Tool Registry

Model

interface ToolEntry {
  name: string         // wire-name, what the LLM emits in tool_calls
  namespace: string    // grouping for `tool_search` (web, kb, mcp:<id>, meta)
  description: string  // one-line summary for `tool_search`
  defer: 'never' | 'always' | 'auto'
  tool: Tool           // AI SDK Tool (schema + execute + needsApproval + toModelOutput)
  applies?(scope): boolean
}

registry (src/main/ai/tools/adapters/aiSdk/registry.ts) is a process-wide singleton. Tool files register at module-import time; the registry is read at request time by buildAgentParams. The Claude Code runtime has a separate tool system — tools/adapters/claudeCode/agentTools.ts builds its descriptors from MCP servers and built-in descriptors directly; it does not consume this aiSdk ToolRegistry.

Tests construct their own new ToolRegistry() to avoid singleton pollution.

Wire-name convention

Double underscore is the segment separator (so internal single _ stays unambiguous):

The built-in wire names live in @shared/ai/builtinTools (single-underscore, e.g. web_search); they are not derived from a __ segment convention like MCP.

Built-in tools

src/main/ai/tools/adapters/aiSdk/builtin/ registers four entries:

• web_search (WebSearchTool.ts → createWebSearchToolEntry) — namespace web. Talks to the configured web-search provider via the renderer-shared search service.
• web_fetch (WebFetchTool.ts → createWebFetchToolEntry) — namespace web. Fetches a URL's content.
• kb_search (KnowledgeSearchTool.ts) — semantic search over the active knowledge base.
• kb_list (KnowledgeListTool.ts) — enumerate available knowledge bases / documents.

Registration happens in builtin/registerBuiltinTools.ts (registerBuiltinTools). Each tool's applies gates on the relevant assistant.settings.* flag (e.g. enableWebSearch).

MCP tools

src/main/ai/tools/adapters/aiSdk/mcp/:

• resolveAssistantMcpToolIds — assistant's enabled MCP servers + per-tool disable list → set of tool ids.
• mcpTools.syncMcpToolsToRegistry({ selectedToolIds }) — reads each selected server's tools from the catalog via McpCatalogService.listTools (cache-only; see Tool catalog reads), registers each as a ToolEntry whose tool.execute proxies through the MCP transport. Scope: only servers owning a selected tool are synced. Because the read is last-known-good cache, a server is only evicted from the registry when its cache is genuinely empty — a transient blip can no longer drop a still-active server's tools.

The sync is idempotent; a stale entry is overwritten on the next sync.

Tool catalog reads never block on MCP

McpCatalogService splits the MCP tool catalog into a read facade and a write/refresh path:

• listTools(serverId) is cache-only — it returns the shared mcp.tools.<serverId> cache and never connects to the upstream MCP server. Every hot path that builds an agent/chat's tool surface uses it: the Claude Code SDK bridge (createSdkMcpServerInstance), buildMcpToolMetadata, the agent tool-policy (agentTools.listMcpDescriptors), and the two AI-SDK adapters above. A dead or slow server therefore cannot block agent/chat startup (issue #16242).
• refreshTools(serverId) (and the private listToolsForServer) is the live path that connects, lists, and writes the cache. It is driven entirely by background warmers: prewarmActiveServerTools (at onReady), the onToolListChanged refresh, the renderer's on-demand refreshTools (via useAgentTools), the server-enable toggle, and restartServer.

listTools also fires a single non-blocking refreshTools the first time it sees a never-warmed server (cache undefined, distinct from a warmed-but-empty []), so headless/cron starts self-warm without re-probing dead servers.

Trade-off: tool availability is eventually consistent. A server whose cache is still cold when a session starts contributes no tools to that session and appears on the next one — the Claude Agent SDK snapshots the tool list per session, so this cannot be made live mid-session.

Meta-tools

src/main/ai/tools/adapters/aiSdk/meta/ defines four tools that turn the registry into a search-then-call interface for the model. Only the first three are injected:

The injected three are added to the tool set by applyDeferExposition when (and only when) the request actually defers tools. See below.

Defer exposition

src/main/ai/tools/adapters/aiSdk/exposition/:

• shouldDefer(entries, contextWindow) — returns the set of names to defer. Two gates above the simple threshold:

  • MIN_AUTO_DEFER_COUNT — the auto pool must be large enough that search-then-invoke beats inlining.
  • META_TOOLS_OVERHEAD_TOKENS — estimated savings must exceed the meta-tools' static prompt cost. Without these gates, small tool sets
    • small-context models trigger defer and pay net-negative tokens.

• applyDeferExposition(tools, registry, contextWindow) — strips the deferred names out of tools, injects tool_search / tool_inspect / tool_invoke, and returns the entries the system-prompt's <DEFERRED_TOOLS> section needs to enumerate (so the model knows what namespaces exist).

Approval-gated tools are never deferred. A force-prompt MCP tool is registered with defer: 'never' — mcp/mcpTools.ts reads isMcpToolForcePromptBySource once to drive both defer and needsApproval — so it stays inline and the SDK's native approval gate fires on it. Deferring it would drop it from the SDK tool-set, so the gate would never fire and it would be reachable only through tool_invoke with no approval card. As a runtime backstop the tool_invoke / tool_exec meta-tools also call isApprovalGated at execution time and refuse a gated tool (covering the registry.getByName(any-name) vector), steering the model to call it inline. See Tool Approval.

tool_exec is not injected by applyDeferExposition — there is no metaTools.exec flag. The injection site (applyDeferExposition.ts:50-53) deliberately leaves it out: its worker_threads + new Function sandbox runs model-authored code with full Node privileges, a privilege-escalation surface vs the renderer's prior restrictions. It is meant to be re-enabled behind an explicit Preference key once there is a concrete need.

applies and tool-call repair

• applies(scope: ToolApplyScope) — per-entry predicate consulted at registry.selectActive. Throws are caught and treated as "inactive" with a warning log.
• createAiRepair(...) (tools/adapters/aiSdk/repair.ts) — passed to AI SDK as experimental_repairToolCall. When the model emits malformed args (InvalidToolInputError), the repair function gets one chance to fix it via a follow-up LLM call. Other failures (e.g. an unknown tool name) are returned unrepaired.

Where to read more

• Code: src/main/ai/tools/adapters/aiSdk/ (Claude Code adapter: src/main/ai/tools/adapters/claudeCode/)
• Tests: tools/adapters/aiSdk/__tests__/, tools/adapters/aiSdk/builtin/__tests__/, tools/adapters/aiSdk/exposition/__tests__/, tools/adapters/aiSdk/mcp/__tests__/, tools/adapters/aiSdk/meta/__tests__/
• Defer rationale, gate thresholds: tools/adapters/aiSdk/exposition/shouldDefer.ts (header doc + tests)
• Approval flow: Tool Approval