agentHost: trim semantic search scope

Author: bhavyaus

src/vs/platform/agentHost/node/copilot/copilotSessionLauncher.ts

@@ -19,7 +19,7 @@ import { CopilotCliConfigKey, copilotCliConfigSchema, normalizeModelFamilyAlias,
 import { IAgentHostOTelService } from '../../common/otel/agentHostOTelService.js';
 import { reasoningEffortLevels, type ReasoningEffortLevel } from '../../common/reasoningEffort.js';
 import { AgentHostSandboxConfigKey, sandboxConfigSchema } from '../../common/sandboxConfigSchema.js';
-import { CLIENT_SEMANTIC_SEARCH_REFERENCE_NAME, SEMANTIC_SEARCH_TOOL_NAME } from '../../common/semanticSearchConstants.js';
+import { SEMANTIC_SEARCH_TOOL_NAME } from '../../common/semanticSearchConstants.js';
 import type { ModelSelection, ToolDefinition } from '../../common/state/protocol/state.js';
 import { RUNTIME_TOOL_SEARCH_TOOL_NAME } from '../../common/toolSearchConstants.js';
 import type { ActiveClientToolSet } from '../activeClientState.js';
@@ -133,15 +133,13 @@ export function clientToolNamesFromSnapshot(snapshot: IActiveClientSnapshot): Re
  * Narrows the names that gate prompt content so the system message never
  * advertises a tool the filters disabled. Client tools are `custom:`-source even
  * when they override a built-in, so bare-name and `custom:` forms match (the
- * aliased tools under either of their names). Routing keeps the unfiltered
+ * tool-search tool under either of its names). Routing keeps the unfiltered
  * set — the runtime is the enforcement point.
  */
 export function filterClientToolNames(names: ReadonlySet<string>, availableTools: readonly string[] | undefined, excludedTools: readonly string[] | undefined): ReadonlySet<string> {
 	if (!availableTools && !excludedTools) {
 		return names;
 	}
-	const sdkAvailableTools = toSdkToolFilterPatterns(availableTools);
-	const sdkExcludedTools = toSdkToolFilterPatterns(excludedTools);
 	const matches = (patterns: readonly string[], name: string) => {
 		const sdkName = toSdkClientToolName(name);
 		return patterns.some(pattern =>
@@ -154,37 +152,33 @@ export function filterClientToolNames(names: ReadonlySet<string>, availableTools
 	};
 	const result = new Set<string>();
 	for (const name of names) {
-		const allowed = !sdkAvailableTools || matches(sdkAvailableTools, name);
-		if (allowed && !(sdkExcludedTools && matches(sdkExcludedTools, name))) {
+		const allowed = !availableTools || matches(availableTools, name);
+		if (allowed && !(excludedTools && matches(excludedTools, name))) {
 			result.add(name);
 		}
 	}
 	return result;
 }
 
-/** Source qualifier the SDK uses for client-contributed tools. */
-const CUSTOM_TOOL_PATTERN_PREFIX = 'custom:';
-
-/** Maps workbench client-tool names to their SDK-registered names. */
+/** The SDK-registered name for a client tool; only the tool-search tool differs. */
 function toSdkClientToolName(name: string): string {
-	switch (name) {
-		case CLIENT_TOOL_SEARCH_REFERENCE_NAME:
-			return RUNTIME_TOOL_SEARCH_TOOL_NAME;
-		case CLIENT_SEMANTIC_SEARCH_REFERENCE_NAME:
-			return SEMANTIC_SEARCH_TOOL_NAME;
-		default:
-			return name;
-	}
+	return name === CLIENT_TOOL_SEARCH_REFERENCE_NAME ? RUNTIME_TOOL_SEARCH_TOOL_NAME : name;
 }
 
 /** Maps Agent Host reference names to the names registered with the SDK. */
 export function toSdkToolFilterPatterns(patterns: readonly string[] | undefined): string[] | undefined {
 	if (!patterns) {
 		return undefined;
 	}
-	return [...new Set(patterns.map(pattern => pattern.startsWith(CUSTOM_TOOL_PATTERN_PREFIX)
