Add provider skill discovery

[juliusmarminge]

Summary

• Adds provider skill presentation and slash-command discovery so composer UI can surface provider capabilities more directly.
• Updates server/provider orchestration and Codex app-server plumbing to carry the new skill metadata through session and turn flows.
• Refactors several effect service tags and ID constructors to the newer ServiceMap/makeUnsafe patterns for consistency.
• Adjusts CLI publish/package metadata and related tests to match the new build outputs and config behavior.

Testing

• Not run
• Existing unit and integration tests were updated alongside the feature changes
• Full project checks still need to pass: bun fmt, bun lint, bun typecheck, and bun run test

---

Note

Medium Risk
Extends the ServerProvider contract and provider health/probing flows to carry new capability metadata, and updates chat composer parsing/rendering and menu ranking logic; regressions could affect provider snapshots or composer input behavior across clients.

Overview
Provider snapshots now include capability metadata: ServerProvider gains slashCommands and skills arrays (with decoding defaults for backward compatibility), and buildServerProvider populates them.

Server-side probing is extended to fill these fields: Claude’s SDK init probe now collects and dedupes slash commands (and caches the combined probe), while Codex switches from account-only probing to probeCodexDiscovery via the app-server JSON-RPC to return both account info and per-CWD skills.

The web composer UI now surfaces provider capabilities: it adds $skill token detection/rendering (Lexical ComposerSkillNode chips with tooltip descriptions), expands slash-command completion to include provider commands, adds skill search/completion, and refactors menu highlighting + ranking using a new shared searchRanking utility (also reused for workspace entry ranking). Tests and fixtures are updated accordingly.

^{Reviewed by Cursor Bugbot for commit 016dbfd. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add provider skill discovery and skill/slash-command support in the composer

• Adds ServerProviderSkill and ServerProviderSlashCommand schemas to the contracts layer; both default to empty arrays when decoding legacy snapshots.
• Probes the Codex provider via a new probeCodexDiscovery API to return discovered skills alongside account info, and parses Claude Agent SDK initialization commands into normalized slash commands.
• Introduces a ComposerSkillNode Lexical node that renders inline skill chips with tooltips, and extends splitPromptTextIntoComposerSegments to parse $<name> tokens as skill segments.
• Extends the composer command menu to show ranked, grouped results for both provider slash commands and skills, with correct /<command> and $<skill> insertion on selection.
• Extracts a shared searchRanking package (scoreQueryMatch, insertRankedSearchResult, etc.) used by workspace entry search, skill search, and slash command search.
• Risk: ComposerPromptEditor now requires a skills prop; any caller that does not pass it will get a type error.

^{Macroscope summarized 016dbfd.}

[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: 1385ec2d-bc74-4b7b-b23e-c104f25cb89c

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/server-provider-skills

---

_{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: Ref assignments moved to useEffect cause stale reads
  • Moved the three ref assignments (composerMenuOpenRef, composerMenuItemsRef, activeComposerMenuItemRef) back to synchronous render phase by removing the useEffect wrapper, ensuring event handlers always read current values.
• ✅ Fixed: Exported probeCodexAccount function is now unused
  • Removed the dead probeCodexAccount function from codexAppServer.ts since it has no remaining imports anywhere in the codebase.

Or push these changes by commenting:

@cursor push 2d23dae3cc

Preview (2d23dae3cc)

diff --git a/apps/server/src/provider/codexAppServer.ts b/apps/server/src/provider/codexAppServer.ts
--- a/apps/server/src/provider/codexAppServer.ts
+++ b/apps/server/src/provider/codexAppServer.ts
@@ -244,16 +244,3 @@
     });
   });
 }
-
-export async function probeCodexAccount(input: {
-  readonly binaryPath: string;
-  readonly homePath?: string;
-  readonly signal?: AbortSignal;
-}): Promise<CodexAccountSnapshot> {
-  return (
-    await probeCodexDiscovery({
-      ...input,
-      cwd: process.cwd(),
-    })
-  ).account;
-}

diff --git a/apps/web/src/components/chat/ChatComposer.tsx b/apps/web/src/components/chat/ChatComposer.tsx
--- a/apps/web/src/components/chat/ChatComposer.tsx
+++ b/apps/web/src/components/chat/ChatComposer.tsx
@@ -820,11 +820,9 @@
       [composerHighlightedItemId, composerMenuItems],
     );
 
-    useEffect(() => {
-      composerMenuOpenRef.current = composerMenuOpen;
-      composerMenuItemsRef.current = composerMenuItems;
-      activeComposerMenuItemRef.current = activeComposerMenuItem;
-    }, [activeComposerMenuItem, composerMenuItems, composerMenuOpen]);
+    composerMenuOpenRef.current = composerMenuOpen;
+    composerMenuItemsRef.current = composerMenuItems;
+    activeComposerMenuItemRef.current = activeComposerMenuItem;
 
     const nonPersistedComposerImageIdSet = useMemo(
       () => new Set(nonPersistedComposerImageIds),

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

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR introduces a significant new feature: provider skill and slash command discovery. It adds new user-facing capabilities including skill autocomplete via $ syntax, skill pills in the composer, provider slash commands in the command menu, and supporting search/ranking infrastructure. New features of this scope warrant human review to validate the design and integration.

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

[cursor]

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

There are 3 total unresolved issues (including 2 from previous reviews).

