td-feature-build — implement one tool, new files only

SkillFiles & storage

Implement ONE tdmcp tool from a spec — the canonical file pattern (Zod schema + …Impl + register… ToolRegistrar), a bridge Python payload via buildPayloadScript, and an offline msw unit test — touching only new files, never shared registries. Use when coding/implementing/writing a tdmcp tool, generator, effect, control, CHOP/TOP/SOP builder, or Python bridge work, especially when several features are built in parallel.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the td-feature-build — implement one tool, new files only skill

What this skill tells your AI

The instructions your AI receives, as published by pantani/tdmcp in .agents/skills/td-feature-build/SKILL.md and read by ahel’s review.

Implement exactly what the spec says, in your own files, green in isolation. The parallel-build contract is simple: new files only; never edit anything shared (layer*/index.ts, tools/index.ts, cli/agent.ts, docs). That isolation is what lets builders run concurrently.

The canonical tool-file pattern

Every tool file exports a pure …Impl and a register…. Mirror a neighbour in the same layer, but the shape is:

import { z } from "zod";
import type { ToolRegistrar } from "../types.js";          // note the .js extension
import { runBuild, errorResult } from "../result.js";

export const fooSchema = z.object({
  name: z.string().default("foo"),
  intensity: z.number().min(0).max(1).default(0.5),
});

export async function fooImpl(ctx: ToolContext, args: z.infer<typeof fooSchema>) {
  // build via the client; never throw — return errorResult / runBuild on failure
}

export const registerFoo: ToolRegistrar = (server, ctx) =>
  server.registerTool(
    "create_foo",
    { title: "…", description: "…", inputSchema: fooSchema.shape },
    (args) => fooImpl(ctx, args),
  );

ctx is the ToolContext ({ client, knowledge, recipes, logger, vault?, allowRawPython }). …Impl is pure and unit-testable with a mocked client.

Hard rules

  • Never throw out of a handler. Validate via the Zod schema; turn TD failures into friendly isError results with errorResult / runBuild / friendlyTdError. Return, don't throw.
  • .js import extensions on every relative import (ESM/NodeNext). noUncheckedIndexedAccess is on — guard indexed access.
  • Biome: 2-space, double quotes, semicolons, trailing commas, 100-col. Check with ./node_modules/.bin/biome check <yourfiles> directly — npm run lint fails with a false ESLint parse error under the RTK proxy.
  • New files only. Your tool file + tests/unit/<feature>.test.ts. Do not touch shared files; report your export names to the integrator instead.

Bridge / Python work

  • Build the script with buildPayloadScript (encodes args as __PAYLOAD_B64__), execute through the client, parse the reply with parsePythonReport.
  • An executePythonScript payload must assign a result variable — the bridge reads it back.
  • If you edit td/ modules: keep TD-globals (op, app, project) inside functions so modules import cleanly, and run python3 -m py_compile on changed files.

TD gotchas that make a "successful" build actually cook

  • Many param/connect calls fail silently. Set the parameter that exists: a Level TOP has no gain — use brightness1. There is no ParMode.
  • No cross-container wires — route a signal out of a container through a Select TOP/CHOP, not a direct wire.
  • GLSL TOPs: declare out vec4 fragColor;; there is no built-in uTime (add your own uniform); avoid preamble #define collisions (F1/F2). KB shader snippets are references, not drop-in.
  • Reactive/time-dependent chains read 0 when the timeline is paused — that's expected, not a bug.

msw unit test

  • Mock the bridge with msw (no live TouchDesigner). Assert the real returned shape — operators created, params set, wiring, and isError paths — not a cast-away generic.
  • Run only yours: npx vitest run tests/unit/<feature>.test.ts. Green + biome-clean = done. You do NOT run the full build (the integrator does after wiring).

Output

The new tool file + test file, plus a note at _workspace/02_build_<feature>.md: files created, exported symbols (<feature>Impl, register<Feature>), target layer index, what the test asserts, and any spec deviation + reason. Then tell td-integrator the export names.

Signals

GitHub stars
41
Forks
9
Last commit
Aug 2026
Advanced
Catalog kind
skill
Gateway key
td-feature-build
Source
github.com/pantani/tdmcp