@neplich/dsh-chat-autofold

@neplich/dsh-chat-autofold

中文版 · English

Web GUI plugin: when an assistant message starts outputting its body text, the plugin automatically collapses every thinking (Think) row and tool-call row that precedes the body within the same turn, and inserts one slim expand bar in their place. Clicking the bar toggles the records between collapsed and expanded; the bar stays mounted and serves as the "collapse again" button after expanding.

Row-internal state is preserved exactly: folding only flips row visibility (display) and hides a Think subtree via CSS — it never rebuilds, replaces, or rewrites the tool-call cards, so an already-expanded tool detail or a user-opened Think disclosure comes back in the same state after expanding.

The strategy is pure DOM over the shipped chat flow: no harness change, no slot takeover, no services. It works against a stock deepseek-harness and follows the shipped rendering structure ([data-chat-flow] rows with data-chat-flow-kind, Think disclosures as [data-variant="think"]).

How it works

• Browser half only (src/client/index.ts): a MutationObserver watches the page for chat flows. When an assistant body row (an assistant-step row whose text outside the Think subtree is non-empty) first appears, the plugin computes the fold plan: every tool-call row and every bodiless think-only assistant row above it, bounded by the nearest user row (a turn). Those rows are hidden, and one expand bar is inserted above the first of them.
• Persistent toggle bar: the bar's click flips collapsed ↔ expanded state (label changes to match). A turn the user expanded is recorded so streaming updates never re-collapse it automatically; the bar remains as the explicit "collapse again" control.
• React-safe: rows are hidden through inline display, which React never manages (the shipped row renderers declare no style), and the Think-hiding uses a data-dsh-hide-think attribute React does not own. The bar is pinned above its first target row on every scan, so React's own list moves cannot strand it.
• Clean lifecycle: on unload the observer disconnects, every hidden row is restored, and all bars and injected styles are removed.
• The fold decision is the exported pure function computeFolds (unit-tested); the DOM mapping lives in apply.

Requires a dsh web profile (it is a browser-only surface plugin; the host half is an empty apply so the row mounts in the Loader).

Internationalization

The expand bar copy (collapsed/expanded counts) ships in zh and en through the chat-autofold locale namespace and re-renders live when dsh's UI language switches (a ctx.locale revision subscription re-paints every mounted bar).

Config

None — no tunables; the plugin always collapses thinking + tool-call records above the body of each turn.

Install

dsh plugin --profile  add @neplich/dsh-chat-autofold   # or a local path: ./packages/feature/dsh-chat-autofold
dsh web --profile 

Then send a message that produces thinking and tool calls: once the body text starts, the records above it collapse behind the expand bar.

Develop

Built with the multi-plugin workspace at the repository root: pnpm run build emits lib/ (tsc), pnpm --filter @neplich/dsh-chat-autofold run bundle emits the browser bundle lib/client.js (tsdown), pnpm run test runs the package-level vitest suites.

Model Experience

No tools, no system-prompt change, no model-visible input. The plugin is pure presentation: it reads only the rendered chat DOM and toggles visibility. The session log, prompts, tool schemas, and outputs are untouched, so nothing about the model's view of the conversation changes.

Known Limitations and Deferred Work

• Rendering-structure coupling — the strategy depends on the shipped DOM markers (data-chat-flow, data-chat-flow-kind, data-variant="think") and on rows not carrying a React-managed style prop; a future harness redesign of the chat flow could break it and would need the selectors updated.
• Web GUI only — TUI/ACP entry points render no chat DOM, so the plugin has no effect there.
• Auto-fold is per body row, not per turn — a turn whose body arrives in several assistant rows folds each body row's own preceding records; the bars stay independent (no merging).
• Running tool rows are folded too — once body text starts, preceding tool rows are hidden even if one is still streaming; the bar's count includes them and expanding shows the live row.
• Expand state is per page load — the "user expanded this turn" memory lives in the page; reopening or refreshing the session chat-autofolds the same turn again (this is the default policy, not a bug).