1;
+		}
 		accountEntry.hours[hour] = (accountEntry.hours[hour] ?? 0) + 1;
 		const accountRow = accountEntry.grid[day];
 		if (accountRow !== void 0) accountRow[hour] = (accountRow[hour] ?? 0) + 1;
@@ -1653,7 +1685,15 @@ async function readRequestHeatmap(days, now = Date.now()) {
 		firstAt,
 		lastAt,
 		models: [...models.values()].sort((a, b) => b.total - a.total || a.model.localeCompare(b.model)),
-		accounts: [...accounts.values()].sort((a, b) => b.total - a.total || a.account.localeCompare(b.account))
+		accounts: [...accounts.values()].map((entry) => {
+			const metrics = accountMetrics.get(entry.account);
+			return {
+				...entry,
+				errors: metrics?.errors ?? 0,
+				avgLatencyMs: metrics !== void 0 && metrics.latencyCount > 0 ? metrics.latencySum / metrics.latencyCount : null,
+				avgTtftMs: metrics !== void 0 && metrics.ttftCount > 0 ? metrics.ttftSum / metrics.ttftCount : null
+			};
+		}).sort((a, b) => b.total - a.total || a.account.localeCompare(b.account))
 	};
 }
 //#endregion
@@ -3036,11 +3076,17 @@ function defaultCooldownPath() {
 	const home = process.env.DSH_HOME?.trim() || join(homedir(), ".dsh");
 	return join(home, "storages", "dsh-opencode", "cooldowns.json");
 }
+function defaultStickyPath() {
+	const home = process.env.DSH_HOME?.trim() || join(homedir(), ".dsh");
+	return join(home, "storages", "dsh-opencode", "sticky.json");
+}
 var OpenCodeAdapter = class extends LlmAdapter {
 	deps;
 	fetchImpl;
 	cooldownPath;
+	stickyPath;
 	cooldownWrites = Promise.resolve();
+	stickyWrites = Promise.resolve();
 	/** key fingerprint → cooldown deadline and the rejection that caused it. */
 	cooldowns = /* @__PURE__ */ new Map();
 	sticky = /* @__PURE__ */ new Map();
@@ -3051,7 +3097,9 @@ var OpenCodeAdapter = class extends LlmAdapter {
 		this.deps = deps;
 		this.fetchImpl = deps.fetchImpl ?? fetch;
 		this.cooldownPath = deps.cooldownPath ?? defaultCooldownPath();
+		this.stickyPath = deps.stickyPath ?? defaultStickyPath();
 		this.loadCooldowns();
+		this.loadSticky();
 	}
 	providerInfo(provider) {
 		return {
@@ -3122,8 +3170,9 @@ var OpenCodeAdapter = class extends LlmAdapter {
 		return this.run(options, false);
 	}
 	async *run(options, fallback) {
-		const record = this.deps.recordRequest ?? ((provider, model, purpose, account, reason) => recordRequest(provider, model, Date.now(), purpose, account, reason));
+		const record = this.deps.recordRequest ?? ((provider, model, purpose, account, reason, outcome) => recordRequest(provider, model, outcome?.at ?? Date.now(), purpose, account, reason, outcome));
 		const config = this.deps.config();
+		const runStartedAt = Date.now();
 		const model = effectiveModel(options.model, config);
 		const sessionId = openCodeSessionId(options);
 		const rawKeys = orderKeysForModel(this.deps.resolveKeys !== void 0 ? await this.deps.resolveKeys(options.provider) : [await this.deps.resolveKey(options.provider)], options.model, config.modelAccountRules);
@@ -3143,7 +3192,6 @@ var OpenCodeAdapter = class extends LlmAdapter {
 			}
 		}
 		const initial = keys.find((entry) => fallback || !this.coolingDown(entry.key)) ?? keys[0];
-		if (initial !== void 0) record(options.provider, options.model, options.purpose, initial.label ?? "default", reason);
 		const url = endpointFor(options.provider, model.transport, options.model, config);
 		let images;
 		if (hasImages(options)) {
@@ -3166,18 +3214,35 @@ var OpenCodeAdapter = class extends LlmAdapter {
 		const body = model.transport === "openai-chat" ? buildChatBody(bodyOptions, model, config, images) : model.transport === "anthropic" ? buildAnthropicBody(bodyOptions, model, config, images) : model.transport === "openai-responses" ? buildResponsesBody(bodyOptions, model, config, images) : buildGoogleBody(bodyOptions, model, config, images);
 		let lastError;
 		let lastAccount = initial?.label ?? "default";
+		let loggedAny = false;
 		for (const [index, entry] of keys.entries()) {
 			if (!fallback && this.coolingDown(entry.key)) continue;
 			lastAccount = entry.label ?? "default";
 			const attemptReason = index === 0 ? reason : "rotation";
+			const attemptStartedAt = Date.now();
 			let started = false;
 			let responseFailure = false;
+			let firstChunkAt;
+			let logged = false;
+			const logAttempt = (ok, errorCode) => {
+				if (logged) return;
+				logged = true;
+				loggedAny = true;
+				record(options.provider, options.model, options.purpose, entry.label ?? "default", attemptReason, {
+					at: attemptStartedAt,
+					ok,
+					latencyMs: Math.max(0, Date.now() - attemptStartedAt),
