rezon-aki/dsh-streamfold

The newest thinking window stays open while the rest of the process folds into one summary; when a turn finishes, the view glides back to your prompt. Folding advances per frame, and folded content stays Ctrl+F searchable.

Why it exists

In a long conversation, tool calls and reasoning drown out the prose. The official Compact mode also stops folding while a session still has unloaded history, so it cannot help until everything is loaded. This plugin enhances the official transcript and folds on its own terms, independent of whether history has finished loading:

• Process no longer drowns the prose: past turns collapse into a single "N rows folded · click to expand" line.
• What you want to watch stays visible: while a turn runs, only the newest thinking window stays open — capped height, scrolling inside itself.
• No scrolling back after a turn: when it finishes, the view glides back to the message that started it.

How it differs from the official "Work details" modes

	Official "Work details"	This plugin's Fold
When it folds	Only once history has loaded; folding stalls while history is still loading	Folds on its own terms, independent of history loading
Thinking while running	Process rows fold too — you see the summary	Only the running turn's newest thinking window stays open (capped, self-scrolling)
After a turn	Stops at the bottom; you scroll back yourself	Glides back to the message that started the turn
Finding folded content	Expand it again	hidden=until-found, so Ctrl+F still hits it
Reading a past turn	Scroll up to find its prompt	A quick navigation stays fixed at the top

What it looks like once installed

• You can watch it think. The newest thinking window of the running turn opens automatically (height configurable); older ones fold back into a one-line summary after a grace period. Superseded windows and past turns no longer fight for the scroll position.
• Folding keeps a rhythm and does not flash. Collapsing runs bottom-up, 3 rows per group with a 70 ms stagger; expanding runs top-down, 3 rows per group. Cross-turn batches (page load, session switch) complete instantly; even a very long single turn advances group by group (only a single turn over 84 rows takes the instant guard rail).
• Answer longer than a screen. Pick Stop and read in the top setting ("When the answer fills the screen"): while the answer streams, scrolling stops automatically once its top reaches the top of the page, leaving you at the answer's start to read it directly. The default is still Settle on the prompt.
• Settle on the prompt. When a turn finishes, the view glides back to the message that started it — or, when "Keep LLM questions unfolded" is on and the turn contains an ask_user_question card, to the turn's last question card, auto-expanded (that is what needs your answer). Skipped if you scrolled away yourself. For 2 seconds after arrival it also corrects late layout shifts from images or code blocks.
• A quick navigation at the top. The prompt of the turn you are reading sits in a bar fixed at the top of the conversation, one line with an ellipsis; it steps aside while the real row is on screen, and clicking it glides back to that message.
• Steady follow. The bottom follow advances per frame; scrolling up stops it, and a "back to bottom" button appears once you are away. Scrolling inside a thinking window or a tool body no longer stops the follow by mistake.
• Search is handled too. Folded content is still findable with Ctrl+F; while searching the plugin releases the scroll and does not pin to the bottom, and returning to the window tail or pressing back-to-bottom resumes the follow.
• Window sparks / forging. Sparks along the bottom edge of the thinking window get brighter and denser with scroll speed; while the answer streams, each new chunk throws a spray of sparks from the writing head. Colour, density, speed and lifetime are configurable, and turning them off means they never run.

Install

Option 1: the DSH plugin panel (recommended) — sidebar → Plugins → Add plugin, paste the line below, then hit “Enable now”:

dsh-streamfold

The panel installs this version (npm latest is this release, with provenance). Option 2 is for tracking the repository's latest code.

Option 2: the command line (three forms):

dsh plugin --profile web add dsh-streamfold                     # npm package (recommended)
dsh plugin --profile web add github:rezon-aki/dsh-streamfold    # track the repository
dsh plugin --profile web add https://github.com/rezon-aki/dsh-streamfold

Restart the profile, then refresh the browser page (client code is injected at page load).

Compatibility (engines.dsh in package.json is the single source of truth; the plugin panel and the market read it to judge compatibility):

Plugin version	Supported DSH
0.7.x (this release)	>=0.2.1-rc.1 <0.3.0; contract-level verified on 0.2.1-rc.4 (rc.2→rc.4 client bundle byte-identical + local host load), 0.2.0-rc.2 / 0.2.1-rc.2 listed as compatible
0.5.0	the 0.1.7 series
0.4.x	0.1.5 and earlier

The official transcript modes are owned by the configForms service; Node >= 20.

Uninstall: remove it in the plugin panel, or

dsh plugin --profile web remove dsh-streamfold

Restart the profile and the official behaviour is back.

Usage

Settings → General → Work details: this row is provided by the plugin (the official row is hidden while ours is available; visibility only, reversible — on 0.2.1 the official "when to collapse work details" row is hidden too in Fold mode and restored in the official modes), with the official four modes plus Fold:

Option	Behaviour
Compact / Standard / Detailed / Verbose	The official modes, with the official labels and behaviour
Fold	This plugin: one window for the running turn, everything else folds into one summary

Choosing Fold parks the official mode on Verbose, so the official side neither folds nor groups anything and folding is done by this plugin alone (no double folding). The labels come from the official dictionary, so they follow the official wording.

Dedicated settings page: Settings → Streamfold (every switch below lives there and applies immediately).

