Skip to content

armgit / mso

Public Imported from GitHub
0

Browser-based graphical shell and control plane for a Linux server you own — real terminal, file manager, live metrics and BYOK AI in one mobile-first workspace. Next.js 16 · React 19 · node-pty · Tailwind v4 · shadcn/ui. Public alpha, self-hosted.

3.4 KB
Raw Blame

Pending-edit preview: design note

Item 6 in docs/AGENT-CODING-PLAN.md: "the TUI cannot show a diff before writing. /diff renders the server's historical artifact diff. There is no preview of a pending edit."

What is actually missing

Two different things, and only the first one is a coding gap:

  1. The agent cannot preview its own edit. fs_edit writes immediately. An agent that wants to check "does this old_string match exactly once, and what exactly changes" has to write and then read back, which is a write it cannot undo.
  2. The owner cannot see what changed. The TUI prints [write] fs_edit and then ✓ fs_edit. The change itself is invisible. That is the reason /diff exists as a separate command, and why it only ever shows historical artifact diffs.

Claude Code and opencode both answer these with the same shape: the edit tool returns the diff it applied, and the UI renders that diff inline.

Design

A pure diff function, not a diff library

lib/host/fs-edit-diff.ts exports editDiff(before, after, { context, maxLines }).

No diff algorithm. fs_edit derives after from before by one substitution, so the change region is exactly the common prefix and the common suffix around it. That is cheap, exact, and cannot misattribute a change the way a heuristic line matcher can. The output is unified-diff shaped so it renders with the colours /diff already uses:

@@ -12,6 +12,7 @@
 context
-removed
+added
 context

maxLines defaults to 40. When the changed region is longer, the hunk is cut and a … N more lines line is appended, with truncated: true so the caller knows the diff is not the whole story. A silently short diff would be worse than no diff.

fs_edit returns the diff it applied, and can dry-run

  • Always: diff, diffTruncated on the result, so the caller sees what happened without a follow-up read.
  • New dry_run: true: compute and validate everything (empty search, uniqueness, size caps, CAS hash), return the diff and the projected sha256, and write nothing. written: false is in the result so a caller cannot mistake a preview for a commit.

The projected hash in a dry run is deliberately the future hash, not the current one. If a caller wrongly passes it as expected_sha256, the real edit fails as stale, which is a safe failure. The reverse (returning the current hash) would silently allow a stale write.

The TUI renders the returned diff

scripts/mso-agent-edit-diff.mjs holds the rendering, so scripts/mso-agent-turn.mjs grows by a handful of lines rather than becoming the next file over the 220 ratchet. It renders under the ✓ fs_edit row, dim context, green additions, red removals, using the same width discipline as the rest of the transcript so a long line does not wrap into a wall.

Bounds

  • Diff text is capped per edit (maxLines), and the tool result already has a byte cap.
  • dry_run performs no write, so it cannot be used to probe write access differently than a real edit: it runs the same guards before returning.

Explicitly not in scope

  • An interactive approve/reject prompt for each edit. Approval already exists at the tool level (requestExactToolApproval with a digest); adding a second gate inside the edit would mean two places decide the same thing.
  • Replacing /diff. That command answers "what did this session change to artifacts over time", which is a different question from "what did this one call change".