Add ACP support with Cursor provider

[juliusmarminge]

Summary

• Adds Cursor as a first-class provider with ACP session lifecycle support, health checks, and adapter wiring in the server.
• Implements Cursor model selection, including fast/plan mode mapping and session restart behavior when model options change.
• Preserves provider/thread model state through orchestration, projection, and turn dispatch paths.
• Updates the web app to surface Cursor traits, provider/model selection, and session drafting behavior.
• Expands runtime ingestion so completed tool events retain structured tool metadata.

Testing

• bun fmt
• bun lint
• bun typecheck
• Added and updated tests across server, contracts, shared, and web layers for Cursor adapter behavior, orchestration routing, session model changes, and UI state handling.
• Not run: bun run test

---

Note

Medium Risk
Adds a new cursor text-generation provider that spawns an external agent acp process and applies model/config options via ACP, which can fail or hang if the binary/protocol or config option mapping is wrong. Also changes git diff capture limits and several long-running integration tests’ timeouts, which could mask regressions or increase CI flakiness.

Overview
Adds Cursor-backed git text generation via ACP (CursorTextGenerationLive) and wires it into RoutingTextGeneration by extending TextGenerationProvider with cursor.

Introduces ACP test tooling (scripts/acp-mock-agent.ts, cursor-acp-model-mismatch-probe.ts) plus a new CursorTextGeneration test suite that validates ACP config-option based model selection and ensures the ACP child process is closed.

Makes a few supporting changes: desktop backend readiness now defaults to /.well-known/t3/environment, effect-acp is added as a dev dependency (and bundled via tsdown), checkpoint git diff output is capped via maxOutputBytes, Claude text generation switches to resolveClaudeApiModelId, and multiple integration/git tests are refactored to use longer, centralized timeouts.

^{Reviewed by Cursor Bugbot for commit e78ff43. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Cursor as a supported AI provider via ACP (Agent Communication Protocol)

• Introduces a new effect-acp package implementing the ACP JSON-RPC protocol layer, including typed client/agent interfaces, error types, schema codegen, and in-process/stdio transport.
• Adds CursorProviderLive and CursorAdapterLive to the server, enabling session management, model selection, streaming runtime events, tool calls, approval flows, and user input handling via Cursor CLI's ACP mode.
• Extends contracts (ProviderKind, ModelSelection, CursorModelOptions, CursorSettings) to formally include cursor as a valid provider with reasoning, fastMode, thinking, and contextWindow options.
• Wires Cursor into the web UI: composer provider registry, TraitsPicker, ProviderModelPicker, settings panel, and composerDraftStore all now handle the cursor provider.
• Adds CursorTextGenerationLive to back git text generation (commit messages, PR content, branch names) via Cursor ACP with a 180s timeout.
• Removes "restart-session" from ProviderSessionModelSwitchMode, so in-session Cursor model switches no longer restart the session.
• Risk: Cursor support depends on the agent binary being installed and accessible; sessions will fail at spawn time if the binary is missing or misconfigured.

^{Macroscope summarized e78ff43.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro

Run ID: 65aea02a-bbd9-41d1-8950-d3c9275fd9bd

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch t3code/greeting

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

Autofix Details

Bugbot Autofix prepared fixes for both issues found in the latest run.

• ✅ Fixed: Merge never clears cached provider model options
  • Replaced ?? fallback with key in incoming check so explicitly-present-but-undefined provider keys now clear cached values, and added cache deletion when merge produces undefined.
• ✅ Fixed: Mock agent test uses strict equal with extra fields
  • Changed toEqual to toMatchObject so the assertion tolerates the extra modes field returned by the mock agent.

Or push these changes by commenting:

@cursor push beb68c40d7

Preview (beb68c40d7)

diff --git a/apps/server/src/orchestration/Layers/ProviderCommandReactor.ts b/apps/server/src/orchestration/Layers/ProviderCommandReactor.ts
--- a/apps/server/src/orchestration/Layers/ProviderCommandReactor.ts
+++ b/apps/server/src/orchestration/Layers/ProviderCommandReactor.ts
@@ -50,20 +50,17 @@
   cached: ProviderModelOptions | undefined,
   incoming: ProviderModelOptions | undefined,
 ): ProviderModelOptions | undefined {
-  if (!cached && !incoming) {
-    return undefined;
+  if (incoming === undefined) return cached;
+  if (cached === undefined) return incoming;
+
+  const providerKeys = ["codex", "claudeAgent", "cursor"] as const;
+  const next: Record<string, unknown> = {};
+  for (const key of providerKeys) {
+    const value = key in incoming ? incoming[key] : cached[key];
+    if (value !== undefined) {
+      next[key] = value;
+    }
   }
-  const next = {
-    ...(incoming?.codex !== undefined || cached?.codex !== undefined
-      ? { codex: incoming?.codex ?? cached?.codex }
-      : {}),
-    ...(incoming?.claudeAgent !== undefined || cached?.claudeAgent !== undefined
-      ? { claudeAgent: incoming?.claudeAgent ?? cached?.claudeAgent }
-      : {}),
-    ...(incoming?.cursor !== undefined || cached?.cursor !== undefined
-      ? { cursor: incoming?.cursor ?? cached?.cursor }
-      : {}),
-  } satisfies Partial<ProviderModelOptions>;
   return Object.keys(next).length > 0 ? (next as ProviderModelOptions) : undefined;
 }
 
