@luan.sh/pi-collapse-transcript  

@luan.sh/pi-collapse-transcript folds runs of tool calls and thinking blocks in Pi's fullscreen transcript into collapsed activity rows, including while the agent is working. Tool output, diffs, tool failures, and detailed thinking stay behind expansion. Assistant prose and request-error notices stay visible. Previous turns keep their fold state. The row shows the latest provider-supplied thinking summary (or tool action when no thinking is available), elapsed time, and a step count. The entire row is muted with a dotted underline that spans the terminal width, including the empty space after the chevron, with the summary italicized: Worked for 6m 56s · Checking index baseline · 12 steps. Click it to expand the run back into the original tool renderers and thinking Markdown. User messages, assistant prose, errors, and other entries stay visible and separate one run from the next. Compaction draws a full-width, muted horizontal divider with its own summary disclosure, not a padded card. Compacting model context does not erase the visible conversation or reset its activity folds. Resuming or reloading renders the full saved branch.

It is a Pi extension, not a model-facing tool. It changes only how existing transcript components are laid out on screen. Session history, tool output, and model-visible content are never rewritten.

Preview

Watch the demo.

Install

pi install npm:@luan.sh/pi-collapse-transcript

That is the only step. The package ships its rendering library and mouse host with it and registers both through package.json; nothing else needs to be installed. It runs inside Pi (@earendil-works/pi-coding-agent with @earendil-works/pi-tui); 0.87.1 is the tested version.

Use it

There are no commands, tools, actions, or side-panel tabs. The extension does its work on session_start:

• In interactive TUI mode it installs a hidden widget (pi-collapse-transcript.host) that mounts a transcript projection over Pi's chat container.
• In fullscreen mode, consecutive thinking and tool entries become one collapsed ActivitySection as they arrive. Its header shows Working for <duration> · summary, and a N steps count. Individual tool failures do not change the group's status.
• agent_start keeps the newest group marked Working between tool calls and model requests, including retries and automatic continuations. agent_settled changes it to Worked; neither event changes the user's expansion choice.
• A queued user request consumed after a final answer starts a fresh activity boundary, even when Pi continues without agent_settled. Steering stays compact too.
• Tool failures remain behind expansion while working and after completion. Expand the group for failure output, partial-edit diffs, and nested calls. Native request-error notices remain visible outside the fold. The header has no status dot or spinner.
• Clicking the row toggles between collapsed and expanded. Expanded content is the original components, so tool renderers, thinking Markdown, and libtui's scrolling and fold controls behave as they do natively.
• Compaction entries become ── Context compacted · N tokens before ── ▸. Click to read the original summary; global tool expansion does not open it.
• Other entries that are not tools or thinking (content entries) are rendered unchanged and end the current run.
• Streaming updates and compaction keep the current fold state and refresh the header summary as new headings arrive. Loading a session rebuilds collapsed sections.
• On session_shutdown, or when the widget is disposed, the projection is released and Pi's native rendering returns.

Regular scrollback mode keeps native rendering: mountTranscriptProjection projects entries only when tui.mode === "fullscreen", because Pi's clickable transcript controls need the fullscreen surface.

Header summary

The group keeps its latest nonempty thinking summary while tools run and after they finish. Groups without thinking use the latest tool action. Elapsed time spans the group's earliest start to its latest completion, so parallel tools are not added together. Saved message timestamps preserve it across reloads and session resumes; active groups refresh once per second. When timing is unavailable, the row omits the duration and the word for. activitySummary picks the header text:

• Tool entries use structured action text with leading punctuation stripped. Reading a summary never renders the tool header or copies its animation, tree branches, or width-dependent truncation into the collapsed row.
• Thinking entries use the last **bold** or # heading in the final 8,000 characters of the thought; without a heading they use the first line of the last paragraph, and finally the literal Thinking. No summary is invented.
• The result is passed through sanitizeTuiFieldPreview, which removes control sequences and caps the text at 240 characters.

Settings and keybindings

The package registers no settings, no actions, and no keybindings. The only input it handles is a mouse press on the activity row, provided by the bundled @luan.sh/pi-libtui mouse host. There is nothing to add to ~/.pi/agent/keybindings.json for this package. Expanded tools keep their own appearance settings.

Library API

import { ActivityTranscript } from "@luan.sh/pi-collapse-transcript" gives a ComponentStack that takes an entry reader, a Pi Theme, and a requestRender callback. It groups TranscriptEntry values from @luan.sh/pi-libtui/tool and owns the fold state of each section. Pass it to mountTranscriptProjection to use it outside this extension. Call beginTurn() when the agent starts and finishTurn() only once the run settles; repeated beginTurn() calls preserve the boundary during automatic continuations. It has no native binary.

Native boundary

@luan.sh/pi-libtui/tool supplies the versioned transcript bridge. It reads Pi 1.0 private transcript fields, guards each node's shape, and fails open: when the document layout does not match, when a node is not a recognised assistant or tool component, or when a projection is already installed, the transcript is left untouched. Unknown nodes and native error notices render as they are. Its separate installTranscriptHistory lease adapts Pi's private display-only history methods: successful compaction appends a marker instead of rebuilding the chat, and history rebuilds use the full active branch. Pi still owns model context, compaction, status cleanup, queued input, cancellation, and retries. The lease fails open on unsupported hosts and restores native methods on disposal.

Layout

| Responsibility | Owner | | --- | --- | | Pi registration, widget lifecycle, session hooks | src/extension.ts | | Grouping entries into sections and fold state | src/activity-transcript.ts (ActivityTranscript, ActivitySection) | | Compaction divider and independent summary disclosure | src/compaction-section.ts | | Header text for a run | activitySummary in src/activity-transcript.ts | | Elapsed time from session history | src/activity-timing.ts | | Public exports | src/index.ts | | Native transcript bridge | mountTranscriptProjection in @luan.sh/pi-libtui/tool | | Display history independent of model context | installTranscriptHistory in @luan.sh/pi-libtui/tool | | Row rendering, motion, folding, mouse | ToolActivity, ComponentStack, and the mouse host in @luan.sh/pi-libtui | | Tool execution | None; existing tools run unchanged |

Develop

Source: https://github.com/luan/agents, directory harnesses/pi/agent/packages/pi-collapse-transcript. Run bun run typecheck and bun test test in that directory.

test/transcript.test.ts drives a real TuiAltScreen with native AssistantMessageComponent and ToolExecutionComponent instances and checks collapse, click expansion, streaming updates, folded tool failures, unmount behaviour, and summary sanitisation. test/compaction.test.ts runs native interactive Pi sessions with an immediate in-process provider and terminal. It checks manual and boundary-hook compaction, retained display history, chronological notices, resumed history, and unchanged model-context compaction.