dominds 1.28.0 → 1.29.0

package/README.md

@@ -375,6 +375,7 @@ Result: fewer bad side effects, higher plan fidelity, and more first‑try succe
 - **[Dialog Persistence](docs/dialog-persistence.md)** — On-disk layout and lifecycle
 - **[Interruption & Resumption](docs/interruption-resumption.md)** — Stop/resume semantics
 - **[Auth](docs/auth.md)** — Authentication and access model
+- **[Codex Provider Auth Policy](docs/codex-provider-auth-policy.md)** — Managed-ChatGPT-OAuth-only boundary and alternatives
 - **[Mottos](docs/mottos.md)** — Short guiding statements
 
 ### Getting Help

package/README.zh.md

@@ -285,6 +285,7 @@ Dominds 面向“长期开发运作（DevOps）”场景设计，基于社会化
 - **[Dialog Persistence](docs/dialog-persistence.zh.md)** — 磁盘持久化存储结构
 - **[Interruption & Resumption](docs/interruption-resumption.zh.md)** — 任务中断与恢复语义
 - **[Auth](docs/auth.zh.md)** — 认证机制与访问模型（WebUI + API）
+- **[Codex Provider 认证策略](docs/codex-provider-auth-policy.zh.md)** — 仅支持 ChatGPT 托管 OAuth 的产品边界与替代路径
 - **[Mottos](docs/mottos.zh.md)** — 警世名言
 
 ## 获取帮助

package/dist/dialog-global-registry.d.ts

@@ -34,6 +34,8 @@ declare class GlobalDialogRegistry {
     private driveTriggerSubChan;
     static getInstance(): GlobalDialogRegistry;
     get(rootId: string): MainDialog | undefined;
+    getRunningDialogsByAgent(agentId: string): Dialog[];
+    private reconcileMainDialogGoalRemindersForAgent;
     register(mainDialog: MainDialog): void;
     unregister(rootId: string): void;
     private enqueueRoot;

package/dist/dialog-global-registry.js

@@ -4,6 +4,7 @@ exports.globalDialogRegistry = void 0;
 const evt_1 = require("@longrun-ai/kernel/evt");
 const dialog_1 = require("./dialog");
 const log_1 = require("./log");
+const main_dialog_goal_reminder_1 = require("./main-dialog-goal-reminder");
 const persistence_1 = require("./persistence");
 const log = (0, log_1.createLogger)('dialog-global-registry');
 class GlobalDialogRegistry {
@@ -33,6 +34,45 @@ class GlobalDialogRegistry {
     get(rootId) {
         return this.entries.get(rootId)?.mainDialog;
     }
+    getRunningDialogsByAgent(agentId) {
+        const dialogs = [];
+        const seenDialogIds = new Set();
+        for (const entry of this.entries.values()) {
+            for (const dialog of entry.mainDialog.getLoadedDialogTreeSnapshot()) {
+                const dialogId = dialog.id.valueOf();
+                if (dialog.status === 'running' &&
+                    dialog.agentId === agentId &&
+                    !seenDialogIds.has(dialogId)) {
+                    seenDialogIds.add(dialogId);
+                    dialogs.push(dialog);
+                }
+            }
+        }
+        return dialogs;
+    }
+    reconcileMainDialogGoalRemindersForAgent(agentId) {
+        const runningDialogs = this.getRunningDialogsByAgent(agentId);
+        if (runningDialogs.length <= 1) {
+            return;
+        }
+        for (const entry of this.entries.values()) {
+            const dialog = entry.mainDialog;
+            if (dialog.status !== 'running' || dialog.agentId !== agentId) {
+                continue;
+            }
+            const changed = (0, main_dialog_goal_reminder_1.refreshFollowTaskdocMainDialogGoalReminderForKnownParallelState)(dialog);
+            if (!changed) {
+                continue;
+            }
+            void dialog.dlgStore.persistReminders(dialog, dialog.reminders).catch((error) => {
+                log.warn('Failed to persist main-dialog goal reminder after parallel dialog detection', error, {
+                    rootId: dialog.id.rootId,
+                    selfId: dialog.id.selfId,
+                    agentId: dialog.agentId,
+                });
+            });
+        }
+    }
     register(mainDialog) {
         // This registry is keyed by the *tree root id*.
         // Only the canonical main dialog (selfId === rootId) should be stored here.
@@ -48,6 +88,7 @@ class GlobalDialogRegistry {
             driveQueued: false,
             activeRunClearedDrivePending: false,
         });
+        this.reconcileMainDialogGoalRemindersForAgent(mainDialog.agentId);
         void (async () => {
             try {
                 const hasPendingNextStepTriggers = await persistence_1.DialogPersistence.hasPendingNextStepTriggers(mainDialog.id);

