[umi008]

Summary

Five focused bug fixes targeting TUI streaming stability and core permission behavior.

1. fix(tui): use markdown renderer for reasoning content (Reasoning/thinking content renders one token per line in TUI during streaming #36037)

   • ReasoningPart was using <code filetype="markdown"> which rendered tokens one-per-line. Changed to <markdown> matching TextPart's approach.

2. fix(tui): normalize carriage returns in question text (TUI: question/ask prompt renders one word per line when model emits carriage returns #36040)

   • Model-emitted question strings containing \r caused single-word-per-line rendering. Added .replace(/\r/g, "") in parseQuestions.

3. fix(core): map DeepSeek cache miss tokens to cache write (DeepSeek provider: prompt_cache_miss_tokens not mapped to cache.write #35974)

   • DeepSeek API returns prompt_cache_miss_tokens in provider metadata but getUsage() never read it. Added fallback to the metadata chain alongside Anthropic/Vertex/Bedrock/Venice.

4. fix(tui): batch streaming delta updates to prevent scroll freeze (TUI scroll freezes for seconds during fast-streaming responses (message.part.delta applied per-chunk with no batching) #36043)

   • Each SSE delta triggered a full store mutation + re-render. Added 16ms buffer with Solid's batch() so rapid delta bursts coalesce into a single render cycle.

5. fix(core): coalesce equivalent pending permission requests (Duplicate equivalent external_directory permission requests are emitted for concurrent tool calls #36055)

   • Multiple tools requesting the same permission type (e.g., fs_write) spawned duplicate dialogs. Now coalesces by sessionID + action so one approval resolves all pending requests of that type.

Testing

• bun typecheck passes from packages/tui and packages/opencode
• Manual verification: reasoning content now renders as flowing markdown, questions display correctly, DeepSeek usage stats show cache tokens, streaming is smooth, permission dialogs don't duplicate

Checklist

• Each fix is atomic and independently reviewable
• No unrelated changes
• Follows existing patterns (no else, const over let, no try/catch)
• All 5 issues referenced: Fixes Reasoning/thinking content renders one token per line in TUI during streaming #36037, Fixes TUI: question/ask prompt renders one word per line when model emits carriage returns #36040, Fixes DeepSeek provider: prompt_cache_miss_tokens not mapped to cache.write #35974, Fixes TUI scroll freezes for seconds during fast-streaming responses (message.part.delta applied per-chunk with no batching) #36043, Fixes Duplicate equivalent external_directory permission requests are emitted for concurrent tool calls #36055

[github-actions]

Hey! Your PR title fix(tui,core): resolve 5 streaming and permission bugs doesn't follow conventional commit format.

Please update it to start with one of:

• feat: or feat(scope): new feature
• fix: or fix(scope): bug fix
• docs: or docs(scope): documentation changes
• chore: or chore(scope): maintenance tasks
• refactor: or refactor(scope): code refactoring
• test: or test(scope): adding or updating tests

Where scope is the package name (e.g., app, desktop, opencode).

See CONTRIBUTING.md for details.

[github-actions]

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

• PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.

[umi008]

Splitting into 5 individual PRs for better review granularity.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Found one potential duplicate:

PR #35979: fix(llm): map DeepSeek cache miss tokens
#35979

This appears to be addressing the same issue as fix #3 in the current PR (mapping DeepSeek cache miss tokens to cache write). However, since the current PR references issue #35974, these may be related but separate efforts. You should verify whether #35979 supersedes or complements the DeepSeek cache fix in PR #36086.

The other 4 fixes in this PR (markdown renderer, carriage returns, streaming batching, and permission coalescing) do not have matching open PRs.