Beam IDE workflows — the oversight cockpit arcs

Date: 2026-07-21 Status: blueprint (mined + sized; not yet built). Directive: “replicate as many Zed/IDE workflows as meaningful into beam otherwise it’s half-baked” — with the filter that decides meaningful: Beam’s job is the supervise–review–intervene loop, never editing. Real IDEs are ACP clients of the same runtime; Beam competing on editing would be a worse editor instead of a better cockpit. OpenHands validates the filter from production: their oversight UI never lets the diff view edit — actual editing is a cleanly separate escaped-to iframe.

Sources: mined 2026-07-21 from OpenHands/frontend, bolt.diy, gitea (diff UI), theia — grounded against what this repo already ships. The sizing surprise: most of the backend already exists.

Arc 1 — Changed-files + diff review (flagship; this IS the state-diff bet)

Already in the tree: runtime/acpsvc/events.go diffContentFromResult() parses every file-write tool result into libacp.ToolCallContent{Type: Diff, Path, OldText, NewText} — the exact {original, modified} shape OpenHands’ diff endpoints serve — flowing per-edit through the session event stream. Missing: per-mission aggregation.

Contract (mirror OpenHands’ two-endpoint shape):

• GET /missions/{id}/changes → [{path, status: added|modified|deleted}]
• GET /missions/{id}/changes/diff?path= → {original, modified} Aggregation: fold the mission session’s diff events by path — first OldText, last NewText; status from create/edit/delete. Cap the list; per-file incomplete suppression flags for huge files (gitea’s pattern).

Rendering: Monaco DiffEditor (@monaco-editor/react) — the proven choice for exactly this use (OpenHands). Collapsed rows by default, lazy per-file diff fetch on expand, old/diff/new toggle. Do NOT hand-roll diffing (bolt.diy’s regret) and do NOT add editing or an editor iframe.

The differentiator: rank the changed-files list by Degree-of-Interest — weight and decay the unit’s interactions per path (reads, edits, tool dwell) from the already-journaled events, so review starts where the agent’s attention concentrated, not alphabetically. Landed-vs-planned joins here (mission plan entries beside the changed files).

Size: Go aggregation + endpoints ~0.5–1d; UI ~1–2d (near-port of OpenHands’ file-diff-viewer.tsx).

Arc 2 — Workspace-wide search

Backend: shell to rg --json under the mission/workspace root, stream matches incrementally (SSE/NDJSON — theia’s push model; never block on full scan). Shape per match: {path, line, column, length, preview}; server- or client-grouped by file; result cap (~500, theia’s default). Frontend: bolt.diy’s Search.tsx anatomy — 300ms debounce, per-file collapsible groups with match-count badges, highlighted substring with context window. Click-through routes into Arc 1’s diff view when the file is a changed file; otherwise inline context in the row — no general file viewer (editor territory). Size: ~0.5d Go, ~1d UI.

Arc 3 — Terminal

Backend is COMPLETE: runtime/terminalservice (real PTY, creack/pty, principal-scoped reattach) + runtime/internal/terminalapi (POST/GET/DELETE sessions, GET /terminal/sessions/{id}/ws, binary frames, in-band JSON {"type":"resize",cols,rows}, token-gated).

Frontend: @xterm/xterm + @xterm/addon-fit against the existing WS — bolt.diy’s bridge nearly verbatim (onData → send; onmessage → write; ResizeObserver → fit → in-band resize frame). Multi-terminal via id-keyed tabs if wanted. Size: ~0.5–1d (+0.5d multi-tab).

DECIDED 2026-07-21 (option b, taken under the land-everything mandate, flippable): the terminal ships read-only by default (“Nur Lesen” badge, xterm disableStdin) with an explicit “Shell übernehmen” affordance that flips the same view interactive behind an honest confirmation naming what it grants (a real shell on the host, bounded by the serve credential). The default lives at TERMINAL_READ_ONLY_DEFAULT in packages/harness/src/lib/hostTerminal.ts with a register-quality comment — flip it there if the posture should change. Same-day fix that unblocked it: terminalapi now authenticates via the injected serverapi gate (raw token OR session JWT — one login, every surface), replacing its raw-only compare that 401’d the browser cookie.

Arc 4 — Layout: inspector tabs, not docking

Theia’s Lumino docking shell is confirmed overkill (multi-week subsystem). Beam’s pages+tabs architecture is already the right shape (OpenHands proves the flat model suffices). Two cheap borrows: OpenHands’ dependency-free useResizablePanels hook (~80 lines, localStorage-persisted) for a two-pane split; react-resizable-panels only if a third pane ever exists. The mission inspector: a flat tab bar on MissionDetailPage hosting Changes / Search / Terminal / (transcript via the adoption affordance) as tab content. Size: hours.

Cross-cutting steal

Inline risk annotation on the action feed (OpenHands renders security_risk beside each bash command as a first-class part of action rendering): when our events carry gate/envelope decisions, render them inline in the transcript and timeline, not in a separate security view.

Build order

Arc 1 (with DOI ranking stubbed as ordering-by-edit-count first), then Arc 4’s inspector shell to host it, then Arc 3 (after the interactivity decision), then Arc 2. Every arc lands behind the same gates as the rest of Beam (typecheck, vitest, build-ui) and each gets a Playwright pass at both widths.