@dan-ai-studio/dshopencodego 0.1.84 → 0.1.86

package/README.en.md

@@ -201,7 +201,7 @@ The gateway's `/v1/models` answers only *which* models exist, never *how* to cal
 - the bundled catalog is **frozen at the plugin's version** and affects only the quality of the "how to call it" decision, never *which* models exist.
 - **fresh installs do not wait**: the bundle carries a directory snapshot from publish time (`src/catalog/seed.snapshot.json`), which answers at once when the disk cache is empty and refreshes behind it; past 24 hours without a refresh counts as stale and refreshes in the background too, whatever the TTL says. A timer re-checks on a live host, so long-running processes stay current without a request to trigger it.
 
-When the gateway is **temporarily unavailable**, the plugin keeps the last listing that succeeded, marks it stale on the page, and retries in the background. The retry rate is governed by a **five-minute back-off** rather than by the TTL, so a gateway that comes back is picked up within five minutes; while it stays down it is asked once, not once per model request. The settings page's Recheck button revalidates at any time and is not bound by that back-off.
+When the gateway is **temporarily unavailable**, the plugin keeps the last listing that succeeded, marks it stale on the page, and retries in the background. The retry rate is governed by a **five-minute back-off** rather than by the TTL, so a gateway that comes back is picked up within five minutes; while it stays down it is asked once, not once per model request. When the settings page opens on a stale reading, it follows up on its own (about every 5 seconds) until the reading turns current — the follow-ups are cheap local reads, and the host's back-off still decides when the gateway is really asked — so there is no need to press Recheck by hand. The settings page's Recheck button revalidates at any time and is not bound by that back-off.
 
 The only thing that waits on a plugin release is a missing wire quirk in the pi-ai catalog (DeepSeek's `thinkingFormat`, say) — the model itself is already usable.
 
@@ -291,7 +291,7 @@ npm run seed        # refresh the bundled catalog seed before a release (see "Re
 
 **Order matters**: `npm run build` must come before `npm test` — `tests/pack-contents.spec.ts` walks every `exports` target, and whether `lib/` exists decides whether that check runs or is silently skipped (it does not fail, it just stops protecting you). `package.json` no longer carries `prepare`, so `npm ci` will not build for you; all three CI workflows run `npm run build` explicitly.
 
-The suite covers configuration validation, the protocol ladder, gateway and online metadata, the catalog projection, the adapter on the wire (all three protocols), Remote contracts and services, session-header invariants, usage windows and the meter, the client controller and usage pill, both halves' mounting and registration, the error mapping for stream idle timeouts and aborts, the exit-code contract of the verification scripts, and several **structural invariants** (inject declarations match the source's reads, peer coverage holds both ways, the READMEs match the manifest, and the READMEs' claims about the upstream match the upstream actually installed). That is **60 test files** — a number `tests/readme-consistency.spec.ts` checks against the `tests/` directory, so it cannot drift.
+The suite covers configuration validation, the protocol ladder, gateway and online metadata, the catalog projection, the adapter on the wire (all three protocols), Remote contracts and services, session-header invariants, usage windows and the meter, the client controller and usage pill, both halves' mounting and registration, the error mapping for stream idle timeouts and aborts, the exit-code contract of the verification scripts, and several **structural invariants** (inject declarations match the source's reads, peer coverage holds both ways, the READMEs match the manifest, and the READMEs' claims about the upstream match the upstream actually installed). That is **61 test files** — a number `tests/readme-consistency.spec.ts` checks against the `tests/` directory, so it cannot drift.
 
 ## Releasing (maintainers)
 

package/README.md

@@ -200,7 +200,7 @@ API Key 通过 Harness 凭证库提供（引用名 `OPENCODE_GO_API_KEY`），
 - 内置目录**冻结**在插件版本里，仅影响"怎么调"的推断质量，不影响"有哪些"。
 - **首次安装不等待**：包里带了一份发布时生成的目录快照（`src/catalog/seed.snapshot.json`，28 个模型上下），磁盘缓存为空时直接用它回答并在后台刷新；过期（超 24 小时未刷新）同样后台刷新，绝不为等网络拦住请求。插件挂载后还会定时触发一次，保证长期运行的进程目录常新。
 
-网关**临时不可用**时，插件保留上次成功获取的列表并在页面上标注陈旧，同时在后台重试：重试由**5 分钟的退避**控制，而不是等满一个 TTL，所以网关恢复后 5 分钟内列表就会更新；网关持续不可用时，它只被问一次，不会每次模型请求都去问一遍。设置页的「重新检查」随时可以强制重验，不受这个退避限制。
+网关**临时不可用**时，插件保留上次成功获取的列表并在页面上标注陈旧，同时在后台重试：重试由**5 分钟的退避**控制，而不是等满一个 TTL，所以网关恢复后 5 分钟内列表就会更新；网关持续不可用时，它只被问一次，不会每次模型请求都去问一遍。设置页打开时若读到的是旧目录，会自动跟进重读直到拿到最新的（约每 5 秒一次读本地快照；何时真问网关由 Host 的 5 分钟退避决定，不会因此多发请求），不用手动点「重新检查」；「重新检查」随时可以强制重验，不受这个退避限制。
 
 只有当 pi-ai 内置目录**缺了某个模型的线怪癖**（如 DeepSeek 的 `thinkingFormat`）时，才需要等插件同步 pi-ai 的新版本——而模型本身早已可用。
 
@@ -290,7 +290,7 @@ npm run seed        # 发版前刷新捆绑的目录种子（见「发布」）
 
 **顺序有讲究**：`npm run build` 要排在 `npm test` 之前——`tests/pack-contents.spec.ts` 会遍历 `exports` 的每个目标，`lib/` 存在与否决定这条检查是执行还是被跳过（不报错，但会少一层保护）。`package.json` 已移除 `prepare`，因此 `npm ci` 不会再替你隐式构建；三个 CI workflow 都显式跑了 `npm run build`。
 
-测试覆盖配置校验、协议阶梯、网关与在线元数据、目录投影、适配器（三种协议的线上请求）、Remote 契约与服务、会话头不变量、用量窗口与计量、客户端控制器与用量组件、两半的挂载与登记、流空闲超时与中止的错误映射、验证脚本的退出码契约、以及几条**结构性不变量**（inject 声明与源码读取一致、peer 覆盖双向、README 与 manifest 一致、README 对上游的断言与实装的上游一致）。共 **60 个测试文件**——这个数字由 `tests/readme-consistency.spec.ts` 对着 `tests/` 目录校验，不会漂移。
+测试覆盖配置校验、协议阶梯、网关与在线元数据、目录投影、适配器（三种协议的线上请求）、Remote 契约与服务、会话头不变量、用量窗口与计量、客户端控制器与用量组件、两半的挂载与登记、流空闲超时与中止的错误映射、验证脚本的退出码契约、以及几条**结构性不变量**（inject 声明与源码读取一致、peer 覆盖双向、README 与 manifest 一致、README 对上游的断言与实装的上游一致）。共 **61 个测试文件**——这个数字由 `tests/readme-consistency.spec.ts` 对着 `tests/` 目录校验，不会漂移。
 
 ## 发布（维护者）
 

package/lib/client.js

@@ -208,6 +208,10 @@ var networkRemote = {
   ]
 };
 
+// src/request-log/store.ts
+var MAX_STREAM_MARKS = 200;
+var MIN_STREAM_MS_FOR_RATE = 100;
+
 // src/request-log/contract.ts
 function count(key, row, optional) {
   const entry = row[key];
@@ -254,6 +258,8 @@ function parseEntry(value) {
     ...row["costUsd"] === void 0 ? {} : { costUsd: rate("costUsd", row) },
     ...row["requestBytes"] === void 0 ? {} : { requestBytes: count("requestBytes", row) },
     ...row["prepMs"] === void 0 ? {} : { prepMs: count("prepMs", row) },
+    ...row["responseMs"] === void 0 ? {} : { responseMs: count("responseMs", row) },
+    ...row["streamMarksMs"] === void 0 ? {} : { streamMarksMs: parseMarks(row["streamMarksMs"]) },
     ...row["imageBytes"] === void 0 ? {} : { imageBytes: count("imageBytes", row) },
     ...row["durationMs"] === void 0 ? {} : { durationMs: count("durationMs", row) },
     ...row["firstTokenMs"] === void 0 ? {} : { firstTokenMs: count("firstTokenMs", row) },
@@ -281,6 +287,15 @@ function parseEntry(value) {
     } : {}
   };
 }
+function parseMarks(value) {
+  if (!Array.isArray(value)) throw new Error("invalid request-log stream marks");
+  const kept = [];
+  for (const mark of value) {
+    if (typeof mark !== "number" || !Number.isFinite(mark) || mark < 0 || kept.length >= MAX_STREAM_MARKS) continue;
+    kept.push(mark);
+  }
+  return kept;
+}
 function rate(key, row) {
   const entry = row[key];
   if (typeof entry !== "number" || !Number.isFinite(entry) || entry < 0) {
@@ -569,6 +584,7 @@ var en = {
   requestLogColPrep: "Prep",
   requestLogColWait: "Wait",
   requestLogTimelineWait: "Waiting for first token",
+  requestLogTimelineFirstByte: "First byte",
   requestLogTimelineStream: "Streaming",
   requestLogTimelineBurst: "Delivered in one burst",
   requestLogColStream: "Content span",
@@ -787,6 +803,7 @@ var zh = {
   requestLogDetailUrl: "\u5730\u5740",
   requestLogDetailRequestBytes: "\u8BF7\u6C42\u4F53",
   requestLogDetailPrep: "\u51C6\u5907\u8017\u65F6",
+  requestLogTimelineFirstByte: "\u9996\u5B57\u8282",
   requestLogDetailContentSplit: "\u5185\u5BB9\u62C6\u5206",
   requestLogDetailText: "\u6587\u672C",
   requestLogDetailImages: "\u56FE\u7247",
@@ -1049,6 +1066,7 @@ var section_default = { "badge": "d7lz0W_badge", "block": "d7lz0W_block", "butto
 var import_jsx_runtime = require("react/jsx-runtime");
 var API_KEY_REF = "OPENCODE_GO_API_KEY";
 var VISIBILITY_WRITE_DELAY_MS = 250;
+var FOLLOW_UP_MS = 5e3;
 var INITIAL = {
   loading: true,
   reading: void 0,
@@ -1099,6 +1117,8 @@ var SectionController = class _SectionController {
   pendingVisibility;
   /** Timer of the coalescing window, when one is armed. */
   visibilityTimer;
+  /** The armed stale-reading follow-up, if one is waiting. */
+  followTimer;
   constructor(services) {
     this.services = services;
     const scope = services.scope;
@@ -1126,6 +1146,10 @@ var SectionController = class _SectionController {
   dispose() {
     this.listeners.clear();
     for (const dispose of this.disposers.splice(0)) dispose();
+    if (this.followTimer !== void 0) {
+      clearTimeout(this.followTimer);
+      this.followTimer = void 0;
+    }
     const scope = this.services.scope;
     if (this.visibilityTimer !== void 0 && scope !== void 0) void this.flushVisibility(scope);
   }
@@ -1140,6 +1164,7 @@ var SectionController = class _SectionController {
       const result = refresh ? await this.services.remote.opencodeGoCatalog.refresh() : await this.services.remote.opencodeGoCatalog.read();
       if (!result.ok) throw result.error;
       this.set({ loading: false, reading: result.value, failure: void 0 });
+      if (result.value.cached === true) this.followReading();
     } catch (error) {
       this.set({
         loading: false,
@@ -1147,6 +1172,35 @@ var SectionController = class _SectionController {
       });
     }
   }
+  /**
+   * Re-read a stale reading until it turns live, then stop.
+   *
+   * Only one follow-up is ever armed. The Host decides the real refresh
+   * rhythm, so this never hammers the gateway — a follow-up is a plain
+   * `read`, and the Host's own back-off decides how often the gateway is
+   * really asked. A failed or still-stale read re-arms; only a live reading
+   * or a disposal lets it lapse.
+   */
+  followReading() {
+    if (this.followTimer !== void 0) return;
+    const timer = setTimeout(() => {
+      this.followTimer = void 0;
+      void (async () => {
+        try {
+          const result = await this.services.remote.opencodeGoCatalog.read();
+          if (result.ok) {
+            this.set({ reading: result.value, failure: void 0 });
+            if (result.value.cached !== true) return;
+          }
+        } catch {
+        }
+        this.followReading();
+      })();
+    }, FOLLOW_UP_MS);
+    const releasable = timer;
+    if (typeof releasable.unref === "function") releasable.unref();
+    this.followTimer = timer;
+  }
   async loadKey() {
     try {
       const described = await this.services.remote.credentials.describe([API_KEY_REF]);
@@ -2153,19 +2207,16 @@ function Section({ controller, t, getLocale, openRequestLog }) {
 var import_react2 = require("react");
 var import_dsh_client_ui_primitives = require("@deepseek-ai/dsh-client-ui-primitives");
 
