feat: queue persistence & recovery, queue management, streaming model/compact, collapsible panel

[windli2018]

What

pi keeps steer/follow-up queues in memory only — a server restart silently loses them. This PR mirrors every queue mutation to a per-session sidecar (<session>.jsonl.queue.json) and surfaces orphaned entries after a restart through a manual recovery dialog. Nothing is ever auto-delivered: the user decides per-entry to re-queue, re-queue & continue, discard, or export.

On top of crash recovery, this PR adds a full set of queue management operations (delete/reorder/recall/requeue), lets users switch models and compact while the agent is running, and ships a collapsible bottom panel for small screens.

How — crash recovery

• lib/queue-store.ts (new): atomic sidecar persistence (tmp+rename) + entry types.
• lib/rpc-manager.ts: queueMirror (live queue w/ images) / queueRecovery (orphans) / pendingQueueHints (FIFO image hints — pi reports queued items by text only and template expansion can rewrite text). queue_update events reconcile the mirror. New commands:
  • resolve_recovery {keep, discard, continueRun} — re-queues via inner.steer()/followUp(); continueRun calls agent.continue() directly (AgentSession has no public continue; events/persistence still flow via its subscriptions)
  • export_queue, import_queue
  • get_state.pendingRecovery
• api/sessions/[id]/queue-recovery (new): lightweight sidecar read that never creates an AgentSession; when a wrapper is alive its in-memory recovery list wins.
• QueueRecoveryDialog (new): checklist with re-queue / re-queue & continue / discard / export .md / .json / import / dismiss. Responsive: desktop single-row footer, mobile 2×2 grid + shorter description; matches the project's existing modal conventions (fixed overlay, var(--bg), radius 10).
• Session DELETE removes the sidecar.

How — queue operations

pi has no single-dequeue / reorder / insert API, so all edits are implemented as clear-all + re-enqueue in the target order (image hints preserved FIFO), mirroring pi's own clearQueue() semantics (allowed mid-run since the loop only pulls entries between turns). New commands:

• move_queue {kind, fromIndex, toIndex} — reorder within a queue.
• recall_queue_item {kind, index} — pull one entry back to the input box for editing (returns its text + images).
• requeue_at {kind, index, text, images} — insert an edited message back at its original position. Sending from the input while a recalledRef is active re-queues instead of prompting.
• remove_queue_item {kind, index} — delete a single entry.
• import_queue now returns the queue snapshot (steering/followUp) so the UI updates immediately without a reload.

UI: each queued row has ↑/↓ (reorder), ↩ (recall to edit), 🗑 (delete), plus HTML5 drag on desktop and touch drag on mobile (pointer: coarse auto-disables native draggable). The queue banner collapses to a first-message preview (… +N) which is clickable to expand. Export/import buttons are always reachable (incl. when the queue is empty) so history can be re-imported at any time, with inline green/red feedback rendered outside the banner so it shows even with an empty queue.

How — model switch & compact while running

• Model: pi's setModel() only overwrites agent.state.model, so the in-flight request still finishes with the old model and the next turn picks up the new one — i.e. it already "applies next turn". This PR simply removes the UI disabled={isStreaming} and adds an amber ⏱ "next turn" badge on the model button that auto-clears when the run ends.
• Compact: compact() rewrites context + the session file, so running it mid-loop is unsafe. handleCompact queues a pendingCompact when agentRunning is true (covers streaming and tool-call execution, where isStreaming is briefly false), and an effect drains it once the agent is truly idle (prompt_done alone is not enough — retries / auto-compaction / queued follow-ups can extend a run). The button shows a three-state cycle: idle → amber "Queued" (click to cancel) → red "Compacting".

How — small-screen bottom panel

ChatInput gains a tri-state bottomMode (full → queueHidden → minimal, persisted in localStorage): a Windows-10-style vertical bar button in the toolbar, plus swipe gestures (swipe down to collapse one step, up to expand) with a hint overlay. Queue-row touch events are isolated (stopPropagation) so dragging to reorder never triggers mode switching.

How — recovery safety

When a wrapper is (re)created with a non-empty pi queue, sidecar restore is skipped and the mirror is rebuilt from the live queue — preventing double-enqueue if the architecture ever changes so pi-web and pi can restart independently.

Why not auto-restore?

Auto-replaying queued messages after a restart risks double-delivery (messages already delivered but still mirrored) and stale-queue resurrection. Manual recovery keeps it explicit.

Notes

• Image-only queued messages never leave pi's _steeringMessages (falsy contentText("")) — they may surface as already-delivered in recovery; users discard manually (documented in AGENTS.md).
• next.config.ts: dev server gets its own build dir (.next-dev) so dev and production can run on separate ports simultaneously. tsconfig.json/.gitignore updated for it.