dsh-verbose

English | 中文

A DeepSeek Harness plugin that gives the transcript a per-chat reading mode, in the shape dsh itself lacks: the session-header 3-dots menu downloads the session log, switches each chat between Normal and Verbose, and defines what Verbose expands — thinking blocks, tool calls, terminal output.

Everything stays in the browser. Two localStorage keys — a per-chat switch and a global category set — no service, no telemetry.

It does four things:

• A per-chat Normal/Verbose switch — in the 3-dots menu, which replaces the built-in Session-log button in place (same seat, same id, lower priority), so the header still has exactly one such control. Applies to the open chat immediately; remembered per session in localStorage.
• A "what Verbose expands" dialog — global switches for Thinking, Tool calls and Terminal output, a Turn footer group and an After a turn fold choice. Changing a category re-applies the open chat at once.
• An expansion scope — by default Verbose opens step rows only as they enter the viewport and folds them again when they scroll out, so a long transcript never builds a wall of open steps; off gives the original "open everything at once" behaviour.
• A turn footer — ⏱ 12.3s · 47 nodes · 9 tools · 84.2k tokens under every turn, in Normal and Verbose, plus a collapsed raw turn JSON with a Copy button.

Install

Requires a working dsh (0.2.1-rc.3 tested) with a web profile.

dsh plugin --profile web add dsh-verbose
dsh --profile web

Or install from GitHub or a local folder instead:

dsh plugin --profile web add github:adithyanraj03/dsh-verbose
dsh plugin --profile web add /path/to/dsh-verbose

Restart the host afterwards — the bundle row and the governor bridge are host-side; the browser half alone is reload-only. The 3-dots menu in the session header gains the Transcript view and Verbose settings… rows.

Confirm the row:

dsh --profile web --dump-config | Select-String "id: verbose"

One line is correct; none means the bundle row was not written.

Why this exists

dsh's transcript has one global presentation knob (transcriptView: compact / standard / detailed / verbose) and no per-chat reading choice at all. That is the wrong shape for the job: how much of a conversation's process you want to see is a property of that conversation, while what "verbose" means is a property of you.

This plugin splits the two:

Scope	Setting	Where it lives
Per chat	Normal or Verbose	localStorage, keyed by session id
Global	Which categories Verbose expands: Thinking, Tool calls, Terminal output	localStorage, one set for every chat

Two scopes on purpose. A per-chat copy of the categories would only drift; a global Normal/Verbose would flip every conversation at once.

What you see

The menu replaces the built-in Session-log button — same seat, same id, lower priority — so the header still has exactly one 3-dots control:

• Session log — downloads the session's log (falls back to the export URL when the host's controller is not mounted).
• Transcript view → Normal / Verbose — applies to the open chat, immediately.
• Verbose settings… — the dialog: the three category switches, Turn metrics / Raw turn JSON under a Turn footer heading, and the fold choice under After a turn. Changing a category re-applies the current chat at once; the footer switches are global and take effect on the next render.
• Compact settings… — only when context-governor is mounted (see Compatibility).

Normal leaves the transcript as dsh renders it. Verbose works the only way available from outside React: it finds every disclosure row dsh already renders ([data-expandable][aria-expanded]), classifies it, and clicks the ones whose category is enabled — then keeps doing so for rows that arrive while an answer streams, because dsh rebuilds them as it grows. The hooks are semantic attributes, never class hashes:

Category	Row it governs	Selector
Thinking	reasoning blocks	closest('[data-variant="think"]')
Terminal	shell samples	closest('[data-sample="bash"]')
Tool calls	everything else	default
Step groups	the nested "Ran commands" rows inside a turn	[data-process-activity]

Thinking is tested first, so a tool row nested inside a reasoning block counts as thinking — the enclosing kind is the one you switched off.

Expansion scope (new in this port)

A switch, Only what is on screen, decides how much work Verbose does:

• on (default) — a step row opens when it enters the viewport and folds again when it scrolls out. The reading position is what matters, so a long transcript never builds a wall of open steps, and scrolling back re-opens what you return to.
• off — everything enabled opens at once, which is what the original plugin did.

Three rules keep the mode stable, and each one exists because of a real failure:

1. Two bands, not one threshold. A row must be clearly on screen (120 px inside) before it opens, and clearly gone (260 px outside) before it closes. The gap between them is a dead zone where the engine does nothing — with a single boundary, expanding a row pushed the next one past it, that row folded, the layout snapped back, and the row qualified again.
2. Folding waits for the reader to stop (600 ms after the last scroll). Nothing collapses out from under a scroll in progress.
3. A pass that clicked anything suppresses itself for 320 ms. The mutations our own clicks cause would otherwise schedule the next decision immediately, which is what made the oscillation self-sustaining rather than a one-off.

Two touches make the remaining motion read as motion rather than as damage: the pass records the row anchoring the reading position and corrects scrollTop by the height it just added or removed (without that, expanding a row above you shoves the page), and each row it toggles gets a 170 ms fade-in. The first pass after switching chats waits one frame, because dsh restores that session's scroll position in the same commit.