While running, only the newest thinking window opens automatically; windows you opened yourself are never auto-collapsed, and a superseded window folds back into a one-line summary after the "superseded grace" (2 s by default). Scrolling up stops the follow, and a "back to bottom" button appears once you are away (the same button serves the thinking windows and the main view).

Safety

• Client-side presentation only: it reads the conversation DOM, makes no network requests, touches no credentials and never modifies DSH source.
• The host half writes nothing: no file I/O, no routes, no exported Config; settings live in browser localStorage only (key dsh-streamfold, scoped to the page origin).
• Official mode changes are reversible: choosing Fold writes the official ui-chat transcriptView to Verbose (avoiding double folding) and hides the official settings rows (on 0.2.1 that includes the "when to collapse" row); switching back to an official mode writes the chosen mode back and restores that row, and uninstalling leaves the official value on Verbose until it is changed in the official settings.
• Clean uninstall: the settings row, window markers, back-to-bottom button and injected styles are all removed, with no leftover inline max-height or hidden attributes.

Settings (22 items)

Setting	Key	Default	Range	Meaning
When the answer fills the screen	settleToPrompt	Settle on the prompt	Settle on the prompt / Stop and read	Settle on the prompt — keep following to the bottom, then glide back to your prompt after the turn (skipped if you scrolled away yourself); Stop and read — while the answer streams, scrolling stops automatically once its top reaches the top of the page, leaving you at the answer's start to read it directly
Work details	mode	Fold	Fold / official four modes	Choosing Fold parks the official mode on Verbose, avoiding double folding
Fold history turns	foldHistory	on	toggle	Off: do not fold past turns on page load (only while a turn runs)
Keep interleaved prose	keepInterleavedText	on	toggle	Keep prose rows while folding and separate them from the final answer with a rule (no capped window)
Keep LLM questions unfolded	keepUserQuestions	on	toggle	The ask_user_question card and its reply node never fold, and "Settle on the prompt" lands on the last question card instead; off folds them with everything else
Follow speed cap	followMaxSpeed	240	60–2000 px/s	Cap on our own catch-up speed; lower is softer and lags more on fast output. Content growth is not capped
Follow acceleration cap	followMaxAccel	10000	2000–120000 px/s²	Cap on speed change; lower is softer at start/stop, higher is more responsive
Window height	windowHeight	360	80–1200 px	Max height of a thinking / interleaved-prose window
Auto-expand thinking while running	autoExpandReasoning	on	toggle	Only the newest block of the running turn; superseded windows fold after the grace period
Superseded grace	supersedeDelay	2	0–60 s	How long an older window waits before folding into a summary line (decimals allowed)
Auto-expand tools while running	autoExpandTools	off	toggle	Expand the newest tool card using the official card's own disclosure/scroll
Return to tail on new message	returnToTail	on	toggle	While stopped mid-transcript, sending a new message glides back to the bottom; off keeps the view where it is until you scroll to the bottom or press back-to-bottom
Smoothing factor	smoothGrow	0.15	0.05–0.9	Share of the remaining gap consumed per frame
Show "back to bottom"	showJumpButton	on	toggle	Shown when away from the bottom (hidden for follow lag, to avoid flicker)
Pin the prompt	pinPrompt	on	toggle	Keep a quick navigation fixed at the top showing the prompt of the turn you are reading: one line with ellipsis, steps aside while the real row is in view, click to glide back
Animated transitions	animations	on	toggle	Off makes folding / expanding / jumping instant
Window sparks	sparks	on	toggle	Sparks along the bottom edge of the running thinking window (off = no canvas, no frames)
Spark colour	sparkColor	#4fa8ff	colour	Spark colour; the core is brightened towards white heat
Forging sparks	forgeSparks	on	toggle	Sparks thrown from the writing head while the answer streams; nothing new written, no hammer
Forging spark speed	forgeSpeed	1	0.2–4 ×	Drift speed multiplier
Forging spark life	forgeLife	1.3	0.2–5 s	How long one spark lives (±30% random)
Spark density	sparkDensity	1	0.2–3 ×	Spark count multiplier, shared by both spark kinds

Settings live in browser localStorage, scoped to the DSH page origin.

Diagnostics

Browser console (__dshStreamfold is the public control/debug surface; fx is the spark test contract; the product has no logging, probes, or frame sampling):

__dshStreamfold.state()          // current settings + official "Work details" snapshot
__dshStreamfold.set({ smoothGrow: 0.1, supersedeDelay: 3 })
__dshStreamfold.scan()           // rescan the current DOM immediately (settings unchanged)
__dshStreamfold.dispose()        // tear down all injections (idempotent; refresh restores native)

When reporting an issue, attach your DSH version and the __dshStreamfold.state() output.

Development and verification

• Hand-written, no build: lib/client.js (browser half, wrapped in window.__ModuleLoader__) and lib/index.js (host half, 12 lines).
• After editing lib/*.js, reload the plugin and refresh the page.
• Row anchors use official semantic attributes only: [data-chat-flow], [data-chat-flow-kind], [data-chat-turn], [data-disclosure-row][aria-expanded], [data-variant=think], [data-tool], [data-sample=bash], [class*=_thinkBody], [class*=_bodyWrap], [data-context-injection-body] — never CSS module hashes.
• Tests: npm test (test/spark.mjs 28 assertions + test/contract.mjs + test/client-apply.mjs), currently EXIT=0.

License

MIT