dcdpr / jp / 24453328630

67%

push

github

feat(md, cli): Fix streaming rendering and LLM output quirks (#553)

Several rendering correctness issues have been fixed together, as they
share the same underlying infrastructure in `jp_md`.

**Nested fenced code blocks**: The `Buffer` parser now tracks a `depth`
counter inside `InFencedCode`. When the LLM emits an inner fence opening
(backticks + language tag), depth increments and the line is emitted as
code content. A bare closing fence only closes the outer block when
depth reaches zero. This prevents inner fences from prematurely
terminating the outer code block.

**List streaming**: Lists were previously buffered in their entirety
until a blank line appeared, causing the stream to stall for the full
duration of list generation. The `Buffer` now flushes at each new
top-level item boundary, so individual list items appear as they arrive
from the LLM.

**Event fixups pipeline**: A new `EventFixup` trait and `FixupChain` sit
between the `Buffer` iterator and the `ChatRenderer`. Two fixups are
registered by default:

- `OrphanedFenceFixup`: detects when an LLM embeds backticks mid-line
(e.g. `text:```lang`) and converts the stray bare closing fence that
follows into a `Block` instead of a code block.
- `FenceEscalationFixup`: rewrites `FencedCodeStart`/`FencedCodeEnd`
events to use at least 5 backticks, so 3-backtick inner fences in the
output render as literal text rather than real fences.

**Streaming code block API**: `Formatter` now exposes `begin_code_block`
→ `CodeBlockState`, `render_code_line`, `render_code_fence`, and
`render_closing_fence`. The internal `SavedHighlightState`,
`CodeHighlighter`, and their methods are now private. `ChatRenderer` and
`ToolRenderer` are updated to use this API.

**Background bleeding fixes**: Inline code background no longer bleeds
to the terminal edge on line-fill (`\x1b[K`). The default background is
now re-asserted before each erase-to-EOL. Syntax- highlighted code
blocks written via `write_raw` now inheri... (continued)

297 of 321 new or added lines in 9 files covered. (92.52%)

22328 of 34991 relevant lines covered (63.81%)

247.05 hits per line