package/dist/dialog.d.ts

@@ -279,6 +279,7 @@ export declare abstract class Dialog {
     private dispatchSharedReminderOwnerUpdateImpacts;
     /**
      * Emits current visible reminders without reconciling tool-owned state or writing persistence.
+     * Dominds-managed fixed dialog reminders may be materialized in memory so display stays complete.
      * This is for read-only dialog display, especially completed/archived dialogs opened from old tabs.
      */
     emitReminderSnapshot(): Promise<ReminderContent[]>;

package/dist/dialog.js

@@ -9,6 +9,7 @@ const time_1 = require("@longrun-ai/kernel/utils/time");
 const util_1 = require("util");
 const evt_registry_1 = require("./evt-registry");
 const log_1 = require("./log");
+const main_dialog_goal_reminder_1 = require("./main-dialog-goal-reminder");
 const async_fifo_mutex_1 = require("./runtime/async-fifo-mutex");
 const driver_messages_1 = require("./runtime/driver-messages");
 const inter_dialog_format_1 = require("./runtime/inter-dialog-format");
@@ -530,6 +531,7 @@ class Dialog {
         this.touchReminders();
     }
     async listVisibleReminderTargets() {
+        await (0, main_dialog_goal_reminder_1.ensureMainDialogGoalReminder)(this);
         const targets = this.reminders.map((reminder, index) => ({
             source: 'dialog',
             index,
@@ -692,9 +694,11 @@ class Dialog {
     }
     /**
      * Emits current visible reminders without reconciling tool-owned state or writing persistence.
+     * Dominds-managed fixed dialog reminders may be materialized in memory so display stays complete.
      * This is for read-only dialog display, especially completed/archived dialogs opened from old tabs.
      */
     async emitReminderSnapshot() {
+        await (0, main_dialog_goal_reminder_1.ensureMainDialogGoalReminder)(this);
         const taskSharedTarget = {
             kind: 'task',
             agentId: this.agentId,
@@ -713,6 +717,7 @@ class Dialog {
      * Returns reminder contents with metadata for the frontend.
      */
     async processReminderUpdates() {
+        await (0, main_dialog_goal_reminder_1.ensureMainDialogGoalReminder)(this);
         const taskSharedTarget = {
             kind: 'task',
             agentId: this.agentId,

