BenStullsBets 94b1a9b0c2 docs: author-colored track-changes design + implementation plan
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-26 13:32:34 -07:00

vscode-cowriting-plugin

Non-shippable proof-of-concept (Feature #2 of Epic #1): a standalone VS Code extension that drives @cline/sdk — validating Approach A (own coauthoring extension on the Cline SDK, no fork).

What it does

Registers one command, Cowriting: Show Cline SDK Info, which loads @cline/sdk and shows the SDK build version plus the agent's builtin tool catalog (a pure, key-free SDK call) in a notification and the "Cowriting (Cline SDK)" output channel.

Features shipped so far: F2 region-anchored threads (Feature #4), F3 live human/Claude attribution (Feature #6), F4 propose/accept diff flow (Feature #12), F5 cross-rung sidecar contract (Feature #14), F6 diff-view toggle (Feature #17), F10 interactive review — write left / review right (Feature #29), and F11 — the preview toolbar as the primary interaction surface (Feature #43).

Architecture

  • CommonJS extension bundled with esbuild (src/extension.tsout/extension.cjs).
  • @cline/sdk is ESM-only (Node ≥22) and uses createRequire(import.meta.url), so it is not bundled — it is shipped in node_modules and loaded at runtime via dynamic import() from the vscode-free src/cline.ts.

Run it (F5)

  1. npm install
  2. npm run build
  3. Press F5 (or Run → "Run Extension") to launch the Extension Development Host. It opens the committed sandbox/ playground as its workspace (not the repo itself — VS Code won't open one folder in two windows, #8); start with sandbox/playground.md.
  4. In the new window: Cmd/Ctrl+Shift+P"Cowriting: Show Cline SDK Info".

F2 — Region-anchored threads (Feature #4)

Attach durable, region-anchored discussion threads to any document. Threads render in the native VS Code Comments gutter and persist as git-native sidecars under .threads/<doc-path>.json (plain, diffable JSON — no server).

  • Create: select text → run "Cowriting: Add Coauthoring Thread on Selection" (or the Comments gutter "+").
  • Reply / Resolve: use the native Comments reply box and the thread's Resolve/Reopen actions.
  • Survives edits, reload, and external change (git pull): a hybrid anchor (durable content fingerprint + live offset tracking) re-resolves the thread. If the anchored text can't be confidently re-found, the thread is shown as orphaned at its last-known line — never silently moved.

Design: vscode-cowriting-plugin-content/specs/coauthoring-inner-loop.md. No live @cline/sdk turn and no credentials are involved in F2.

F3 — Live human/Claude attribution (Feature #6)

As you and Claude coauthor, every span in the buffer carries an author: human edits render with a subtle left border, Claude-authored spans with a background tint. Text that predates tracking stays plain — the honest record. Edits that touch a boundary (split, merge, partial overwrite) are handled char-precisely.

Attribution is persisted git-natively in the same .threads/ sidecars (attributions[], sharing the same anchors fingerprints as F2 threads). On reload, fingerprints re-resolve spans against the current document; spans that can't be confidently re-found are orphaned (status-bar count + "Cowriting Attribution" output channel) rather than silently moved or discarded.

Commands

  • Cowriting: Ask Claude to Edit Selection — select text → enter an instruction → a live @cline/sdk turn runs on the built-in claude-code provider (rides your local Claude Code Pro/Max login; the extension stores no credentials). As of F4 the turn ends in a proposal (see below); accepted text lands as a Claude-attributed span.
  • Cowriting: Toggle Attribution — show/hide attribution decorations.
  • cowriting.applyAgentEdit (palette-hidden) — the single machine-edit ingress seam. Tests drive this directly so CI requires no LLM.

Design: vscode-cowriting-plugin-content/specs/coauthoring-attribution.md.

F4 — Propose/accept diff flow (Feature #12)

Claude's edits arrive as pending proposals — propose-by-default, the document never changes without your say-so. A proposal renders two ways at once: an amber tint on the target range, and a "Claude proposes" comment thread showing a fenced diff of current → proposed text with two actions:

  • ✓ Accept Proposal — applies the replacement through the applyAgentEdit seam, so it lands Claude-attributed (F3) — and the proposal disappears.
  • ✗ Reject Proposal — the document is untouched; the proposal disappears.

Pending proposals persist git-natively in the same sidecar (proposals[], sharing the F2/F3 anchors fingerprints), survive reload, and re-anchor as surrounding text changes. If the target text itself changes, the proposal goes stale (status-bar count, accept disabled — never applied by guess); undo the change and it becomes decidable again. cowriting.proposeAgentEdit (palette-hidden) is the programmatic propose ingress E2E drives — no LLM in CI.

Design: vscode-cowriting-plugin-content/specs/coauthoring-propose-accept.md. Live smoke: docs/MANUAL-SMOKE-F4.md.

F5 — Cross-rung format + round-trip (Feature #14)

The .threads/ sidecar is the published cross-rung contract: any rung of the ladder (this editor → Gitea substrate → rfc-app) reads and writes the same record per the contract; git push/pull is the transport, no re-homing ever.

