How the widget keeps a meaningful caption next to a session row across the whole approval-cycle dance — and what determines whether you see the task you asked for or the thing happening right now.

Two fields, two questions

Each session row carries two pieces of human-readable text on AgentSession (src-tauri/src/state.rs):

Field	Question it answers	Lifetime
original_prompt	“What is this session working on?”	Captured at task start, sticky until the next task starts
label	“What’s the latest thing that just happened?”	Overwritten by every event

label is the transient now: it jitters with whatever the agent or user just did. original_prompt is the persistent task identity: pinned to what the user asked for at the top of a task, surviving every approval round-trip until a new task explicitly begins.

Both fields exist because Claude Code emits a flurry of events during a single user task — the user’s prompt, every permission ask, every “y”, every internal Stop with a clarifying question — and each one carries its own snippet of text. Showing only the most recent text would mean staring at "y" or "needs approval: Bash" instead of remembering what you actually asked for; showing only the original prompt would hide the in-the-moment information you need (like the question being asked or an error message).

How label is set

label is set by the per-client adapter on every event. The Claude adapter’s full list of (event, status, label) mappings — including how box-drawing chrome gets stripped from messages and how the 60-character truncation works — lives in Classification.

One detail relevant to the state machine: when the adapter emits None for the label, the state layer keeps the prior label value rather than blanking it (src-tauri/src/label_policy.rs:49). This is how Stop and Notification events that carry no label of their own (just a status change) leave the previous text in place.

How original_prompt is set

The state layer (src-tauri/src/state.rs::apply_set via src-tauri/src/label_policy.rs::select) decides what happens to original_prompt on every set event. The decision depends on whether the session already exists, the prior status, and the new status:

Existing session?	Prior status	New status	Action on original_prompt
no	—	working	set to incoming label
no	—	anything	leave None
yes	done / idle / working	working	re-capture to incoming label (a new task is starting); also reset working_accumulated_ms = 0 and state_entered_at = now — unless the incoming label is a continuation prompt, in which case the boundary is suppressed and the row is treated as if it were an approval cycle
yes	awaiting	working	leave pinned (approval cycle: agent asked, user answered)
yes	any other	any	leave pinned

The third row is the task boundary: a transition into working from any status except awaiting counts as a new task.

• done / idle → working is the natural case: the agent has finished (or is freshly seeded) and the user is starting something new.
• working → working covers the cancellation case: the user hit Esc mid-task and submitted a fresh prompt before the agent could emit a Stop. Without this rule the row would still display the cancelled prompt, which is misleading.
• awaiting → working is the only transition into working that’s not a boundary. It’s the canonical approval cycle (agent asks → user answers → agent resumes), so typing y doesn’t clobber the original prompt.

If the new event has label: None on a task boundary, the prior original_prompt survives unchanged.

Restore from disk

When a session is re-created from prompt_history.json — after an app restart, or when /clear ends and immediately re-starts the session under the same cwd-derived id — apply_set seeds original_prompt from the persisted value, so an in-flight task survives a restart. The exception: if the restored dialog ends in a separator (the boundary marker /clear and /compact leave behind), the conversation has been cleared, so the row starts clean — original_prompt and task_started_at are dropped while the dialog history is kept for the History window. A working prompt arriving on the same event still takes precedence and starts a real task.

Cancelled turns revert to the prior status

A turn cancelled with Esc fires no lifecycle hook. The transcript watcher (the [Request interrupted by user] marker) and, on Windows, idle_probe (the terminal idle-prompt read) both call state::revert_cancelled_turn, which settles the row back to AgentSession::status_before_working — the status captured on the last non-working → working transition — rather than blanket Idle. The cancelled prompt produced nothing, so the row should look as if it never landed: a reply aborted mid-question reverts to Awaiting, and the user’s real answer is then an awaiting → working approval-cycle reply (no task boundary), so original_prompt survives. Gated by detect_cancelled_turns.