package/dist/docs/codex-provider-auth-policy.md

@@ -0,0 +1,59 @@
+# Codex Provider Authentication Policy
+
+## Decided product boundary
+
+The built-in Dominds `apiType: codex` provider **supports only managed ChatGPT OAuth file authentication**. Its runtime accepts only a final `auth_mode: chatgpt` state backed by refreshable `id_token`, `access_token`, and `refresh_token` values plus a ChatGPT account ID.
+
+This product boundary is intentionally narrower than `codex-rs`; it is not a backlog of modes that should be enabled automatically. `codex-auth` may continue recognizing the complete `codex-rs` authentication contract so it can produce accurate diagnostics. Recognizing an authentication method does not authorize the Dominds Codex provider to send requests with it.
+
+Current behavior is:
+
+| Detected authentication method                          | Dominds Codex provider behavior                                     |
+| ------------------------------------------------------- | ------------------------------------------------------------------- |
+| Complete, refreshable managed `chatgpt` OAuth file auth | Accepted                                                            |
+| `chatgpt` from the ephemeral auth store                 | Rejected                                                            |
+| `chatgpt` missing a required token or account ID        | Rejected; Codex ChatGPT file login must be run again                |
+| `chatgptAuthTokens`, whether refreshable or not         | Rejected                                                            |
+| `apikey`                                                | Rejected                                                            |
+| External `headers`                                      | Rejected                                                            |
+| `agentIdentity`                                         | Rejected                                                            |
+| `personalAccessToken`                                   | Rejected                                                            |
+| `bedrockApiKey`                                         | Rejected                                                            |
+| `CODEX_ACCESS_TOKEN` containing a PAT or Agent Identity | Rejected; this variable overrides persistent auth and must be unset |
+
+Unsupported startup auth is rejected before request-client creation. If credentials change to another mode, come from the ephemeral auth store, or remain `chatgpt` but become incomplete during refresh or 401 recovery, they are rejected again before any retry request. These local policy failures use the stable `DOMINDS_CODEX_PROVIDER_AUTH_POLICY` error code and are non-retriable. The provider must not fall back based on an upstream error, repeat the same authentication path, or silently select another credential.
+
+## When another authentication method is required
+
+If the target service exposes the OpenAI Responses API, configure a custom `apiType: openai` provider with its own authentication environment variable. For example:
+
+```yaml
+providers:
+  my_openai_responses:
+    name: My OpenAI Responses API
+    apiType: openai
+    baseUrl: https://api.openai.com/v1
+    apiKeyEnvVar: MY_OPENAI_API_KEY
+    models:
+      gpt-5.6-sol:
+        name: GPT-5.6 Sol
+        optimal_max_tokens: 600000
+        critical_max_tokens: 922000
+        caution_remediation_cadence_generations: 10
+        context_length: 1050000
+        input_length: 1050000
+        output_length: 128000
+        context_window: '1.05M'
+```
+
+If `apiType: openai` cannot cover the required authentication flow, submit a feature request through [Dominds issues](https://github.com/longrun-ai/dominds/issues). Do not opportunistically add the new authentication method to the Codex provider.
+
+## Default rule for future `codex-rs` alignment
+
+Future synchronization with the `codex-rs` authentication contract must preserve this policy by default:
+
+1. Update `codex-auth` parsing, types, and diagnostics for new contracts and modes.
+2. Continue accepting only complete, refreshable, file-backed managed `chatgpt` OAuth at the Dominds Codex provider request boundary.
+3. Reject every new mode loudly before any HTTP request, retaining guidance for a custom OpenAI Responses API provider or a Dominds feature request.
+4. Update code comments, this document, and regression tests covering every authentication branch.
+5. Expand the Codex provider authentication surface only through an explicit Dominds product decision. Upstream support alone is not sufficient justification.

package/dist/docs/codex-provider-auth-policy.zh.md

@@ -0,0 +1,59 @@
+# Codex Provider 认证策略
+
+## 确定的产品边界
+
+Dominds 内置 `apiType: codex` provider **只支持 ChatGPT 托管 OAuth 文件认证**。运行时最终只放行 `auth_mode: chatgpt`，并要求 `auth.json` 具有可刷新的 `id_token`、`access_token`、`refresh_token` 和 ChatGPT account ID。
+
+这是有意窄于 `codex-rs` 的产品边界，不是尚未完成的自动兼容清单。`codex-auth` 可以继续识别 `codex-rs` 的完整认证契约，以便给出准确诊断；但识别某种认证方式，不代表 Dominds Codex provider 可以使用它发送请求。
+
+当前处理如下：
+
+| 检测到的认证方式                              | Dominds Codex provider 行为                  |
+| --------------------------------------------- | -------------------------------------------- |
+| 完整、可刷新的 `chatgpt` 托管 OAuth 文件认证  | 放行                                         |
+| 临时认证仓中的 `chatgpt`                      | 拒绝                                         |
+| 缺少必要 token 或 account ID 的 `chatgpt`     | 拒绝；要求重新执行 Codex ChatGPT 文件登录    |
+| `chatgptAuthTokens`（无论能否刷新）           | 拒绝                                         |
+| `apikey`                                      | 拒绝                                         |
+| 外部 `headers`                                | 拒绝                                         |
+| `agentIdentity`                               | 拒绝                                         |
+| `personalAccessToken`                         | 拒绝                                         |
+| `bedrockApiKey`                               | 拒绝                                         |
+| `CODEX_ACCESS_TOKEN`（PAT 或 Agent Identity） | 拒绝；该变量会覆盖持久化认证，必须先取消设置 |
+
+启动时检测到不支持的认证，会在创建请求客户端之前拒绝。运行期间若凭证在刷新或 401 恢复中变成其他模式、来自临时认证仓，或仍为 `chatgpt` 但已不完整，会在任何重试请求发出前再次拒绝。此类本地认证策略错误使用稳定错误码 `DOMINDS_CODEX_PROVIDER_AUTH_POLICY`，属于不可重试失败；不允许根据上游错误回退、反复重试同一认证路径，也不允许静默改用其他凭证。
+
+## 需要其他认证方式时
+
+如果目标服务提供 OpenAI Responses API，请配置自定义 `apiType: openai` provider，并使用它自己的认证环境变量。例如：
+
+```yaml
+providers:
+  my_openai_responses:
+    name: My OpenAI Responses API
+    apiType: openai
+    baseUrl: https://api.openai.com/v1
+    apiKeyEnvVar: MY_OPENAI_API_KEY
+    models:
+      gpt-5.6-sol:
+        name: GPT-5.6 Sol
+        optimal_max_tokens: 600000
+        critical_max_tokens: 922000
+        caution_remediation_cadence_generations: 10
+        context_length: 1050000
+        input_length: 1050000
+        output_length: 128000
+        context_window: '1.05M'
+```
+
+如果 `apiType: openai` 不能覆盖所需认证流程，请在 [Dominds issues](https://github.com/longrun-ai/dominds/issues) 提交 feature request；不要把新认证方式顺手接入 Codex provider。
+
+## 与 `codex-rs` 对齐时的默认规则
+
+以后同步 `codex-rs` 认证契约时，默认必须保留本策略：
+
+1. 更新 `codex-auth` 对新契约和新认证方式的解析、类型及诊断。
+2. 在 Dominds Codex provider 的请求边界继续只放行完整、可刷新且来自文件的托管 `chatgpt` OAuth。
+3. 对任何新认证方式在发送 HTTP 请求前响亮拒绝，并保留自定义 OpenAI Responses API provider / feature request 指引。
+4. 同步更新代码注释、本文档和覆盖全部认证分支的回归测试。
+5. 只有经过明确的 Dominds 产品决策，才能扩大 Codex provider 的认证范围；“上游已经支持”本身不是扩大范围的理由。