-		? `${CUSTOM_TOOL_PATTERN_PREFIX}${toSdkClientToolName(pattern.slice(CUSTOM_TOOL_PATTERN_PREFIX.length))}`
-		: toSdkClientToolName(pattern)))];
+	return [...new Set(patterns.map(pattern => {
+		if (pattern === CLIENT_TOOL_SEARCH_REFERENCE_NAME) {
+			return toSdkClientToolName(pattern);
+		}
+		if (pattern === `custom:${CLIENT_TOOL_SEARCH_REFERENCE_NAME}`) {
+			return `custom:${toSdkClientToolName(CLIENT_TOOL_SEARCH_REFERENCE_NAME)}`;
+		}
+		return pattern;
+	}))];
 }
 
 export interface ICopilotSessionRuntime {

src/vs/platform/agentHost/node/copilot/prompts/AGENTS.md

@@ -71,14 +71,13 @@ const TOOL_INSTRUCTION_LINES: readonly ToolInstructionLine[] = [largeOutputToolI
 ```
 
 **Caveat — `hasTool` sees CLIENT tools only.** It is `context.hasClientTool`,
-which knows only the forwarded workbench tools, addressed by the `name` each one
-is published under — normally the camelCase `toolReferenceName`
-(e.g. `openBrowserPage`, `runTask`, `getTaskOutput`), but a tool that overrides
-an SDK built-in is republished under the built-in's name (`codebase` arrives as
-`semantic_search`; see "Semantic search override"). NOT shell / server-SDK / MCP
-tools (MCP is discovered dynamically and isn't in the launch snapshot). A line
-gated on a name that is never a client tool silently never renders. Client-tool
-membership is controlled by the Chat Customizations tools enablement.
+which knows only the forwarded workbench tools, addressed by their **camelCase
+`toolReferenceName`** (e.g. `openBrowserPage`, `runTask`, `getTaskOutput`) — NOT
+the extension's snake_case ids, and NOT shell / server-SDK / MCP tools (MCP is
+discovered dynamically and isn't in the launch snapshot). A
+line gated on a name that is never a client tool silently never renders. The
+default client-tool allowlist is `chat.agentHost.clientTools` (see
+`chat.shared.contribution.ts`). Broadening this context is a known follow-up.
 
 These lines compose with a per-model `tool_instructions` override (see
 `composeToolInstructions`), so Lever 1 and Lever 2 stack.
@@ -119,28 +118,6 @@ This B-inject bridge is intentionally interim. A follow-up moves tool-search
 registration and ranking into VS Code core so Agent Host no longer depends on
 the Copilot extension's tool implementation or embeddings plumbing.
 
-## Semantic search override
-
-`chat.copilot.semanticSearch.enabled` controls the complete Copilot
-semantic-search surface, for local and remote Copilot CLI session types alike
-(`isCopilotCliSessionType`). When off, the launcher excludes the SDK's built-in
-`semantic_search` and the client publishes no `codebase` tool at all. When on,
-`AgentHostActiveClientService` republishes the Copilot extension's `codebase`
-tool under the name `semantic_search`; the session registers it as a
-non-deferred, permission-free built-in override. The launcher's
-`toSdkToolFilterPatterns` rewrite of `codebase` → `semantic_search` assumes that
-republishing, so the two must stay gated on the same predicate.
-
-The tool is addressed by its `codebase` reference name, like every other client
-tool, and routing resolves that same name via `getToolByName` — so publishing
-and routing cannot disagree about which tool owns the slot. Any *other* tool
-claiming `codebase` or `semantic_search` is dropped for the session, because two
-client tools cannot share one SDK registration.
-
-The tool-gated prompt reminders keep queries focused and require switching to
-`grep` or `glob` instead of retrying when the workspace index is unavailable,
-updating, empty, or otherwise unhelpful.
