Tool Registry Boundary Skill

You keep the 5,000+-tool executable registry out of module graphs that don't execute tools.

The rule

Client-reachable code reads tool metadata. Only code that actually executes a tool imports the registry.

@/tools/registry is a 10,000+-line barrel importing every tool. External ToolConfig entries mix plain data (params, outputs, name) with request/response closures, while InternalToolConfig entries contain semantic input projection and load their server implementation through lib/internal/tool-operations/registry.server.ts. Request closures can still reach SDK clients, API helpers, and parsers, which is what makes the executable barrel expensive: reaching it costs roughly 4,700 additional modules (measured; re-measure with --verbose).

getTool() returns the whole ToolConfig, so a single getTool import anywhere in a client-reachable file drags all of it in.

Which module to import

Three modules, cheapest first. Ids are their own artifact because resolution and existence checks need only the key set; outputs are their own because they are the larger half of the data with a single consumer. @/tools/metadata and @/tools/metadata-outputs both resolve ids through @/tools/tool-ids, which is what keeps them independent of each other — do not "helpfully" re-export one from another, or every caller pays for all three.

All lookups guard with Object.hasOwn. JSON.parse yields an object with the normal prototype, so a bare bracket lookup returns inherited members: a bare bracket lookup would answer getToolMetadata('constructor') with a function typed as tool metadata.

The generated artifacts

apps/sim/tools/generated/tool-ids.ts, tool-metadata.ts and tool-outputs.ts are produced by scripts/sync-tool-metadata.ts:

bun run tool-metadata:generate   # after adding/changing a tool
bun run tool-metadata:check      # what CI runs; fails if stale

Never hand-edit them. If you add a tool or change a tool's params/outputs, regenerate and commit the result, or CI fails.

Three non-obvious properties, each of which was measured and is easy to undo by accident:

• The data is a JSON string parsed at runtime, not an imported .json and not an object literal. With resolveJsonModule (which this repo enables), a .json import makes TypeScript infer a literal type for all 5,000+ entries and takes tsc --noEmit from 12.6s to 8m07s — a 38x regression. An ambient declare module does not short-circuit it, and an object literal costs the same. A single string literal is one cheap token for both the compiler and the bundler, and JSON.parse beats evaluating the equivalent literal at runtime. Do not "clean this up" into a .json import.
• The generator refuses to emit function values. If you add a field to METADATA_FIELDS that contains a closure, generation fails loudly rather than shipping executable config to the client. hosting and schemaEnrichment are excluded for exactly this reason (hosting.enabled, pricing, and enrichSchema are functions) — they are server-only.
• Empty param entries are stripped. The registry contains one (stt_deepgram_v2), which crashes callers that read param.type while iterating.
• Lookups resolve versions. getTool maps an unversioned name onto the newest version, and a few hundred tools are versioned. A plain key lookup would silently report them missing — a quiet correctness bug, not a crash. resolveToolId reproduces that against the id set and is differentially tested against the original.

Testing code that reads tool metadata

Mock the module the code under test actually reads. vi.mock('@/tools/utils', () => toolsUtilsMock) only controls getTool; code that reads params/outputs/name goes through @/tools/metadata, so mocking tools/utils there is a no-op that still passes — because the real generated artifacts happen to agree with the mock fixtures. The test looks green while controlling nothing.

import { blocksMock, toolsMetadataMock, toolsUtilsMock } from '@sim/testing/mocks'

vi.mock('@/tools/utils', () => toolsUtilsMock)      // executable lookup
vi.mock('@/tools/metadata', () => toolsMetadataMock) // params / outputs / name

Both are backed by the same mockToolConfigs, so mocking both gives one consistent tool universe. If you are unsure whether a mock is load-bearing, change a fixture value to a sentinel and confirm the test fails.

The guard

bun run check:tool-registry-boundary (CI: "Tool registry client-boundary audit") walks the module graph from each workspace route and fails if @/tools/registry is reachable, printing the exact import chain that reintroduced it.

If it fails, do not add the entry to an allowlist — there isn't one. Find the symbol the offending file actually needs and move it to a registry-free module, exactly as mergeToolParameters and formatParameterLabel were.

Run it with --verbose to print per-route module counts, which is also the quickest way to see whether a change moved the graph.

The same command also ratchets those counts against check-tool-registry-boundary.baseline.json. --check (what CI runs) fails when an entry exceeds its baseline by more than max(25 modules, 2%), naming the import chain responsible. This catches bloat the registry rule misses — a prefetch importing listTables cost the Tables page 444 modules without ever touching @/tools/registry.

Re-record with --update-baseline and commit the JSON when growth is deliberate. A shrink passes but is reported — re-record then too, or the win is silently spendable again.

How to verify an edge actually got cut

Do not eyeball imports — the registry is reached through several redundant paths, so cutting one buys nothing while another survives. Walk the graph:

1. From the entry you care about, follow import and export … from (skipping import type), resolving @/ against apps/sim.
2. Check whether apps/sim/tools/registry.ts is in the reachable set, and print the parent chain if it is.
3. Compare the reachable module count before and after.

Reference points measured on this repo:

The canvas route reached the registry through four redundant edges — providers/utils (via tools/params), lib/workflows/blocks/block-outputs, lib/workflows/sanitization/validation, and serializer/index. Cutting any one alone moved the module count by ~1. They all had to go before anything improved; measure the route, not the file you edited.

When adding a new caller

Ask what the caller does with the config. If it reads params, outputs, name, description or just checks existence, it belongs on @/tools/metadata — no exceptions, even on a path you believe is server-only today, because a future client import will silently re-attach the registry to the graph.

If it genuinely executes — builds an external request, transforms a response, or dispatches a registered internal operation — use getTool, and keep that file off client-reachable paths.