<!-- Authored per the the-loop:writing skill. -->

Bugfix spec: the trace panel hides the agent's turn inside its own bookkeeping

Phase 1 of 4 (bugfix → design → testing plan → tasks). Source: issue-419.

Summary

The control plane's work-item trace renders every row it holds at the same weight, and the rows a human can act on are the minority. A work item that has merely been polled for a few days shows dozens of consecutive bus.published, poll.comment_settled, reaction.added and dispatch.dropped lines; a long agent turn is followed by a stack of system (system) meta rows carrying no text at all. The agent's prose, the human's reply, the parked gate and the open question — the four things the panel exists to show — are scrolled past to reach.

The panel already has a Tool calls switch, and it is the answer: it is the one control a reader uses to say "show me more". Today it defaults on and governs one row type.

Steps to reproduce

1. Open the control plane on a work item whose session serves no transcript, so the panel falls back to the event-log trail.
2. Scroll the trace: consecutive bus.published (event_type=comment.agent posted=0 channels=[]), poll.comment_settled, reaction.added rows, with no human-readable content between them.
3. Open a work item whose session does serve a transcript, after a long agent turn.
4. Scroll to the end of that turn: several one-line system (system) meta rows, each a label with no content.

Expected vs actual

	Expected	Actual
the default stream	what a human can read or act on	every row the projection emits, at equal weight
a polled-but-quiet work item	says so in one line	dozens of bookkeeping rows
harness bookkeeping entries	hidden unless asked for	one empty system (system) row each
a failed dispatch (level=error)	always visible	visible, but buried among the info rows it looks identical to
the switch	governs the whole stream's verbosity	governs the tool groups only, and defaults on

Root cause

There is no notion of verbosity in the trace — only a notion of tool groups.

TranscriptView's showTools (ui/src/components/Transcript.tsx) is threaded into ThreadRowView and consulted at exactly one place: whether to render ToolGroup. Every other row type — meta, malformed, tool result — renders unconditionally, and EventLine has no switch at all: WorkItemDetail's fallback branch maps the last 40 events straight to rows.

So the panel has no answer to "is this row worth a human's attention?", and its one control is scoped to a single row type and defaults to the noisier setting. Nothing is broken in the projection — transcriptThread is doing exactly what issue-230 asked of it, which is to never drop a line. The defect is that the rendering inherited that "never drop" rule, where it does not belong.

Requirements

Requirement 1 — the default stream is what a human can read

User story: As an operator opening a work item, I want the trace to show the agent's prose, my own replies, gates and questions, so that I can read what happened without scrolling past bookkeeping.

Acceptance criteria (EARS)

1. WHEN the trace panel loads THEN the system SHALL render the stream with the verbosity switch off.
2. WHEN the verbosity switch is off THEN the system SHALL NOT render tool groups (Used n tools), tool-output rows, transcript meta rows carrying no text, or bookkeeping event rows.
3. WHEN the verbosity switch is off AND a row is an assistant turn with prose or thinking, a human reply, a malformed line, or a meta row carrying text THEN the system SHALL render it.
4. WHEN the verbosity switch is off AND an assistant turn carries no prose and no thinking — its only content having been tool calls — THEN the system SHALL render no row for it.

Requirement 2 — nothing actionable is hidden

User story: As an operator, I want failures to reach me whatever the switch says, so that turning the noise down never turns a problem off.

Acceptance criteria (EARS)

1. WHEN an event carries level=error THEN the system SHALL render it regardless of the switch.
2. WHEN an event's type is not a known bookkeeping type THEN the system SHALL render it regardless of the switch.
3. WHEN a gate banner or an agent question is present THEN the system SHALL render it regardless of the switch.

Requirement 3 — the reader can always get everything back

User story: As an operator debugging a delivery, I want one control that restores the full trail, so that the panel stays the place I look when I need the detail.

Acceptance criteria (EARS)

1. WHEN the verbosity switch is turned on THEN the system SHALL render every row it renders today, unchanged.
2. WHEN the switch is rendered THEN its visible copy and its aria-label SHALL describe the whole stream's verbosity and SHALL NOT read Tool calls.
3. WHEN the switch is operated with the mouse, the Enter key or the space key THEN the system SHALL toggle it, as today.

Requirement 4 — hiding never leaves a blank panel

User story: As an operator, I want a panel with nothing left to show to say so, so that a quiet work item does not look like a broken one.

Acceptance criteria (EARS)

1. WHEN hiding leaves the event-log trail with no visible row THEN the system SHALL render one line naming how many rows are hidden and the switch that reveals them.
2. WHEN hiding leaves the transcript with no visible row THEN the system SHALL render the same line.
3. WHEN the trail or the transcript is empty for its own reasons THEN the system SHALL keep today's empty state, which says the source holds nothing rather than that rows were hidden.

Security considerations

No new data reaches the browser and none leaves it. The change is a render-time predicate over rows the panel already holds: the same GET /api/v1/sessions/transcript and GET /api/v1/events responses, the same fields. No request, route, payload or credential changes, and the switch's state is component state — nothing is persisted and nothing is sent.

One risk is real and is what Requirement 2 exists to bound: hiding a row a human needed to see. Two rules answer it — level=error is always rendered, and classification is a deny-list over named bookkeeping types, so an event type nobody has classified stays visible. A future event type is therefore noisy at worst, never silent. The switch restores everything, and the empty state names its own existence, so a reader is never left believing they have seen the whole trail when they have not.

Untrusted text keeps the treatment it already has: every row still renders through React's escaping (Transcript.test.tsx asserts attacker-shaped tool text renders as text), and the classification reads event.event and event.level only — never rendered content.