package/dist/docs/llm-provider-isolation.md

@@ -12,7 +12,7 @@ This means:
 - `apiType: anthropic` owns official Anthropic Messages semantics, including object-shaped `model_params.anthropic.thinking`.
 - `apiType: anthropic-compatible` owns Anthropic-compatible gateway semantics, including boolean `model_params.anthropic-compatible.thinking` mapped to provider `enabled` / `disabled` request objects.
 
-Some providers expose OpenAI-compatible or Anthropic-compatible endpoints while still requiring explicit provider quirks. Volcano Engine Ark Coding Plan now uses the OpenAI-compatible Chat Completions shape with its dedicated `/api/coding/v3` endpoint; the historical Anthropic-compatible Volcano tool-call quirks have been removed. Kimi Code also uses the OpenAI-compatible Chat Completions shape, but must use the built-in `kimi-code` provider's dedicated `/coding/v1` endpoint, versioned `KimiCLI/Dominds/<version>` User-Agent, `prompt_cache_key`, and Kimi-specific `thinking`/`reasoning_effort` request shaping. See [`volcengine-coding-plan-openai-compatible.zh.md`](./volcengine-coding-plan-openai-compatible.zh.md) for the Volcano migration design record.
+Some provider / API paths expose similar endpoints or model capabilities while still requiring explicit provider quirk profiles. The built-in Codex (ChatGPT) provider enables `codex-anti-early-finalization` by default; it appends a work-language-specific reasoning-completion guard only to Codex provider `instructions` to reduce suspected early reasoning truncation / premature finalization on the ChatGPT/Codex path, and it does not apply to `apiType: openai`. Volcano Engine Ark Coding Plan now uses the OpenAI-compatible Chat Completions shape with its dedicated `/api/coding/v3` endpoint; the historical Anthropic-compatible Volcano tool-call quirks have been removed. Kimi Code also uses the OpenAI-compatible Chat Completions shape, but must use the built-in `kimi-code` provider's dedicated `/coding/v1` endpoint, versioned `KimiCLI/Dominds/<version>` User-Agent, `prompt_cache_key`, and Kimi-specific `thinking`/`reasoning_effort` request shaping. See [`volcengine-coding-plan-openai-compatible.zh.md`](./volcengine-coding-plan-openai-compatible.zh.md) for the Volcano migration design record.
 
