Streaming transcript & tool cards

The transcript layer folds the session event stream into items and renders them. This page describes what you see; the event-to-item rules are identical for live streams and snapshot replay.

Message items

• User messages — ❯ gutter bubbles (roleUser color); images render inline where the terminal supports it (12-line cap), otherwise as [image] placeholders.
• Assistant messages — ordinary-sized output is rendered as Markdown while streaming; oversized output uses a bounded safe tail both live and after close, while the complete raw text remains in the session record. Messages stay separated by blank lines, with the first line bulleted ● and continuation indented two columns.
• Thinking blocks — reasoning streams as its own block above the body. Live it shows only the ✻ marker over the reasoning's last two lines, with no caption or clock: the activity row says Thinking and owns the elapsed time, token count, and rate. Once the phase ends it settles into one row, ✻ Thought for 6s, previewing the first line (Detailed and Verbose, where reasoning shows). Ctrl-O opens the full body.

Work details

The transcript speaks in the past tense and the bottom activity row in the present: everything live — the spinner, the turn's elapsed time, the current step, and throughput — appears only in the activity row, so no transcript row ticks or says what is running. mayfly.transcriptView keeps the Harness Chat mode names; each step up shows more of the work:

Mode	Running turn	Completed turn	Stopped / failed turn	Activity-row detail
Compact	messages only	header + final answer	past-tense titles	—
Standard (default)	messages, file-change cards, past-tense titles	header + file-change cards + final answer	titles + file-change cards	command, path, query, or reasoning
Detailed	every card, live	header + file-change cards + final answer	titles + file-change cards	—
Verbose	every card, live	stays open	every card	—

• Turn header — a turn with process work gets its header row when it ends: Took 38s, Stopped, or Failed, followed by its tool-call and subagent counts (a subagent spawn counts as a subagent, not as a tool call). A running turn has no header; the activity row carries its elapsed time. A completed turn folds its process and interim replies behind this row; a stopped, failed, or steered turn never folds.
• Process groups — reasoning and tool calls between two replies form one group. As a title, a group summarizes only work that has finished: Read files and searched code, Ran commands · 1 failed. A group whose work is still running, or that holds only reasoning, shows no title until something settles.
• File changes — from Standard up, any call whose presenter declares the diff card (Write <path>, Edit <path>, apply_patch) renders as its diff card, splits the groups around it, and stays visible when the turn folds, so the code changes read next to the final answer. Compact counts them with the rest of the work.
• Ctrl-O — opens the fold, the groups, and the card bodies of the most recent 3 turns (expandTurns). Hints name the key only where it reaches; older turns say how much is hidden without promising the key.

Tool cards

• Terminal card — $ command with an exit N or signal pill (failures show ✗) and the run time; the body is the description over the output's last three rows; Ctrl-O shows the complete bounded output.
• Diff card — the presenter's title (Edit src/auth.ts · +1 −1) over per-file unified diffs with LCS line coloring; the collapsed card caps at 12 rows.
• Presenter cards — any other tool with a presenter shows its own call title and its result preview (three rows for text).
• Grouped calls — consecutive file reads, code searches, and commands within a step merge into one tree card (Read 3 files, Searched 2 patterns, Ran 3 commands); a partial failure shows ◐ and a per-member ✗, a non-zero exit included.
• One-row calls — delegation, plan and goal updates, questions, web retrieval, skills, and agent messages render as one row (✓ Searched the web · vitest vi.fn · 5 sources, ? Also fix signup? → No); Ctrl-O opens the bounded result.
• Fallback card — tools without presenters show Used name (key argument) (MCP tools read server › tool) with a result preview; JSON results pretty-print when expanded.
• A call still unanswered when its turn ended shows ⊘ cancelled.

Long-session window

Only the latest 15 turns stay mounted (windowTurns); older turns leave the render tree while their events stay complete in the session record.

Special cases

• Delegation (subagent, subagent_*) and todo_write calls appear as one-row process members; the agents and todo panes stay their live surfaces (see Bottom panes).
• Synthetic injected messages (workspace instructions, context snapshots) render as nothing — the content still reaches the model, the stream stays clean (see the FAQ).