feat(web): render transcript text as markdown and payloads as JSON

[JeremyFunk]

Follow-up to #624: the Transcript view rendered every captured body as plain pre-wrapped text. This gives each kind of body the reading it deserves, while keeping the captured bytes one click away.

What changed

Markdown for message text. User, assistant, prompt and (expanded) system bodies render through the existing MessageResponse pipeline (Streamdown + Shiki/KaTeX/Mermaid), so replies with code fences, lists and inline code read the way the chat surfaces already do. The collapsed system preview stays the plain first line — a one-line elision of markdown source reads better raw.

JSON detection for payloads. Tool arguments/results, subagent task prompts/returns and failed-call status messages go through a shared useJsonPayload hook: the same tryParseJson object/array test the log body uses, then pretty-print + sugar-high highlighting on a hit. Everything else stays verbatim — including an emitter-truncated prefix, which fails the parse on purpose rather than being dressed up as a whole document. Each payload gets a copy button that copies what is displayed.

A rendered ↔ raw selector on every rendered body. Rendering is a reading of the capture, and a reading can hide things — whitespace, key order, a literal **. Each rendered body carries a small segment pair (md | raw on message text, json | raw on payloads) where the selected segment is the view you are in; switching to raw shows the captured text verbatim and points the copy button at it. The pair only appears where the two views actually differ, and the choice lives in the existing openRows set so it survives the row scrolling out of the virtualizer.

Tool payloads start collapsed. The "Tool payloads" chip now defaults off; every card opens on its one-line size summary.

ClampedText gained html (highlighted markup) and body (rendered node) modes so all of the above still clamps and "Show full"s identically; text remains the measurement source. The lab fixture's closing replies and investigation system prompt are now markdown so /lab/agent-session exercises the new paths.

Verification

• All 250 agent-session tests and tsc pass.
• Verified in the lab harness: markdown replies render and switch to raw source, JSON arguments pretty-print/highlight and switch to the captured one-liner, plain-text payloads show no selector, the failed-call error renders with its copy button, and virtualized rows re-measure cleanly through every switch.

Known trade-off

JSON pretty-printing goes through JSON.parse/stringify — the same accepted pattern as the log body viewer — so an integer above 2^53 inside a payload displays and copies with precision loss. The raw view is the escape hatch; a lexeme-preserving re-indenter is the fix if this bites in practice.

🤖 Generated with Claude Code

[github-actions]

🍁 Maple PR preview

Warning

Preview cleanup could not be confirmed. The Alchemy teardown outcome was skipped.

Final commit b92eb3a · View workflow run