release: v1.2.6 — 全量代码审计修复、滚动与流式渲染优化、MCP 协议升级

CLAUDE.md

@@ -51,7 +51,7 @@ Bumping the version requires editing `releaseVersionName` and `defaultConfig.ver
 - `ANTHROPIC_MESSAGES` → `AnthropicMessagesProtocol`
 - `LOCAL_GGUF` → `LocalGgufProtocol`
 
-Messages flow via `cn.lineai.ai.message.*` (`SystemModelMessage`, `UserModelMessage`, `AssistantModelMessage`, `ToolModelMessage`). `ModelClient` is the high-level entry point used by `MainCoordinator`. Tool-call text inside a stream is parsed by `ToolCallTextParser`; reasoning by `ThinkTagParser`. System prompts are assembled by `cn.lineai.ai.prompt.SystemPromptProvider` from `app/src/main/assets/prompts/*.txt` templates (`system-prompt-template.txt`, tone variants, `context-compaction-template.txt`, `memory-extraction-template.txt`, `skill-extraction-template.txt`, `work-directory-template.txt`, `learning-context-template.txt`, `model-identity-template.txt`) — modify templates there, not by hardcoding in Java. `SystemPromptProvider.build(..., ModelConfig model)` renders the model identity block (modelId / name / provider / protocol) from the currently selected model and injects it as `{{MODEL_IDENTITY}}` so the assistant can answer capability questions against the actual model id.
+Messages flow via `cn.lineai.ai.message.*` (`SystemModelMessage`, `UserModelMessage`, `AssistantModelMessage`, `ToolModelMessage`). `ModelClient` is the high-level entry point used by `MainCoordinator`. Tool-call text inside a stream is parsed by `ToolCallTextParser`; reasoning by `ThinkTagParser`. System prompts are assembled by `cn.lineai.ai.prompt.SystemPromptProvider` from `app/src/main/assets/prompts/*.txt` templates (`system-prompt-template.txt`, tone variants, `context-compaction-template.txt`, `work-directory-template.txt`, `learning-context-template.txt`, `model-identity-template.txt`) — modify templates there, not by hardcoding in Java. `SystemPromptProvider.build(..., ModelConfig model)` renders the model identity block (modelId / name / provider / protocol) from the currently selected model and injects it as `{{MODEL_IDENTITY}}` so the assistant can answer capability questions against the actual model id.
 
 ### Model config & context size
 `ModelConfig` (`:core-model`) carries `contextSize` (tokens) as a first-class field, parsed by `ContextSizeParser` from user input like `128K`, `1m`, `128000`. UI uses a plain text field (`R.string.model_field_context_size`); the legacy `{id}[{size}]` suffix in `modelId` is still tolerated for old DB rows via `ModelContextParser.parse(ModelConfig)` / `apiModelId(ModelConfig)`. Protocols must call `apiModelId(ModelConfig)` rather than reading `modelId` directly, so the `[size]` suffix is stripped when needed. The `model_configs` SQLite table has a `context_size` column; `ModelRepository.ensureModelConfigColumns` auto-`ALTER`s old databases.
@@ -68,7 +68,7 @@ Messages flow via `cn.lineai.ai.message.*` (`SystemModelMessage`, `UserModelMess
 - Compaction calls are dedicated model requests using the codex-style handoff-summary template (`assets/prompts/context-compaction-template.txt` + `context-compaction-summary-prefix.txt`), with retry + backoff (`MAX_COMPACT_RETRIES`) and usage recorded into `TokenUsageTracker` so the post-compact context becomes the new baseline.
 - Transcript is built segment-by-segment (`List<String>`, each ≤ `TRANSCRIPT_SEGMENT_MAX_CHARS = 256KB`) to avoid single-`StringBuilder` OOM. **Do not reintroduce per-message or total-length truncation** — earlier `MAX_MESSAGE_CONTENT_CHARS` / `MAX_TRANSCRIPT_CHARS` limits were removed as harmful. `ContextCompactionController.startSoftContextCompaction` / `finishSoftContextCompaction` drive the soft flow; `ChatInteractionController` checks the soft trigger before each request when the hard trigger has not fired and the switch is enabled. `OutOfMemoryError` is caught in both the service and the controller as a last-resort fallback. `TokenUsageTracker` is reset on conversation switch (`ContextCompactionController.onConversationChanged`).
 - `AndroidManifest.xml` sets `android:largeHeap="true"` to support large transcripts.
-- `MemoryExtractionService` and `LearningContextRepository` feed durable knowledge back into the system prompt.
+- Durable knowledge (memories saved via the `memory_update` tool or the memory management screen, plus conversation/skill indexes) is reinjected into the system prompt by `LearningContextRepository` / `LearningContextService` when Learning Mode is enabled.
 