  • Normative contract (INV-14): vscode-cowriting-plugin-content/specs/coauthoring-sidecar-contract.md — format changes land there first, then schema, then code.
  • Machine-checkable half: schemas/coauthoring-sidecar.schema.json; validate any sidecar from any rung with node scripts/validate-sidecar.mjs <file>. The unit suite validates every serialized artifact — contract drift fails CI.
  • Identity crosses rungs: Provenance.email (git's own join key — populated from the workspace git config when available) and agent.onBehalfOf (who the machine acted for).
  • Writers preserve unknown fields (INV-15): rewriting a sidecar never destroys another rung's data — unknown keys survive, after known keys, sorted.
  • Newer-major sidecars are read-only (INV-16): the editor renders what it understands, warns once, and writes nothing (the store refuses as a backstop).
  • Deterministic merge (INV-17): src/mergeArtifacts.ts — union-by-id, documented tie-breaks, every resolved divergence surfaced in conflicts.
  • The round-trip is proven, not asserted: scripts/crossrung-reply.mjs is a self-contained conforming foreign writer (the Gitea-rung stand-in); host E2E drives editor-thread → stand-in reply → external-change → the reply renders in the thread.

Design: vscode-cowriting-plugin-content/specs/coauthoring-cross-rung-format.md.

F6 — Diff-view toggle (Feature #17, #19)

Ctrl+Alt+D (the same chord on macOS — not Cmd; or Cowriting: Toggle Diff View) flips the focused document into a native vscode.diff against a coauthoring baseline — the readonly baseline on the left, your live, editable document on the right (so you keep writing inside the diff; toggling again closes it). The diff answers "what did I change?" in one keystroke instead of git archaeology.

  • Works on any file (#19): any document you can edit — a file inside or outside the workspace folder, or an untitled scratch buffer. Only a non-text-editor focus warns. (Untitled buffers diff in-memory; saving makes the baseline persist.)
  • Machine-factored baseline (INV-18): the baseline initializes when a doc is first seen and advances automatically at every machine landing (every successful applyAgentEdit seam apply — INV-9). So text Claude landed never shows as a change; everything the diff shows is operator-authored by construction — no attribution filtering.
  • Pin on demand: Cowriting: Pin Diff Baseline to Now resets the baseline to the current buffer for a deliberate "review my next pass" epoch; the diff tab title names the epoch (opened / Claude landed / pinned).
  • Pure view, repo-free (INV-19): the baseline snapshot lives in VS Code global extension storage, keyed by a hash of the document URI, never the repo — .threads/, the cross-rung contract (INV-14..17), and SCHEMA_VERSION are untouched. Storage-unavailable degrades to in-memory baselines + one warning.
  • No LLM in CI: host E2E (test/e2e/suite/diffView.test.ts) drives the same programmatic seam ingress (propose + accept) the F4 suite uses.

Design: vscode-cowriting-plugin-content/specs/coauthoring-diff-view.md. Live smoke: docs/MANUAL-SMOKE-F6.md.

F7 — Rendered track-changes preview (Feature #21)

Ctrl+Alt+R (or Cowriting: Open Track-Changes Preview) opens a read-only webview beside a Markdown editor that renders the document and marks what changed since the F6 baseline — the "track changes" / "suggesting mode" altitude rather than a raw-text split-diff:

  • Prose additions are highlighted (<ins>), deletions struck (<del>), refined to the word.
  • Code and mermaid fences are diffed whole (atomic, INV-23): a changed or added one renders fully with a small "changed" badge; a removed one renders struck. Mermaid fences render as diagrams (mermaid runs in the webview). Intra-diagram node/edge diffing is deferred (#22).
  • It updates live as you and Claude edit (debounced), and re-bases when Claude lands an edit (baseline advances, INV-18) or you pin — so accepted text drops its marks. It reuses the F6 baseline and adds no persistence (pure read-only, INV-20).
  • The webview is sealed (INV-21): local bundled assets only, strict CSP with a per-load nonce, no network/CDN, no LLM. Mermaid is bundled into the webview asset only, never the extension-host bundle.
  • The render engine (src/trackChangesModel.ts) is a pure, vscode-free function (INV-22), unit-tested with no editor and no webview; host E2E (test/e2e/suite/trackChangesPreview.test.ts) drives the same programmatic propose/accept seam with no LLM.

F7 is markdown-only; for any other file (incl. code), use F6's diff toggle (Ctrl+Alt+D). The webview's visual rendering (mermaid, theming) is verified by the manual smoke, not the sealed-sandbox E2E.

Design: vscode-cowriting-plugin-content/specs/coauthoring-rendered-preview.md. Live smoke: docs/MANUAL-SMOKE-F7.md.

F9 — Authorship view in the preview (Feature ~#27)

The rendered preview (F7) gains a second mode, switched by a [ Track changes | Authorship ] toggle in its header. Authorship mode re-renders the current document with each span colored by its F3 author — Claude (blue) vs you (green), inline and char-precise — with a legend. Unlike track-changes (which diffs against the F6 baseline, and so hides Claude's text once the baseline advances past a landing), authorship reads F3 attribution directly, so Claude's contributions stay visible. Code/mermaid fences carry a block-level author badge (atomic). Read-only, sealed webview, no new persistence (INV-26..28).

