@dan-ai-studio/dshopencodego 0.1.8 → 0.1.10

package/README.en.md

@@ -76,10 +76,21 @@ Configuration lives in the profile's `cordis.patch.yml`:
         contextWindow: 262144
     modelProtocols:                           # last resort: per-model protocol override
       some-new-model: openai-responses
+    retryPolicy:                              # optional: this route's retry policy, executed by dsh-llm-retry
+      mode: normal
+      maxRetries: 2
+      backoff:
+        initialDelayMs: 500
 ```
 
 The API key comes from the Harness credential store (reference name `OPENCODE_GO_API_KEY`), or from `export OPENCODE_GO_API_KEY=...`.
 
+**Output cap**: a caller's own `maxTokens` travels as given, still clamped by a `modelLimits` ceiling. When a caller names none, the host materializes the route's ceiling — the catalog's output capacity, or whatever `modelLimits` overrode it with.
+
+**Tool declarations**: every request carries the complete current tool list. The session-folded history the host offers (`toolHistory`) is deliberately **not** projected: projecting it requires the route to declare a `toolUpdate` mode, and models.dev states no such mode for this gateway's models (only that tools are callable). Claiming one would silently change what the model sees.
+
+**Retries**: `retryPolicy` is optional — omit it and the host's own default applies. When set, the optional `dsh-llm-retry` plugin executes it, and a malformed policy fails where it is written rather than at the first failure.
+
 ## How a model is resolved
 
 Four evidence levels decide how a model is called, first hit wins:
@@ -114,12 +125,31 @@ The "OpenCode Go" section shows the live gateway catalog (42+ models) with, per
 
 - all satisfied → the plugin loads;
 - any mismatch → **that plugin row is disabled with a printed reason** (`Plugin … is incompatible with dsh …`), leaving other plugins untouched;
-- to take the risk anyway, grant an **exact-version exemption**: `dsh plugin allow-version <pkg>@<version> --dsh-version <exact> --accept-risk` (applies only to that package and that exact runtime version).
+- to take the risk anyway, grant an **exact-version exemption**: `dsh plugin --profile <profile> allow-version <pkg>@<version> --dsh-version <exact> --accept-risk` (applies only to that package and that exact runtime version); `dsh plugin --profile <profile> version-exemptions` lists what a profile holds.
 
 The `engines.dsh` field is informational for readers and package managers; **DSH's compatibility gate reads only `peerDependencies`**. `@deepseek-ai/cordis` is declared separately as `4.0.2 || 4.0.3 || 4.0.4`.
 
 **0.1.8+ or older releases**: the plugin is refused. If the seam is compatible, widen the peer range in your own build (never in DSH core), or wait for a plugin release.
 
+### Version-binding notes
+
+| Binding layer | Declared | Enforced by |
+| --- | --- | --- |
+| `peerDependencies` (15 `@deepseek-ai/dsh-*` packages) | all `>=0.1.7-alpha.1 <0.1.8` | **DSH's gate, at plugin load time** |
+| `@deepseek-ai/cordis` | `4.0.2 \|\| 4.0.3 \|\| 4.0.4` | the same gate |
+| `engines.dsh` | `>=0.1.7-alpha.1 <0.1.8` | informational only; DSH never reads it |
+| `engines.node` | `^22.19.0 \|\| >=24.0.0` | the package manager |
+| bundled `@earendil-works/pi-ai` | pinned to exactly `0.87.1` | an independent coupling: a change in pi-ai's request construction changes this plugin's wire behaviour |
+| bundled `@deepseek-ai/schemastery` | `^3.18.3` | ordinary semver |
+
+Things to keep in mind when maintaining this:
+
+- **Required versus optional peers**: six of the fifteen are marked `optional` (`dsh-api-remotes`, `dsh-client-locale`, `dsh-client-store`, `dsh-client-ui-model-selection`, `dsh-client-ui-settings`, `dsh-client-ui-slots`). The nine that actually block loading are `dsh-llm`, `dsh-typert-protocol`, `dsh-attachment`, `dsh-brand`, `dsh-credentials`, `dsh-fs`, `dsh-launch-environment`, `dsh-settings`, and `dsh-timeout`. `dsh-llm` is the heaviest coupling — more than twenty imports across the source.
+- **Declared range ≠ verified range**: `0.1.7-alpha.1` and `alpha.2` fall inside the declaration but were never verified here. What was verified is `0.1.7-rc.1` (the dev-dependency baseline) and `0.1.7-rc.2` (daily use).
+- **The upper bound is a tracking line**: `<0.1.8` means that the moment DSH ships `0.1.8`, this plugin must ship a new release in the same window or every user loses the plugin. After changing a range, `npm test` is the verification (local mock gateway, no network, no tokens) — ranges follow what the interfaces declare, and are never "measured" with live probes.
+- **Peer coverage is now guarded both ways**: `dsh-typert-registry`, `dsh-client-ui-conversation`, and `dsh-client-ui-renderer` were imported by the source but absent from `peerDependencies`, invisible to the gate; they are declared now, and `tests/peer-coverage.spec.ts` asserts both directions — anything the source imports must be declared, and anything declared but unreached must be removed (`dsh-settings` went that way: the settings form is actually provided by `dsh-client-ui-settings`).
+- **Exemptions are per profile**: only a profile that ran `dsh plugin allow-version` has one, and no record means the plugin must stay inside the declared range. `dsh plugin version-exemptions` lists a profile's exemptions.
+
 ## Troubleshooting
 
 | Symptom | Cause and fix |

package/README.md

@@ -76,10 +76,21 @@ dsh plugin --profile web remove @dan-ai-studio/dshopencodego   # 卸载
         contextWindow: 262144
     modelProtocols:                           # 最后兜底：逐模型协议覆盖
       some-new-model: openai-responses
+    retryPolicy:                              # 可选：本路由的重试策略（由 dsh-llm-retry 执行）
+      mode: normal
+      maxRetries: 2
+      backoff:
+        initialDelayMs: 500
 ```
 
