Skip to content

Node-based terminal manager for AI coding agents — tmux-backed terminals and parallel agent sessions as draggable nodes on an infinite pan/zoom canvas. macOS, Linux, and a browser Server Edition.

improvement(agents): leave the stale decay outside the auto-hide setting

Shogunad06feeffedb
diff --git a/CLAUDE.md b/CLAUDE.md
index 75823f596..ba821852b 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -1846,10 +1846,15 @@ else, and its context links must keep classifying across restarts).
late `finish()` is the no-op it already was and the next turn boundary takes it.
**A third removal path is opt-in:** `settings.autoHideFinishedSubagentCards` (default OFF, and
off reproduces the two paths above exactly) mirrors into the store as `autoHideFinished`, and
- with it on a card is dropped the moment its subagent reports done, by `finish()` and by the
- decay alike, with no turn boundary. Only a DONE card is ever dropped, so #547's rule (a new turn
- keeps the cards of subagents still running) and Eco's `liveSubagents` are untouched. Renderer
- only: Desktop and Server Edition identical, Mobile N/A.
+ with it on a card is dropped the moment its subagent REPORTS done, with no turn boundary. Only a
+ DONE card is ever dropped, so #547's rule (a new turn keeps the cards of subagents still running)
+ and Eco's `liveSubagents` are untouched. **The decay is deliberately NOT one of the paths it
+ gates**: `sweepStaleWorking` fires precisely because the end never ARRIVED, which is the opposite
+ of what the setting promises, and it is the one case where a visible card carries the most
+ information — drop it and a fan-out whose subagents died silently leaves nothing on the canvas
+ saying one was ever launched. It costs Eco nothing either way, since an absent card and a `done`
+ card are the same answer to `liveSubagents`. Renderer only: Desktop and Server Edition identical,
+ Mobile N/A.
(Subagents share the parent's process — no PTY.) Each card shows
duration/tokens/tool-uses and **expands** (click) to a **live transcript**:
`core/subagent-tail.ts` resolves the subagent's own transcript file
diff --git a/src/renderer/canvas/Canvas.tsx b/src/renderer/canvas/Canvas.tsx
index fdaaf98a1..c90c2c569 100644
--- a/src/renderer/canvas/Canvas.tsx
+++ b/src/renderer/canvas/Canvas.tsx
@@ -7751,7 +7751,6 @@ export function Canvas() {
.setAutoHideFinished(settings.autoHideFinishedSubagentCards === true)
}, [settings.autoHideFinishedSubagentCards])
-
/** ids to act on for a node menu: the whole selection if the node is part of it, else just it. */
const targetIds = useCallback((node: Node): string[] => {
const selected = nodesRef.current.filter((n) => n.selected).map((n) => n.id)
diff --git a/src/renderer/state/agentNodes.ts b/src/renderer/state/agentNodes.ts
index 0a1c786af..c441401ba 100644
--- a/src/renderer/state/agentNodes.ts
+++ b/src/renderer/state/agentNodes.ts
@@ -62,10 +62,18 @@ interface AgentNodesState {
* `settings.autoHideFinishedSubagentCards`, mirrored here by Canvas so the store stays free of
* the settings store.
*
- * On, a card is DROPPED at the moment its subagent reports done (by `finish()` and by
- * `sweepStaleWorking`'s decay alike) rather than left `done` for the next turn boundary to
- * take. Only a finished card is ever dropped, so #547's rule (a new turn must keep the cards of
- * subagents that are still running) and Eco's `liveSubagents` are untouched.
+ * On, a card is DROPPED at the moment its subagent REPORTS done — `finish()`, and only
+ * `finish()` — rather than left `done` for the next turn boundary to take. Only a finished card
+ * is ever dropped, so #547's rule (a new turn must keep the cards of subagents that are still
+ * running) and Eco's `liveSubagents` are untouched.
+ *
+ * **`sweepStaleWorking`'s decay is deliberately NOT gated on this.** That sweep fires precisely
+ * because the end never ARRIVED (crashed CLI, killed pane, slept machine), which is the opposite
+ * of the setting's promise, and it is the one case where a visible card carries the most