 Similar field names across wrappers do not imply compatibility. For example, `reasoning_effort`, `verbosity`, `parallel_tool_calls`, and web search controls may look similar but can still differ in accepted values, payload shape, lifecycle events, validation rules, and runtime meaning.
 

package/dist/docs/llm-provider-isolation.zh.md

@@ -12,7 +12,7 @@ Dominds 把每个 LLM provider wrapper 视为独立的协议适配器，而不
 - `apiType: anthropic` 只负责 Anthropic 官方 Messages 语义，包括 object 形态的 `model_params.anthropic.thinking`。
 - `apiType: anthropic-compatible` 负责 Anthropic 兼容网关语义，包括 boolean 形态的 `model_params.anthropic-compatible.thinking`，并映射为 provider 请求里的 `enabled` / `disabled` object。
 
-某些 provider 虽然暴露 OpenAI-compatible 或 Anthropic-compatible endpoint，但仍可能需要明确的 provider quirk profile。火山方舟 Coding Plan 现在走 OpenAI-compatible Chat Completions 形态，并使用专属 `/api/coding/v3` endpoint；历史 Anthropic-compatible 火山工具调用 quirk 已取消。Kimi Code 也走 OpenAI-compatible Chat Completions 形态，但必须使用内置 `kimi-code` provider 的专属 `/coding/v1` endpoint、带 Dominds 版本的 `KimiCLI/Dominds/<version>` User-Agent、`prompt_cache_key` 和 Kimi 专用 `thinking`/`reasoning_effort` 请求整形。火山迁移设计记录见 [`volcengine-coding-plan-openai-compatible.zh.md`](./volcengine-coding-plan-openai-compatible.zh.md)。
+某些 provider / API 路径虽然暴露相似 endpoint 或模型能力，但仍可能需要明确的 provider quirk profile。内置 Codex (ChatGPT) provider 默认启用 `codex-anti-early-finalization`，只在 Codex provider 的 `instructions` 里按当前工作语言追加推理完成检查提示，用于规避 ChatGPT/Codex 路径疑似思考过早截断/提前收尾的问题；这个 quirk 不应用到 `apiType: openai`。火山方舟 Coding Plan 现在走 OpenAI-compatible Chat Completions 形态，并使用专属 `/api/coding/v3` endpoint；历史 Anthropic-compatible 火山工具调用 quirk 已取消。Kimi Code 也走 OpenAI-compatible Chat Completions 形态，但必须使用内置 `kimi-code` provider 的专属 `/coding/v1` endpoint、带 Dominds 版本的 `KimiCLI/Dominds/<version>` User-Agent、`prompt_cache_key` 和 Kimi 专用 `thinking`/`reasoning_effort` 请求整形。火山迁移设计记录见 [`volcengine-coding-plan-openai-compatible.zh.md`](./volcengine-coding-plan-openai-compatible.zh.md)。
 
 不同 wrapper 下看起来同名的字段，不代表它们可以互相兼容。比如 `reasoning_effort`、`verbosity`、`parallel_tool_calls`、web search 相关开关，名字可能相似，但可接受值、请求载荷形状、流事件生命周期、校验规则和运行时含义都可能不同。
 

package/dist/llm/api-quirks.d.ts

@@ -6,6 +6,7 @@ export type LlmFailureKind = 'retriable' | 'rejected' | 'fatal';
 export declare const KIMI_CODE_API_QUIRK = "kimi-code";
 export declare const MINIMAX_REASONING_DETAILS_API_QUIRK = "minimax-reasoning-details";
 export declare const MINIMAX_THINKING_TYPE_API_QUIRK = "minimax-thinking-type";
+export declare const CODEX_ANTI_EARLY_FINALIZATION_API_QUIRK = "codex-anti-early-finalization";
 export declare const SAME_CONTEXT_EMPTY_RESPONSE_API_QUIRK = "same-context-empty-response";
 export declare const VOLCENGINE_INVALID_PARAMETER_AGGRESSIVE_RETRY_API_QUIRK = "volcengine-invalid-parameter-aggressive-retry";
 export type LlmFailureSummary = {