Design: docs/superpowers/specs/2026-06-11-authorship-preview-design.md. Live smoke: docs/MANUAL-SMOKE-F9.md.

F8 — Out-of-workspace authoring (Feature #25)

"Ask Claude to Edit Selection" (and F2 threads / F3 attribution / F4 propose-accept) now work on any document the editor shows — saved in the workspace folder, saved outside it, or untitled — matching the already-universal F6 diff and F7 preview. Authoring is no longer gated to in-workspace files, and the commands are live even with no folder open.

Persistence is hybrid (one SidecarStore abstraction, routed per-document):

  • an in-workspace file keeps its committable .threads/<repo-rel>.json sidecar, byte-for-byte unchanged (INV-2) — the only home the F5 cross-rung contract ever sees;
  • an out-of-workspace file or untitled buffer stores its coauthoring artifact in VS Code global storage keyed by sha256(uri) (the same home and key F6's baseline uses, INV-19/24); untitled buffers are in-memory only (lost on reload/save).

A global-storage artifact is not a committed file, so it is never cross-rung-shareable (INV-25), and renaming/moving the file orphans its artifact (the key is the URI hash) — both stated as design contract, not discovered later. Routing leaves the in-workspace path untouched, so rollback is a plain PR revert with zero data migration.

Design: vscode-cowriting-plugin-content/specs/coauthoring-out-of-workspace.md. Live smoke: docs/MANUAL-SMOKE-F8.md.

F10 — Interactive review: write left / review right (Feature #29)

A clean, zero-annotation editor on the left; the rendered preview on the right as the single interactive review surface. The editor carries no attribution tint, no in-editor proposal threads, and no diff — all review lives in the preview, toggled by the Annotations switch in its header (on by default). Open it via Ctrl+Alt+R, the editor title-bar button, or right-click a markdown file in the Explorer / its editor tab → "Open Cowriting Review Panel" (#41) — the right-click acts on the clicked document, opening it first if needed.

In the on-state the preview shows green = human / blue = Claude / strikethrough = deleted, and surfaces each of Claude's pending F4 proposals as a blue cw-proposal block with ✓ / ✗ buttons: accepts (the replacement lands Claude-attributed via the seam and the baseline advances past it), rejects (the block vanishes, the document untouched). With no preview open, a status-bar indicator shows the pending-proposal count and opens the review when clicked. Toggle Annotations off for clean rendered markdown. Read-only, sealed webview, no new persistence (INV-32..34).

F6's two-pane diff and F9's authorship view are retained only as data layers (the baseline the preview diffs against; the F3 attribution that colors it) — they are no longer separate user surfaces.

Design: vscode-cowriting-plugin-content/specs/coauthoring-interactive-review.md. Live smoke: docs/MANUAL-SMOKE-F10.md.

F11 — Preview toolbar as the primary interaction surface (Feature #43)

The review preview's header toolbar becomes the cockpit for the inner loop. Beside the existing Annotations switch it gains two controls:

  • ⌖ Pin baseline — pins the previewed document's review baseline to now and clears the change-marks (homes the previously-orphaned pinDiffBaseline command, which is reachable from the palette again too).
  • ✦ Ask Claude… — one adaptive button. Its label flips on the preview's selection: Edit Selection when text is selected in the rendered preview, Edit Document otherwise. Clicking it opens a host input box for the instruction (the LLM turn and prompt stay host-side — the sealed webview gains no LLM/credential surface), then surfaces the result as F4 proposals — one for a selection (mapped back to its source block-union), or one per changed hunk for a whole-document rewrite (diffToHunks), each independently ✓/✗-able.

The pure render layer emits data-src-start/data-src-end on every block in both modes (INV-36); the webview's only mapping duty is walking a selection to its nearest data-src ancestor. Right-clicking a markdown tab → Open Review Preview is the minimal gateway making the surface reachable end to end (#41/#42 expand it). Edit controls are disabled on a non-authorable doc (reading stays allowed). No new model, no new persistence — pin via the F6 store, edits via the F4 propose/accept seam with F3 attribution (INV-35..37).

Design: docs/superpowers/specs/2026-06-12-f11-preview-toolbar-interaction-surface.md. Live smoke: docs/MANUAL-SMOKE-F11.md.

Develop

  • npm run watch — rebuild on change.
  • npx vitest run — unit suite (SDK driver, schema, store, anchorer, thread mutations, attribution split/merge).
  • npm run test:e2e@vscode/test-electron host E2E (create → reply → resolve → persist → reload → re-anchor → orphan; drives cowriting.applyAgentEdit directly — no LLM required).
  • npm run smoke:live — scripted live-turn smoke test for F3; requires Claude Code installed and signed in. See docs/MANUAL-SMOKE-F3.md.
  • npm run typecheck — type-check without emit.
S
Description
Plugin for collaborative writing with machines
Readme 3.4 MiB
Languages
TypeScript 97.7%
JavaScript 1.9%
CSS 0.4%