 API Key 通过 Harness 凭证库提供（引用名 `OPENCODE_GO_API_KEY`），也可以直接 `export OPENCODE_GO_API_KEY=...`。
 
+**输出上限**：调用方给了 `maxTokens` 就用它（仍受 `modelLimits` 的逐模型上限约束）；没给时由宿主按路由上限物化——该上限就是目录里的输出容量，或 `modelLimits` 覆盖后的值。
+
+**工具声明**：每个请求都发送当前完整的工具列表。宿主提供的会话折叠历史（`toolHistory`）**故意不投影**：投影需要路由声明 `toolUpdate` 模式，而 models.dev 对本网关模型只声明"可调用工具"、没有声明该模式；凭空声明会让模型看到的东西悄悄改变。
+
+**重试**：`retryPolicy` 是可选的，不写就用宿主默认。写了就由可选的 `dsh-llm-retry` 插件执行，配置错误在写入时报错而不是等到失败发生。
+
 ## 模型是怎么判定的
 
 一个模型"怎么调"由四级证据决定，先命中先赢：
@@ -114,12 +125,31 @@ API Key 通过 Harness 凭证库提供（引用名 `OPENCODE_GO_API_KEY`），
 
 - 全部满足 → 正常加载；
 - 任一不满足 → **该插件行被禁用并打印原因**（`Plugin … is incompatible with dsh …`），其他插件不受影响；
-- 需要冒险时可用**精确版本豁免**：`dsh plugin allow-version <包>@<版本> --dsh-version <精确版本> --accept-risk`（只对该包与该精确版本生效）。
+- 需要冒险时可用**精确版本豁免**：`dsh plugin --profile <profile> allow-version <包>@<版本> --dsh-version <精确版本> --accept-risk`（只对该包与该精确版本生效）；`dsh plugin --profile <profile> version-exemptions` 可列出当前 profile 的豁免。
 
 `package.json` 里的 `engines.dsh` 是本插件为读者/包管理器写的信息字段；**DSH 的兼容门禁只读 `peerDependencies`**，请以它为准。`@deepseek-ai/cordis` 单独声明为 `4.0.2 || 4.0.3 || 4.0.4`。
 
 **0.1.8+ 或更早版本**：会被拒绝加载。如果 DSH 侧接口兼容，可自行放宽该插件的 peer 范围并重新构建（不改 DSH 核心）；否则请等插件跟进发版。
 
+### 版本绑定的注意事项
+
+| 绑定层 | 声明 | 谁在把关 |
+| --- | --- | --- |
+| `peerDependencies`（15 个 `@deepseek-ai/dsh-*`） | 均为 `>=0.1.7-alpha.1 <0.1.8` | **DSH 加载时的真门禁** |
+| `@deepseek-ai/cordis` | `4.0.2 \|\| 4.0.3 \|\| 4.0.4` | 同一门禁 |
+| `engines.dsh` | `>=0.1.7-alpha.1 <0.1.8` | 仅信息字段，DSH 不读 |
+| `engines.node` | `^22.19.0 \|\| >=24.0.0` | 包管理器 |
+| 自带依赖 `@earendil-works/pi-ai` | 精确锁 `0.87.1` | 独立耦合：pi-ai 改请求构造即影响本插件的线上行为 |
+| 自带依赖 `@deepseek-ai/schemastery` | `^3.18.3` | 常规 semver |
+
+维护时需要注意：
+
+- **必选与可选之分**：15 个 peer 中 6 个标了 `optional`（`dsh-api-remotes`、`dsh-client-locale`、`dsh-client-store`、`dsh-client-ui-model-selection`、`dsh-client-ui-settings`、`dsh-client-ui-slots`）；真正卡住加载的是 9 个——`dsh-llm`、`dsh-typert-protocol`、`dsh-attachment`、`dsh-brand`、`dsh-credentials`、`dsh-fs`、`dsh-launch-environment`、`dsh-settings`、`dsh-timeout`。其中 `dsh-llm` 是最重的一处耦合（源码 import 二十余处）。
+- **名义范围 ≠ 实测范围**：`0.1.7-alpha.1`、`alpha.2` 落在声明范围内，但本项目没有验证过；实际验证过的是 `0.1.7-rc.1`（开发依赖基线）与 `0.1.7-rc.2`（日常使用）。
+- **上界就是跟版线**：`<0.1.8` 意味着 DSH 一旦发布 `0.1.8`，本插件必须同批发新版，否则所有用户加载失败。调整范围后用 `npm test` 验证即可（本地 mock 网关，零网络零 token）——范围以接口声明为准，不要用线上探测来"测"出一个范围。
+- **peer 覆盖已双向守住**：`dsh-typert-registry`、`dsh-client-ui-conversation`、`dsh-client-ui-renderer` 曾被源码 import 却未声明、门禁管不到，现已补进 `peerDependencies`；`tests/peer-coverage.spec.ts` 双向断言——源码用到的必须声明，声明了却没人用的必须删（`dsh-settings` 就这样被移除：设置表单实际由 `dsh-client-ui-settings` 提供）。
+- **豁免是按 profile 记的**：只有显式执行过 `dsh plugin allow-version` 的 profile 才享有豁免；没有记录就代表必须落在声明范围内。用 `dsh plugin version-exemptions` 可查看当前 profile 的豁免。
+
 ## 故障排查
 