-// src/request-log/store.ts
-var MIN_STREAM_MS_FOR_RATE = 100;
-
 // plugin-css:/home/runner/work/dshopencodego/dshopencodego/src/client/request-log.module.css
 var id2 = "@dan-ai-studio/dshopencodego/request-log.module.css";
 if (!document.querySelector("style[data-plugin-css=" + JSON.stringify(id2) + "]")) {
   const style = document.createElement("style");
   style.dataset.plugin = "@dan-ai-studio/dshopencodego";
   style.dataset.pluginCss = id2;
-  style.textContent = ".Lx7MOq_page{box-sizing:border-box;background:var(--dsw-alias-bg-base);height:100%;min-height:0;color:var(--dsw-alias-label-primary);flex-direction:column;display:flex}.Lx7MOq_pageLoose{height:auto}.Lx7MOq_pageLoose .Lx7MOq_tableWrap{overflow:visible}.Lx7MOq_pageLoose .Lx7MOq_stuck{z-index:3;background:var(--dsw-alias-bg-base);position:sticky;top:0}.Lx7MOq_pageLoose .Lx7MOq_table th{top:var(--request-log-stuck,0px)}.Lx7MOq_pageLoose .Lx7MOq_foot{z-index:2;background:var(--dsw-alias-bg-base);position:sticky;bottom:0}.Lx7MOq_head{flex-wrap:wrap;flex:none;align-items:center;gap:12px;padding:20px clamp(16px,3vw,32px) 12px;display:flex}.Lx7MOq_headText{min-width:0}.Lx7MOq_title{margin:0;font-size:20px;font-weight:500;line-height:28px}.Lx7MOq_subtitle{color:var(--dsw-alias-label-secondary);margin:4px 0 0;font-size:13px;line-height:20px}.Lx7MOq_headActions{flex:none;align-items:center;gap:8px;margin-left:auto;display:flex}.Lx7MOq_button{border:1px solid var(--dsw-alias-border-l2);border-radius:var(--dsw-radius-sm);background:var(--dsw-alias-bg-layer-1);height:28px;color:var(--dsw-alias-label-primary);font:inherit;cursor:pointer;align-items:center;gap:6px;padding:0 12px;font-size:13px;display:inline-flex}.Lx7MOq_button:hover:not(:disabled){background:var(--dsw-alias-interactive-bg-hover)}.Lx7MOq_button:disabled{cursor:default;opacity:.5}.Lx7MOq_summary{flex-wrap:wrap;flex:none;gap:8px;padding:0 clamp(16px,3vw,32px) 12px;display:flex}.Lx7MOq_card{border:1px solid var(--dsw-alias-border-l2);border-radius:var(--dsw-radius-md);background:var(--dsw-alias-bg-layer-1);flex-direction:column;gap:2px;min-width:104px;padding:8px 12px;display:flex}.Lx7MOq_cardLabel{color:var(--dsw-alias-label-tertiary);font-size:12px;line-height:16px}.Lx7MOq_cardValue{font-variant-numeric:tabular-nums;font-size:16px;font-weight:500;line-height:22px}.Lx7MOq_cardError .Lx7MOq_cardValue{color:var(--dsw-alias-state-error-primary)}.Lx7MOq_cardWarn .Lx7MOq_cardValue{color:var(--dsw-alias-state-warn-primary)}.Lx7MOq_toolbar{flex-wrap:wrap;flex:none;align-items:center;gap:8px 12px;padding:0 clamp(16px,3vw,32px) 12px;display:flex}.Lx7MOq_field{color:var(--dsw-alias-label-secondary);align-items:center;gap:6px;font-size:13px;line-height:20px;display:inline-flex}.Lx7MOq_select,.Lx7MOq_input{border:1px solid var(--dsw-alias-border-l2);border-radius:var(--dsw-radius-sm);background:var(--dsw-alias-bg-layer-1);height:28px;color:var(--dsw-alias-label-primary);font:inherit;padding:0 8px;font-size:13px}.Lx7MOq_input{width:160px}.Lx7MOq_inputNumber{width:72px}.Lx7MOq_checkbox{accent-color:var(--dsw-alias-brand-primary)}.Lx7MOq_sessionLink{color:var(--dsw-alias-link);font:inherit;text-align:left;cursor:pointer;white-space:nowrap;background:0 0;border:none;padding:0}.Lx7MOq_sessionLink:hover{text-decoration:underline}.Lx7MOq_sessionText{text-overflow:ellipsis;vertical-align:bottom;max-width:340px;display:inline-block;overflow:hidden}.Lx7MOq_modelText{text-overflow:ellipsis;vertical-align:bottom;max-width:220px;display:inline-block;overflow:hidden}.Lx7MOq_tableWrap{border:1px solid var(--dsw-alias-border-l2);border-radius:var(--dsw-radius-md);background:var(--dsw-alias-bg-layer-1);flex:1;min-height:0;margin:0 clamp(16px,3vw,32px);overflow:auto}.Lx7MOq_table{border-collapse:separate;border-spacing:0;width:100%;min-width:940px;font-size:13px;line-height:20px}.Lx7MOq_table th{z-index:1;border-bottom:1px solid var(--dsw-alias-border-l2);background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-label-tertiary);text-align:left;white-space:nowrap;padding:8px 10px;font-size:12px;font-weight:500;position:sticky;top:0}.Lx7MOq_thSortable{cursor:pointer;user-select:none}.Lx7MOq_thSortable:hover{color:var(--dsw-alias-label-primary)}.Lx7MOq_table td{border-bottom:1px solid var(--dsw-alias-border-l1);white-space:nowrap;font-variant-numeric:tabular-nums;padding:6px 10px}.Lx7MOq_table tbody tr:last-child td{border-bottom:none}.Lx7MOq_row{cursor:pointer}.Lx7MOq_row:hover td{background:var(--dsw-alias-interactive-bg-hover)}.Lx7MOq_rowOk td:first-child{box-shadow:inset 3px 0 0 var(--dsw-alias-border-l3)}.Lx7MOq_rowSlow td:first-child{box-shadow:inset 3px 0 0 var(--dsw-alias-state-warn-primary)}.Lx7MOq_rowFail td:first-child{box-shadow:inset 3px 0 0 var(--dsw-alias-state-error-primary)}.Lx7MOq_rowRunning td:first-child{box-shadow:inset 3px 0 0 var(--dsw-alias-brand-primary)}.Lx7MOq_rowFail td:last-child,.Lx7MOq_rowSlow td:last-child{color:var(--dsw-alias-state-error-primary)}.Lx7MOq_group td{border-bottom:1px solid var(--dsw-alias-border-l2);color:var(--dsw-alias-label-secondary);padding-top:12px;font-weight:500}.Lx7MOq_detail td{border-bottom:1px solid var(--dsw-alias-border-l1);white-space:normal;background:var(--dsw-alias-bg-layer-2);padding:12px 10px 16px}.Lx7MOq_detailGrid{grid-template-columns:repeat(auto-fit,minmax(220px,1fr));gap:6px 24px;display:grid}.Lx7MOq_timeline{flex-direction:column;grid-column:1/-1;gap:6px;padding:4px 0 8px;display:flex}.Lx7MOq_timelineTrackWrap{padding-bottom:18px;position:relative}.Lx7MOq_timelineTrack{background:var(--dsw-alias-bg-layer-1);border-radius:999px;height:10px;position:relative;overflow:hidden}.Lx7MOq_timelineMidTick{background:var(--dsw-alias-label-tertiary);opacity:.7;width:1px;position:absolute;top:0;bottom:0;left:50%}.Lx7MOq_timelineMid{color:var(--dsw-alias-label-tertiary);white-space:nowrap;font-variant-numeric:tabular-nums;font-size:11px;line-height:16px;position:absolute;top:100%;transform:translate(-50%)}.Lx7MOq_timelineSeg{min-width:2px;position:absolute;top:0;bottom:0}.Lx7MOq_timelinePrep{background:var(--dsw-alias-brand-primary)}.Lx7MOq_timelineWait{background:var(--dsw-alias-state-warn-primary)}.Lx7MOq_timelineStream{background:var(--dsw-alias-brand-primary);opacity:.45}.Lx7MOq_timelineTick{background:var(--dsw-alias-label-primary);width:2px;margin-left:-2px;position:absolute;top:-2px;bottom:-2px;left:100%}.Lx7MOq_timelineLegend{color:var(--dsw-alias-label-secondary);font-variant-numeric:tabular-nums;flex-wrap:wrap;gap:4px 16px;font-size:12px;display:flex}.Lx7MOq_timelineItem{white-space:nowrap;align-items:center;gap:6px;display:inline-flex}.Lx7MOq_timelineSwatch{border-radius:2px;width:8px;height:8px}.Lx7MOq_timelineTickSwatch{background:var(--dsw-alias-label-primary)}.Lx7MOq_detailItem{gap:8px;min-width:0;font-size:13px;line-height:20px;display:flex}.Lx7MOq_detailLabel{color:var(--dsw-alias-label-tertiary);flex:none}.Lx7MOq_detailValue{overflow-wrap:anywhere;min-width:0;color:var(--dsw-alias-label-primary)}.Lx7MOq_mono{font-family:ui-monospace,SFMono-Regular,Menlo,Consolas,monospace;font-size:12px}.Lx7MOq_empty{text-align:center;color:var(--dsw-alias-label-tertiary);padding:48px 16px}.Lx7MOq_foot{color:var(--dsw-alias-label-tertiary);flex-wrap:wrap;flex:none;align-items:center;gap:12px;padding:12px clamp(16px,3vw,32px) 20px;font-size:12px;display:flex}.Lx7MOq_pager{align-items:center;gap:8px;margin-left:auto;display:flex}.Lx7MOq_notice{color:var(--dsw-alias-label-secondary);padding:0 clamp(16px,3vw,32px) 12px;font-size:13px;line-height:20px}.Lx7MOq_noticeError{color:var(--dsw-alias-state-error-primary)}";
+  style.textContent = ".Lx7MOq_page{box-sizing:border-box;background:var(--dsw-alias-bg-base);height:100%;min-height:0;color:var(--dsw-alias-label-primary);flex-direction:column;display:flex}.Lx7MOq_pageLoose{height:auto}.Lx7MOq_pageLoose .Lx7MOq_tableWrap{overflow:visible}.Lx7MOq_pageLoose .Lx7MOq_stuck{z-index:3;background:var(--dsw-alias-bg-base);position:sticky;top:0}.Lx7MOq_pageLoose .Lx7MOq_table th{top:var(--request-log-stuck,0px)}.Lx7MOq_pageLoose .Lx7MOq_foot{z-index:2;background:var(--dsw-alias-bg-base);position:sticky;bottom:0}.Lx7MOq_head{flex-wrap:wrap;flex:none;align-items:center;gap:12px;padding:20px clamp(16px,3vw,32px) 12px;display:flex}.Lx7MOq_headText{min-width:0}.Lx7MOq_title{margin:0;font-size:20px;font-weight:500;line-height:28px}.Lx7MOq_subtitle{color:var(--dsw-alias-label-secondary);margin:4px 0 0;font-size:13px;line-height:20px}.Lx7MOq_headActions{flex:none;align-items:center;gap:8px;margin-left:auto;display:flex}.Lx7MOq_button{border:1px solid var(--dsw-alias-border-l2);border-radius:var(--dsw-radius-sm);background:var(--dsw-alias-bg-layer-1);height:28px;color:var(--dsw-alias-label-primary);font:inherit;cursor:pointer;align-items:center;gap:6px;padding:0 12px;font-size:13px;display:inline-flex}.Lx7MOq_button:hover:not(:disabled){background:var(--dsw-alias-interactive-bg-hover)}.Lx7MOq_button:disabled{cursor:default;opacity:.5}.Lx7MOq_summary{flex-wrap:wrap;flex:none;gap:8px;padding:0 clamp(16px,3vw,32px) 12px;display:flex}.Lx7MOq_card{border:1px solid var(--dsw-alias-border-l2);border-radius:var(--dsw-radius-md);background:var(--dsw-alias-bg-layer-1);flex-direction:column;gap:2px;min-width:104px;padding:8px 12px;display:flex}.Lx7MOq_cardLabel{color:var(--dsw-alias-label-tertiary);font-size:12px;line-height:16px}.Lx7MOq_cardValue{font-variant-numeric:tabular-nums;font-size:16px;font-weight:500;line-height:22px}.Lx7MOq_cardError .Lx7MOq_cardValue{color:var(--dsw-alias-state-error-primary)}.Lx7MOq_cardWarn .Lx7MOq_cardValue{color:var(--dsw-alias-state-warn-primary)}.Lx7MOq_toolbar{flex-wrap:wrap;flex:none;align-items:center;gap:8px 12px;padding:0 clamp(16px,3vw,32px) 12px;display:flex}.Lx7MOq_field{color:var(--dsw-alias-label-secondary);align-items:center;gap:6px;font-size:13px;line-height:20px;display:inline-flex}.Lx7MOq_select,.Lx7MOq_input{border:1px solid var(--dsw-alias-border-l2);border-radius:var(--dsw-radius-sm);background:var(--dsw-alias-bg-layer-1);height:28px;color:var(--dsw-alias-label-primary);font:inherit;padding:0 8px;font-size:13px}.Lx7MOq_input{width:160px}.Lx7MOq_inputNumber{width:72px}.Lx7MOq_checkbox{accent-color:var(--dsw-alias-brand-primary)}.Lx7MOq_sessionLink{color:var(--dsw-alias-link);font:inherit;text-align:left;cursor:pointer;white-space:nowrap;background:0 0;border:none;padding:0}.Lx7MOq_sessionLink:hover{text-decoration:underline}.Lx7MOq_sessionText{text-overflow:ellipsis;vertical-align:bottom;max-width:340px;display:inline-block;overflow:hidden}.Lx7MOq_modelText{text-overflow:ellipsis;vertical-align:bottom;max-width:220px;display:inline-block;overflow:hidden}.Lx7MOq_tableWrap{border:1px solid var(--dsw-alias-border-l2);border-radius:var(--dsw-radius-md);background:var(--dsw-alias-bg-layer-1);flex:1;min-height:0;margin:0 clamp(16px,3vw,32px);overflow:auto}.Lx7MOq_table{border-collapse:separate;border-spacing:0;width:100%;min-width:940px;font-size:13px;line-height:20px}.Lx7MOq_table th{z-index:1;border-bottom:1px solid var(--dsw-alias-border-l2);background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-label-tertiary);text-align:left;white-space:nowrap;padding:8px 10px;font-size:12px;font-weight:500;position:sticky;top:0}.Lx7MOq_thSortable{cursor:pointer;user-select:none}.Lx7MOq_thSortable:hover{color:var(--dsw-alias-label-primary)}.Lx7MOq_table td{border-bottom:1px solid var(--dsw-alias-border-l1);white-space:nowrap;font-variant-numeric:tabular-nums;padding:6px 10px}.Lx7MOq_table tbody tr:last-child td{border-bottom:none}.Lx7MOq_row{cursor:pointer}.Lx7MOq_row:hover td{background:var(--dsw-alias-interactive-bg-hover)}.Lx7MOq_rowOk td:first-child{box-shadow:inset 3px 0 0 var(--dsw-alias-border-l3)}.Lx7MOq_rowSlow td:first-child{box-shadow:inset 3px 0 0 var(--dsw-alias-state-warn-primary)}.Lx7MOq_rowFail td:first-child{box-shadow:inset 3px 0 0 var(--dsw-alias-state-error-primary)}.Lx7MOq_rowRunning td:first-child{box-shadow:inset 3px 0 0 var(--dsw-alias-brand-primary)}.Lx7MOq_rowFail td:last-child,.Lx7MOq_rowSlow td:last-child{color:var(--dsw-alias-state-error-primary)}.Lx7MOq_group td{border-bottom:1px solid var(--dsw-alias-border-l2);color:var(--dsw-alias-label-secondary);padding-top:12px;font-weight:500}.Lx7MOq_detail td{border-bottom:1px solid var(--dsw-alias-border-l1);white-space:normal;background:var(--dsw-alias-bg-layer-2);padding:12px 10px 16px}.Lx7MOq_detailGrid{grid-template-columns:repeat(auto-fit,minmax(220px,1fr));gap:6px 24px;display:grid}.Lx7MOq_timeline{flex-direction:column;grid-column:1/-1;gap:6px;padding:4px 0 8px;display:flex}.Lx7MOq_timelineTrackWrap{padding-bottom:18px;position:relative}.Lx7MOq_timelineTrack{background:var(--dsw-alias-bg-layer-1);border-radius:999px;height:10px;position:relative;overflow:hidden}.Lx7MOq_timelineMidTick{background:var(--dsw-alias-label-tertiary);opacity:.7;width:1px;position:absolute;top:0;bottom:0;left:50%}.Lx7MOq_timelineMid{color:var(--dsw-alias-label-tertiary);white-space:nowrap;font-variant-numeric:tabular-nums;font-size:11px;line-height:16px;position:absolute;top:100%;transform:translate(-50%)}.Lx7MOq_timelineSeg{min-width:2px;position:absolute;top:0;bottom:0}.Lx7MOq_timelinePrep{background:var(--dsw-alias-brand-primary)}.Lx7MOq_timelineWait{background:var(--dsw-alias-state-warn-primary)}.Lx7MOq_timelineStream{background:var(--dsw-alias-brand-primary);opacity:.45}.Lx7MOq_timelineStreamAlt{background:var(--dsw-alias-brand-primary);opacity:.9}.Lx7MOq_timelineRespTick{background:var(--dsw-alias-label-primary);opacity:.7;width:1px;position:absolute;top:-2px;bottom:-2px}.Lx7MOq_timelineTick{background:var(--dsw-alias-label-primary);width:2px;margin-left:-2px;position:absolute;top:-2px;bottom:-2px;left:100%}.Lx7MOq_timelineLegend{color:var(--dsw-alias-label-secondary);font-variant-numeric:tabular-nums;flex-wrap:wrap;gap:4px 16px;font-size:12px;display:flex}.Lx7MOq_timelineItem{white-space:nowrap;align-items:center;gap:6px;display:inline-flex}.Lx7MOq_timelineSwatch{border-radius:2px;width:8px;height:8px}.Lx7MOq_timelineTickSwatch{background:var(--dsw-alias-label-primary)}.Lx7MOq_detailItem{gap:8px;min-width:0;font-size:13px;line-height:20px;display:flex}.Lx7MOq_detailLabel{width:88px;color:var(--dsw-alias-label-tertiary);flex:none}.Lx7MOq_detailValue{overflow-wrap:anywhere;min-width:0;color:var(--dsw-alias-label-primary)}.Lx7MOq_mono{font-family:ui-monospace,SFMono-Regular,Menlo,Consolas,monospace;font-size:12px}.Lx7MOq_empty{text-align:center;color:var(--dsw-alias-label-tertiary);padding:48px 16px}.Lx7MOq_foot{color:var(--dsw-alias-label-tertiary);flex-wrap:wrap;flex:none;align-items:center;gap:12px;padding:12px clamp(16px,3vw,32px) 20px;font-size:12px;display:flex}.Lx7MOq_pager{align-items:center;gap:8px;margin-left:auto;display:flex}.Lx7MOq_notice{color:var(--dsw-alias-label-secondary);padding:0 clamp(16px,3vw,32px) 12px;font-size:13px;line-height:20px}.Lx7MOq_noticeError{color:var(--dsw-alias-state-error-primary)}";
   document.head.appendChild(style);
 }
-var request_log_default = { "button": "Lx7MOq_button", "card": "Lx7MOq_card", "cardError": "Lx7MOq_cardError", "cardLabel": "Lx7MOq_cardLabel", "cardValue": "Lx7MOq_cardValue", "cardWarn": "Lx7MOq_cardWarn", "checkbox": "Lx7MOq_checkbox", "detail": "Lx7MOq_detail", "detailGrid": "Lx7MOq_detailGrid", "detailItem": "Lx7MOq_detailItem", "detailLabel": "Lx7MOq_detailLabel", "detailValue": "Lx7MOq_detailValue", "empty": "Lx7MOq_empty", "field": "Lx7MOq_field", "foot": "Lx7MOq_foot", "group": "Lx7MOq_group", "head": "Lx7MOq_head", "headActions": "Lx7MOq_headActions", "headText": "Lx7MOq_headText", "input": "Lx7MOq_input", "inputNumber": "Lx7MOq_inputNumber", "modelText": "Lx7MOq_modelText", "mono": "Lx7MOq_mono", "notice": "Lx7MOq_notice", "noticeError": "Lx7MOq_noticeError", "page": "Lx7MOq_page", "pageLoose": "Lx7MOq_pageLoose", "pager": "Lx7MOq_pager", "row": "Lx7MOq_row", "rowFail": "Lx7MOq_rowFail", "rowOk": "Lx7MOq_rowOk", "rowRunning": "Lx7MOq_rowRunning", "rowSlow": "Lx7MOq_rowSlow", "select": "Lx7MOq_select", "sessionLink": "Lx7MOq_sessionLink", "sessionText": "Lx7MOq_sessionText", "stuck": "Lx7MOq_stuck", "subtitle": "Lx7MOq_subtitle", "summary": "Lx7MOq_summary", "table": "Lx7MOq_table", "tableWrap": "Lx7MOq_tableWrap", "thSortable": "Lx7MOq_thSortable", "timeline": "Lx7MOq_timeline", "timelineItem": "Lx7MOq_timelineItem", "timelineLegend": "Lx7MOq_timelineLegend", "timelineMid": "Lx7MOq_timelineMid", "timelineMidTick": "Lx7MOq_timelineMidTick", "timelinePrep": "Lx7MOq_timelinePrep", "timelineSeg": "Lx7MOq_timelineSeg", "timelineStream": "Lx7MOq_timelineStream", "timelineSwatch": "Lx7MOq_timelineSwatch", "timelineTick": "Lx7MOq_timelineTick", "timelineTickSwatch": "Lx7MOq_timelineTickSwatch", "timelineTrack": "Lx7MOq_timelineTrack", "timelineTrackWrap": "Lx7MOq_timelineTrackWrap", "timelineWait": "Lx7MOq_timelineWait", "title": "Lx7MOq_title", "toolbar": "Lx7MOq_toolbar" };
+var request_log_default = { "button": "Lx7MOq_button", "card": "Lx7MOq_card", "cardError": "Lx7MOq_cardError", "cardLabel": "Lx7MOq_cardLabel", "cardValue": "Lx7MOq_cardValue", "cardWarn": "Lx7MOq_cardWarn", "checkbox": "Lx7MOq_checkbox", "detail": "Lx7MOq_detail", "detailGrid": "Lx7MOq_detailGrid", "detailItem": "Lx7MOq_detailItem", "detailLabel": "Lx7MOq_detailLabel", "detailValue": "Lx7MOq_detailValue", "empty": "Lx7MOq_empty", "field": "Lx7MOq_field", "foot": "Lx7MOq_foot", "group": "Lx7MOq_group", "head": "Lx7MOq_head", "headActions": "Lx7MOq_headActions", "headText": "Lx7MOq_headText", "input": "Lx7MOq_input", "inputNumber": "Lx7MOq_inputNumber", "modelText": "Lx7MOq_modelText", "mono": "Lx7MOq_mono", "notice": "Lx7MOq_notice", "noticeError": "Lx7MOq_noticeError", "page": "Lx7MOq_page", "pageLoose": "Lx7MOq_pageLoose", "pager": "Lx7MOq_pager", "row": "Lx7MOq_row", "rowFail": "Lx7MOq_rowFail", "rowOk": "Lx7MOq_rowOk", "rowRunning": "Lx7MOq_rowRunning", "rowSlow": "Lx7MOq_rowSlow", "select": "Lx7MOq_select", "sessionLink": "Lx7MOq_sessionLink", "sessionText": "Lx7MOq_sessionText", "stuck": "Lx7MOq_stuck", "subtitle": "Lx7MOq_subtitle", "summary": "Lx7MOq_summary", "table": "Lx7MOq_table", "tableWrap": "Lx7MOq_tableWrap", "thSortable": "Lx7MOq_thSortable", "timeline": "Lx7MOq_timeline", "timelineItem": "Lx7MOq_timelineItem", "timelineLegend": "Lx7MOq_timelineLegend", "timelineMid": "Lx7MOq_timelineMid", "timelineMidTick": "Lx7MOq_timelineMidTick", "timelinePrep": "Lx7MOq_timelinePrep", "timelineRespTick": "Lx7MOq_timelineRespTick", "timelineSeg": "Lx7MOq_timelineSeg", "timelineStream": "Lx7MOq_timelineStream", "timelineStreamAlt": "Lx7MOq_timelineStreamAlt", "timelineSwatch": "Lx7MOq_timelineSwatch", "timelineTick": "Lx7MOq_timelineTick", "timelineTickSwatch": "Lx7MOq_timelineTickSwatch", "timelineTrack": "Lx7MOq_timelineTrack", "timelineTrackWrap": "Lx7MOq_timelineTrackWrap", "timelineWait": "Lx7MOq_timelineWait", "title": "Lx7MOq_title", "toolbar": "Lx7MOq_toolbar" };
 
