telegram-copilot-qa

Verify that Telegram control of TouchDesigner is correct, bounded, and safe. This feature crosses external chat, local LLM, tdmcp tools, and a live TD bridge, so most bugs live at boundaries.

Required checks

Security boundaries

• Bot token is read from config/env and never printed.
• chat_id / user allowlist is enforced before LLM invocation.
• Non-allowlisted input cannot trigger LLM calls, tool calls, or bridge calls.
• safe mode exposes read-only tools only.
• Mutating tools require confirmation unless the implementation explicitly

documents a narrower trusted mode.

• Confirmation is bound to the exact pending action and expires/cancels.
• No local server (9980, Ollama, chat UI) is exposed publicly by default.

Telegram boundary

• getUpdates offset handling prevents duplicate execution.
• sendMessage handles long/failing responses gracefully.
• Unsupported update types do not crash the loop.
• /cancel, /status, tier commands, /approve, and /panic are covered.

LLM/tool boundary

• Mock an OpenAI-compatible/Ollama response that emits a tool call.
• Assert tool-tier selection maps to the expected resolveTools output.
• Assert invalid tool-call args produce a safe user-visible failure.
• Assert conversation history is bounded.

TouchDesigner boundary

• Offline tests must pass without TD.
• If bridge is reachable, run a live smoke:

1. bridge health;
2. read-only prompt;
3. confirmed mutating prompt in a scratch parent;
4. node errors after cook;
5. layout coordinates inspected for created nodes.

• If bridge is offline, mark live checks UNVERIFIED - pending bridge.

Gate commands

Run the relevant focused tests first, then the repo gates before final PASS:

npm run typecheck
npm run build
./node_modules/.bin/biome check .
npm test
npm run test:bridge

If docs or recipes changed, also run:

npm run validate:recipes
npm run docs:build

Report format

Write _workspace/telegram-copilot/04_qa.md with:

• PASS: checks completed with evidence;
• FAIL: file:line, observed behavior, owner, smallest acceptable fix;
• UNVERIFIED: service/hardware unavailable or live check intentionally held;
• commands run and exact outcomes;
• residual security risks.

Trigger validation

Should trigger:

• "QA do bot Telegram do tdmcp"
• "validar segurança do Telegram + Ollama"
• "testar se chat_id não autorizado chega no LLM"
• "checar confirmações antes de mutar TouchDesigner"
• "rodar gates do telegram-copilot"
• "verificar live control pelo Telegram"

Should not trigger:

• generic TD feature QA unrelated to Telegram/Ollama;
• release docs-only updates;
• cookbook examples;
• Connectors Directory QA;
• broad coverage campaign work.