流式文本显示太慢了，上一个版本有此问题但问题不明显

[AnonymousUser443]

Description

模型流式回答的时候，终端的文字吐字速度明显跟不上模型的响应速度，尤其是deep seek V-flash，本来是非常快的输出的，但是终端中的打字速度明显慢得多，而且经常模型输出完毕了突然给把后面一大块没来得及输出的内容“咻”地一下子全弹出来了。

Steps to reproduce

正常的模型输出就是

Expected behavior

Actual behavior

Impact

Environment

• OS:Windows11家庭中文版 Pwsh7终端
• codewhale version:v0.8.67
• Install method:安装器安装
• codewhale doctor summary:
• Model/provider:deepseek v4 flash
• Terminal app:pwsh7
• Shell:pwsh7

Logs, screenshots, or recordings

【codewhale doctor 】
Version Information:
codewhale-tui: 0.8.67 (737ac98)
rust: rustc 1.96.0 (ac68faa20 2026-05-25)

Updates:
· current: v0.8.67
✓ latest: v0.8.67
Already up to date.

Configuration:
✓ config.toml found at ~.codewhale\config.toml
workspace: ~
search_provider: duckduckgo (source: default; set [search] provider = "bing" | "tavily" | "bocha" to switch)

State Root:
active: ~.codewhale
dual roots: ~.codewhale (primary) + ~.deepseek (legacy)
! legacy state needs review:
! ~.deepseek\tasks exists but ~.codewhale\tasks is missing
! ~.deepseek\skills exists alongside primary ~.codewhale\skills; legacy data may still need review
! ~.deepseek\config.toml exists alongside primary ~.codewhale\config.toml; legacy data may still need review
Start CodeWhale once to trigger safe migration where available, then rerun codewhale doctor.

Setup State:
· source: persisted
✓ first-run: ready
✓ update checkpoint 0.8.67: ready
! operate/fleet: needs action
· constitution autonomy: unspecified (guidance only)
· runtime posture: default_mode=yolo (settings), approval_policy=never (config), allow_shell=true (interactive default), sandbox=danger-full-access (config), network.default=prompt (default), trust=workspace trusted
· next actions: /constitution (standing law), /setup report (readiness), /setup provider or /provider setup (provider credentials), /model (route), /config (runtime posture), /setup fleet (Operate/Fleet readiness), /fleet setup (explicit profile authoring), /setup hotbar (optional shortcuts), /setup tools (Tools/MCP readiness), /setup remote (remote runtime on-ramp), /setup persistence (path review)
· language: verified (required, inherited, no result)
· provider_model: verified (required, inherited, no result)
· trust_sandbox: verified (required, inherited, no result)
· constitution: verified (required, 0.8.67, checkpoint deferred; bundled applies)
· operate_fleet: not_started (optional, unversioned, no result)
· hotbar: not_started (optional, unversioned, no result)
· tools_mcp: not_started (optional, unversioned, no result)
· remote_runtime: not_started (optional, unversioned, no result)
· persistence: not_started (optional, unversioned, no result)
· verification: not_started (optional, unversioned, no result)

API Keys:
✓ deepseek: env=no, config=yes
· deepseek-anthropic: env=no, config=no
· nvidia-nim: env=no, config=no
· openai: env=no, config=no
· atlascloud: env=no, config=no
· wanjie-ark: env=no, config=no
· volcengine: env=no, config=no
· openrouter: env=no, config=no
· xiaomi-mimo: env=no, config=no
· novita: env=no, config=no
· fireworks: env=no, config=no
· siliconflow: env=no, config=no
· arcee: env=no, config=no
· siliconflow-CN: env=no, config=no
· moonshot: env=no, config=no
· sglang: env=no, config=no
· vllm: env=no, config=no
· ollama: env=no, config=no
· huggingface: env=no, config=no
· together: env=no, config=no
· qianfan: env=no, config=no
· openai-codex: env=no, config=no
· anthropic: env=no, config=no
· openmodel: env=no, config=no
· zai: env=no, config=no
· stepfun: env=no, config=no
· minimax: env=no, config=no
· deepinfra: env=no, config=no
· sakana: env=no, config=no
· longcat: env=no, config=no
· custom: env=no, config=no
· credential precedence: ~/.codewhale/config.toml, OS keyring, then env
✓ active provider key resolved from config.toml

API Connectivity:
· provider: deepseek
· base_url: https://api.deepseek.com/beta
· model: deepseek-v4-pro
· strict_tool_mode: disabled
✓ API connection successful

MCP Servers:
✓ MCP feature flag enabled
· MCP config not found at ~.codewhale\mcp.json
· Project MCP config not found at \?\C:\Users\12929.codewhale\mcp.json
· 0 merged server(s) configured
Run codewhale mcp init or add .codewhale/mcp.json.

Skills:
· local skills dir not found at ~\skills
✓ .agents skills dir found at ~.agents\skills (3 items)
✓ global .agents skills dir found at ~.agents\skills (3 items)
✓ global skills dir found at ~.codewhale\skills (14 items)
· selected skills dir: ~.agents\skills

Tools:
· tools dir not found at ~.codewhale\tools
Run codewhale setup --tools to scaffold a starter dir.

Plugins:
· plugins dir not found at ~.codewhale\plugins
Run codewhale setup --plugins to scaffold a starter dir.

Storage:
✓ tool-output spillover at ~.deepseek\tool_outputs (2036 files)
· composer stash empty (Ctrl+S in the composer to park a draft)

Tool Dependencies:
✓ Python: python → code_execution tool registered
✓ Node.js: present → js_execution tool registered
· pandoc: not found (optional)
pandoc_convert tool is NOT advertised to the model. Install pandoc to enable:
winget install JohnMacFarlane.Pandoc
· tesseract: not found (optional)
image_ocr tool is NOT advertised to the model. Install tesseract to enable:
winget install UB-Mannheim.TesseractOCR
✓ pdftotext: available (optional — pure-Rust extractor is the default in v0.8.32)
Set prefer_external_pdftotext = true in settings.toml for column-heavy PDFs.

Terminal Quirks:
· no env-driven terminal-quirk overrides active

Platform:
OS: windows
Arch: x86_64
! sandbox not available (commands run best-effort)

All checks complete!

[github-actions]

Thanks @AnonymousUser443 for the report.

This issue is staying open for maintainer triage. CodeWhale gets better because people bring us real edge cases from real machines, providers, regions, and workflows.

If you can add a reproduction, logs, version output, screenshots, or the provider/model involved, that makes it much easier for us to verify and harvest the fix. Maintainers may comment /lgtmi to mark recurring issue reporters as approved so this intake note is skipped next time.

[AnonymousUser443]

[Hmbown]

已在 v0.8.68 候选版修复。fd68e7f2 取消了低动态模式下人为逐字输出，3cdfa522 又把队列单位从单个字形改为模型原始流式增量；普通和低动态模式现在都会按 33 ms 显示时钟批量提交已到达内容，不会在结束时突然补出大段文本。当前流式回归测试 38/38 通过，包括 DeepSeek 这类高速突发输出、低动态模式、最终刷新和内容无丢失。感谢 @AnonymousUser443 提供 Windows/DeepSeek 的完整诊断信息。v0.8.67 仍包含旧行为，请用 v0.8.68 测试。