Traycer parity epic: gap analysis of capabilities Overdeck lacks #2566

PRD: drafts/pan-2566.md on overdeck-state

Summary

Epic: gap analysis of Traycer (traycerai/traycer, Apache-2.0) against Overdeck, derived from a full sweep of their open-source clients + protocol and docs (2026-07-10). This epic tracks only genuine gaps — capabilities Traycer has that Overdeck lacks. Each item pursued gets its own child issue; this epic is the triage list, not an implementation plan.

Excluded because Overdeck already has an equivalent (verified against main): multi-harness support, mid-conversation model/harness switching (switch-strategy.ts tiers + summary-fork.ts seeding ≈ Traycer's session re-seeding), conversation forking, voice dictation (VoiceWidget + useVoiceTranscription), image paste/drop attachments (ComposerFooter + composerStore), context-window meter, command palette, cost tracking, worktree-per-workspace isolation, planning/review/verification pipeline (deeper than Traycer's), desktop shell with tray/auto-update/deep-links/autostart (apps/desktop/src/), TTS, MCP wiring for agents.

Major gaps (each is a candidate child issue)

1. Multi-agent task surface + agent-to-agent messaging

Tracked separately — see PAN-2565. Listed here for completeness; it is the largest gap.

2. Checkpoints / restore / revert

Traycer records per-turn checkpoint manifests and offers restore-checkpoint and revert-on-edit flows from the chat timeline (checkpoint-manifests.ts, restore-checkpoint-dialog.tsx). Overdeck has git commits (per-bead in the pipeline) but no conversation-level "rewind the files this turn touched" affordance for interactive conversations.

3. Live collaborative artifacts with comments

Traycer's Epic mode stores specs/tickets/stories/reviews as live collaborative documents (Yjs, per-artifact rooms) with comment threads (create/reply/resolve), @mentions, statuses, hierarchy, and human assignment. Overdeck artifacts (PRDs, vBRIEFs) are git-committed files — durable and reviewable, but with no comments, no co-editing, no in-dashboard artifact editing. Related: PAN-277. The PAN-659-lineage "collaborative markdown editing" followup to PAN-658 is the natural home for the co-editing transport.

4. Real-time co-editing + persistent roles (owner/editor/viewer) + orgs/teams — GATED, filed as PAN-2608

Traycer collaboration is cloud-CRDT: persistent per-task roles, org/team-level grants, ticket assignment to humans, presence including cross-device "agent working" indicators. PAN-658 deliberately targets the live session model (lobby, ephemeral membership); the persistent-roles / team-workspace model is a different product surface and needs its own decision: do we want durable multi-user projects, or only live shared sessions?

5. Cross-device sync of conversations/tasks — PURSUED, filed as PAN-2609 (PRD: drafts/pan-2609.md)

Traycer syncs durable task state (transcripts, artifacts, agent metadata — not PTY buffers) across devices via cloud CRDT rooms. Overdeck is single-machine: overdeck-state syncs pipeline state per-project via git, but conversations/JSONL/tmux are machine-local. A design here must reconcile with the no-content-on-our-servers stance PAN-658 locks in (options: git-backed sync, E2E-encrypted relay, self-hosted room server).

6. Message queueing UI

Traycer queues prompts sent while the agent is mid-turn, with drag-reorder and edit-before-send of queued drafts. Overdeck typing during a turn just lands in the PTY input. Small but high-touch UX. (Also intersects PAN-658's contributor-queue open question — same queue primitive.)

7. Worktree safety classification + bulk cleanup UI

Traycer's worktrees settings classify each worktree by evidence tier (Merged / At base / Unreferenced / Review / Orphaned / In use) with bulk delete. Overdeck has close-out teardown + the reaper, but no operator-facing classification/cleanup surface for accumulated workspaces (a recurring real pain — disk pressure from stale workspaces).

8. Per-provider rate-limit usage views

Traycer surfaces provider rate-limit windows/usage in settings (host.getRateLimitUsage). Overdeck tracks cost but not rate-limit headroom per provider/subscription.

Minor gaps (batch into a polish issue if pursued)

• Customizable keybindings (chord capture, per-action rebinding) — dashboard has fixed shortcuts.
• MCP server management UI at user/org scope with auth status — Overdeck wires MCP per-agent internally, no management surface.
• In-dashboard notification center (bell + turn-completion/comment notifications) — Overdeck has the activity feed, needs-you queue, and TTS; no unified bell/inbox.
• Resource monitor UI chip (host CPU/mem) — the memory governor exists server-side with no frontend surface.
• Terminal appearance settings (font family/size, cursor style/blink) per user.
• Find-in-terminal / find-in-panel (dashboard has conversation search; no within-pane find).
• Onboarding flow (guided first-run: detect agents, theme, first task) — pan install/pan new-project are CLI-only.
• Prevent-sleep while agents run (desktop).
• Agent selection guide (editable global + per-workspace instructions steering which child agent/model to spawn) — partially covered by context layers; revisit alongside the multi-agent issue.

Non-goals

• Cloning Traycer's product shape. Overdeck's pipeline-first, git-durable, local-first architecture is a deliberate difference; parity items are adopted only where they serve it.
• Anything already excluded above as "have it".

Process

1. Triage each major gap: pursue / defer / reject, recorded here.
2. Pursued items get child issues (linked back here) with their own PRDs.
3. This epic closes when every major gap has a disposition.