 // src/client/RequestLogPage.tsx
 var import_jsx_runtime2 = require("react/jsx-runtime");
@@ -2254,13 +2305,15 @@ function timelineOf(entry, t) {
   const first = entry.firstTokenMs;
   if (prep === void 0 && first === void 0) return void 0;
   const start = prep ?? 0;
-  const slices = [];
+  const blocks = [];
   if (prep !== void 0 && prep > 0) {
-    slices.push({ kind: "prep", label: t("requestLogDetailPrep"), from: 0, to: Math.min(prep, end) });
+    blocks.push({ kind: "prep", label: t("requestLogDetailPrep"), from: 0, to: Math.min(prep, end) });
   }
+  const responseAt = entry.responseMs === void 0 ? void 0 : Math.min(Math.max(entry.responseMs, 0), end);
+  let streamTotal;
   if (first !== void 0) {
     if (first > start) {
-      slices.push({
+      blocks.push({
         kind: "wait",
         label: t("requestLogTimelineWait"),
         from: Math.min(start, end),
@@ -2269,14 +2322,30 @@ function timelineOf(entry, t) {
     }
     if (entry.streamMs !== void 0 && entry.streamMs > 0) {
       const last = Math.min(first + entry.streamMs, end);
-      if (last > first) slices.push({ kind: "stream", label: t("requestLogTimelineStream"), from: first, to: last });
+      if (last > first) {
+        const cuts = [first, ...(entry.streamMarksMs ?? []).filter((mark) => mark > first && mark < last), last];
+        for (let index = 0; index + 1 < cuts.length; index += 1) {
+          const from = cuts[index];
+          const to = cuts[index + 1];
+          if (from === void 0 || to === void 0) continue;
+          blocks.push({
+            kind: index % 2 === 0 ? "stream" : "streamAlt",
+            label: t("requestLogTimelineStream"),
+            from,
+            to
+          });
+        }
+        streamTotal = last - first;
+      }
     }
   }
-  if (slices.length === 0) return void 0;
+  if (blocks.length === 0) return void 0;
   return {
-    slices,
+    blocks,
     end,
-    burst: !running && first !== void 0 && (entry.streamMs === void 0 || entry.streamMs < MIN_STREAM_MS_FOR_RATE)
+    burst: !running && first !== void 0 && (entry.streamMs === void 0 || entry.streamMs < MIN_STREAM_MS_FOR_RATE),
+    responseAt,
+    streamTotal
   };
 }
 function percent(value) {
@@ -2491,44 +2560,69 @@ function RequestLogPage({ readLog, enabled, canOpenSession, openSession, t, getL
             (() => {
               const timeline = timelineOf(entry, t);
               if (timeline === void 0) return null;
-              const kindClass = (kind) => kind === "prep" ? request_log_default.timelinePrep : kind === "wait" ? request_log_default.timelineWait : request_log_default.timelineStream;
-              const spoken = [
-                ...timeline.slices.map((slice) => `${slice.label} ${duration(slice.to - slice.from)}`),
-                `${t("requestLogColLatency")} ${duration(timeline.end)}`
-              ].join(", ");
+              const kindClass = (kind) => kind === "prep" ? request_log_default.timelinePrep : kind === "wait" ? request_log_default.timelineWait : kind === "stream" ? request_log_default.timelineStream : request_log_default.timelineStreamAlt;
+              const legend = [];
+              const legendKey = (kind) => kind === "streamAlt" ? "stream" : kind;
+              for (const block of timeline.blocks) {
+                const key = legendKey(block.kind);
+                const seen = legend.some((item) => item.key === key);
+                if (!seen) {
+                  legend.push({
+                    key,
+                    swatch: kindClass(key),
+                    text: `${block.label} ${duration(
+                      timeline.blocks.filter((other) => legendKey(other.kind) === key).reduce((sum, other) => sum + (other.to - other.from), 0)
+                    )}`
+                  });
+                }
+              }
+              if (timeline.responseAt !== void 0) {
+                legend.push({
+                  key: "response",
+                  swatch: request_log_default.timelineTickSwatch,
+                  text: `${t("requestLogTimelineFirstByte")} ${duration(timeline.responseAt)}`
+                });
+              }
+              legend.push({
+                key: "total",
+                swatch: request_log_default.timelineTickSwatch,
+                text: `${t("requestLogColLatency")} ${duration(timeline.end)}`
+              });
+              const spoken = legend.map((item) => item.text).join(", ");
+              const percent2 = (at) => `${(at / timeline.end * 100).toFixed(2)}%`;
               return /* @__PURE__ */ (0, import_jsx_runtime2.jsxs)("div", { className: request_log_default.timeline, role: "img", "aria-label": spoken, children: [
                 /* @__PURE__ */ (0, import_jsx_runtime2.jsxs)("div", { className: request_log_default.timelineTrackWrap, children: [
                   /* @__PURE__ */ (0, import_jsx_runtime2.jsxs)("div", { className: request_log_default.timelineTrack, children: [
-                    timeline.slices.map((slice) => /* @__PURE__ */ (0, import_jsx_runtime2.jsx)(
+                    timeline.blocks.map((block, index) => /* @__PURE__ */ (0, import_jsx_runtime2.jsx)(
                       "div",
                       {
-                        className: `${request_log_default.timelineSeg} ${kindClass(slice.kind)}`,
+                        className: `${request_log_default.timelineSeg} ${kindClass(block.kind)}`,
                         style: {
-                          left: `${(slice.from / timeline.end * 100).toFixed(2)}%`,
-                          width: `${Math.max(0, (slice.to - slice.from) / timeline.end * 100).toFixed(2)}%`
+                          left: percent2(block.from),
+                          width: percent2(Math.max(0, block.to - block.from))
                         },
-                        title: `${slice.label} ${duration(slice.to - slice.from)}`
+                        title: `${block.label} ${duration(block.to - block.from)}`
                       },
-                      slice.kind
+                      `${block.kind}-${index}`
                     )),
+                    timeline.responseAt !== void 0 && /* @__PURE__ */ (0, import_jsx_runtime2.jsx)(
+                      "div",
+                      {
+                        className: request_log_default.timelineRespTick,
+                        style: { left: percent2(Math.min(timeline.responseAt, timeline.end)) },
+                        title: `${t("requestLogTimelineFirstByte")} ${duration(timeline.responseAt)}`
+                      }
+                    ),
                     /* @__PURE__ */ (0, import_jsx_runtime2.jsx)("div", { className: request_log_default.timelineMidTick }),
                     /* @__PURE__ */ (0, import_jsx_runtime2.jsx)("div", { className: request_log_default.timelineTick, title: duration(timeline.end) })
                   ] }),
-                  /* @__PURE__ */ (0, import_jsx_runtime2.jsx)("div", { className: request_log_default.timelineMid, title: duration(timeline.end / 2), children: duration(timeline.end / 2) })
+                  /* @__PURE__ */ (0, import_jsx_runtime2.jsx)("div", { className: request_log_default.timelineMid, style: { left: "50%" }, title: duration(timeline.end / 2), children: duration(timeline.end / 2) })
                 ] }),
                 /* @__PURE__ */ (0, import_jsx_runtime2.jsxs)("div", { className: request_log_default.timelineLegend, children: [
-                  timeline.slices.map((slice) => /* @__PURE__ */ (0, import_jsx_runtime2.jsxs)("span", { className: request_log_default.timelineItem, children: [
-                    /* @__PURE__ */ (0, import_jsx_runtime2.jsx)("span", { className: `${request_log_default.timelineSwatch} ${kindClass(slice.kind)}` }),
-                    slice.label,
-                    " ",
-                    duration(slice.to - slice.from)
-                  ] }, slice.kind)),
-                  /* @__PURE__ */ (0, import_jsx_runtime2.jsxs)("span", { className: request_log_default.timelineItem, children: [
-                    /* @__PURE__ */ (0, import_jsx_runtime2.jsx)("span", { className: `${request_log_default.timelineSwatch} ${request_log_default.timelineTickSwatch}` }),
-                    t("requestLogColLatency"),
-                    " ",
-                    duration(timeline.end)
-                  ] }),
+                  legend.map((item) => /* @__PURE__ */ (0, import_jsx_runtime2.jsxs)("span", { className: request_log_default.timelineItem, children: [
+                    /* @__PURE__ */ (0, import_jsx_runtime2.jsx)("span", { className: `${request_log_default.timelineSwatch} ${item.swatch ?? ""}` }),
+                    item.text
+                  ] }, item.key)),
                   timeline.burst && /* @__PURE__ */ (0, import_jsx_runtime2.jsx)("span", { className: request_log_default.timelineItem, children: t("requestLogTimelineBurst") })
                 ] })
               ] });