 | 现象 | 原因与处置 |

package/lib/client.js

@@ -440,7 +440,7 @@ if (!document.querySelector("style[data-plugin-css=" + JSON.stringify(id) + "]")
   style.textContent = ".d7lz0W_root{color:var(--dsw-alias-label-primary);flex-direction:column;gap:1.25rem;display:flex}.d7lz0W_title{margin:0;font-size:1.05rem}.d7lz0W_hint{color:var(--dsw-alias-label-secondary);margin:0;font-size:.8rem}.d7lz0W_warn{color:var(--dsw-alias-state-warn-primary);margin:0;font-size:.8rem}.d7lz0W_ok{color:var(--dsw-alias-state-success-primary);font-size:.8rem}.d7lz0W_block{border:1px solid var(--dsw-alias-border-l1);background:var(--dsw-alias-bg-layer-1);border-radius:.6rem;flex-direction:column;gap:.5rem;padding:.85rem;display:flex}.d7lz0W_row{flex-wrap:wrap;align-items:center;gap:.5rem;display:flex}.d7lz0W_row>strong{margin-right:auto}.d7lz0W_input{border:1px solid var(--dsw-alias-border-l1);background:var(--dsw-alias-bg-base);min-width:10rem;color:var(--dsw-alias-label-primary);font:inherit;border-radius:.35rem;flex:14rem;padding:.35rem .5rem;font-size:.85rem}.d7lz0W_button{border:1px solid var(--dsw-alias-border-l1);background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-label-primary);font:inherit;cursor:pointer;border-radius:.35rem;padding:.35rem .7rem;font-size:.8rem}.d7lz0W_button:disabled{color:var(--dsw-alias-state-idle-primary);cursor:default}.d7lz0W_number{border:1px solid var(--dsw-alias-border-l1);background:var(--dsw-alias-bg-base);width:5rem;color:var(--dsw-alias-label-primary);font:inherit;border-radius:.35rem;padding:.25rem .4rem;font-size:.8rem}.d7lz0W_select{border:1px solid var(--dsw-alias-border-l1);background:var(--dsw-alias-bg-base);width:auto;max-width:100%;color:var(--dsw-alias-label-primary);font:inherit;border-radius:.35rem;padding:.25rem .4rem;font-size:.8rem}.d7lz0W_segmented{border:1px solid var(--dsw-alias-border-l1);border-radius:.35rem;display:inline-flex;overflow:hidden}.d7lz0W_segment,.d7lz0W_segmentActive{background:var(--dsw-alias-bg-base);color:var(--dsw-alias-label-secondary);font:inherit;cursor:pointer;border:0;padding:.25rem .6rem;font-size:.8rem}.d7lz0W_segmentActive{background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-label-primary)}.d7lz0W_tableWrap{overflow-x:auto}.d7lz0W_table{border-collapse:collapse;width:100%;min-width:62rem;font-size:.78rem}.d7lz0W_table th{color:var(--dsw-alias-label-secondary);text-align:left;white-space:nowrap;padding:.25rem .45rem;font-weight:500}.d7lz0W_table td{border-top:1px solid var(--dsw-alias-border-l1);vertical-align:middle;white-space:nowrap;padding:.28rem .45rem}.d7lz0W_toggle{width:2rem}.d7lz0W_toggle input{accent-color:var(--dsw-alias-brand-primary)}.d7lz0W_list{flex-direction:column;display:flex}.d7lz0W_item{border-top:1px solid var(--dsw-alias-border-l1);align-items:flex-start;gap:.5rem;padding:.45rem .15rem;display:flex}.d7lz0W_item input[type=checkbox]{accent-color:var(--dsw-alias-brand-primary);margin-top:.15rem}.d7lz0W_itemBody{flex-direction:column;gap:.15rem;min-width:0;display:flex}.d7lz0W_itemMain{flex-wrap:wrap;align-items:baseline;gap:.6rem;display:flex}.d7lz0W_itemMeta{color:var(--dsw-alias-label-secondary);flex-wrap:wrap;gap:.75rem;font-size:.75rem;display:flex}.d7lz0W_name{color:var(--dsw-alias-label-primary)}.d7lz0W_mono{color:var(--dsw-alias-label-secondary);font-family:ui-monospace,SFMono-Regular,Menlo,monospace}.d7lz0W_badge{color:var(--dsw-alias-brand-primary)}";
   document.head.appendChild(style);
 }
-var section_default = { "ok": "d7lz0W_ok", "hint": "d7lz0W_hint", "number": "d7lz0W_number", "root": "d7lz0W_root", "item": "d7lz0W_item", "segmentActive": "d7lz0W_segmentActive", "segmented": "d7lz0W_segmented", "itemMeta": "d7lz0W_itemMeta", "segment": "d7lz0W_segment", "tableWrap": "d7lz0W_tableWrap", "input": "d7lz0W_input", "select": "d7lz0W_select", "name": "d7lz0W_name", "toggle": "d7lz0W_toggle", "button": "d7lz0W_button", "mono": "d7lz0W_mono", "warn": "d7lz0W_warn", "row": "d7lz0W_row", "itemBody": "d7lz0W_itemBody", "block": "d7lz0W_block", "table": "d7lz0W_table", "badge": "d7lz0W_badge", "list": "d7lz0W_list", "itemMain": "d7lz0W_itemMain", "title": "d7lz0W_title" };
+var section_default = { "block": "d7lz0W_block", "table": "d7lz0W_table", "ok": "d7lz0W_ok", "row": "d7lz0W_row", "button": "d7lz0W_button", "itemBody": "d7lz0W_itemBody", "mono": "d7lz0W_mono", "title": "d7lz0W_title", "badge": "d7lz0W_badge", "list": "d7lz0W_list", "segment": "d7lz0W_segment", "warn": "d7lz0W_warn", "segmentActive": "d7lz0W_segmentActive", "segmented": "d7lz0W_segmented", "tableWrap": "d7lz0W_tableWrap", "toggle": "d7lz0W_toggle", "root": "d7lz0W_root", "input": "d7lz0W_input", "select": "d7lz0W_select", "item": "d7lz0W_item", "hint": "d7lz0W_hint", "itemMain": "d7lz0W_itemMain", "itemMeta": "d7lz0W_itemMeta", "name": "d7lz0W_name", "number": "d7lz0W_number" };
 
