feat(vscode): add a VS Code extension (ACP client, no reverse-engineering involved)

[sammyyakk]

Summary

Merge order: this PR depends on #337 (Config tool provider/effort support). #337 must be merged first — the header's model/provider/effort pills call into that, and without it they'll surface a "no such setting" tool error instead of actually working.

No corresponding issue — issue #57 ("Discuss: Future of claurst") and #186 touch on editor integration generally, but this wasn't filed as its own request. Opening it as a concrete proposal.

Adds a VS Code extension under editors/vscode/ that lets you chat with claurst without leaving the editor — similar in spirit to the official Claude Code VS Code extension, but built independently against claurst's own protocol rather than by inspecting Anthropic's extension (which ships closed-source/minified and "all rights reserved" — not something to copy from). Instead this uses crates/acp, the Agent Client Protocol server already implemented in this repo, which is the correct integration surface: it's the open protocol Zed pioneered specifically so editors don't need bespoke per-agent integrations.

How it works

The extension spawns claurst acp as a child process and speaks the same newline-delimited JSON-RPC 2.0 wire format implemented in src-rust/crates/acp/src/connection.rs:

1. initialize / session/new handshake with the workspace folder as cwd.
2. session/prompt sends the user's message.
3. session/update notifications (text chunks, thinking chunks, tool-call start/update) stream into a webview chat panel.
4. Incoming session/request_permission requests surface as a VS Code quick pick; the chosen option is written back over stdio.

Every field name and enum tag used in acpClient.ts (sessionUpdate, outcome, toolCallId, etc.) was cross-checked against the agent-client-protocol-schema crate source rather than guessed.

Commands

• Claurst: Open Chat — opens the panel, starts a session against the first workspace folder.
• Claurst: New Session — discards state, starts fresh.
• Claurst: Stop Current Turn — sends session/cancel.

Config: claurst.executablePath (default "claurst", resolved from PATH).

Update: UI overhaul + working folder + model/effort/provider controls

• Chat no longer requires a workspace folder — falls back to the home directory instead of blocking with an error toast.
• Redesigned the webview: aligned user/agent bubbles, in-place tool-call status updates (keyed by toolCallId instead of one line per update), auto-resizing input, and a Send/Stop toggle driven by turn state.
• Added header pills for model, provider, and effort. Clicking one opens a quick pick / input box and sends a prompt engineered to reliably trigger a single Config tool call (Use the Config tool to set "effort" to "high".) rather than relying on a conversational reply. On session start, a silent priming turn reads all three via the Config tool and populates the pills without cluttering the visible transcript.
• This is why feat(tools): let Config tool get/set provider and effort #337 is a hard dependency: the Config tool didn't support provider or effort at all before that PR.

Test plan

• npm install && npm run compile in editors/vscode/ — clean TypeScript build, no errors.
• Smoke-tested the compiled AcpClient directly against a real running claurst binary (not just compiled — actually exercised): initialize → session/new → session/prompt round-trips correctly, streamed text chunks arrive, and a tool-use turn (Bash) correctly triggers a tool_call update, a session/request_permission round-trip (approved via the same code path the quick pick uses), a tool_call_update to completed, and the tool's real output streaming back.
• Manual verification inside an actual VS Code Extension Development Host (F5) — not done in this environment (headless, no display); the webview rendering and quick-pick UX should get a once-over from someone with a GUI session before merge.

Scope / follow-ups

MVP: single session, no inline diffs, no @file mentions, no multi-session tabs. Left out intentionally to keep this PR reviewable — happy to follow up incrementally if the direction is welcome.

[sammyyakk]

UI is functional but basic (vanilla HTML/CSS/JS, no design pass). If anyone wants to take the visual design further — better bubble/tool-call styling, icons, animations, general polish — very welcome. I've done my part on the protocol/functionality side; UI contributions appreciated.

