Find AI Agent Failures in Seconds

You are Claude Code working directly inside my repository.

Goal:
Install ZappyBee tracing as fast as possible with zero product regressions.

Critical constraints:
1) Detect framework(s), language(s), runtime(s), and package manager(s) automatically.
2) Keep existing app behavior intact. No breaking changes.
3) Do NOT hardcode secrets. Use env vars only.
4) Do NOT commit or push anything.
5) At the end, output:
   - changed files
   - exact run command
   - exact smoke test command

Configuration:
- SDK package name: zappybee
- API key env var: ZAPPYBEE_API_KEY
- Base URL env var: ZAPPYBEE_BASE_URL
- Base URL value: https://tokencat-api-riuvb.ondigitalocean.app

Implementation tasks:
1) Detect stack
- Identify whether this repo uses TypeScript/JavaScript, Python, or both.
- Detect package manager preference from lockfiles/config:
  pnpm-lock.yaml > package-lock.json > yarn.lock > bun.lockb
- For Python detect toolchain:
  uv > poetry > pipenv > pip

2) Install dependencies
- If TS/JS exists, install zappybee with the detected package manager.
- If Python exists, install zappybee with the detected Python tool.
- If OpenAI or Anthropic SDKs are already present, keep using them.
- Do not remove existing dependencies.

3) Configure environment
- Add/update env examples so they include:
  ZAPPYBEE_API_KEY=
  ZAPPYBEE_BASE_URL=https://tokencat-api-riuvb.ondigitalocean.app
- If there is an app env loader, wire both vars there.
- Never place a real API key in source files.

4) Add SDK initialization
- Add one shared initialization location per runtime (TS and/or Python).
- Initialize with:
  apiKey from ZAPPYBEE_API_KEY
  baseUrl from ZAPPYBEE_BASE_URL (fallback to https://tokencat-api-riuvb.ondigitalocean.app if needed)

5) Add automatic tracing where possible
- If Anthropic/OpenAI clients are found, apply ZappyBee.wrap(client).
- Keep existing client options and behavior unchanged.

6) Add manual tracing for unsupported/custom flows
- Add one traced execution path with:
  startTrace -> startStep -> step.end -> trace.end
- Include model, prompt/completion tokens, status, output/error.
- Use try/catch/finally (or equivalent) so traces close on failures.

7) Add smoke test path
- Add a minimal script or command that triggers one traced run.
- Script should work locally without exposing secrets in code.
- Add clear comment on how to run it.

8) Validate
- Run existing build/typecheck/lint commands where available.
- Fix only issues introduced by your changes.

9) Final report
- Print:
  a) what was detected (framework/language/package manager)
  b) exact files changed
  c) install command(s) used
  d) run command
  e) smoke test command
  f) where to check traces in UI

Now apply the changes directly.