+ * information: drop it and a fan-out whose subagents died silently leaves nothing on the canvas
+ * that says a subagent was ever launched. The decay keeps marking `done` and the turn boundary
+ * takes the card, exactly as it did before this setting existed. It costs Eco nothing — an
+ * absent card and a `done` card are the same answer to `liveSubagents`.
*
* Off is the default and reproduces the previous behavior exactly.
*/
@@ -282,7 +290,6 @@ export const useAgentNodes = create<AgentNodesState>((set) => ({
(id) => s.byId[id].state === 'working' && now - s.byId[id].startedAt > staleMs
)
if (!stale.length) return s
- if (s.autoHideFinished) return dropCards(s, stale)
const byId = { ...s.byId }
for (const id of stale) {
const prev = byId[id]
diff --git a/src/renderer/state/agentNodes.turn-boundary.test.ts b/src/renderer/state/agentNodes.turn-boundary.test.ts
index a2a398c70..1d3df2365 100644
--- a/src/renderer/state/agentNodes.turn-boundary.test.ts
+++ b/src/renderer/state/agentNodes.turn-boundary.test.ts
@@ -200,7 +200,11 @@ describe('autoHideFinished: dropping a card the moment its subagent finishes', (
expect(useAgentNodes.getState().byId['tu1'].state).toBe('done')
})
- it('drops the stale sweep’s cards while the setting is on', () => {
+ it('does NOT drop the stale sweep’s cards, even while the setting is on', () => {
+ // The sweep fires because the end never ARRIVED, which is the opposite of what this setting
+ // promises. It is also the case where a visible card carries the most information: drop it and
+ // a fan-out whose subagents died silently leaves nothing behind saying one was ever launched.
+ // So the decay keeps marking `done` and the turn boundary takes the card, as it always did.
const s = useAgentNodes.getState()
s.setAutoHideFinished(true)
s.start('tu-stale', { parentNodeId: 'n1' })
@@ -214,8 +218,18 @@ describe('autoHideFinished: dropping a card the moment its subagent finishes', (
}))
useAgentNodes.getState().sweepStaleWorking(now)
const { byId } = useAgentNodes.getState()
- expect(byId['tu-stale']).toBeUndefined()
- expect(byId['tu-young']).toBeDefined()
+ expect(byId['tu-stale'].state).toBe('done')
+ expect(byId['tu-young'].state).toBe('working')
+ })
+
+ it('and the turn boundary then takes that decayed card, as it does with the setting off', () => {
+ const s = useAgentNodes.getState()
+ s.setAutoHideFinished(true)
+ s.start('tu-stale', { parentNodeId: 'n1' })
+ const startedAt = useAgentNodes.getState().byId['tu-stale'].startedAt
+ useAgentNodes.getState().sweepStaleWorking(startedAt + WORKING_STALE_MS + 1)
+ useAgentNodes.getState().clearFinishedForParent('n1')
+ expect(useAgentNodes.getState().byId['tu-stale']).toBeUndefined()
})
it('a late finish on a dropped card is a no-op', () => {
diff --git a/src/shared/types.ts b/src/shared/types.ts
index c381a8fda..0fa1bff0d 100644
--- a/src/shared/types.ts
+++ b/src/shared/types.ts
@@ -1606,10 +1606,10 @@ export interface Settings {
* hook holds briefly for a phone/canvas Approve/Deny before falling through to the normal
* interactive prompt. Off ⇒ the env var is absent ⇒ exact legacy behavior. Claude-only. */
hookReplyApprovals: boolean
- /** Remove a subagent's ephemeral canvas card the moment that subagent finishes, instead of
- * keeping it until the parent agent's next turn clears it. Default off ⇒ exact legacy
- * behavior. A card whose subagent is still running is never removed by this. Settings →
- * Agents. */
+ /** Remove a subagent's ephemeral canvas card the moment that subagent REPORTS finished, instead
+ * of keeping it until the parent agent's next turn clears it. Default off ⇒ exact legacy
+ * behavior. A card whose subagent is still running is never removed by this, and neither is one
+ * the stale decay marked done because its end never arrived. Settings → Agents. */
autoHideFinishedSubagentCards: boolean
/** Hold an idle-sleep power assertion while a LOCAL agent node is working, so long runs
* survive an unattended laptop. Released when the last one stops (or goes stale). Cannot