 // src/client/Section.tsx
 var import_jsx_runtime = require("react/jsx-runtime");
@@ -996,7 +996,7 @@ if (!document.querySelector("style[data-plugin-css=" + JSON.stringify(id2) + "]"
   style.textContent = ".mMNkkW_root{align-items:center;display:inline-flex;position:relative}.mMNkkW_trigger{border:1px solid var(--dsw-alias-border-l1);background:var(--dsw-alias-bg-layer-1);color:var(--dsw-alias-label-secondary);font:inherit;cursor:pointer;white-space:nowrap;border-radius:999px;padding:.35rem .6rem;font-size:.8rem;line-height:1}.mMNkkW_trigger:hover{color:var(--dsw-alias-label-primary);border-color:var(--dsw-alias-border-l2)}.mMNkkW_panel{z-index:20;border:1px solid var(--dsw-alias-border-l1);background:var(--dsw-alias-bg-overlay);width:18rem;color:var(--dsw-alias-label-primary);border-radius:.6rem;flex-direction:column;gap:.6rem;padding:.75rem .85rem;display:flex;position:absolute;bottom:calc(100% + .5rem);right:0;box-shadow:0 8px 24px #0000002e}.mMNkkW_window{flex-direction:column;gap:.25rem;display:flex}.mMNkkW_row{justify-content:space-between;gap:.5rem;font-size:.85rem;display:flex}.mMNkkW_hint{color:var(--dsw-alias-label-secondary);margin:0;font-size:.75rem}.mMNkkW_progress{width:100%;height:.35rem;accent-color:var(--dsw-alias-brand-primary)}.mMNkkW_limited{color:var(--dsw-alias-state-warn-primary);font-size:.75rem}.mMNkkW_warning{border:1px solid var(--dsw-alias-state-warn-primary);color:var(--dsw-alias-label-primary);border-radius:.4rem;padding:.4rem .5rem;font-size:.75rem}.mMNkkW_retry{border:1px solid var(--dsw-alias-border-l1);background:var(--dsw-alias-bg-layer-1);color:var(--dsw-alias-label-primary);font:inherit;cursor:pointer;border-radius:.35rem;align-self:flex-start;padding:.25rem .5rem;font-size:.75rem}.mMNkkW_local{border-top:1px solid var(--dsw-alias-border-l1);flex-direction:column;gap:.2rem;padding-top:.5rem;display:flex}";
   document.head.appendChild(style);
 }
-var pill_default = { "trigger": "mMNkkW_trigger", "window": "mMNkkW_window", "row": "mMNkkW_row", "root": "mMNkkW_root", "limited": "mMNkkW_limited", "local": "mMNkkW_local", "panel": "mMNkkW_panel", "retry": "mMNkkW_retry", "hint": "mMNkkW_hint", "progress": "mMNkkW_progress", "warning": "mMNkkW_warning" };
+var pill_default = { "root": "mMNkkW_root", "window": "mMNkkW_window", "limited": "mMNkkW_limited", "progress": "mMNkkW_progress", "row": "mMNkkW_row", "retry": "mMNkkW_retry", "local": "mMNkkW_local", "hint": "mMNkkW_hint", "panel": "mMNkkW_panel", "warning": "mMNkkW_warning", "trigger": "mMNkkW_trigger" };
 
 // src/client/UsagePill.tsx
 var import_jsx_runtime2 = require("react/jsx-runtime");

package/lib/index.js

@@ -57,7 +57,8 @@ import {
   contentHasImage as contentHasImage2,
   LlmAdapter,
   LlmError as LlmError7,
-  ReasoningEffortId
+  ReasoningEffortId,
+  resolveRetryPolicy
 } from "@deepseek-ai/dsh-llm";
 import { idleWatchdog, timeoutOf } from "@deepseek-ai/dsh-timeout";
 
@@ -223,17 +224,12 @@ function rates(value) {
   };
   return { input: rate("input"), output: rate("output"), cacheRead: rate("cache_read"), cacheWrite: rate("cache_write") };
 }
-var DISABLES_THINKING_WHEN_UNSET = /* @__PURE__ */ new Set(["deepseek", "zai", "qwen", "qwen-chat-template"]);
-var UNDOCUMENTED_EFFORTS = ["low", "medium", "high"];
 function thinkingLevels(metadata, known) {
   const options = metadata["reasoning_options"];
   if (!Array.isArray(options)) return known?.thinkingLevelMap;
   const map = Object.fromEntries(LEVELS.map((level) => [level, null]));
   for (const item of options) {
     const option = record(item);
-    if (option["type"] === "toggle" || option["type"] === "budget_tokens") {
-      map.high ??= "high";
-    }
     if (option["type"] !== "effort" || !Array.isArray(option["values"])) continue;
     for (const value of option["values"]) {
       const level = value === "none" ? "off" : value;
@@ -242,17 +238,6 @@ function thinkingLevels(metadata, known) {
       }
     }
   }
-  const format = known?.compat?.thinkingFormat;
-  if (known?.reasoning === true && format !== void 0 && DISABLES_THINKING_WHEN_UNSET.has(format) && known.thinkingLevelMap?.off !== null) {
-    map.off ??= known.thinkingLevelMap?.off ?? "off";
-  }
-  if (options.length === 0) {
-    const installed = known?.thinkingLevelMap;
-    if (installed !== void 0 && LEVELS.some((level) => typeof installed[level] === "string")) return installed;
-    if (format === void 0) {
-      for (const level of UNDOCUMENTED_EFFORTS) map[level] = level;
-    }
-  }
   return map;
 }
 function familySibling(id, api, documentFamily, entries, builtin) {
@@ -621,6 +606,7 @@ async function discoverCatalogModels(catalog) {
 }
 