package/lib/index.js

@@ -73672,7 +73672,7 @@ import dc from "node:diagnostics_channel";
 import { credentialRef } from "@deepseek-ai/dsh-credentials";
 import { launchEnvironmentOf } from "@deepseek-ai/dsh-launch-environment";
 import * as llm2 from "@deepseek-ai/dsh-llm";
-import { LlmError as LlmError13 } from "@deepseek-ai/dsh-llm";
+import { LlmError as LlmError14 } from "@deepseek-ai/dsh-llm";
 
 // src/adapter.ts
 init_models();
@@ -73682,7 +73682,7 @@ import {
   contentHasImage as contentHasImage2,
   isTokenDelta,
   LlmAdapter,
-  LlmError as LlmError9,
+  LlmError as LlmError10,
   ReasoningEffortId,
   resolveRetryPolicy
 } from "@deepseek-ai/dsh-llm";
@@ -77887,6 +77887,74 @@ function sortModels(models, now = Date.now()) {
   return [...models].sort((left, right) => rank(left) - rank(right) || (isNewModel(left, now) && isNewModel(right, now) ? right.releaseDate.localeCompare(left.releaseDate) : 0));
 }
 
+// src/context-fit.ts
+import { CONTEXT_WINDOW_EXCEEDED_CODE as CONTEXT_WINDOW_EXCEEDED_CODE2, LlmError as LlmError9 } from "@deepseek-ai/dsh-llm";
+var CONTEXT_FIT_RESERVE_TOKENS = 2048;
+var CONTEXT_FIT_MIN_OUTPUT_TOKENS = 256;
+var CHARS_PER_TOKEN2 = 3;
+var TOKENS_PER_MESSAGE = 8;
+var TOKENS_PER_IMAGE = 2e3;
+var IMAGE_BYTES_PER_TOKEN = 2e3;
+var TOKENS_PER_FILE = 500;
+var FILE_BYTES_PER_TOKEN = 4e3;
+function estimateInputTokens(options) {
+  let chars = 0;
+  let images = 0;
+  let imageBytes = 0;
+  let files = 0;
+  let fileBytes = 0;
+  if (typeof options.system === "string") chars += options.system.length;
+  for (const message of options.messages) {
+    for (const block of message.content) {
+      switch (block.type) {
+        case "text":
+        case "reasoning":
+          chars += block.text.length;
+          break;
+        case "tool-call":
+          chars += block.name.length + block.arguments.length;
+          break;
+        case "image":
+          if (block.offloaded === true) {
+            chars += 200;
+          } else {
+            images += 1;
+            imageBytes += block.attachment.bytes;
+          }
+          break;
+        case "file":
+          files += 1;
+          fileBytes += block.attachment.bytes;
+          chars += block.attachment.name.length;
+          break;
+        default:
+          break;
+      }
+    }
+  }
+  if (options.tools !== void 0 && options.tools.length > 0) {
+    chars += JSON.stringify(options.tools).length;
+  }
+  const estimate = Math.ceil(chars / CHARS_PER_TOKEN2) + options.messages.length * TOKENS_PER_MESSAGE + images * TOKENS_PER_IMAGE + Math.ceil(imageBytes / IMAGE_BYTES_PER_TOKEN) + files * TOKENS_PER_FILE + Math.ceil(fileBytes / FILE_BYTES_PER_TOKEN);
+  return Number.isFinite(estimate) && estimate > 0 ? Math.floor(estimate) : 0;
+}
+function resolveFittedMaxTokens(fit) {
+  const { modelId, contextWindow, routeMaxTokens, requestedMaxTokens, estimatedInputTokens } = fit;
+  if (!Number.isFinite(contextWindow) || contextWindow <= 0) return requestedMaxTokens;
+  const input = Number.isFinite(estimatedInputTokens) && estimatedInputTokens > 0 ? Math.floor(estimatedInputTokens) : 0;
+  const allowed = Math.floor(contextWindow - input - CONTEXT_FIT_RESERVE_TOKENS);
+  const over = (what) => new LlmError9(
+    `opencode-go model "${modelId}" ${what} (~${input} input tokens) leaves no room inside context window ${contextWindow} (reserve ${CONTEXT_FIT_RESERVE_TOKENS}); compact the conversation or lower maxTokens`,
+    CONTEXT_WINDOW_EXCEEDED_CODE2
+  );
+  if (allowed < CONTEXT_FIT_MIN_OUTPUT_TOKENS) throw over("input");
+  const effective = requestedMaxTokens ?? routeMaxTokens;
+  if (effective === void 0) return void 0;
+  if (!Number.isFinite(effective) || effective <= 0) return requestedMaxTokens;
+  if (effective <= allowed) return requestedMaxTokens;
+  return allowed;
+}
+
 // src/session-header.ts
 import { randomUUID as randomUUID2 } from "node:crypto";
 import { attributionHeaders as attributionHeaders4 } from "@deepseek-ai/dsh-llm";
@@ -78008,7 +78076,7 @@ function withModelLimit(model, limits) {
 function stopSequencesFor(model, stop) {
   if (stop === void 0) return void 0;
   if (model.api === "openai-completions") return stop;
-  throw new LlmError9(
+  throw new LlmError10(
     `opencode-go model "${model.id}" speaks ${model.api}, which has no stop-sequence parameter this adapter can reach; drop GenerateOptions.stop for this model`,
     "UNSUPPORTED_OPTION"
   );
@@ -78128,7 +78196,7 @@ var OpencodeGoAdapter = class extends LlmAdapter {
     const config = this.options.config();
     const snapshot2 = await this.catalogOf(config).forModel(model);
     const facts = snapshot2.facts.get(model);
-    if (facts === void 0) throw new LlmError9(`opencode-go has no model "${model}"`, "UNKNOWN_MODEL");
+    if (facts === void 0) throw new LlmError10(`opencode-go has no model "${model}"`, "UNKNOWN_MODEL");
     return this.modelInfo(withModelLimit(toPiModel(facts, config.baseURL), config.modelLimits));
   }
   /** Describe one model: capacities plus the reasoning levels it really offers. */
@@ -78174,7 +78242,7 @@ var OpencodeGoAdapter = class extends LlmAdapter {
     if (effort === void 0 || effort === "off") return void 0;
     const supported = getSupportedThinkingLevels(model);
     if (supported.some((level) => level === effort)) return effort;
-    throw new LlmError9(
+    throw new LlmError10(
       `opencode-go model "${model.id}" does not support reasoning effort "${effort}"`,
       "UNSUPPORTED_REASONING_EFFORT"
     );
@@ -78227,7 +78295,7 @@ var OpencodeGoAdapter = class extends LlmAdapter {
       hasImage: false,
       toolCount: 0,
       headers: {}
-    })?.finish(false, error instanceof LlmError9 ? error.code : void 0);
+    })?.finish(false, error instanceof LlmError10 ? error.code : void 0);
   }
   async prepareCall(_provider, model, _signal) {
     const config = this.options.config();
@@ -78241,7 +78309,7 @@ var OpencodeGoAdapter = class extends LlmAdapter {
     }
     const facts = snapshot2.facts.get(model);
     if (facts === void 0) {
-      const error = new LlmError9(`opencode-go has no model "${model}"`, "UNKNOWN_MODEL");
+      const error = new LlmError10(`opencode-go has no model "${model}"`, "UNKNOWN_MODEL");
       this.logUnstartedFailure(config, model, error);
       throw error;
     }
@@ -78262,7 +78330,7 @@ var OpencodeGoAdapter = class extends LlmAdapter {
     try {
       snapshot2 = await this.catalogOf(config).forModel(options.model);
     } catch (error) {
-      span?.finish(false, error instanceof LlmError9 ? error.code : void 0);
+      span?.finish(false, error instanceof LlmError10 ? error.code : void 0);
       throw error;
     }
     yield* this.streamWithSnapshot(options, config, snapshot2, span);
@@ -78272,7 +78340,7 @@ var OpencodeGoAdapter = class extends LlmAdapter {
     try {
       const span = outerSpan ?? this.startSpan(options, config);
       const failSpan = (error) => {
-        span?.finish(false, error instanceof LlmError9 ? error.code : void 0);
+        span?.finish(false, error instanceof LlmError10 ? error.code : void 0);
         throw error;
       };
       try {
@@ -78282,7 +78350,7 @@ var OpencodeGoAdapter = class extends LlmAdapter {
       }
       const facts = snapshot2.facts.get(options.model);
       if (facts === void 0) {
-        const error = new LlmError9(`opencode-go has no model "${options.model}"`, "UNKNOWN_MODEL");
+        const error = new LlmError10(`opencode-go has no model "${options.model}"`, "UNKNOWN_MODEL");
         span?.finish(false, error.code);
         throw error;
       }
@@ -78290,10 +78358,22 @@ var OpencodeGoAdapter = class extends LlmAdapter {
       const rates2 = usableRates(facts.cost);
       const stop = stopSequencesFor(model, options.stop);
       const ceiling = config.modelLimits[model.id]?.maxTokens;
-      const maxTokens = options.maxTokens === void 0 || ceiling === null || ceiling === void 0 ? options.maxTokens : Math.min(options.maxTokens, ceiling);
+      const requested = options.maxTokens === void 0 || ceiling === null || ceiling === void 0 ? options.maxTokens : Math.min(options.maxTokens, ceiling);
+      let maxTokens;
+      try {
+        maxTokens = resolveFittedMaxTokens({
+          modelId: model.id,
+          contextWindow: model.contextWindow,
+          routeMaxTokens: model.maxTokens,
+          requestedMaxTokens: requested,
+          estimatedInputTokens: estimateInputTokens(options)
+        });
+      } catch (error) {
+        failSpan(error);
+      }
       const apiKey = await this.options.resolveApiKey();
       if (apiKey === void 0 || apiKey.length === 0) {
-        const error = new LlmError9("dshopencodego: no credential resolved for the opencode-go route", "MISSING_CREDENTIAL");
+        const error = new LlmError10("dshopencodego: no credential resolved for the opencode-go route", "MISSING_CREDENTIAL");
         span?.finish(false, error.code);
         throw error;
       }
@@ -78323,14 +78403,14 @@ var OpencodeGoAdapter = class extends LlmAdapter {
       };
       try {
         if (containsImage && !model.input.includes("image")) {
-          throw new LlmError9(`opencode-go model "${model.id}" does not support image input`, "UNSUPPORTED_CONTENT");
+          throw new LlmError10(`opencode-go model "${model.id}" does not support image input`, "UNSUPPORTED_CONTENT");
         }
         let imageRequest;
         if (containsImage) {
           const access2 = this.options.imageAccess;
           const store = access2?.resolveAttachments();
           if (access2 === void 0 || store === void 0) {
-            throw new LlmError9("dshopencodego image input requires the durable attachment service", "UNSUPPORTED_CONTENT");
+            throw new LlmError10("dshopencodego image input requires the durable attachment service", "UNSUPPORTED_CONTENT");
           }
           imageRequest = {
             attachments: store,
@@ -78374,7 +78454,7 @@ var OpencodeGoAdapter = class extends LlmAdapter {
           while (true) {
             const result = await watchdog.next(iterator);
             if (timeoutOf(watchdog.signal, "LLM_STREAM_IDLE_TIMEOUT") !== void 0) {
-              throw new LlmError9("opencode-go stream idle timeout", "TIMEOUT");
+              throw new LlmError10("opencode-go stream idle timeout", "TIMEOUT");
             }
             if (result.done) {
               exhausted = true;
@@ -78427,13 +78507,13 @@ var OpencodeGoAdapter = class extends LlmAdapter {
       } catch (error) {
         if (timedOut || timeoutOf(watchdog.signal, "LLM_STREAM_IDLE_TIMEOUT") !== void 0) {
           span?.finish(false, "TIMEOUT");
-          throw new LlmError9("opencode-go stream idle timeout", "TIMEOUT", { cause: error });
+          throw new LlmError10("opencode-go stream idle timeout", "TIMEOUT", { cause: error });
         }
         if (options.signal?.aborted) {
           span?.finish(false, "ABORTED");
-          throw new LlmError9("opencode-go request aborted by caller", "ABORTED", { cause: error });
+          throw new LlmError10("opencode-go request aborted by caller", "ABORTED", { cause: error });
         }
-        span?.finish(false, error instanceof LlmError9 ? error.code : void 0);
+        span?.finish(false, error instanceof LlmError10 ? error.code : void 0);
         throw error;
       } finally {
         consumer.abort("opencode-go stream consumer stopped");
@@ -78618,6 +78698,199 @@ var networkRemote = {
   ]
 };
 