-
 ## Lever 2 — per-model contributor (`promptRegistry.ts` + `allPrompts.ts`)
 
 Guidance scoped to a model or family. Implement `IAgentHostPrompt` and register

src/vs/platform/agentHost/node/copilot/prompts/toolInstructions.ts

@@ -6,7 +6,6 @@
 import type { SectionOverride } from '@github/copilot-sdk';
 import { coalesce } from '../../../../../base/common/arrays.js';
 import { BrowserChatToolReferenceName, browserChatToolReferenceNames } from '../../../../browserView/common/browserChatToolReferenceNames.js';
-import { SEMANTIC_SEARCH_TOOL_NAME } from '../../../common/semanticSearchConstants.js';
 import { CLIENT_TOOL_SEARCH_REFERENCE_NAME } from '../../../common/toolSearchConstants.js';
 
 /**
@@ -27,9 +26,8 @@ import { CLIENT_TOOL_SEARCH_REFERENCE_NAME } from '../../../common/toolSearchCon
  */
 
 /**
- * A single tool-instructions entry. Returns its content (one or more
- * newline-separated sentences, with no leading or trailing newline) when it
- * applies, or `undefined` to contribute nothing.
+ * A single tool-instructions line. Returns its content (a single sentence, no
+ * surrounding newlines) when it applies, or `undefined` to contribute nothing.
  * Mirrors one `<>…</>` fragment in the extension's `toolUseInstructions` block.
  *
  * @param hasTool predicate for whether a tool name is available in the session.
@@ -64,18 +62,10 @@ const browserToolInstructions: ToolInstructionLine = hasTool => {
 	return `Use the browser tools (${BrowserChatToolReferenceName.OpenBrowserPage}, ${companion}, etc.) when beneficial for front-end tasks, such as when visualizing or validating UI changes.`;
 };
 
-/** Scope guidance, then what to do when the workspace index cannot answer. */
-export const COPILOT_AGENT_HOST_SEMANTIC_SEARCH_INSTRUCTION = [
-	`Use \`${SEMANTIC_SEARCH_TOOL_NAME}\` only for focused natural-language codebase queries when the exact text or symbol is unknown; prefer \`grep\` or \`glob\` when you know what to look for, run it alone rather than in parallel with other searches, and keep queries narrow because broad queries can return oversized results.`,
-	`If \`${SEMANTIC_SEARCH_TOOL_NAME}\` reports that the index is unavailable or updating, or returns no results, treat the outcome as inconclusive rather than as proof the code does not exist: do not retry or rephrase the query; switch to \`grep\` or \`glob\` instead.`,
-].join('\n');
-const semanticSearchToolInstructions: ToolInstructionLine = hasTool =>
-	hasTool(SEMANTIC_SEARCH_TOOL_NAME) ? COPILOT_AGENT_HOST_SEMANTIC_SEARCH_INSTRUCTION : undefined;
-
 /**
  * The registered tool-instruction lines, in render order.