package/dist/llm/api-quirks.js

@@ -1,6 +1,6 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.XCODE_BEST_STREAM_INTERNAL_ERROR_CODE = exports.VOLCENGINE_INVALID_PARAMETER_AGGRESSIVE_RETRY_API_QUIRK = exports.SAME_CONTEXT_EMPTY_RESPONSE_API_QUIRK = exports.MINIMAX_THINKING_TYPE_API_QUIRK = exports.MINIMAX_REASONING_DETAILS_API_QUIRK = exports.KIMI_CODE_API_QUIRK = void 0;
+exports.XCODE_BEST_STREAM_INTERNAL_ERROR_CODE = exports.VOLCENGINE_INVALID_PARAMETER_AGGRESSIVE_RETRY_API_QUIRK = exports.SAME_CONTEXT_EMPTY_RESPONSE_API_QUIRK = exports.CODEX_ANTI_EARLY_FINALIZATION_API_QUIRK = exports.MINIMAX_THINKING_TYPE_API_QUIRK = exports.MINIMAX_REASONING_DETAILS_API_QUIRK = exports.KIMI_CODE_API_QUIRK = void 0;
 exports.normalizeProviderApiQuirks = normalizeProviderApiQuirks;
 exports.createLlmFailureQuirkHandlerSession = createLlmFailureQuirkHandlerSession;
 const persistence_errors_1 = require("../persistence-errors");
@@ -8,6 +8,7 @@ exports.KIMI_CODE_API_QUIRK = 'kimi-code';
 exports.MINIMAX_REASONING_DETAILS_API_QUIRK = 'minimax-reasoning-details';
 // MiniMax's `/v1` API rejects `thinking.type: 'enabled'`; rewrites the request `thinking` form to use `adaptive` instead.
 exports.MINIMAX_THINKING_TYPE_API_QUIRK = 'minimax-thinking-type';
+exports.CODEX_ANTI_EARLY_FINALIZATION_API_QUIRK = 'codex-anti-early-finalization';
 exports.SAME_CONTEXT_EMPTY_RESPONSE_API_QUIRK = 'same-context-empty-response';
 exports.VOLCENGINE_INVALID_PARAMETER_AGGRESSIVE_RETRY_API_QUIRK = 'volcengine-invalid-parameter-aggressive-retry';
 const DOMINDS_LLM_EMPTY_RESPONSE_ERROR_CODE = 'DOMINDS_LLM_EMPTY_RESPONSE';

package/dist/llm/client.d.ts

@@ -56,8 +56,9 @@ export type ProviderModelParamOptions = {
     'openai-compatible'?: Record<string, ModelParamOption>;
     anthropic?: Record<string, ModelParamOption>;
     'anthropic-compatible'?: Record<string, ModelParamOption>;
+    google?: Record<string, ModelParamOption>;
 };
-export type ProviderApiType = 'codex' | 'anthropic' | 'anthropic-compatible' | 'mock' | 'openai' | 'openai-compatible';
+export type ProviderApiType = 'mock' | 'anthropic' | 'anthropic-compatible' | 'openai' | 'openai-compatible' | 'codex' | 'google';
 export type ProviderConfig = {
     name: string;
     apiType: ProviderApiType;