+// src/request-log/store.ts
+var MAX_POOL_EVENTS = 20;
+var STREAM_MARK_EVERY = 10;
+var MAX_STREAM_MARKS = 200;
+var MIN_STREAM_MS_FOR_RATE = 100;
+var RequestLog = class {
+  gates;
+  entries = [];
+  poolEvents = [];
+  seq = 0;
+  totalRecords = 0;
+  droppedRecords = 0;
+  newConnections = 0;
+  poolKey;
+  constructor(gates) {
+    this.gates = gates;
+  }
+  /**
+   * Count one accepted socket. Called from the diagnostics channel, not the
+   * request path, so its cost never lands on a request.
+   */
+  connectionOpened() {
+    this.newConnections += 1;
+  }
+  /**
+   * Record one pool lifecycle event. Pools live for minutes, so this is rare
+   * by construction; still gated on debug so the common log stays request-only.
+   */
+  poolEvent(type, key) {
+    if (!this.gates.enabled() || !this.gates.debug()) return;
+    this.poolKey = type === "create" ? key : this.poolKey;
+    if (type === "close" && this.poolKey === key) this.poolKey = void 0;
+    this.poolEvents = [{ atMs: Date.now(), type, key }, ...this.poolEvents].slice(0, MAX_POOL_EVENTS);
+  }
+  /**
+   * Begin one request's span, or nothing when the log is disabled.
+   *
+   * This is the single gate on the request path: a disabled log costs one
+   * method call and one boolean read, and constructs nothing. When enabled,
+   * the running row is appended before the span is returned, so the viewer
+   * sees the request while it runs; `request` and `status` mutate that same
+   * visible row as the wire facts arrive.
+   */
+  start(detail) {
+    if (!this.gates.enabled()) return void 0;
+    const debug = this.gates.debug();
+    const startedAt = Date.now();
+    const session = detail.sessionId ?? "";
+    const record2 = {
+      seq: ++this.seq,
+      atMs: startedAt,
+      model: detail.model,
+      ...detail.protocol === void 0 ? {} : { protocol: detail.protocol },
+      session,
+      method: detail.method ?? "POST",
+      url: detail.url ?? "",
+      running: true,
+      ok: false,
+      shape: { messages: detail.messageCount, hasImage: detail.hasImage, tools: detail.toolCount },
+      ...detail.imageBytes === void 0 ? {} : { imageBytes: detail.imageBytes },
+      ...!debug ? {} : {
+        headers: {
+          ...detail.headers.session === void 0 ? {} : { session: detail.headers.session },
+          ...detail.headers.userAgent === void 0 ? {} : { userAgent: detail.headers.userAgent }
+        }
+      }
+    };
+    this.insert(record2, true);
+    let firstTokenAt;
+    let lastTokenAt;
+    let tokenDeltas = 0;
+    const marks = [];
+    let reported = false;
+    let finished = false;
+    return {
+      identify: (protocol) => {
+        if (record2.protocol === void 0) record2.protocol = protocol;
+      },
+      token: () => {
+        const now = Date.now();
+        if (firstTokenAt === void 0) {
+          firstTokenAt = now;
+          record2.firstTokenMs = Math.max(0, now - startedAt);
+        }
+        lastTokenAt = now;
+        tokenDeltas += 1;
+        if (tokenDeltas % STREAM_MARK_EVERY === 0 && marks.length < MAX_STREAM_MARKS) {
+          marks.push(Math.max(0, now - startedAt));
+          record2.streamMarksMs = marks;
+        }
+      },
+      request: (nextMethod, nextUrl, nextRequestBytes) => {
+        if (!reported) {
+          reported = true;
+          record2.method = nextMethod;
+          record2.url = nextUrl;
+          if (nextRequestBytes !== void 0) record2.requestBytes = nextRequestBytes;
+          record2.prepMs = Math.max(0, Date.now() - startedAt);
+        }
+      },
+      status: (value) => {
+        record2.status = value;
+        if (record2.responseMs === void 0) record2.responseMs = Math.max(0, Date.now() - startedAt);
+      },
+      finish: (ok, code, usage, costUsd) => {
+        if (finished) return;
+        finished = true;
+        const seq = record2.seq;
+        const endedAt = Date.now();
+        const decodeMs = firstTokenAt === void 0 ? void 0 : Math.max(0, endedAt - firstTokenAt);
+        const streamMs = firstTokenAt === void 0 || lastTokenAt === void 0 ? void 0 : Math.max(0, lastTokenAt - firstTokenAt);
+        const tps = decodeMs === void 0 || decodeMs <= 0 || usage === void 0 || streamMs === void 0 || streamMs < MIN_STREAM_MS_FOR_RATE ? void 0 : usage.outputTokens / (decodeMs / 1e3);
+        const entry = {
+          seq,
+          atMs: startedAt,
+          model: detail.model,
+          // The protocol may have arrived after the row opened, through
+          // `identify` — the running row and the completion read the same
+          // field, so they never disagree.
+          ...record2.protocol === void 0 ? {} : { protocol: record2.protocol },
+          session,
+          method: record2.method,
+          url: record2.url,
+          ...record2.status === void 0 ? {} : { status: record2.status },
+          ...record2.prepMs === void 0 ? {} : { prepMs: record2.prepMs },
+          ...record2.responseMs === void 0 ? {} : { responseMs: record2.responseMs },
+          ...record2.streamMarksMs === void 0 ? {} : { streamMarksMs: [...record2.streamMarksMs] },
+          ...usage === void 0 ? {} : {
+            inputTokens: usage.inputTokens,
+            outputTokens: usage.outputTokens,
+            ...usage.cacheReadTokens === void 0 || usage.cacheReadTokens <= 0 ? {} : { cacheReadTokens: usage.cacheReadTokens },
+            ...usage.cacheWriteTokens === void 0 || usage.cacheWriteTokens <= 0 ? {} : { cacheWriteTokens: usage.cacheWriteTokens },
+            ...usage.reasoningTokens === void 0 || usage.reasoningTokens <= 0 ? {} : { reasoningTokens: usage.reasoningTokens }
+          },
+          ...costUsd === void 0 ? {} : { costUsd },
+          ...record2.requestBytes === void 0 ? {} : { requestBytes: record2.requestBytes },
+          ...detail.imageBytes === void 0 ? {} : { imageBytes: detail.imageBytes },
+          durationMs: Math.max(0, endedAt - startedAt),
+          ...firstTokenAt === void 0 ? {} : { firstTokenMs: Math.max(0, firstTokenAt - startedAt) },
+          ...decodeMs === void 0 ? {} : { decodeMs },
+          ...streamMs === void 0 ? {} : { streamMs },
+          ...tps === void 0 ? {} : { tps },
+          ok,
+          ...code === void 0 ? {} : { code },
+          // The pool in force when the request finished. Pool switches are
+          // rare enough that attributing the finishing pool is honest; the
+          // create/close events carry the exact history.
+          ...this.poolKey === void 0 ? {} : { poolKey: this.poolKey },
+          shape: { messages: detail.messageCount, hasImage: detail.hasImage, tools: detail.toolCount },
+          ...!debug ? {} : {
+            headers: {
+              ...detail.headers.session === void 0 ? {} : { session: detail.headers.session },
+              ...detail.headers.userAgent === void 0 ? {} : { userAgent: detail.headers.userAgent }
+            }
+          }
+        };
+        const at = this.entries.findIndex((candidate) => candidate.seq === seq);
+        if (at === -1) {
+          this.insert(entry, false);
+        } else {
+          this.entries[at] = entry;
+        }
+      }
+    };
+  }
+  insert(entry, count2) {
+    if (count2) this.totalRecords += 1;
+    const bound = Math.max(1, Math.floor(this.gates.maxEntries()));
+    const next = [entry, ...this.entries].slice(0, bound);
+    this.droppedRecords += this.entries.length + 1 - next.length;
+    this.entries = next;
+  }
+  /** Shrink the ring to a new bound immediately; growing backfills nothing. */
+  setMaxEntries(bound) {
+    if (!Number.isSafeInteger(bound) || bound < 1) return;
+    if (this.entries.length > bound) {
+      this.droppedRecords += this.entries.length - bound;
+      this.entries = this.entries.slice(0, bound);
+    }
+  }
+  /** The current reading, newest first. Computed lazily: nothing on the hot path. */
+  snapshot() {
+    return {
+      entries: this.entries,
+      poolEvents: this.poolEvents,
+      totalRecords: this.totalRecords,
+      droppedRecords: this.droppedRecords,
+      newConnections: this.newConnections,
+      ...this.poolKey === void 0 ? {} : { poolKey: this.poolKey }
+    };
+  }
+};
+
 // src/request-log/contract.ts
 function count(key, row, optional) {
   const entry = row[key];
@@ -78664,6 +78937,8 @@ function parseEntry(value) {
     ...row["costUsd"] === void 0 ? {} : { costUsd: rate("costUsd", row) },
     ...row["requestBytes"] === void 0 ? {} : { requestBytes: count("requestBytes", row) },
     ...row["prepMs"] === void 0 ? {} : { prepMs: count("prepMs", row) },
+    ...row["responseMs"] === void 0 ? {} : { responseMs: count("responseMs", row) },
+    ...row["streamMarksMs"] === void 0 ? {} : { streamMarksMs: parseMarks(row["streamMarksMs"]) },
     ...row["imageBytes"] === void 0 ? {} : { imageBytes: count("imageBytes", row) },
     ...row["durationMs"] === void 0 ? {} : { durationMs: count("durationMs", row) },
     ...row["firstTokenMs"] === void 0 ? {} : { firstTokenMs: count("firstTokenMs", row) },
@@ -78691,6 +78966,15 @@ function parseEntry(value) {
     } : {}
   };
 }
+function parseMarks(value) {
+  if (!Array.isArray(value)) throw new Error("invalid request-log stream marks");
+  const kept = [];
+  for (const mark of value) {
+    if (typeof mark !== "number" || !Number.isFinite(mark) || mark < 0 || kept.length >= MAX_STREAM_MARKS) continue;
+    kept.push(mark);
+  }
+  return kept;
+}
 function rate(key, row) {
   const entry = row[key];
   if (typeof entry !== "number" || !Number.isFinite(entry) || entry < 0) {
@@ -78906,7 +79190,7 @@ function registerRemotes(ctx) {
 
