fix(tui): follow new content only from the bottom, and steady the chat

.github/workflows/ci.yml

@@ -35,8 +35,8 @@ jobs:
           python-version-file: .python-version
           allow-prereleases: true
       - run: uv sync
-      - run: uv run ruff check src/ tests/ scripts/chrys_test.py scripts/check_wheel.py scripts/calibrate_gc_freeze.py scripts/gc_freeze_calibration_math.py
-      - run: uv run ruff format --check src/ tests/ scripts/chrys_test.py scripts/check_wheel.py scripts/calibrate_gc_freeze.py scripts/gc_freeze_calibration_math.py
+      - run: uv run ruff check src/ tests/ scripts/chrys_test.py scripts/check_wheel.py scripts/calibrate_gc_freeze.py scripts/gc_freeze_calibration_math.py scripts/chat_panel_workloads.py
+      - run: uv run ruff format --check src/ tests/ scripts/chrys_test.py scripts/check_wheel.py scripts/calibrate_gc_freeze.py scripts/gc_freeze_calibration_math.py scripts/chat_panel_workloads.py
 
   test_linux:
     name: Unix shard (${{ matrix.label }})

.github/workflows/gc-freeze-calibration.yml

@@ -15,7 +15,9 @@ on:
     paths:
       - ".github/workflows/gc-freeze-calibration.yml"
       - "scripts/calibrate_gc_freeze.py"
+      - "scripts/chat_panel_workloads.py"
       - "scripts/gc_freeze_calibration_math.py"
+      - "scripts/profile_chat_panel.py"
       - "src/chrys/app/tui/support/gc_freeze.py"
       - "tests/app/tui/support/test_gc_freeze_calibration.py"
 