Autofix Details

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

• ✅ Fixed: skills/list failure rejects entire account probe
  • Changed the skills/list error handler to gracefully fall back to an empty skills array instead of calling fail(), so account probing remains resilient when skills/list is unsupported or errors.

Or push these changes by commenting:

@cursor push 26dfbbe0ac

Preview (26dfbbe0ac)

diff --git a/apps/server/src/provider/codexAppServer.ts b/apps/server/src/provider/codexAppServer.ts
--- a/apps/server/src/provider/codexAppServer.ts
+++ b/apps/server/src/provider/codexAppServer.ts
@@ -205,12 +205,7 @@
 
       if (response.id === 2) {
         const errorMessage = readErrorMessage(response);
-        if (errorMessage) {
-          fail(new Error(`skills/list failed: ${errorMessage}`));
-          return;
-        }
-
-        skills = parseCodexSkillsResult(response.result, input.cwd);
+        skills = errorMessage ? [] : parseCodexSkillsResult(response.result, input.cwd);
         maybeResolve();
         return;
       }

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

[juliusmarminge]

@cursor push 26dfbbe

[cursor]

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

Autofix Details

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

• ✅ Fixed: Inconsistent field lookup for displayName vs shortDescription
  • Updated displayName lookup to check skill?.displayName first, falling back to display?.displayName, matching the same pattern used for shortDescription.

Or push these changes by commenting:

@cursor push 6b028d659a

Preview (6b028d659a)

diff --git a/apps/server/src/provider/codexAppServer.ts b/apps/server/src/provider/codexAppServer.ts
--- a/apps/server/src/provider/codexAppServer.ts
+++ b/apps/server/src/provider/codexAppServer.ts
@@ -64,8 +64,11 @@
           ? { description: nonEmptyTrimmed(skill?.description) }
           : {}),
         ...(nonEmptyTrimmed(skill?.scope) ? { scope: nonEmptyTrimmed(skill?.scope) } : {}),
-        ...(nonEmptyTrimmed(display?.displayName)
-          ? { displayName: nonEmptyTrimmed(display?.displayName) }
+        ...(nonEmptyTrimmed(skill?.displayName) || nonEmptyTrimmed(display?.displayName)
+          ? {
+              displayName:
+                nonEmptyTrimmed(skill?.displayName) ?? nonEmptyTrimmed(display?.displayName),
+            }
           : {}),
         ...(nonEmptyTrimmed(skill?.shortDescription) || nonEmptyTrimmed(display?.shortDescription)
           ? {

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

[cursor]

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

Autofix Details

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

• ✅ Fixed: Callback recreated on every keystroke causing unnecessary re-renders
  • Replaced the prompt state variable with promptRef.current in the setPromptFromTraits callback and removed prompt from the dependency array, restoring the stable ref-based pattern that prevents callback recreation on every keystroke.

Or push these changes by commenting:

@cursor push aadef2980e

Preview (aadef2980e)

diff --git a/apps/web/src/components/chat/ChatComposer.tsx b/apps/web/src/components/chat/ChatComposer.tsx
--- a/apps/web/src/components/chat/ChatComposer.tsx
+++ b/apps/web/src/components/chat/ChatComposer.tsx
@@ -868,7 +868,8 @@
     // ------------------------------------------------------------------
     const setPromptFromTraits = useCallback(
       (nextPrompt: string) => {
-        if (nextPrompt === prompt) {
+        const currentPrompt = promptRef.current;
+        if (nextPrompt === currentPrompt) {
           scheduleComposerFocus();
           return;
         }
@@ -879,7 +880,7 @@
         setComposerTrigger(detectComposerTrigger(nextPrompt, nextPrompt.length));
         scheduleComposerFocus();
       },
-      [composerDraftTarget, prompt, promptRef, scheduleComposerFocus, setComposerDraftPrompt],
+      [composerDraftTarget, promptRef, scheduleComposerFocus, setComposerDraftPrompt],
     );
 
     const providerTraitsMenuContent = renderProviderTraitsMenuContent({

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

[cursor]

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

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

• ✅ Fixed: Double-lowercasing in shared scoring utility is redundant
  • Removed the redundant .trim().toLowerCase() calls inside scoreQueryMatch since all callers already pass pre-normalized inputs, replacing them with a direct destructure and a JSDoc contract documenting the pre-normalization requirement.

Or push these changes by commenting:

@cursor push 20f5a4b15c

Preview (20f5a4b15c)

diff --git a/packages/shared/src/searchRanking.ts b/packages/shared/src/searchRanking.ts
--- a/packages/shared/src/searchRanking.ts
+++ b/packages/shared/src/searchRanking.ts
@@ -77,6 +77,12 @@
   return bestIndex;
 }
 
+/**
+ * Scores how well `value` matches `query` using tiered match strategies.
+ *
+ * **Expects pre-normalized inputs**: both `value` and `query` must already be
+ * trimmed and lowercased (e.g. via {@link normalizeSearchQuery}).
+ */
 export function scoreQueryMatch(input: {
   value: string;
   query: string;
@@ -87,8 +93,7 @@
   fuzzyBase?: number;
   boundaryMarkers?: readonly string[];
 }): number | null {
-  const value = input.value.trim().toLowerCase();
-  const query = input.query.trim().toLowerCase();
+  const { value, query } = input;
 
   if (!value || !query) {
     return null;

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

^{Reviewed by Cursor Bugbot for commit b0488e2. Configure here.}

[juliusmarminge]

@cursor push 20f5a4b