gujieye · Nov 3, 2026 · Nov 3, 2026 · Nov 3, 2026 · Nov 3, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -6,6 +6,34 @@ The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and
 
 [中文版](CHANGELOG.zh.md) · [README](README.md) · [Contributing](CONTRIBUTING.md)
 
+## [2.7.3] - 2026-11-03
+
+### Added
+
+- **Custom model import**: Add `bl cmodel base-models/import/get/list/watch/delete` to import Bailian base models as custom models.
+  - `cmodel base-models` lists base models eligible for import (console API; import is currently available in the Beijing region only).
+  - `cmodel import` submits an import job with weight type, display name, and optional bucket/object-key, validating the base model against the console-supported import list before submission.
+  - `cmodel watch --job-id <id>` tracks import job progress with `--follow`, polling interval, and timeout options.
+  - `cmodel get/list/delete` query and remove imported custom models.
+
+### Changed
+
+- **Unified output contract**: Route every command family through one invocation output contract (result / progress / info / warning / debug). Progress and informational messages go to stderr and are hidden in quiet mode; warnings and errors are always preserved. In JSON mode, warnings and errors are emitted as single-line JSON on stderr. Quiet mode no longer trims result fields and gains a `DASHSCOPE_QUIET=1` environment variable.
+- **Advisor thinking disabled**: `bl advisor recommend` LLM calls no longer use thinking mode; intent analysis was the only thinking stage. `--no-thinking` remains accepted as a deprecated no-op for existing scripts.
+
+### Fixed
+
+- **Root help configuration**: Reuse the loaded configuration when rendering root help, preserving parsed flags, Profile selection, and credential precedence.
+- **Duplicate error messages**: Stop repeating the redundant normalization cause in text errors while preserving explicit causes, error codes, JSON details, and verbose stacks.
+- **SDK warnings and empty ASR results**: Preserve SDK warnings and empty ASR task results in managed-agent workflows and `bl speech recognize`.
+- **Finetune fee diagnostics**: `bl finetune fee` now emits gateway diagnostics for fee queries in verbose mode.
+
+### Internal
+
+- Instrument advisor LLM requests for telemetry.
+- Align all command tests and guidance with the output contract; optimize CLI test suites and shard CI test jobs.
+- Clarify quiet output behavior in the CLI and Knowledge Studio READMEs.
+
 ## [2.7.2] - 2026-10-30
 
 ### Added

diff --git a/CHANGELOG.zh.md b/CHANGELOG.zh.md
@@ -6,6 +6,34 @@
 
 [English](CHANGELOG.md) · [README](README.zh.md) · [参与贡献](CONTRIBUTING.zh.md)
 
+## [2.7.3] - 2026-11-03
+
+### 新增
+
+- **自定义模型导入**：新增 `bl cmodel base-models/import/get/list/watch/delete`，支持将百炼基础模型导入为自定义模型。
+  - `cmodel base-models` 列出可导入的基础模型（控制台接口动态返回；导入当前仅北京地域开放）。
+  - `cmodel import` 提交导入任务，支持权重类型、展示名称与可选 bucket/object-key，提交前按控制台可导入清单校验基础模型。
+  - `cmodel watch --job-id <id>` 跟踪导入任务进度，支持 `--follow`、轮询间隔与超时配置。
+  - `cmodel get/list/delete` 查询与删除已导入的自定义模型。
+
+### 变更
+
+- **统一输出契约**：全部命令统一走本次调用的输出契约（result / progress / info / warning / debug）。进度与过程提示写 stderr（quiet 下隐藏），警告与错误始终保留；JSON 模式下警告/错误以 stderr 单行 JSON 输出。quiet 不再裁剪结果字段，并新增 `DASHSCOPE_QUIET=1` 环境变量。
+- **Advisor 关闭思考模式**：`bl advisor recommend` 的 LLM 调用始终关闭思考模式（意图分析此前为唯一开启思考的阶段）；`--no-thinking` 仍被接受但已无效果，为兼容旧脚本保留。
+
+### 修复
+
+- **根帮助配置复用**：渲染根帮助时复用已加载的配置，保留 flag 解析、Profile 选择与凭证优先级。
+- **重复错误信息**：文本错误不再重复输出归一化 cause，同时保留显式 cause、错误码、JSON 详情与 verbose 堆栈。
+- **SDK 警告与空 ASR 结果**：managed-agent 工作流与 `bl speech recognize` 保留 SDK 警告与空的 ASR 任务结果。
+- **微调费用诊断**：`bl finetune fee` 在 verbose 模式下输出费用查询的 gateway 诊断信息。
+
+### 内部
+
+- Advisor LLM 调用接入请求埋点。
+- 全部命令测试与指引对齐输出契约；优化 CLI 测试套件并对 CI 测试任务分片。
+- 澄清 CLI 与 Knowledge Studio README 中 quiet 行为说明。
+
 ## [2.7.2] - 2026-10-30
 
 ### 新增

diff --git a/docs/agents/auth-change.md b/docs/agents/auth-change.md
@@ -29,6 +29,10 @@ defineCommand({ auth }) → runtime/authStage → ctx.client → command.run(ctx
 - `openapi` — 阿里云 OpenAPI 签名域,用 AccessKey ID/Secret 调用 Token Plan 等 OpenAPI
 - `none` — 本地命令、登录/配置类命令、无需 credential 的命令
 
+### 可选 console 副凭证
+
+`auth: "apiKey"` 的命令可用 `optionalConsole: true` 声明 best-effort console 副凭证(例:`cmodel import` 提交前用 console 目录预检基础模型)。runtime 在注入 apiCred 后尽力 `resolveConsole`;解析失败静默跳过,**不得**让 apiKey 主流程失败。此类命令的 help 会额外展示 console 域 flag(`--console-region` / `--console-site` / `--console-switch-agent`),命令捕获 `client.console` 的 AUTH 错误后降级并 warning。纯 console 请求仍必须声明 `auth: "console"`。
+
 ### 多凭证并存
 
 `~/.bailian/config.json` 可同时保存 `api_key`、`access_token` 与 `access_key_*`。登录任一种方式不得删除另一种:
@@ -130,6 +134,7 @@ defineCommand({ auth }) → runtime/authStage → ctx.client → command.run(ctx
   - `globalFlags` 与 `ownFlags` 分流正确
 - [ ] `packages/runtime/src/middleware.ts:authStage`:
   - 根据 `command.auth` 解析 credential 并注入 `ctx.client`
+  - `optionalConsole` 只做 best-effort 副注入;console 解析失败不得影响 apiKey 主流程
   - `settings.dryRun` 下是否允许缺 credential 的策略明确
 - [ ] `packages/runtime/src/error-handler.ts`:
   - AUTH hint 增强使用 `binName`,不要硬编码 `bl`

diff --git a/docs/agents/command-add-remove.md b/docs/agents/command-add-remove.md
@@ -67,6 +67,7 @@ packages/commands/src/index.ts
 - [ ] 新建/删除/移动对应的 `packages/commands/src/commands/<...>.ts`
 - [ ] `defineCommand` 字段使用当前 schema:
   - `auth: "apiKey" | "console" | "openapi" | "none"`
+  - `optionalConsole: true` 仅用于 `auth: "apiKey"` 但需要 best-effort console 预检的命令:runtime 会尽力注入 console 凭证并展示 console 域 flag;console 会话缺失/过期不得阻断主流程,命令必须自行捕获 `client.console` 的 AUTH 错误
   - `flags`(camelCase key,由 runtime 渲染为 kebab-case)
   - `usageArgs`(不含 bin/path 前缀)
   - `exampleArgs`(不含 bin/path 前缀)