package/dist/llm/defaults.yaml

@@ -6,8 +6,8 @@
 # - llm_retry_conservative_delay_ms: base delay for provider-classified conservative retries (default 30000; fixed for first 10 retries, then ramps by x1.5 with the normal max-delay cap).
 # - llm_retry_backoff_multiplier: exponential factor between retries (default 1.5).
 # - llm_retry_max_delay_ms: upper bound for retry delay (default 1800000 / 30m).
-# - apiQuirks: optional provider/gateway quirk profile(s) for non-standard transport behavior
-#   and provider-specific failure handling that can override the driver's default retry disposition.
+# - apiQuirks: optional provider/gateway quirk profile(s) for non-standard transport behavior,
+#   request shaping, and provider-specific failure handling that can override the driver's default retry disposition.
 #   Example: `apiQuirks: xcode.best` currently covers both keepalive/heartbeat stream quirks and
 #   repeated empty-response deadlock detection for the same unchanged dialog context, plus conservative retry
 #   classification for gateway-returned HTML 502 Bad Gateway pages and
@@ -17,13 +17,109 @@
 #   `apiQuirks: volcengine-invalid-parameter-aggressive-retry` classifies Volcano Ark Coding Plan
 #   transient `400 InvalidParameter` failures as aggressive retry when the shape matches that
 #   provider's replay-successful instability.
+#   `apiQuirks: codex-anti-early-finalization` is Codex-provider-only request shaping that appends
+#   a work-language-specific reasoning-completion guard to ChatGPT/Codex instructions to reduce
+#   early-finalization / suspiciously short reasoning runs. It is intentionally not applied to the
+#   OpenAI API provider.
 # - tool_result_max_chars: optional transport-level cap for a single tool-result text payload
 #   before Dominds projects it into the provider request. Use this when a provider/gateway enforces
 #   a stricter per-item string limit than Dominds' built-in defaults.
 providers:
+  google:
+    name: Google Gemini
+    apiType: google
+    baseUrl: https://generativelanguage.googleapis.com
+    apiKeyEnvVar: GEMINI_API_KEY
+    tech_spec_url: https://ai.google.dev/gemini-api/docs
+    api_mgmt_url: https://aistudio.google.com/
+    model_param_options:
+      google:
+        thinking:
+          prominent: true
+          type: boolean
+          description: Enable or disable model internal reasoning/thinking.
+        thinking_budget:
+          type: integer
+          min: 0
+          description: Budget for thinking tokens. Leave unset to use the model's default adaptive/dynamic thinking.
+        temperature:
+          type: number
+          min: 0
+          max: 2
+          description: Sampling temperature (0–2).
+        top_p:
+          type: number
+          min: 0
+          max: 1
+          description: Nucleus sampling probability (0–1).
+        google_search_tool:
+          prominent: true
+          type: boolean
+          default: true
+          description: Enable Gemini Google Search Grounding to access real-time internet data.
+    models:
+      gemini-3.5-flash:
+        name: Gemini 3.5 Flash
+        optimal_max_tokens: 300000
+        supports_thinking: true
+        default_thinking: true
+        context_length: 1048576
+        input_length: 1048576
+        output_length: 65536
+        context_window: '1M'
+      gemini-3.1-pro:
+        name: Gemini 3.1 Pro
+        optimal_max_tokens: 500000
+        supports_thinking: true
+        default_thinking: true
+        context_length: 2097152
+        input_length: 2097152
+        output_length: 65536
+        context_window: '2M'
+      gemini-2.5-flash:
+        name: Gemini 2.5 Flash
+        optimal_max_tokens: 300000
+        supports_thinking: true
+        default_thinking: true
+        context_length: 1048576
+        input_length: 1048576
+        output_length: 65536
+        context_window: '1M'
+      gemini-2.5-pro:
+        name: Gemini 2.5 Pro
+        optimal_max_tokens: 500000
+        supports_thinking: true
+        default_thinking: true
+        context_length: 2097152
+        input_length: 2097152
+        output_length: 65536
+        context_window: '2M'
+      gemini-1.5-flash:
+        name: Gemini 1.5 Flash
+        optimal_max_tokens: 300000
+        supports_thinking: false
+        default_thinking: false
+        context_length: 1048576
+        input_length: 1048576
+        output_length: 8192
+        context_window: '1M'
+      gemini-1.5-pro:
+        name: Gemini 1.5 Pro
+        optimal_max_tokens: 500000
+        supports_thinking: false
+        default_thinking: false
+        context_length: 2097152
+        input_length: 2097152
+        output_length: 8192
+        context_window: '2M'
   codex:
     name: Codex (ChatGPT)
     apiType: codex
+    # Product auth policy: this provider accepts only managed, refreshable ChatGPT OAuth file
+    # auth. Do not widen it automatically when syncing codex-rs auth modes; direct other auth
+    # methods to a custom `apiType: openai` Responses provider or a Dominds feature request.
+    apiQuirks:
+      - codex-anti-early-finalization
     baseUrl: https://chatgpt.com/backend-api/
     apiKeyEnvVar: CODEX_HOME
     tech_spec_url: https://platform.openai.com/docs/api-reference/responses
@@ -41,10 +137,10 @@ providers:
           prominent: true
           default: high
           # Trade latency/cost for quality on reasoning-capable models.
-          # Allowed: none | minimal | low | medium | high | xhigh
+          # Allowed: none | minimal | low | medium | high | xhigh | max | ultra
           type: enum
-          values: [none, minimal, low, medium, high, xhigh]
-          description: Reasoning effort level (when supported by the model).
+          values: [none, minimal, low, medium, high, xhigh, max, ultra]
+          description: Reasoning effort level. For GPT-5.6 Sol/Terra, ultra is a Codex orchestration preset and is sent to inference as max, matching codex-rs.
         reasoning_summary:
           type: enum
           default: auto
@@ -95,6 +191,30 @@ providers:
           type: boolean
           description: Force JSON object output mode.
     models:
+      gpt-5.6-sol:
+        name: GPT-5.6 Sol
+        optimal_max_tokens: 200000
+        caution_remediation_cadence_generations: 10
+        context_length: 272000
+        input_length: 272000
+        output_length: 32768
+        context_window: '272K'
+      gpt-5.6-terra:
+        name: GPT-5.6 Terra
+        optimal_max_tokens: 200000
+        caution_remediation_cadence_generations: 10
+        context_length: 272000
+        input_length: 272000
+        output_length: 32768
+        context_window: '272K'
+      gpt-5.6-luna:
+        name: GPT-5.6 Luna
+        optimal_max_tokens: 200000
+        caution_remediation_cadence_generations: 10
+        context_length: 272000
+        input_length: 272000
+        output_length: 32768
+        context_window: '272K'
       gpt-5.5:
         name: GPT-5.5
         optimal_max_tokens: 200000
@@ -923,9 +1043,9 @@ providers:
           prominent: true
           default: high
           # Trade latency/cost for quality on reasoning-capable models.
-          # Allowed: none | minimal | low | medium | high | xhigh
+          # Allowed: none | minimal | low | medium | high | xhigh | max
           type: enum
-          values: [none, minimal, low, medium, high, xhigh]
+          values: [none, minimal, low, medium, high, xhigh, max]
           description: Reasoning effort level (when supported by the model).
         reasoning_summary:
           type: enum
@@ -994,6 +1114,42 @@ providers:
           type: boolean
           description: Include `web_search_call.action.sources` in the response payload.
     models:
+      gpt-5.6:
+        name: GPT-5.6 Sol (alias)
+        optimal_max_tokens: 600000
+        critical_max_tokens: 922000
+        caution_remediation_cadence_generations: 10
+        context_length: 1050000
+        input_length: 1050000
+        output_length: 128000
+        context_window: '1.05M'
+      gpt-5.6-sol:
+        name: GPT-5.6 Sol
+        optimal_max_tokens: 600000
+        critical_max_tokens: 922000
+        caution_remediation_cadence_generations: 10
+        context_length: 1050000
+        input_length: 1050000
+        output_length: 128000
+        context_window: '1.05M'
+      gpt-5.6-terra:
+        name: GPT-5.6 Terra
+        optimal_max_tokens: 600000
+        critical_max_tokens: 922000
+        caution_remediation_cadence_generations: 10
+        context_length: 1050000
+        input_length: 1050000
+        output_length: 128000
+        context_window: '1.05M'
+      gpt-5.6-luna:
+        name: GPT-5.6 Luna
+        optimal_max_tokens: 600000
+        critical_max_tokens: 922000
+        caution_remediation_cadence_generations: 10
+        context_length: 1050000
+        input_length: 1050000
+        output_length: 128000
+        context_window: '1.05M'
       gpt-5.5:
         name: GPT-5.5
         optimal_max_tokens: 200000