@@ -405,8 +402,12 @@
       threadModelOptions.get(input.threadId),
       input.modelOptions,
     );
-    if (mergedModelOptions !== undefined) {
-      threadModelOptions.set(input.threadId, mergedModelOptions);
+    if (input.modelOptions !== undefined) {
+      if (mergedModelOptions !== undefined) {
+        threadModelOptions.set(input.threadId, mergedModelOptions);
+      } else {
+        threadModelOptions.delete(input.threadId);
+      }
     }
     const normalizedInput = toNonEmptyProviderInput(input.messageText);
     const normalizedAttachments = input.attachments ?? [];

diff --git a/apps/server/src/provider/acp/AcpJsonRpcConnection.test.ts b/apps/server/src/provider/acp/AcpJsonRpcConnection.test.ts
--- a/apps/server/src/provider/acp/AcpJsonRpcConnection.test.ts
+++ b/apps/server/src/provider/acp/AcpJsonRpcConnection.test.ts
@@ -32,7 +32,7 @@
         cwd: process.cwd(),
         mcpServers: [],
       });
-      expect(newResult).toEqual({ sessionId: "mock-session-1" });
+      expect(newResult).toMatchObject({ sessionId: "mock-session-1" });
 
       const promptResult = yield* conn.request("session/prompt", {
         sessionId: "mock-session-1",

[macroscopeapp]

Approvability

Verdict: Needs human review

10 blocking correctness issues found. Diff is too large for automated approval analysis. A human reviewer should evaluate this PR.

^{You can customize Macroscope's approvability policy. Learn more.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 2 total unresolved issues (including 1 from previous review).

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: Default path overridden by subsequent options spread
  • Swapped the order so ...options is spread first and the path default (using ??) comes after, ensuring the computed default wins when options.path is absent or undefined.

Or push these changes by commenting:

@cursor push 4298a93567

Preview (4298a93567)

diff --git a/apps/desktop/src/main.ts b/apps/desktop/src/main.ts
--- a/apps/desktop/src/main.ts
+++ b/apps/desktop/src/main.ts
@@ -375,8 +375,8 @@
 
   try {
     await waitForHttpReady(baseUrl, {
+      ...options,
       path: options?.path ?? "/.well-known/t3/environment",
-      ...options,
       signal: controller.signal,
     });
   } finally {

_{You can send follow-ups to the cloud agent here.}

^{Reviewed by Cursor Bugbot for commit 6c36109. Configure here.}

[cursor]

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: Child process leaked - runtime close never called
  • Wrapped makeCursorAcpRuntime with Effect.acquireRelease so that runtime.close is called as a scope finalizer when Effect.scoped exits, properly terminating the spawned child process.

Or push these changes by commenting:

@cursor push 0f4c368c4c

Preview (0f4c368c4c)

diff --git a/apps/server/src/git/Layers/CursorTextGeneration.ts b/apps/server/src/git/Layers/CursorTextGeneration.ts
--- a/apps/server/src/git/Layers/CursorTextGeneration.ts
+++ b/apps/server/src/git/Layers/CursorTextGeneration.ts
@@ -126,12 +126,15 @@
       ).pipe(Effect.catch(() => Effect.undefined));
 
       const outputRef = yield* Ref.make("");
-      const runtime = yield* makeCursorAcpRuntime({
-        cursorSettings,
-        childProcessSpawner: commandSpawner,
-        cwd,
-        clientInfo: { name: "t3-code-git-text", version: "0.0.0" },
-      });
+      const runtime = yield* Effect.acquireRelease(
+        makeCursorAcpRuntime({
+          cursorSettings,
+          childProcessSpawner: commandSpawner,
+          cwd,
+          clientInfo: { name: "t3-code-git-text", version: "0.0.0" },
+        }),
+        (rt) => rt.close,
+      );
 
       yield* runtime.handleSessionUpdate((notification) => {
         const update = notification.update;

_{You can send follow-ups to the cloud agent here.}

[macroscopeapp]

🟠 High

t3code/apps/server/src/orchestration/Layers/ProviderCommandReactor.test.ts

Line 269 in e78ff43

const runtime = ManagedRuntime.make(layer);

Line 269 declares const runtime inside the createHarness function, which shadows the module-level let runtime declared on lines 73-76. The outer runtime remains null forever, so the cleanup in afterEach (lines 86-88) never calls runtime.dispose(). This causes ManagedRuntime instances to leak between tests.

🚀 Reply "fix it for me" or copy this AI Prompt for your agent:

In file apps/server/src/orchestration/Layers/ProviderCommandReactor.test.ts around line 269:

Line 269 declares `const runtime` inside the `createHarness` function, which shadows the module-level `let runtime` declared on lines 73-76. The outer `runtime` remains `null` forever, so the cleanup in `afterEach` (lines 86-88) never calls `runtime.dispose()`. This causes `ManagedRuntime` instances to leak between tests.

Evidence trail:
apps/server/src/orchestration/Layers/ProviderCommandReactor.test.ts lines 73-76 (module-level `let runtime = null`), lines 86-88 (`afterEach` cleanup checking `if (runtime) { await runtime.dispose(); }`), line 269 (`const runtime = ManagedRuntime.make(layer);` inside `createHarness`), lines 303-319 (return object from `createHarness` does not include `runtime`).