 // src/config.ts
+import { RetryPolicySchema } from "@deepseek-ai/dsh-llm";
 import { MAX_TIMER_DELAY_MS } from "@deepseek-ai/dsh-timeout";
 import z from "@deepseek-ai/schemastery";
 var DEFAULT_API_KEY_ENV = "OPENCODE_GO_API_KEY";
@@ -643,14 +629,18 @@ var fields = {
     contextWindow: z.union([z.const(null), z.number().step(1).min(1).max(Number.MAX_SAFE_INTEGER)]),
     maxTokens: z.union([z.const(null), z.number().step(1).min(1).max(Number.MAX_SAFE_INTEGER)])
   })])).default({}),
-  modelProtocols: z.dict(z.string()).default({})
+  modelProtocols: z.dict(z.string()).default({}),
+  retryPolicy: RetryPolicySchema
 };
 var PlainConfig = z.object(fields);
 var Config = z.object(Object.fromEntries(
   Object.entries(fields).map(([key, schema]) => [key, schema.volatile()])
 ));
 function readConfig(config) {
-  return Object.fromEntries(Object.keys(fields).map((key) => [key, config[key].get()]));
+  return Object.fromEntries(Object.keys(fields).flatMap((key) => {
+    const field = config[key];
+    return field?.get === void 0 ? [] : [[key, field.get()]];
+  }));
 }
 function assertBaseURL(raw) {
   let url;
@@ -1018,8 +1008,7 @@ import {
   EMPTY_RESPONSE_CODE,
   isContextWindowExceededError,
   isQuotaExceededError,
-  LlmError as LlmError6,
-  QUOTA_EXCEEDED_CODE
+  LlmError as LlmError6
 } from "@deepseek-ai/dsh-llm";
 import { isContextOverflow } from "@earendil-works/pi-ai";
 function mapUsage(usage) {
@@ -1031,9 +1020,10 @@ function mapUsage(usage) {
     ...usage.cacheWrite > 0 ? { cacheWriteTokens: usage.cacheWrite } : {}
   };
 }
+var ACCOUNT_QUOTA_CODE = "ACCOUNT_QUOTA";
 function classifyPiAiError(message) {
   if (/\b(?:401|403)\b/.test(message)) return "AUTH";
-  if (isQuotaExceededError(message)) return QUOTA_EXCEEDED_CODE;
+  if (isQuotaExceededError(message)) return ACCOUNT_QUOTA_CODE;
   if (/\b429\b|rate.?limit/i.test(message)) return "RATE_LIMIT";
   if (/\b413\b|payload too large|request body too large|length limit exceeded/i.test(message)) return "INVALID_REQUEST";
   if (/\b400\b|invalid.?request/i.test(message)) return "INVALID_REQUEST";
@@ -1264,6 +1254,7 @@ function providerHeaders(sessionId) {
 }
 
 // src/adapter.ts
+var DISABLES_THINKING_WHEN_UNSET = /* @__PURE__ */ new Set(["deepseek", "zai", "qwen", "qwen-chat-template"]);
 function withModelLimit(model, limits) {
   const limit = limits[model.id];
   if (limit === null || limit === void 0) return model;
@@ -1347,9 +1338,23 @@ var OpencodeGoAdapter = class extends LlmAdapter {
       name: model.name,
       inputModalities: [...model.input],
       context: { contextWindow: model.contextWindow },
+      // The route's own output ceiling, so the host materializes it when a
+      // caller names none and a caller that does name one still wins.
+      defaultMaxTokens: model.maxTokens,
       ...reasoning
     };
   }
+  /**
+   * This route's retry policy, resolved from configuration on every read so a
+   * profile edit reaches the next failed step. Absent configuration leaves the
+   * host's own default in force.
+   * @param _provider - unused; the mount owns exactly one route.
+   * @returns the resolved policy, or undefined when none is configured.
+   */
+  providerRetryPolicy(_provider) {
+    const configured = this.options.config().retryPolicy;
+    return configured === void 0 ? void 0 : resolveRetryPolicy(configured, "dshopencodego: retryPolicy");
+  }
   /** Validate an explicit effort against the model's own levels, without clamping. */
   resolveReasoningLevel(model, effort) {
     if (effort === void 0 || effort === "off") return void 0;
@@ -1371,8 +1376,8 @@ var OpencodeGoAdapter = class extends LlmAdapter {
       const facts = snapshot.facts.get(options.model);
       if (facts === void 0) throw new LlmError7(`opencode-go has no model "${options.model}"`, "UNKNOWN_MODEL");
       const model = withModelLimit(toPiModel(facts, config.baseURL), config.modelLimits);
-      const outputLimit = config.modelLimits[model.id]?.maxTokens;
-      const maxTokens = outputLimit === null || outputLimit === void 0 ? options.maxTokens : Math.min(options.maxTokens ?? outputLimit, outputLimit);
+      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 apiKey = await this.options.resolveApiKey();
       if (apiKey === void 0 || apiKey.length === 0) {
         throw new LlmError7("dshopencodego: no credential resolved for the opencode-go route", "MISSING_CREDENTIAL");

package/lib/types/adapter.d.ts

@@ -13,7 +13,7 @@
  * @module @dan-ai-studio/dshopencodego/adapter
  */
 import { attributionHeaders, LlmAdapter } from '@deepseek-ai/dsh-llm';
-import type { GenerateOptions, ImageAttachmentAccess, LlmModelInfo, LlmResolvedModelInfo, StreamChunk, TokenUsage } from '@deepseek-ai/dsh-llm';
+import type { GenerateOptions, ImageAttachmentAccess, LlmModelInfo, LlmResolvedModelInfo, ResolvedRetryPolicy, StreamChunk, TokenUsage } from '@deepseek-ai/dsh-llm';
 import type { AttachmentStore, ImageAttachmentRef } from '@deepseek-ai/dsh-attachment';
 import { OpencodeGoCatalog } from './catalog/index.ts';
 import type { OpencodeGoConfig } from './config.ts';
@@ -76,6 +76,14 @@ export declare class OpencodeGoAdapter extends LlmAdapter {
     resolveModel(_provider: string, model: string, _signal?: AbortSignal): Promise<LlmResolvedModelInfo>;
     /** Describe one model: capacities plus the reasoning levels it really offers. */
     private modelInfo;
+    /**
+     * This route's retry policy, resolved from configuration on every read so a
+     * profile edit reaches the next failed step. Absent configuration leaves the
+     * host's own default in force.
+     * @param _provider - unused; the mount owns exactly one route.
+     * @returns the resolved policy, or undefined when none is configured.
+     */
+    providerRetryPolicy(_provider: string): ResolvedRetryPolicy | undefined;
     /** Validate an explicit effort against the model's own levels, without clamping. */
     private resolveReasoningLevel;
     stream(options: GenerateOptions): AsyncIterable<StreamChunk>;