 ### Image attachments in chat
 `ComposerView` exposes a dedicated image button (right of the attach `+`) that opens the system image picker (`ACTION_OPEN_DOCUMENT`, `image/*`). Selected images are compressed (long edge ≤ 1568px, JPEG q=85, >3.5MB further downgraded) and base64-encoded. Send path goes through `ChatInteractionController.sendMessageWithImage` → `ImageInputPayload.rawInputJson(prompt, mimeType, base64)` → `ChatMessage.responseInputItemJson`. Protocols detect `ImageInputPayload.KIND` in `message.getRawInputJson()` and emit the correct shape per provider (OpenAI `image_url`, Anthropic `image.source.base64`, Codex `input_image`). No protocol changes are needed when adding new image entry points — reuse `ImageInputPayload`.
@@ -144,7 +144,7 @@ The chat list renders Markdown through `:markdown` (`cn.lineai.ui.markdown.*`),
 - 所有用户可见字符串必须使用 R.string.* 资源，禁止在 Java 代码中硬编码中文或英文字符串
 - 英文资源在 values/strings.xml，中文资源在 values-zh/strings.xml，俄文资源在 values-ru/strings.xml
 - 警惕 \uXXXX 转义的中文字符，这些绕过 grep 搜索但不违反规则
-- AI 提示词中的中文关键词（如 MemoryExtractionService 的记忆匹配词、SkillFileManager 的 Skill 模板）不属于用户可见字符串，不需要 i18n
+- AI 提示词中的中文关键词（如 SkillFileManager 的 Skill 模板）不属于用户可见字符串，不需要 i18n
 - Repository 层通过 ResourceProvider.getString() 获取字符串资源，不直接使用 Context
 - Agent 提示词 fallback 通过 Context.getString() 读取资源，无 Context 时回退到硬编码英文
 

README.md

@@ -50,8 +50,8 @@ The application id is `cn.lineai` and the project is a multi-module Gradle proje
 - Streaming chat with **multiple model protocols** in the same UI: OpenAI-compatible HTTP APIs, Anthropic Messages, OpenAI Codex Responses, and a local GGUF runtime.
 - Reasoning blocks (`<think>…</think>`) are extracted and rendered separately from the final answer.
 - Tool-call text inside a stream is parsed and dispatched by `ToolCallTextParser`; everything the model asks to do is shown to you before it runs.
-- System prompts are assembled from `feature-model/src/main/assets/prompts/*.txt` — tone variants (chat / coding), context-compaction, memory-extraction, skill-extraction, work-directory, learning-context, and model-identity templates. You can override the tone, the work directory, the identity block, and the prompt template from settings.
-- Long conversations are summarised in the background by `ContextCompactionService` with **dynamic compaction** (50% soft trigger + 80% hard trigger) using the active model itself; durable knowledge is extracted by `MemoryExtractionService` and reinjected next session by `LearningContextRepository`.
+- System prompts are assembled from `feature-model/src/main/assets/prompts/*.txt` — tone variants (chat / coding), context-compaction, work-directory, learning-context, and model-identity templates. You can override the tone, the work directory, the identity block, and the prompt template from settings.
+- Long conversations are summarised in the background by `ContextCompactionService` with **dynamic compaction** (50% soft trigger + 80% hard trigger) using the active model itself; durable knowledge saved via the `memory_update` tool or the memory screen is reinjected next session by `LearningContextRepository`.
 - **Image attachments** — pick an image from the system picker; it is compressed, base64-encoded, and sent in the protocol-specific format (OpenAI `image_url`, Anthropic `image.source.base64`, Codex `input_image`).
 
 ### Tool execution

README_CN.md

@@ -50,8 +50,8 @@ LineCode 不是一个轻量聊天客户端，它是一个**完整的编程工作
 - 同一个聊天界面下支持**多种模型协议**：OpenAI 兼容 HTTP API、Anthropic Messages、OpenAI Codex Responses、本地 GGUF 推理。
 - 推理块（`<think>…</think>`）会被 `ThinkTagParser` 单独抽出来，和最终回答分块渲染。
 - 流中的工具调用文本由 `ToolCallTextParser` 解析并派发；模型请求做的每件事，在真正执行前都会先展示给你看。
-- 系统提示由 `feature-model/src/main/assets/prompts/*.txt` 中的模板拼装：语气（聊天 / 编程）变体、上下文压缩、记忆抽取、技能抽取、工作目录、学习上下文、模型身份。你可以在设置里覆盖语气、工作目录、身份块、提示模板。
-- 长对话由 `ContextCompactionService` 后台用**当前模型本身**做**动态压缩**（50% 软触发 + 80% 硬触发）；`MemoryExtractionService` 抽取长期知识，`LearningContextRepository` 在下一次会话中喂回上下文。
+- 系统提示由 `feature-model/src/main/assets/prompts/*.txt` 中的模板拼装：语气（聊天 / 编程）变体、上下文压缩、工作目录、学习上下文、模型身份。你可以在设置里覆盖语气、工作目录、身份块、提示模板。
+- 长对话由 `ContextCompactionService` 后台用**当前模型本身**做**动态压缩**（50% 软触发 + 80% 硬触发）；通过 `memory_update` 工具或记忆管理界面保存的长期知识，由 `LearningContextRepository` 在下一次会话中喂回上下文。
 - **图片附件** - 在输入框右侧点图片按钮打开系统选择器；选中的图片经压缩后 base64 编码，按协议格式输出（OpenAI `image_url`、Anthropic `image.source.base64`、Codex `input_image`）。
 
 ### 工具执行

app/build.gradle.kts

@@ -5,7 +5,7 @@ plugins {
     alias(libs.plugins.android.application)
 }
 
-val releaseVersionName = "1.2.6-rc.2"
+val releaseVersionName = "1.2.6"
 val releaseApkName = "LineCode Pro $releaseVersionName.APK"
 val releaseIdsigName = "$releaseApkName.idsig"
 val releaseSigningProperties = Properties()
@@ -103,7 +103,7 @@ android {
         applicationId = "cn.lineai"
         minSdk = 26
         targetSdk = 37
-        versionCode = 30
+        versionCode = 31
         versionName = releaseVersionName
     }