 // src/catalog/service.ts
 import { RemoteError, TypertRemoteService } from "@deepseek-ai/dsh-typert-protocol";
-import { LlmError as LlmError10 } from "@deepseek-ai/dsh-llm";
+import { LlmError as LlmError11 } from "@deepseek-ai/dsh-llm";
 
 // src/catalog/reading.ts
 function catalogReading(snapshot2, visibility, listingFailure, now) {
@@ -78987,7 +79271,7 @@ var OpencodeGoCatalogService = class extends TypertRemoteService {
       const reading = catalogReading(
         snapshot2,
         visibility,
-        failure === void 0 ? void 0 : failure instanceof LlmError10 || failure instanceof Error ? failure.message : String(failure)
+        failure === void 0 ? void 0 : failure instanceof LlmError11 || failure instanceof Error ? failure.message : String(failure)
       );
       this.reportUnadvertised(snapshot2, visibility);
       return reading;
@@ -79213,187 +79497,6 @@ var OpencodeGoRequestLogService = class extends TypertRemoteService3 {
   }
 };
 
-// src/request-log/store.ts
-var MAX_POOL_EVENTS = 20;
-var MIN_STREAM_MS_FOR_RATE = 100;
-var RequestLog = class {
-  gates;
-  entries = [];
-  poolEvents = [];
-  seq = 0;
-  totalRecords = 0;
-  droppedRecords = 0;
-  newConnections = 0;
-  poolKey;
-  constructor(gates) {
-    this.gates = gates;
-  }
-  /**
-   * Count one accepted socket. Called from the diagnostics channel, not the
-   * request path, so its cost never lands on a request.
-   */
-  connectionOpened() {
-    this.newConnections += 1;
-  }
-  /**
-   * Record one pool lifecycle event. Pools live for minutes, so this is rare
-   * by construction; still gated on debug so the common log stays request-only.
-   */
-  poolEvent(type, key) {
-    if (!this.gates.enabled() || !this.gates.debug()) return;
-    this.poolKey = type === "create" ? key : this.poolKey;
-    if (type === "close" && this.poolKey === key) this.poolKey = void 0;
-    this.poolEvents = [{ atMs: Date.now(), type, key }, ...this.poolEvents].slice(0, MAX_POOL_EVENTS);
-  }
-  /**
-   * Begin one request's span, or nothing when the log is disabled.
-   *
-   * This is the single gate on the request path: a disabled log costs one
-   * method call and one boolean read, and constructs nothing. When enabled,
-   * the running row is appended before the span is returned, so the viewer
-   * sees the request while it runs; `request` and `status` mutate that same
-   * visible row as the wire facts arrive.
-   */
-  start(detail) {
-    if (!this.gates.enabled()) return void 0;
-    const debug = this.gates.debug();
-    const startedAt = Date.now();
-    const session = detail.sessionId ?? "";
-    const record2 = {
-      seq: ++this.seq,
-      atMs: startedAt,
-      model: detail.model,
-      ...detail.protocol === void 0 ? {} : { protocol: detail.protocol },
-      session,
-      method: detail.method ?? "POST",
-      url: detail.url ?? "",
-      running: true,
-      ok: false,
-      shape: { messages: detail.messageCount, hasImage: detail.hasImage, tools: detail.toolCount },
-      ...detail.imageBytes === void 0 ? {} : { imageBytes: detail.imageBytes },
-      ...!debug ? {} : {
-        headers: {
-          ...detail.headers.session === void 0 ? {} : { session: detail.headers.session },
-          ...detail.headers.userAgent === void 0 ? {} : { userAgent: detail.headers.userAgent }
-        }
-      }
-    };
-    this.insert(record2, true);
-    let firstTokenAt;
-    let lastTokenAt;
-    let reported = false;
-    let finished = false;
-    return {
-      identify: (protocol) => {
-        if (record2.protocol === void 0) record2.protocol = protocol;
-      },
-      token: () => {
-        const now = Date.now();
-        if (firstTokenAt === void 0) {
-          firstTokenAt = now;
-          record2.firstTokenMs = Math.max(0, now - startedAt);
-        }
-        lastTokenAt = now;
-      },
-      request: (nextMethod, nextUrl, nextRequestBytes) => {
-        if (!reported) {
-          reported = true;
-          record2.method = nextMethod;
-          record2.url = nextUrl;
-          if (nextRequestBytes !== void 0) record2.requestBytes = nextRequestBytes;
-          record2.prepMs = Math.max(0, Date.now() - startedAt);
-        }
-      },
-      status: (value) => {
-        record2.status = value;
-      },
-      finish: (ok, code, usage, costUsd) => {
-        if (finished) return;
-        finished = true;
-        const seq = record2.seq;
-        const endedAt = Date.now();
-        const decodeMs = firstTokenAt === void 0 ? void 0 : Math.max(0, endedAt - firstTokenAt);
-        const streamMs = firstTokenAt === void 0 || lastTokenAt === void 0 ? void 0 : Math.max(0, lastTokenAt - firstTokenAt);
-        const tps = decodeMs === void 0 || decodeMs <= 0 || usage === void 0 || streamMs === void 0 || streamMs < MIN_STREAM_MS_FOR_RATE ? void 0 : usage.outputTokens / (decodeMs / 1e3);
-        const entry = {
-          seq,
-          atMs: startedAt,
-          model: detail.model,
-          // The protocol may have arrived after the row opened, through
-          // `identify` — the running row and the completion read the same
-          // field, so they never disagree.
-          ...record2.protocol === void 0 ? {} : { protocol: record2.protocol },
-          session,
-          method: record2.method,
-          url: record2.url,
-          ...record2.status === void 0 ? {} : { status: record2.status },
-          ...record2.prepMs === void 0 ? {} : { prepMs: record2.prepMs },
-          ...usage === void 0 ? {} : {
-            inputTokens: usage.inputTokens,
-            outputTokens: usage.outputTokens,
-            ...usage.cacheReadTokens === void 0 || usage.cacheReadTokens <= 0 ? {} : { cacheReadTokens: usage.cacheReadTokens },
-            ...usage.cacheWriteTokens === void 0 || usage.cacheWriteTokens <= 0 ? {} : { cacheWriteTokens: usage.cacheWriteTokens },
-            ...usage.reasoningTokens === void 0 || usage.reasoningTokens <= 0 ? {} : { reasoningTokens: usage.reasoningTokens }
-          },
-          ...costUsd === void 0 ? {} : { costUsd },
-          ...record2.requestBytes === void 0 ? {} : { requestBytes: record2.requestBytes },
-          ...detail.imageBytes === void 0 ? {} : { imageBytes: detail.imageBytes },
-          durationMs: Math.max(0, endedAt - startedAt),
-          ...firstTokenAt === void 0 ? {} : { firstTokenMs: Math.max(0, firstTokenAt - startedAt) },
-          ...decodeMs === void 0 ? {} : { decodeMs },
-          ...streamMs === void 0 ? {} : { streamMs },
-          ...tps === void 0 ? {} : { tps },
-          ok,
-          ...code === void 0 ? {} : { code },
-          // The pool in force when the request finished. Pool switches are
-          // rare enough that attributing the finishing pool is honest; the
-          // create/close events carry the exact history.
-          ...this.poolKey === void 0 ? {} : { poolKey: this.poolKey },
-          shape: { messages: detail.messageCount, hasImage: detail.hasImage, tools: detail.toolCount },
-          ...!debug ? {} : {
-            headers: {
-              ...detail.headers.session === void 0 ? {} : { session: detail.headers.session },
-              ...detail.headers.userAgent === void 0 ? {} : { userAgent: detail.headers.userAgent }
-            }
-          }
-        };
-        const at = this.entries.findIndex((candidate) => candidate.seq === seq);
-        if (at === -1) {
-          this.insert(entry, false);
-        } else {
-          this.entries[at] = entry;
-        }
-      }
-    };
-  }
-  insert(entry, count2) {
-    if (count2) this.totalRecords += 1;
-    const bound = Math.max(1, Math.floor(this.gates.maxEntries()));
-    const next = [entry, ...this.entries].slice(0, bound);
-    this.droppedRecords += this.entries.length + 1 - next.length;
-    this.entries = next;
-  }
-  /** Shrink the ring to a new bound immediately; growing backfills nothing. */
-  setMaxEntries(bound) {
-    if (!Number.isSafeInteger(bound) || bound < 1) return;
-    if (this.entries.length > bound) {
-      this.droppedRecords += this.entries.length - bound;
-      this.entries = this.entries.slice(0, bound);
-    }
-  }
-  /** The current reading, newest first. Computed lazily: nothing on the hot path. */
-  snapshot() {
-    return {
-      entries: this.entries,
-      poolEvents: this.poolEvents,
-      totalRecords: this.totalRecords,
-      droppedRecords: this.droppedRecords,
-      newConnections: this.newConnections,
-      ...this.poolKey === void 0 ? {} : { poolKey: this.poolKey }
-    };
-  }
-};
-
 // src/update/check.ts
 var PACKAGE_NAME = "@dan-ai-studio/dshopencodego";
 var REGISTRY_LATEST_URL = `https://registry.npmjs.org/${encodeURIComponent(PACKAGE_NAME)}/latest`;
@@ -79487,7 +79590,7 @@ function createUpdateChecker(deps) {
 }
 