package/lib/types/adapter.js

@@ -65,14 +65,20 @@ var __disposeResources = (this && this.__disposeResources) || (function (Suppres
     return e.name = "SuppressedError", e.error = error, e.suppressed = suppressed, e;
 });
 import { getSupportedThinkingLevels, normalizeContext } from '@earendil-works/pi-ai';
-import { attributionHeaders, contentHasImage, LlmAdapter, LlmError, ReasoningEffortId, } from '@deepseek-ai/dsh-llm';
+import { attributionHeaders, contentHasImage, LlmAdapter, LlmError, ReasoningEffortId, resolveRetryPolicy, } from '@deepseek-ai/dsh-llm';
 import { idleWatchdog, timeoutOf } from '@deepseek-ai/dsh-timeout';
 import { DISPLAY_NAME, PROVIDER_ID, OpencodeGoCatalog } from "./catalog/index.js";
-import { DISABLES_THINKING_WHEN_UNSET, toPiModel } from "./catalog/metadata.js";
+import { toPiModel } from "./catalog/metadata.js";
 import { assertBaseURL } from "./config.js";
 import { toPiContext, toStreamChunks } from "./conversion/index.js";
 import { isModelEnabled, recommendedIds } from "./models.js";
 import { providerHeaders } from "./session-header.js";
+/**
+ * pi-ai thinking formats that answer an unset effort with an explicit disable.
+ * Every other format omits the parameter and lets the provider decide, so only
+ * these need a default effort to avoid silently turning thinking off.
+ */
+const DISABLES_THINKING_WHEN_UNSET = new Set(['deepseek', 'zai', 'qwen', 'qwen-chat-template']);
 /** Apply one request's configured capacities without touching the catalog. */
 function withModelLimit(model, limits) {
     const limit = limits[model.id];
@@ -177,9 +183,25 @@ export class OpencodeGoAdapter extends LlmAdapter {
             name: model.name,
             inputModalities: [...model.input],
             context: { contextWindow: model.contextWindow },
+            // The route's own output ceiling, so the host materializes it when a
+            // caller names none and a caller that does name one still wins.
+            defaultMaxTokens: model.maxTokens,
             ...reasoning,
         };
     }
+    /**
+     * This route's retry policy, resolved from configuration on every read so a
+     * profile edit reaches the next failed step. Absent configuration leaves the
+     * host's own default in force.
+     * @param _provider - unused; the mount owns exactly one route.
+     * @returns the resolved policy, or undefined when none is configured.
+     */
+    providerRetryPolicy(_provider) {
+        const configured = this.options.config().retryPolicy;
+        return configured === undefined
+            ? undefined
+            : resolveRetryPolicy(configured, 'dshopencodego: retryPolicy');
+    }
     /** Validate an explicit effort against the model's own levels, without clamping. */
     resolveReasoningLevel(model, effort) {
         // "off" is not a level to validate: it is the absence of a request, exactly
@@ -203,10 +225,13 @@ export class OpencodeGoAdapter extends LlmAdapter {
             if (facts === undefined)
                 throw new LlmError(`opencode-go has no model "${options.model}"`, 'UNKNOWN_MODEL');
             const model = withModelLimit(toPiModel(facts, config.baseURL), config.modelLimits);
-            const outputLimit = config.modelLimits[model.id]?.maxTokens;
-            const maxTokens = outputLimit === null || outputLimit === undefined
+            // `defaultMaxTokens` already carries the route's ceiling for a caller that
+            // names none, so only a caller's own cap travels, clamped by a configured
+            // per-model ceiling when one exists.
+            const ceiling = config.modelLimits[model.id]?.maxTokens;
+            const maxTokens = options.maxTokens === undefined || ceiling === null || ceiling === undefined
                 ? options.maxTokens
-                : Math.min(options.maxTokens ?? outputLimit, outputLimit);
+                : Math.min(options.maxTokens, ceiling);
             const apiKey = await this.options.resolveApiKey();
             if (apiKey === undefined || apiKey.length === 0) {
                 throw new LlmError('dshopencodego: no credential resolved for the opencode-go route', 'MISSING_CREDENTIAL');

package/lib/types/catalog/metadata.d.ts

@@ -49,13 +49,6 @@ export interface ModelDefaults {
     readonly maxTokens: number;
     readonly input: readonly ('text' | 'image')[];
 }
-/**
- * Transports whose request branch spells "thinking off" itself — DeepSeek's
- * separate thinking flag, for example — so an installed `off` marker is a
- * promise they can keep. Everywhere else the marker would ride the wire as
- * `reasoning_effort: "off"`.
- */
-export declare const DISABLES_THINKING_WHEN_UNSET: ReadonlySet<string>;
 /** The base URL one protocol's SDK expects for this gateway. */
 export declare function modelBaseURL(api: WireProtocol, baseURL: string): string;
 /** Inputs the ladder needs beyond the document itself. */