  */
-const TOOL_INSTRUCTION_LINES: readonly ToolInstructionLine[] = [largeOutputToolInstructions, browserToolInstructions, semanticSearchToolInstructions];
+const TOOL_INSTRUCTION_LINES: readonly ToolInstructionLine[] = [largeOutputToolInstructions, browserToolInstructions];
 
 /** Tool-search guidance mirrored from the Copilot extension prompt. */
 const toolSearchToolInstructions: ToolInstructionLine = hasTool =>

src/vs/platform/agentHost/test/node/copilotSessionLauncher.test.ts

@@ -19,9 +19,9 @@ import type { IByokLmBridgeConnection, IByokLmChatRequest, IByokLmChatResult, IB
 import { AgentHostByokModelsEnabledConfigKey, type SchemaValues } from '../../common/agentHostSchema.js';
 import type { IAgentHostManagedSettingsPermissions } from '../../common/agentHostManagedSettings.js';
 import { CopilotCliConfigKey, copilotCliConfigSchema } from '../../common/copilotCliConfig.js';
-import { CLIENT_SEMANTIC_SEARCH_REFERENCE_NAME, SEMANTIC_SEARCH_TOOL_NAME } from '../../common/semanticSearchConstants.js';
 import type { IAgentHostOTelService } from '../../common/otel/agentHostOTelService.js';
 import { reasoningEffortLevels } from '../../common/reasoningEffort.js';
+import { SEMANTIC_SEARCH_TOOL_NAME } from '../../common/semanticSearchConstants.js';
 import { CustomizationType, McpServerStatus, type ModelSelection } from '../../common/state/protocol/state.js';
 import { CLIENT_TOOL_SEARCH_REFERENCE_NAME, RUNTIME_TOOL_SEARCH_TOOL_NAME } from '../../common/toolSearchConstants.js';
 import { ActiveClientToolSet } from '../../node/activeClientState.js';
@@ -519,7 +519,7 @@ suite('CopilotSessionLauncher shared session config', () => {
 				resumeManagedSettings: { permissions: managedSettingsPermissions },
 				ephemeralMcpServers: {},
 				ephemeralDisabledMcpServers: ['azure', 'disabled-workspace-server', 'github', 'native-plugin-server', 'synced-server'],
-				ephemeralExcludedTools: ['task'],
+				ephemeralExcludedTools: ['task', `builtin:${SEMANTIC_SEARCH_TOOL_NAME}`],
 			});
 		} finally {
 			sessions.dispose();
@@ -932,22 +932,6 @@ suite('filterClientToolNames', () => {
 			]
 		);
 	});
-
-	test('keeps workbench and SDK semantic-search names consistent', () => {
-		const names = new Set([SEMANTIC_SEARCH_TOOL_NAME]);
-		assert.deepStrictEqual(
-			[
-				[...filterClientToolNames(names, [CLIENT_SEMANTIC_SEARCH_REFERENCE_NAME], undefined)],
-				[...filterClientToolNames(names, undefined, [`custom:${CLIENT_SEMANTIC_SEARCH_REFERENCE_NAME}`])],
-				toSdkToolFilterPatterns([CLIENT_SEMANTIC_SEARCH_REFERENCE_NAME, `custom:${CLIENT_SEMANTIC_SEARCH_REFERENCE_NAME}`, 'builtin:*']),
-			],
-			[
-				[SEMANTIC_SEARCH_TOOL_NAME],
-				[],
-				[SEMANTIC_SEARCH_TOOL_NAME, `custom:${SEMANTIC_SEARCH_TOOL_NAME}`, 'builtin:*'],
-			]
-		);
-	});
 });
 
 /**
@@ -1030,51 +1014,20 @@ suite('CopilotSessionLauncher resume config', () => {
 		return (launcher as unknown as { _buildSessionConfig(plan: unknown, runtime: unknown): Promise<{ model?: string; reasoningEffort?: string; contextTier?: string; availableTools?: string[]; excludedTools?: string[]; modelCapabilities?: Record<string, unknown>; toolSearch?: { enabled: boolean } }> })._buildSessionConfig(plan, runtime);
 	}
 
-	test('excludes the built-in semantic search unless the client override is enabled', async () => {
-		const store = new DisposableStore();
-		// Same configured filters on both arms, so the snapshot is the only variable.