 // src/update/version.ts
-var PLUGIN_VERSION = true ? "0.1.84" : "0.0.0-dev";
+var PLUGIN_VERSION = true ? "0.1.86" : "0.0.0-dev";
 
 // src/update/service.ts
 import { TypertRemoteService as TypertRemoteService4 } from "@deepseek-ai/dsh-typert-protocol";
@@ -79516,7 +79619,7 @@ var OpencodeGoUpdateService = class extends TypertRemoteService4 {
 };
 
 // src/usage/windows.ts
-import { attributionHeaders as attributionHeaders6, LlmError as LlmError11 } from "@deepseek-ai/dsh-llm";
+import { attributionHeaders as attributionHeaders6, LlmError as LlmError12 } from "@deepseek-ai/dsh-llm";
 var USAGE_MAX_BYTES = 1024 * 1024;
 var USAGE_TIMEOUT_MS = 1e4;
 function parseGoUsage(value) {
@@ -79534,7 +79637,7 @@ function parseGoUsage(value) {
 }
 async function readUsageWindows(options) {
   if (options.apiKey === void 0 || options.apiKey.length === 0) {
-    throw new LlmError11("dshopencodego: no credential is configured for the opencode-go route", "USAGE_UNAVAILABLE");
+    throw new LlmError12("dshopencodego: no credential is configured for the opencode-go route", "USAGE_UNAVAILABLE");
   }
   const url = `${options.baseURL.replace(/\/+$/, "")}/usage`;
   const timeout = AbortSignal.timeout(USAGE_TIMEOUT_MS);
@@ -79551,12 +79654,12 @@ async function readUsageWindows(options) {
       signal: options.signal === void 0 ? timeout : AbortSignal.any([options.signal, timeout])
     });
   } catch (error) {
-    throw new LlmError11(`could not reach ${url}`, "USAGE_UNAVAILABLE", { cause: error });
+    throw new LlmError12(`could not reach ${url}`, "USAGE_UNAVAILABLE", { cause: error });
   }
   if (!response.ok) {
     await response.body?.cancel().catch(() => {
     });
-    throw new LlmError11(`${url} answered HTTP ${response.status}`, "USAGE_UNAVAILABLE");
+    throw new LlmError12(`${url} answered HTTP ${response.status}`, "USAGE_UNAVAILABLE");
   }
   const body = await readBoundedJson(response, url, USAGE_MAX_BYTES);
   try {
@@ -79565,16 +79668,16 @@ async function readUsageWindows(options) {
       ...options.source === void 0 ? {} : { source: options.source }
     };
   } catch (error) {
-    throw new LlmError11(`${url} returned an invalid usage document`, "USAGE_UNAVAILABLE", { cause: error });
+    throw new LlmError12(`${url} returned an invalid usage document`, "USAGE_UNAVAILABLE", { cause: error });
   }
 }
 
 // src/usage/service.ts
 import { randomUUID as randomUUID3 } from "node:crypto";
 import { RemoteError as RemoteError2, TypertRemoteService as TypertRemoteService5 } from "@deepseek-ai/dsh-typert-protocol";
-import { LlmError as LlmError12 } from "@deepseek-ai/dsh-llm";
+import { LlmError as LlmError13 } from "@deepseek-ai/dsh-llm";
 function usageFailure(error, source) {
-  const missing = error instanceof LlmError12 && error.code === "MISSING_CREDENTIAL";
+  const missing = error instanceof LlmError13 && error.code === "MISSING_CREDENTIAL";
   return new RemoteError2(
     "dshopencodego/usage-unavailable",
     error instanceof Error ? error.message : "OpenCode Go usage is unavailable",
@@ -79601,7 +79704,7 @@ var OpencodeGoUsageService = class extends TypertRemoteService5 {
     }
     if (key === void 0 || key.length === 0) {
       this.identity = void 0;
-      throw usageFailure(new LlmError12("No OpenCode Go API key is configured", "MISSING_CREDENTIAL"), void 0);
+      throw usageFailure(new LlmError13("No OpenCode Go API key is configured", "MISSING_CREDENTIAL"), void 0);
     }
     if (this.identity?.baseURL !== baseURL || this.identity.key !== key) {
       this.identity = { baseURL, key, source: randomUUID3() };
@@ -79650,7 +79753,7 @@ function apply(ctx, raw) {
     const credentials = ctx.get("credentials");
     const hit = credentials !== void 0 ? (await credentials.resolve(credentialRef(ref)))?.value : launchEnvironmentOf(ctx).get(ref)?.value;
     if (hit !== void 0 && hit.length > 0) return llm2.assertUsableApiKey(hit, name, ref);
-    throw new LlmError13(
+    throw new LlmError14(
       `dshopencodego: no credential; the profile resolves ${ref}, which is not set \u2014 store ${ref} through the credentials service (the Web Models page writes it) or export it`,
       "MISSING_CREDENTIAL"
     );
@@ -79796,7 +79899,7 @@ function apply(ctx, raw) {
     settingsNs,
     async (request, signal) => {
       if (request.provider !== PROVIDER_ID && !(request.baseURL ?? "").includes("opencode.ai")) {
-        throw new LlmError13(
+        throw new LlmError14(
           "dshopencodego discovers only OpenCode Go/Console endpoints; enter this provider's models by hand",
           "DISCOVERY_UNSUPPORTED"
         );

package/lib/types/context-fit.d.ts

@@ -0,0 +1,58 @@
+/**
+ * 发送前按上下文窗口装配一次请求。
+ *
+ * 宿主拥有压缩会话的职责，插件拥有不上报谎言的职责：网关按
+ * `input + max_tokens <= contextWindow` 做严格校验（`muse-spark`
+ * 系 `1048576 / 131072` 即为一例），而宿主在 `input` 已占 `87%`
+ * 时仍可能按满额输出请求。本模块在发网关之前做两件事：
+ *
+ * - 输入本身已超窗：直接抛 `CONTEXT_WINDOW_EXCEEDED`，让宿主走压缩/
+ *   重试分支，而不是等网关 `400 context_length_exceeded` 再失败；
+ * - 输入加请求输出会超窗：把本次 `maxTokens` 收缩到剩余窗口内。
+ *   收缩是保底成功，不是替代压缩：历史照旧原样发送，不截断、不总结。
+ *
+ * 估算是有意保守的上界：网关分词器未知，按 `3 字符 ≈ 1 token`
+ * 加固定开销估算，多估只会多收一点输出，不会把能过的请求误杀。
+ * 图片/文件按固定开销加字节折算，视觉 token 的真实计费未知时宁可高估。
+ *
+ * @module @dan-ai-studio/dshopencodego/context-fit
+ */
+import type { GenerateOptions } from '@deepseek-ai/dsh-llm';
+/** 为分词器漂移和线上传输开销预留的窗口。 */
+export declare const CONTEXT_FIT_RESERVE_TOKENS = 2048;
+/** 收缩后仍值得发一次请求的最小输出。低于它就直接失败，逼宿主压缩。 */
+export declare const CONTEXT_FIT_MIN_OUTPUT_TOKENS = 256;
+/** 估算所需的请求子集：只读文本与附件元数据，不读字节。 */
+export interface FittableRequest {
+    readonly messages: GenerateOptions['messages'];
+    readonly system?: GenerateOptions['system'];
+    readonly tools?: GenerateOptions['tools'];
+}
+/**
+ * 保守估算一次请求的输入 token 数（上界，非精确计数）。
+ * @param options - 本次请求的消息、系统提示与工具声明。
+ * @returns 估算的输入 token 数，至少为 0 的有限整数。
+ */
+export declare function estimateInputTokens(options: FittableRequest): number;
+/** `resolveFittedMaxTokens` 的输入。 */
+export interface FitRequest {
+    /** 展示给宿主看的模型 id，只用于报错文本。 */
+    readonly modelId: string;
+    /** 该模型的上下文窗口（含本次已应用的覆盖）。 */
+    readonly contextWindow: number;
+    /** 路由自身的输出上限（含覆盖），未知时传 undefined。 */
+    readonly routeMaxTokens?: number | undefined;
+    /** 调用方本次请求的输出上限，未指定时传 undefined。 */
+    readonly requestedMaxTokens?: number | undefined;
+    /** `estimateInputTokens` 的输出。 */
+    readonly estimatedInputTokens: number;
+}
+/**
+ * 按剩余窗口装配本次的 `maxTokens`。
+ * @param fit - 窗口、路由上限、调用方请求与输入估算。
+ * @returns 发往网关的 `maxTokens`：不需要收缩时与调用方请求一致
+ *   （调用方未指定则保持 `undefined`，不断言网关默认值）。
+ * @throws {LlmError} `CONTEXT_WINDOW_EXCEEDED`：输入本身已超窗，或剩余
+ *   窗口连最小输出都放不下。调用方应据此走压缩/截断分支。
+ */
+export declare function resolveFittedMaxTokens(fit: FitRequest): number | undefined;

package/lib/types/request-log/store.d.ts

@@ -91,6 +91,20 @@ export interface RequestLogEntry {
      * reached the wire (rejected up front).
      */
     readonly prepMs?: number;
+    /**
+     * Wall time from request start to the response headers' arrival, in
+     * milliseconds. Splits the first-token wait in two: up to here the
+     * gateway is queuing, after here the model is thinking. Absent when the
+     * response never arrived.
+     */
+    readonly responseMs?: number;
+    /**
+     * Elapsed milliseconds of every tenth accepted token delta, capped.
+     * Cuts the streaming phase into visible blocks, so a stall mid-output
+     * reads as a gap rather than a uniformly slow bar. Absent when fewer
+     * than ten deltas arrived — then the output slice needs no subdivision.
+     */
+    readonly streamMarksMs?: readonly number[];
     /** Wall time from request start to terminal event, in milliseconds. */
     readonly durationMs?: number;
     /** Wall time from request start to the first token delta, in milliseconds. */
@@ -177,6 +191,13 @@ export interface RequestLogGates {
     /** How many entries to keep; the ring trims to it on every append. */
     readonly maxEntries: () => number;
 }
+/**
+ * How many streaming timestamps one row keeps at most. Past the cap the
+ * trail simply stops growing: the newest marks matter more than the oldest
+ * for reading a stall, and the oldest are still bracketed by the first
+ * token and the last one.
+ */
+export declare const MAX_STREAM_MARKS = 200;
 /**
  * Minimum content span for a meaningful rate, in milliseconds.
  *

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dan-ai-studio/dshopencodego",
-  "version": "0.1.84",
+  "version": "0.1.86",
   "description": "OpenCode Go provider for DeepSeek Harness: live gateway catalog, per-conversation session header, usage",
   "type": "module",
   "license": "MIT",