package/lib/types/catalog/metadata.js

@@ -41,26 +41,17 @@ function rates(value) {
     };
     return { input: rate('input'), output: rate('output'), cacheRead: rate('cache_read'), cacheWrite: rate('cache_write') };
 }
-/**
- * Transports whose request branch spells "thinking off" itself — DeepSeek's
- * separate thinking flag, for example — so an installed `off` marker is a
- * promise they can keep. Everywhere else the marker would ride the wire as
- * `reasoning_effort: "off"`.
- */
-export const DISABLES_THINKING_WHEN_UNSET = new Set(['deepseek', 'zai', 'qwen', 'qwen-chat-template']);
-/**
- * The efforts the gateway accepts for a reasoning model whose document names
- * none. Measured 2026-09-25 on every such model that was available:
- * low/medium/high all stream, while "minimal" and "off" answer 400.
- */
-const UNDOCUMENTED_EFFORTS = ['low', 'medium', 'high'];
 /**
  * Reasoning levels a model actually offers.
  *
- * An absent `reasoning_options` list is not "no levels": it means the document
- * does not describe them, so an installed entry's map is kept when there is
- * one. Levels the model does not offer stay `null`, which is what stops the
- * seam from offering a control the provider would ignore.
+ * Only what a source states becomes a level. An absent `reasoning_options`
+ * list is not "no levels": it means the document does not describe them, so an
+ * installed entry's map is kept when there is one. A `toggle` or
+ * `budget_tokens` option says the model can think, not how the wire spells a
+ * level, so it contributes none — inventing one would put a value on the wire
+ * that the gateway may reject. Levels the model does not offer stay `null`,
+ * which is what stops the seam from offering a control the provider would
+ * ignore.
  */
 function thinkingLevels(metadata, known) {
     const options = metadata['reasoning_options'];
@@ -69,13 +60,6 @@ function thinkingLevels(metadata, known) {
     const map = Object.fromEntries(LEVELS.map(level => [level, null]));
     for (const item of options) {
         const option = record(item);
-        if (option['type'] === 'toggle' || option['type'] === 'budget_tokens') {
-            // A switch exists but the document names no wire value; "high" is the
-            // one "on" spelling measured to work wherever this gateway offers such
-            // a switch. "off" is deliberately not invented here: some of these
-            // models answer `reasoning_effort: "off"` with a 400.
-            map.high ??= 'high';
-        }
         if (option['type'] !== 'effort' || !Array.isArray(option['values']))
             continue;
         for (const value of option['values']) {
@@ -85,26 +69,6 @@ function thinkingLevels(metadata, known) {
             }
         }
     }
-    const format = known?.compat?.thinkingFormat;
-    // Only a transport that expresses "off" in its own convention may keep the
-    // marker; one that would send it as `reasoning_effort` must not promise a
-    // switch it cannot keep (measured: qwen3.7-max, qwen3.8-max and qwen3.8-flash
-    // all answer 400 to that value).
-    if (known?.reasoning === true && format !== undefined && DISABLES_THINKING_WHEN_UNSET.has(format) && known.thinkingLevelMap?.off !== null) {
-        map.off ??= known.thinkingLevelMap?.off ?? 'off';
-    }
-    if (options.length === 0) {
-        // "Reasoning, options undocumented" — the MiMo family and friends.
-        const installed = known?.thinkingLevelMap;
-        if (installed !== undefined && LEVELS.some(level => typeof installed[level] === 'string'))
-            return installed;
-        // Without a format of its own the transport speaks plain reasoning_effort,
-        // so it offers exactly the measured trio.
-        if (format === undefined) {
-            for (const level of UNDOCUMENTED_EFFORTS)
-                map[level] = level;
-        }
-    }
     return map;
 }
 /**

package/lib/types/config.d.ts

@@ -11,6 +11,7 @@
  *
  * @module @dan-ai-studio/dshopencodego/config
  */
+import type { RetryPolicyConfig } from '@deepseek-ai/dsh-llm';
 import z from '@deepseek-ai/schemastery';
 /** Credential reference resolving the OpenCode Go API key. */
 export declare const DEFAULT_API_KEY_ENV = "OPENCODE_GO_API_KEY";
@@ -70,6 +71,11 @@ export interface OpencodeGoConfig {
      * The escape hatch for a model whose family rule guessed wrong.
      */
     modelProtocols: Record<string, string>;
+    /**
+     * Retry policy for this route, owned here and executed by the optional
+     * `dsh-llm-retry` plugin. Absent leaves the host's own default in force.
+     */
+    retryPolicy?: RetryPolicyConfig;
 }
 /** Plain resolved values used by the adapter. */
 export declare const PlainConfig: z<OpencodeGoConfig>;
@@ -81,7 +87,12 @@ export type LiveConfig = {
 };
 /** Runtime schema for {@link OpencodeGoConfig}; every field stays live. */
 export declare const Config: z<Partial<OpencodeGoConfig>, LiveConfig>;
-/** Keep the Loader's references: reparsing them would detach live updates. */
+/**
+ * Keep the Loader's references: reparsing them would detach live updates.
+ *
+ * An optional field a profile never wrote has no live reference to read, so it
+ * stays absent instead of failing the operation that asked for the config.
+ */
 export declare function readConfig(config: LiveConfig): OpencodeGoConfig;
 /**
  * Accept only an http(s) base without a query or fragment. Runs at load for the

package/lib/types/config.js

@@ -11,6 +11,7 @@
  *
  * @module @dan-ai-studio/dshopencodego/config
  */
+import { RetryPolicySchema } from '@deepseek-ai/dsh-llm';
 import { MAX_TIMER_DELAY_MS } from '@deepseek-ai/dsh-timeout';
 import z from '@deepseek-ai/schemastery';
 import { DEFAULT_BASE_URL } from "./catalog/constants.js";
@@ -41,15 +42,24 @@ const fields = {
             maxTokens: z.union([z.const(null), z.number().step(1).min(1).max(Number.MAX_SAFE_INTEGER)]),
         })])).default({}),
     modelProtocols: z.dict(z.string()).default({}),
+    retryPolicy: RetryPolicySchema,
 };
 /** Plain resolved values used by the adapter. */
 export const PlainConfig = z.object(fields);
 /** Runtime schema for {@link OpencodeGoConfig}; every field stays live. */
 export const Config = z.object(Object.fromEntries(Object.entries(fields).map(([key, schema]) => [key, schema.volatile()])));
-/** Keep the Loader's references: reparsing them would detach live updates. */
+/**
+ * Keep the Loader's references: reparsing them would detach live updates.
+ *
+ * An optional field a profile never wrote has no live reference to read, so it
+ * stays absent instead of failing the operation that asked for the config.
+ */
 export function readConfig(config) {
     return Object.fromEntries(Object.keys(fields)
-        .map(key => [key, config[key].get()]));
+        .flatMap(key => {
+        const field = config[key];
+        return field?.get === undefined ? [] : [[key, field.get()]];
+    }));
 }
 /**
  * Accept only an http(s) base without a query or fragment. Runs at load for the