Keyboard focus

dsh's own disclosure rows call focus() inside their click handler, which moves the scrollport to the focused element. Two things follow from that: a synthetic click used to park focus on a row (the next Enter re-toggled it instead of sending), and — during a live generation, where every arriving tool call triggers a pass — clicking a row above the reader yanked the transcript up to it.

The fix attacks the cause and then the symptom: for the duration of each programmatic click the row's focus is shadowed with a no-op, so focus never moves and nothing scrolls; afterwards any focus that still changed is restored with preventScroll.

The turn footer (new in this port)

Two switches that dsh's renderer has no toggle for, so they are this plugin's own node in conversation.chat.turnTail:

• Turn metrics — ⏱ 12.3s · 47 nodes · 9 tools · 84.2k tokens under every turn, in Normal and Verbose. Everything is measured from the Turn object the slot hands over; a field that is not present is omitted rather than printed as a zero.
• Raw turn JSON — a collapsed copy of that same object with a Copy button, capped so a giant turn cannot freeze the page.

They are deliberately not engine categories: the engine classifies disclosure rows, and a "metrics" row does not exist — listing them as categories would make it look for rows that can never be found. They live in the same dsh-verbose.categories document, and a dsh-verbose:prefs event keeps every mounted footer in step with the dialog.

After a turn (new in this port)

The third group decides what happens to a finished answer:

Choice	Effect
At answer end	writes dsh's own collapseTiming: completion
On next message	writes dsh's own collapseTiming: next-input
Keep expanded	flips this chat to Verbose — "never fold" is the per-chat toggle, not a global value dsh has

The first two go through ctx.configForms.get("ui-chat"), the same seam dsh's own Settings row uses, so the shipped renderer reacts exactly as if you had used it. The third deliberately reuses the per-chat switch instead of storing a separate flag: two sources of truth for "is this chat expanded" would eventually disagree, and the disagreement would look like a bug in one of them.

Screenshots

Taken from a live dsh session, not mocked:

The shipped menu, extended. Download session log and Feedback are untouched; our two rows are appended below them.

Per-chat Normal/Verbose. The tick shows this chat's current view.

The dialog. What Verbose expands, the expansion scope, the turn footer, and the fold choice.

A Verbose chat. Step groups and reasoning open as the reader scrolls.

Compatibility

• dsh 0.2 requires create() on every strict codec. The original plugin predates that rule, so its CG_CODEC gains create: () => CG_CODEC.schema. Without it the loader refuses the whole client plugin — not just the remote.
• The slot shadow is intentional. Registration uses the built-in's own entry id (session-log-download) with priority: -1; a list slot projects one entry per id, lowest priority first. That is what removes the duplicate button rather than adding a second one.
• context-governor is optional. The host half republishes its service over a Typert Remote for the Compact panel. In a profile without that plugin the bridge fails inside a try, every method answers available: false, and the menu simply omits the item.
• dsh 0.2 groups its process rows. 0.1.5 marked every transcript row [data-expandable][aria-expanded]; 0.2 collapses a turn's rows into one button[data-turn-process] and only the bash toolview still marks itself expandable (as data-variant="bash"). The engine therefore matches [data-turn-process][aria-expanded], [data-expandable][aria-expanded] — the ported 0.1.5 selector matched nothing here, which looked exactly like "Verbose does nothing". Consequence worth knowing: inside a collapsed group there is one button, so switching Thinking off individually cannot hide reasoning while the group is closed — the group is expanded or it is not.
• Storage keys are namespaced to this plugin (dsh-verbose.verbose, dsh-verbose.categories), so installing the original alongside it cannot share state.

Troubleshooting

Symptom	Cause	Fix
Entry fails to load (dsh-verbose: failed)	an undeclared service in apply	it declares slots + remote; re-check against the error text
Two 3-dots buttons	another plugin also shadows session-log-download	keep one; both use the same id
Verbose does nothing	the chat has no disclosure rows yet	send a message; rows appear as the turn streams
Compact settings missing	context-governor not mounted	expected — it is optional
Choices reset on reload	private mode / blocked storage	the plugin needs localStorage

Development

npm test

• tests/client.test.mjs — evaluates the real bundle against a shim and asserts the two things that decide whether it mounts: the declared services, and the slot shadow (name, built-in id, priority, registrant). It also renders the menu one level and checks the trigger exists and the modal is closed.
• tests/classify.test.mjs — the engine's classification spec: a fake transcript with a reasoning wrapper, a generic tool row, a bash row that carries both attributes on one element, and an already-open thinking block.

Requirements

• dsh 0.2 (0.2.1-rc.3 tested), with a web profile
• Node 22+
• localStorage — in private mode, or with storage blocked, the choices reset on reload

Credits

The implementation is adapted from session-view-menu 1.1.0 (MIT, same author), retargeted at dsh 0.2 with the codec rule and namespacing above.

Licence

MIT — see the LICENSE file.