package/dist/llm/gen/codex.d.ts

@@ -5,7 +5,7 @@
  * Isolation principle: this wrapper owns Codex-native request/stream semantics and must not reuse
  * OpenAI Responses parameter namespaces or event interpretations.
  */
-import type { ChatGptResponsesRequest, ChatGptTool } from '@longrun-ai/codex-auth';
+import type { ChatGptReasoningEffort, ChatGptResponsesRequest, ChatGptTool } from '@longrun-ai/codex-auth';
 import type { Team } from '../../team';
 import type { FuncTool } from '../../tool';
 import type { ChatMessage, ProviderConfig } from '../client';
@@ -13,7 +13,8 @@ import type { LlmBatchResult, LlmFailureDisposition, LlmGenerator, LlmRequestCon
 type CodexToolChoice = 'none' | 'auto' | 'required';
 export declare function resolveCodexToolChoice(tools: readonly ChatGptTool[], requestContext: LlmRequestContext): CodexToolChoice;
 export declare function resolveCodexServiceTier(serviceTier: ChatGptResponsesRequest['service_tier'] | undefined): Exclude<NonNullable<ChatGptResponsesRequest['service_tier']>, 'default'> | undefined;
-export declare function resolveCodexInstructions(systemPrompt: string): string;
+export declare function resolveCodexInstructions(systemPrompt: string, providerConfig?: ProviderConfig): string;
+export declare function resolveCodexReasoningEffortForRequest(model: string, reasoningEffort: NonNullable<Team.ModelParams['codex']>['reasoning_effort']): ChatGptReasoningEffort | undefined;
 export declare class CodexGen implements LlmGenerator {
     get apiType(): string;
     classifyFailure(error: unknown): LlmFailureDisposition | undefined;