package/lib/types/conversion/context.js

@@ -42,6 +42,16 @@ function assertSupportedImageRoles(messages) {
         }
     }
 }
+/**
+ * Every tool declaration this request carries.
+ *
+ * `GenerateOptions.toolHistory` is deliberately not projected: projecting it
+ * requires the route to declare a `toolUpdate` mode, and no source states one
+ * for this gateway's models (models.dev describes tools only as callable). A
+ * route that claims a mode the provider does not implement would silently
+ * change what the model sees, so every request sends the complete current tool
+ * list — the documented behavior when a route declares no mode.
+ */
 function toolsOf(options) {
     return options.tools?.map(tool => ({
         name: tool.name,

package/lib/types/conversion/stream.d.ts

@@ -23,6 +23,11 @@ export declare function mapUsage(usage: PiUsage): TokenUsage;
  * so the actionable detail is only available as text. The patterns are ordered
  * from most specific to least, and anything unrecognized stays `PI_AI_ERROR`
  * rather than being guessed into a retryable class.
+ *
+ * Exhausted quota is reported as `ACCOUNT_QUOTA` rather than the generic
+ * `QUOTA`: this gateway meters a prepaid balance (the usage window the plugin
+ * shows is the same meter), so the remedy is topping up, not retrying — the
+ * same upgrade `llm-deepseek-account` applies to its own prepaid route.
  */
 export declare function classifyPiAiError(message: string): string;
 /**

package/lib/types/conversion/stream.js

@@ -8,7 +8,7 @@
  * @module @dan-ai-studio/dshopencodego/conversion/stream
  */
 import { brandString } from '@deepseek-ai/dsh-brand';
-import { CONTEXT_WINDOW_EXCEEDED_CODE, EMPTY_RESPONSE_CODE, isContextWindowExceededError, isQuotaExceededError, LlmError, QUOTA_EXCEEDED_CODE, } from '@deepseek-ai/dsh-llm';
+import { CONTEXT_WINDOW_EXCEEDED_CODE, EMPTY_RESPONSE_CODE, isContextWindowExceededError, isQuotaExceededError, LlmError, } from '@deepseek-ai/dsh-llm';
 import { isContextOverflow } from '@earendil-works/pi-ai';
 import { toPiReplayState } from "./replay.js";
 /**
@@ -26,6 +26,13 @@ export function mapUsage(usage) {
         ...usage.cacheWrite > 0 ? { cacheWriteTokens: usage.cacheWrite } : {},
     };
 }
+/**
+ * The account-quota failure class, written out rather than imported:
+ * `dsh-llm` exports it as `ACCOUNT_QUOTA_EXCEEDED_CODE` from `0.1.7-rc.2`, and
+ * the spelling is the stable protocol value the host routes on, so spelling it
+ * here keeps the route correct on every host this plugin supports.
+ */
+const ACCOUNT_QUOTA_CODE = 'ACCOUNT_QUOTA';
 /**
  * Classify a provider error string into a Harness error code.
  *
@@ -33,12 +40,17 @@ export function mapUsage(usage) {
  * so the actionable detail is only available as text. The patterns are ordered
  * from most specific to least, and anything unrecognized stays `PI_AI_ERROR`
  * rather than being guessed into a retryable class.
+ *
+ * Exhausted quota is reported as `ACCOUNT_QUOTA` rather than the generic
+ * `QUOTA`: this gateway meters a prepaid balance (the usage window the plugin
+ * shows is the same meter), so the remedy is topping up, not retrying — the
+ * same upgrade `llm-deepseek-account` applies to its own prepaid route.
  */
 export function classifyPiAiError(message) {
     if (/\b(?:401|403)\b/.test(message))
         return 'AUTH';
     if (isQuotaExceededError(message))
-        return QUOTA_EXCEEDED_CODE;
+        return ACCOUNT_QUOTA_CODE;
     if (/\b429\b|rate.?limit/i.test(message))
         return 'RATE_LIMIT';
     if (/\b413\b|payload too large|request body too large|length limit exceeded/i.test(message))

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dan-ai-studio/dshopencodego",
-  "version": "0.1.8",
+  "version": "0.1.10",
   "description": "OpenCode Go provider for DeepSeek Harness: live gateway catalog, per-conversation session header, usage",
   "type": "module",
   "license": "MIT",
@@ -72,16 +72,18 @@
     "@deepseek-ai/dsh-brand": ">=0.1.7-alpha.1 <0.1.8",
     "@deepseek-ai/dsh-client-locale": ">=0.1.7-alpha.1 <0.1.8",
     "@deepseek-ai/dsh-client-store": ">=0.1.7-alpha.1 <0.1.8",
+    "@deepseek-ai/dsh-client-ui-conversation": ">=0.1.7-alpha.1 <0.1.8",
     "@deepseek-ai/dsh-client-ui-model-selection": ">=0.1.7-alpha.1 <0.1.8",
+    "@deepseek-ai/dsh-client-ui-renderer": ">=0.1.7-alpha.1 <0.1.8",
     "@deepseek-ai/dsh-client-ui-settings": ">=0.1.7-alpha.1 <0.1.8",
     "@deepseek-ai/dsh-client-ui-slots": ">=0.1.7-alpha.1 <0.1.8",
     "@deepseek-ai/dsh-credentials": ">=0.1.7-alpha.1 <0.1.8",
     "@deepseek-ai/dsh-fs": ">=0.1.7-alpha.1 <0.1.8",
     "@deepseek-ai/dsh-launch-environment": ">=0.1.7-alpha.1 <0.1.8",
     "@deepseek-ai/dsh-llm": ">=0.1.7-alpha.1 <0.1.8",
-    "@deepseek-ai/dsh-settings": ">=0.1.7-alpha.1 <0.1.8",
     "@deepseek-ai/dsh-timeout": ">=0.1.7-alpha.1 <0.1.8",
-    "@deepseek-ai/dsh-typert-protocol": ">=0.1.7-alpha.1 <0.1.8"
+    "@deepseek-ai/dsh-typert-protocol": ">=0.1.7-alpha.1 <0.1.8",
+    "@deepseek-ai/dsh-typert-registry": ">=0.1.7-alpha.1 <0.1.8"
   },
   "peerDependenciesMeta": {
     "@deepseek-ai/dsh-api-remotes": {