AGENTS.md

@@ -12,8 +12,8 @@ uv run icode acp                     # ACP stdio server; also: serve, agent, mod
 uv run python scripts/chrys_test.py --smart --paths <changed files...>   # default verification; both modes run below normal CPU priority
 uv run python scripts/chrys_test.py --full        # complete non-integration suite; only when asked
 uv run pytest tests/x/test_y.py::test_fn -n0      # debugging only
-uv run ruff check src/ tests/ scripts/chrys_test.py scripts/check_wheel.py scripts/calibrate_gc_freeze.py scripts/gc_freeze_calibration_math.py   # pyproject's fix = true EDITS files (--no-fix inspects); CI auto-fixes too, so only unfixable errors fail there
-uv run ruff format --check src/ tests/ scripts/chrys_test.py scripts/check_wheel.py scripts/calibrate_gc_freeze.py scripts/gc_freeze_calibration_math.py
+uv run ruff check src/ tests/ scripts/chrys_test.py scripts/check_wheel.py scripts/calibrate_gc_freeze.py scripts/gc_freeze_calibration_math.py scripts/chat_panel_workloads.py   # pyproject's fix = true EDITS files (--no-fix inspects); CI auto-fixes too, so only unfixable errors fail there
+uv run ruff format --check src/ tests/ scripts/chrys_test.py scripts/check_wheel.py scripts/calibrate_gc_freeze.py scripts/gc_freeze_calibration_math.py scripts/chat_panel_workloads.py
 uv run ty check --error-on-warning --python-platform linux src/chrys   # repeat for darwin and win32; the src/chrys argument is required
 uv run python scripts/i18n.py extract   # then update, translate, compile, check (see i18n)
 ```

docs/en/guides/configuration/agents.md

@@ -161,7 +161,7 @@ These settings work as follows:
 
 During a task, the agent uses each sub-agent's tool name and description to decide whether to delegate work to it. You can also request a specific sub-agent in the input field, for example: `Call <tool name> to complete <task>`. A successful result from the corresponding sub-agent tool confirms that the configuration has taken effect.
 
-A sub-agent returns its final reply to the agent that called it. Text it writes between tool calls appears on its card but is not returned. The card shows each piece of text once the sub-agent has finished writing it. A model request that fails before any of its text has appeared on the card is retried as in the main conversation; once some has appeared, the card keeps it and shows the error. In `icode run`, ACP clients and workflows, a sub-agent's text is not shown as it is written, so text it has already written does not by itself stop such a retry. If the final reply has no text, all the text the sub-agent wrote is returned, in order. If the sub-agent fails, its card shows the reason. An external ACP agent returns what its **Result** setting selects; see [Set the result and timeouts](../extensions/external-acp-agents.md#set-the-result-and-timeouts).
+A sub-agent returns its final reply to the agent that called it. Text it writes between tool calls appears on its card but is not returned. The card shows each piece of text once the sub-agent has finished writing it. A model request that fails before any of its text has appeared on the card is retried as in the main conversation; once some has appeared, the card keeps it and shows the error. In `icode run`, ACP clients and workflows, a sub-agent's text is not shown as it is written, so text it has already written does not by itself stop such a retry. If the final reply has no text, all the text the sub-agent wrote is returned, in order. If the sub-agent fails, its card shows the reason. An external ACP agent returns what its **Result** setting selects; see [Set the result and timeouts](../extensions/external-acp-agents.md#set-the-result-and-timeouts). Click **View details** on the card to see everything the sub-agent did: the view opens at its newest content and keeps showing new tool calls and text as they arrive while you stay at the bottom; scroll up to read earlier content and it stays where you are until you scroll back to the bottom.
 
 **Notes:**
 

docs/en/guides/running/workflows.md

@@ -18,7 +18,7 @@ Alternatively, click “APP MODE: Chat” in the upper left and select “Workfl
 
 After selecting the example, explore these tabs:
 
-- **Workflow**: Shows the execution graph, including nodes, connections, and execution states. When the graph is larger than the panel, drag it with the mouse to move around. After a run starts, click a node to view its input and output. Tabs under each one switch between “Markdown” (the default), “Plain text” (the text unformatted), “Data” (structured data as fields) and “Progress messages” (what the node reported while it ran); only the tabs with content appear. Agent nodes also have a “Transcript” tab for model responses and tool calls; each piece of the agent's text appears there once the model has finished writing it.
+- **Workflow**: Shows the execution graph, including nodes, connections, and execution states. When the graph is larger than the panel, drag it with the mouse to move around. After a run starts, click a node to view its input and output. Tabs under each one switch between “Markdown” (the default), “Plain text” (the text unformatted), “Data” (structured data as fields) and “Progress messages” (what the node reported while it ran); only the tabs with content appear. Agent nodes also have a “Transcript” tab for model responses and tool calls; each piece of the agent's text appears there once the model has finished writing it. The transcript opens at its newest content and keeps showing new content as it arrives while you stay at the bottom; scroll up to read earlier content and it stays where you are until you scroll back to the bottom.
 - **Info**: Shows the workflow name, description, script location, execution environment, and node configuration, including agents and models.
 - **Source**: Shows the Python code that defines the workflow. Both “Workflow” and “Source” are read-only; to change a workflow, edit its Python source file.
 - **Input**: Shows the input submitted for the run after it starts, with Markdown formatting. Click “copy” to copy the input exactly as it was submitted.

docs/zh-Hans/guides/configuration/agents.md

@@ -161,7 +161,7 @@ iCode 默认提供的 Code Agent 和 Q&A Agent 可以作为主智能体，Explor
 
 智能体在执行任务时，会根据子智能体的工具名称和工具描述判断是否向其分派任务。也可以在输入框中指定使用某个子智能体，例如：`调用 <工具名称> 完成 <任务>`。对应的子智能体工具成功返回结果，说明配置已经生效。
 
-子智能体将最终回复返回给调用它的智能体。子智能体在工具调用之间输出的文本显示在其卡片中，但不会返回。卡片中的每段文本在子智能体写完后才显示。模型请求失败时，若其文本尚未在卡片中显示，按与主对话相同的规则重试；若已显示部分文本，卡片保留该文本并显示错误。`icode run`、ACP 客户端和工作流不会实时显示子智能体的文本，因此子智能体已输出的文本本身不会阻止上述重试。最终回复无文本时，将子智能体输出的全部文本按顺序拼接后返回。子智能体失败时，其卡片显示失败原因。外部 ACP 智能体返回的内容由其“结果”设置决定，参阅[设置结果和超时](../extensions/external-acp-agents.md#设置结果和超时)。
+子智能体将最终回复返回给调用它的智能体。子智能体在工具调用之间输出的文本显示在其卡片中，但不会返回。卡片中的每段文本在子智能体写完后才显示。模型请求失败时，若其文本尚未在卡片中显示，按与主对话相同的规则重试；若已显示部分文本，卡片保留该文本并显示错误。`icode run`、ACP 客户端和工作流不会实时显示子智能体的文本，因此子智能体已输出的文本本身不会阻止上述重试。最终回复无文本时，将子智能体输出的全部文本按顺序拼接后返回。子智能体失败时，其卡片显示失败原因。外部 ACP 智能体返回的内容由其“结果”设置决定，参阅[设置结果和超时](../extensions/external-acp-agents.md#设置结果和超时)。单击卡片上的“查看详情”可查看子智能体执行的全部内容：打开时显示最新内容；停留在底部时，新的工具调用和文本到来会随之显示；向上滚动查看之前的内容时，视图停在原处，直到滚回底部。
 
 **注意**：
 

docs/zh-Hans/guides/running/workflows.md

@@ -18,7 +18,7 @@
 
 选择示例后，可以通过以下页签了解工作流：
 
-- **工作流**：展示执行图，包括节点、连接关系和执行状态。执行图超出面板时，可以按住鼠标拖动来移动视图。运行后，点击图中的节点可以查看该节点的输入、输出，下方的子页签可切换“Markdown”（默认）、“纯文本”（未经排版的原文）、“数据”（结构化数据按字段展示）和“进度消息”（节点运行时报告的进度），只显示有内容的页签；智能体节点还提供“对话记录”，用于查看模型回复和工具调用；智能体的每段文本在模型写完后才出现在其中。
+- **工作流**：展示执行图，包括节点、连接关系和执行状态。执行图超出面板时，可以按住鼠标拖动来移动视图。运行后，点击图中的节点可以查看该节点的输入、输出，下方的子页签可切换“Markdown”（默认）、“纯文本”（未经排版的原文）、“数据”（结构化数据按字段展示）和“进度消息”（节点运行时报告的进度），只显示有内容的页签；智能体节点还提供“对话记录”，用于查看模型回复和工具调用；智能体的每段文本在模型写完后才出现在其中。对话记录打开时显示最新内容；停留在底部时，新内容到来会随之显示；向上滚动查看之前的内容时，视图停在原处，直到滚回底部。
 - **信息**：展示工作流名称、描述、脚本位置、执行环境和节点配置，包括智能体及模型。
 - **源代码**：工作流通过 Python 代码定义，此处显示对应的源代码。“工作流”和“源代码”页均为只读；如需修改工作流，需要编辑对应的 Python 源代码文件。
 - **输入**：启动后按 Markdown 格式展示本次运行提交的输入内容。点击“复制”可复制提交时的原始内容。