+					...firstChunkAt === void 0 ? {} : { ttftMs: Math.max(0, firstChunkAt - attemptStartedAt) },
+					...errorCode === void 0 ? {} : { error: errorCode }
+				});
+			};
 			try {
 				const response = await this.send(url, entry.key, sessionId, model.transport, body, config, options.signal);
 				if (!response.ok) {
 					const text = await response.text().catch(() => "");
 					const error = classifyResponseError(response.status, text, response.headers);
 					lastError = error;
+					logAttempt(false, error.code);
 					const canRotate = index < keys.length - 1;
 					if (error.code === "INVALID_CREDENTIAL" || error.code === "RATE_LIMIT" || error.code === "QUOTA") {
 						if (!fallback) this.markCooldown(entry.key, response.status, response.headers, error.code);
@@ -3197,13 +3262,16 @@ var OpenCodeAdapter = class extends LlmAdapter {
 					ok: true
 				});
 				for await (const chunk of this.parse(response, model.transport, config, options.signal)) {
+					if (firstChunkAt === void 0) firstChunkAt = Date.now();
 					started = true;
 					yield chunk;
 				}
+				logAttempt(true);
 				return;
 			} catch (error) {
-				if (started || options.signal?.aborted === true) throw error;
 				const classified = error instanceof LlmError ? error : new LlmError(error instanceof Error ? error.message : "OpenCode request failed", "TRANSPORT", { cause: error });
+				logAttempt(false, classified.code);
+				if (started || options.signal?.aborted === true) throw error;
 				lastError = classified;
 				const transient = this.isTransient(classified);
 				if (transient && !responseFailure && !fallback) this.markCooldown(entry.key, 0);
@@ -3214,6 +3282,12 @@ var OpenCodeAdapter = class extends LlmAdapter {
 			}
 		}
 		const finalError = lastError ?? new LlmError("llm-opencode: every configured account is cooling down after provider rejections", "RATE_LIMIT");
+		if (!loggedAny) record(options.provider, options.model, options.purpose, lastAccount, reason, {
+			at: runStartedAt,
+			ok: false,
+			latencyMs: 0,
+			error: finalError.code
+		});
 		this.pushTrace({
 			at: Date.now(),
 			provider: options.provider,
@@ -3249,6 +3323,55 @@ var OpenCodeAdapter = class extends LlmAdapter {
 			model: options.model,
 			at: Date.now()
 		});
+		this.persistSticky();
+	}
+	/** Await pending sticky persistence writes (tests and shutdown). */
+	async flushSticky() {
+		await this.stickyWrites;
+	}
+	loadSticky() {
+		try {
+			const entries = JSON.parse(readFileSync(this.stickyPath, "utf8")).sticky;
+			if (typeof entries !== "object" || entries === null || Array.isArray(entries)) return;
+			for (const [sessionId, value] of Object.entries(entries)) {
+				if (typeof value !== "object" || value === null || Array.isArray(value)) continue;
+				const record = value;
+				const fingerprint = record.fingerprint;
+				const label = record.label;
+				const provider = record.provider;
+				const model = record.model;
+				const at = record.at;
+				if (typeof fingerprint !== "string" || !/^[0-9a-f]{16}$/.test(fingerprint)) continue;
+				if (typeof label !== "string" || label === "") continue;
+				if (typeof provider !== "string" || provider === "") continue;
+				if (typeof model !== "string" || model === "") continue;
+				if (typeof at !== "number" || !Number.isFinite(at)) continue;
+				this.sticky.set(sessionId, {
+					fingerprint,
+					label,
+					provider,
+					model,
+					at
+				});
+			}
+			this.pruneSticky();
+		} catch {}
+	}
+	pruneSticky() {
+		const cutoff = Date.now() - 12096e5;
+		for (const [sessionId, entry] of this.sticky) if (entry.at < cutoff) this.sticky.delete(sessionId);
+		const overflow = this.sticky.size - 500;
+		if (overflow <= 0) return;
+		const oldest = [...this.sticky.entries()].sort((a, b) => a[1].at - b[1].at).slice(0, overflow);
+		for (const [sessionId] of oldest) this.sticky.delete(sessionId);
+	}
+	persistSticky() {
+		this.pruneSticky();
+		const payload = JSON.stringify({ sticky: Object.fromEntries(this.sticky) });
+		this.stickyWrites = this.stickyWrites.then(async () => {
+			await mkdir(dirname(this.stickyPath), { recursive: true });
+			await writeFile(this.stickyPath, payload, "utf8");
+		}).catch(() => {});
 	}
 	pushTrace(entry) {
 		this.routing.unshift(entry);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-opencode-provider",
-  "version": "0.5.4",
+  "version": "0.5.6",
   "description": "OpenCode / OpenCode Go LLM provider plugin for DeepSeek Harness (dsh).",
   "license": "MIT",
   "type": "module",