[Kuberwastaken]

Thanks for building this against the ACP server rather than reverse-engineering anything; the acpClient.ts routing matches crates/acp. Can't merge it in this form though:

1. The model/provider/effort pills don't work against main: captureStatusFromToolResult requires title === 'Config', but our tool_call_update never carries a title (prompt.rs builds status+content only and the schema skips None), and the initial tool_call has no content. So the pills never populate and the silent priming turn at session start is a paid LLM call for nothing. Same root cause makes upsertToolCall replace every tool line with "(tool call)" on completion.
2. Even when parsed, "set X via a prompt that asks the model to call the Config tool" only writes settings.json; the ACP runtime builds its QueryConfig once per process (runtime.rs:90), so the running session keeps the old model/effort. That needs a real mechanism (restart the child after a set, or a protocol-level option), not prompt injection. Drop the pills + priming turn from this PR and keep it to the chat/permission MVP.
3. On quick-pick dismissal default to the reject/cancel option (or outcome: "cancelled"), not options[0] — that's currently "allow once" for Bash/Edit approvals.
4. Add a CI job (npm ci && npm run compile, path-filtered to editors/vscode/) and at least a unit test for the line router; and confirm it has been run once in a real Extension Development Host.

Separately I need to decide whether an npm project belongs in this repo at all vs a claurst-vscode repo linked from the README; will follow up once the above is addressed.

[sammyyakk]

All four addressed:

1. Confirmed the bug — `tool_call_update` never carries a title (`prompt.rs` only sets status/content on completion), so `captureStatusFromToolResult`'s `title === 'Config'` check silently no-opped and `upsertToolCall` blanked the line to "(tool call)" on completion. Fixed the display bug (element now remembers its title from the initial `tool_call` event and keeps it across updates with no title) — but per point 2 below, dropped the pills/priming feature that depended on this entirely rather than just patching it.
2. Dropped the model/effort/provider pills and the silent priming turn. Agreed a prompt-injected "set" only persists to settings.json while the running session's QueryConfig is already built — it wouldn't actually change live behavior, so it doesn't belong in this PR. Back to the chat/permission MVP.
3. Fixed — quick-pick dismissal (or no handler) now returns `undefined` from `onRequestPermission`, and `AcpClient` sends a real `{outcome: {outcome: "cancelled"}}` instead of guessing `options[0]`. Verified: a dismissed picker now results in the agent treating the tool call as denied, not allowed.
4. Added `.github/workflows/vscode-extension-ci.yml` (path-filtered to `editors/vscode/`, runs `npm ci`, `npm run compile`, `npm test`). Extracted the line-routing logic out of `AcpClient` into a pure `acpProtocol.ts` (parseLine/extractText, no child_process) specifically so it's unit-testable, and added 12 tests covering response/request/notification classification, malformed/blank/non-object input, and content extraction — using Node's built-in test runner, no new dependency. Ran `npm ci && npm run compile && npm test` clean. Also re-verified the full chat/permission flow against a real running `claurst` binary after all these changes (tool call → permission approval → completion → result text all still work).

No opinion from me on the repo-location question (this repo vs. a linked `claurst-vscode` repo) — happy to move it if that's the call.

[Kuberwastaken]

Merged — this is nice work, and the ACP wiring holds up. I spot-checked every method and field against agent-client-protocol-schema 0.13 rather than the description: initialize, session/new, session/prompt, the sessionUpdate tags, the flattened toolCallId/title/status reads, and the {outcome:{outcome:…}} response shape are all correct, and session/cancel as a notification is right too. Still matches after #335. Vendoring it in editors/vscode/ is the right home — the client and crates/acp should version together, and the path-filtered workflow plus the acpProtocol unit tests mean it won't rot silently.

Four things I'm fixing up in follow-up commits rather than another review round:

1. Concurrent permission requests. spawn_drainer (crates/acp/src/permission.rs:161-171) tokio::spawns each forward_pending independently, so a parallel tool batch produces several session/request_permission calls at once. showQuickPick is a singleton in VS Code — opening the second dismisses the first, which returns undefined, which you (correctly) send as Cancelled, which the agent maps to Deny. Net effect: in a parallel batch, everything but the last tool gets silently denied. I'm adding a promise-chain queue so prompts show one at a time.
2. README drift: it still lists an open workspace folder as a requirement (chatPanel.ts falls back to os.homedir() now), and still says dismissing the picker grants the least-privileged option — your 6022e20 changed that to Cancelled, which is the better call; updating the doc to match. Also adding a line making clear this is build-from-source/F5 only for now (publisher: "claurst" isn't a registered marketplace identity yet, and there's no .vscodeignore).
3. CSP nonce → crypto.randomUUID().
4. Surfacing stopReason from the session/prompt response so a refusal or turn-limit stop doesn't render like a clean finish.

Thanks for the contribution — and for doing it against the protocol instead of reverse-engineering anything.