Continuation prompts

Some replies look like new prompts but are really “keep going with what you were doing” — "go", "continue", "proceed" — or a bare approval like "yes" / "y" / "ok". These bite when the row is genuinely Done / Idle with no pending question to revert to: the agent finishes without a ? (no Notification of type permission_prompt / plan_approval / idle_prompt), or its closing question slips past the question detector, and the user replies "y". From Done / Idle a one-word follow-up would otherwise look like a fresh task and clobber original_prompt plus reset the working timer — the recurring “the row now shows y as the task” bug. Treating approvals as continuations catches that. (The other route — a reply aborted mid-question — is handled by the revert-to-Awaiting above; the two together cover both ways the row can leave Awaiting before the real answer arrives.)

To avoid that, apply_set checks the incoming label against Config::continuation_prompts (defaults: ["go", "continue", "proceed", "yes", "y", "yeah", "yep", "yup", "ok", "okay", "sure", "go ahead", "do it"]). If the trimmed label matches any phrase exactly (case-insensitive), the task boundary is suppressed:

• original_prompt stays pinned to the prior task.
• working_accumulated_ms is preserved (the timer continues from where it left off).
• label is still updated to the incoming text (e.g. "go"), but with status = working displayLabel falls back to original_prompt anyway, so the user keeps seeing the real task on screen.

Match is exact after trim, not substring or starts-with — "go" matches "go" and "Go" and " go ", but not "go ahead" or "google something". If you want phrases like "go ahead" to count, add them to the list verbatim.

This rule only fires on what would otherwise be a task boundary (transitions into working from done / idle / working). On an awaiting → working transition the row is already in an approval cycle, so the rule is a no-op there.

What the widget actually shows

The frontend’s displayLabel (src/lib/types.ts:58-61) chooses between the two fields based on the row’s current status:

Status	Widget shows
awaiting	label — the agent’s question or permission request
error	label — the error message
working / done / idle	original_prompt if set, else label

The principle: when the agent is blocked, surface what’s blocking it (the transient label). When the agent is acting on or finished with a task, surface the task itself (original_prompt).

Walk-through

A typical task with one approval cycle and a clarifying question, then a brand-new task on the same row:

Step	Hook fires	Status	label	original_prompt	Widget shows
1	UserPromptSubmit “fix foo.py”	working	"fix foo.py"	"fix foo.py" (idle → working: captured)	"fix foo.py"
2	Notification permission	awaiting	"needs approval: Bash"	"fix foo.py" (pinned)	"needs approval: Bash"
3	UserPromptSubmit “y”	working	"y"	"fix foo.py" (awaiting → working: pinned)	"fix foo.py"
4	Stop with question	awaiting	"has a question"	"fix foo.py" (pinned)	"has a question"
5	UserPromptSubmit follow-up	working	"the follow-up text"	"fix foo.py" (still pinned)	"fix foo.py"
6	Stop, task done	done	"the follow-up text" (preserved; Stop emits no label, so the prior label from step 5 stays)	"fix foo.py" (pinned)	"fix foo.py"
7	UserPromptSubmit “add tests”	working	"add tests"	"add tests" (done → working: re-captured)	"add tests"

Step 7 is the only point after step 1 where original_prompt gets re-captured: the prior status was done, so the table’s third row fires. Every other transition into working (steps 3 and 5) had awaiting as the prior status, falling under “leave pinned.”

Implementation pointers

• The state machine is enforced by src-tauri/src/state.rs::apply_set, which delegates the (label, original_prompt) decision to src-tauri/src/label_policy.rs::select. Every mutation to session state — from HTTP events, the transcript watcher, or Tauri commands — funnels through apply_set so the rules are applied in exactly one place.
• The transcript watcher (src-tauri/src/log_watcher.rs::apply_watcher_update) is allowed to upgrade status to working, update model / token counts, and upsert the latest Assistant dialog text, but it cannot touch label or original_prompt — those stay hook-authoritative.
• See Data flow for how apply_set fits into the full event pipeline, and Classification for how the per-event (status, label) pair is computed before reaching the state layer.