-		const overrides = { modelCapabilityOverrides: { '*': { excludedTools: ['mcp:*'] } } };
-		const disabled = await buildResumeConfig(
-			createLauncher(store, overrides),
-			undefined,
-			{ tools: [], plugins: [], mcpServers: {} },
-		);
-		const enabled = await buildResumeConfig(
-			createLauncher(store, overrides),
-			undefined,
-			{ tools: [{ name: SEMANTIC_SEARCH_TOOL_NAME }], plugins: [], mcpServers: {} },
-		);
-
-		assert.deepStrictEqual(
-			[disabled.excludedTools, enabled.excludedTools],
-			[['mcp:*', `builtin:${SEMANTIC_SEARCH_TOOL_NAME}`], ['mcp:*']],
-		);
-		store.dispose();
-	});
-
-	test('does not fall back to built-in semantic search when filters remove the client override', async () => {
+	test('exposes only the client semantic-search override', async () => {
 		const store = new DisposableStore();
 		const snapshot = { tools: [{ name: SEMANTIC_SEARCH_TOOL_NAME }], plugins: [], mcpServers: {} };
-		const builtinOnly = await buildResumeConfig(
-			createLauncher(store, { modelCapabilityOverrides: { '*': { availableTools: [`builtin:${SEMANTIC_SEARCH_TOOL_NAME}`] } } }),
-			undefined,
-			snapshot,
-		);
-		const customExcluded = await buildResumeConfig(
+		const disabled = await buildResumeConfig(createLauncher(store, {}), undefined, { tools: [], plugins: [], mcpServers: {} });
+		const enabled = await buildResumeConfig(createLauncher(store, {}), undefined, snapshot);
+		const filtered = await buildResumeConfig(
 			createLauncher(store, { modelCapabilityOverrides: { '*': { excludedTools: [`custom:${SEMANTIC_SEARCH_TOOL_NAME}`] } } }),
 			undefined,
 			snapshot,
 		);
 
 		assert.deepStrictEqual(
-			[
-				[builtinOnly.availableTools, builtinOnly.excludedTools],
-				[customExcluded.availableTools, customExcluded.excludedTools],
-			],
-			[
-				[[`builtin:${SEMANTIC_SEARCH_TOOL_NAME}`], [`builtin:${SEMANTIC_SEARCH_TOOL_NAME}`]],
-				[undefined, [`custom:${SEMANTIC_SEARCH_TOOL_NAME}`, `builtin:${SEMANTIC_SEARCH_TOOL_NAME}`]],
-			]
+			[disabled.excludedTools, enabled.excludedTools, filtered.excludedTools],
+			[[`builtin:${SEMANTIC_SEARCH_TOOL_NAME}`], undefined, [`custom:${SEMANTIC_SEARCH_TOOL_NAME}`, `builtin:${SEMANTIC_SEARCH_TOOL_NAME}`]],
 		);
 		store.dispose();
 	});

src/vs/platform/agentHost/test/node/toolInstructions.test.ts

@@ -5,8 +5,7 @@
 
 import assert from 'assert';
 import type { SectionOverride } from '@github/copilot-sdk';
-import { COPILOT_AGENT_HOST_LARGE_OUTPUT_TOOL_INSTRUCTION, COPILOT_AGENT_HOST_SEMANTIC_SEARCH_INSTRUCTION, resolveToolInstructionsOverride, toolSearchInstructionLines, universalToolInstructions } from '../../node/copilot/prompts/toolInstructions.js';
-import { SEMANTIC_SEARCH_TOOL_NAME } from '../../common/semanticSearchConstants.js';
+import { COPILOT_AGENT_HOST_LARGE_OUTPUT_TOOL_INSTRUCTION, resolveToolInstructionsOverride, toolSearchInstructionLines, universalToolInstructions } from '../../node/copilot/prompts/toolInstructions.js';
 import { CLIENT_TOOL_SEARCH_REFERENCE_NAME } from '../../common/toolSearchConstants.js';
 import { ensureNoDisposablesAreLeakedInTestSuite } from '../../../../base/test/common/utils.js';
 
@@ -60,16 +59,6 @@ suite('toolInstructions', () => {
 				]
 			);
 		});
-
-		test('adds focused-query and fallback reminders only when semantic search is available', () => {
-			assert.deepStrictEqual([
-				universalToolInstructions(hasTools(SEMANTIC_SEARCH_TOOL_NAME)),
-				universalToolInstructions(hasTools()),
-			], [
-				`${LARGE_OUTPUT_LINE}\n${COPILOT_AGENT_HOST_SEMANTIC_SEARCH_INSTRUCTION}`,
-				LARGE_OUTPUT_LINE,
-			]);
-		});
 	});
 
 	// `composeToolInstructions` is module-private; its composition/spacing