[tau-bench] Make example runnable offline: stub user provider + robust tool parsing (parse_tools signature compatibility)

[Fengzdadi]

What happened

I’m running examples/tau-bench/generate_with_tau.py to validate the episode logging pipeline on a local machine (no GPU / no triton).
The run previously aborted due to:

• user simulation calling external LLM (litellm/gemini) → 429 / API key issues
• tool parser API mismatch (parse_tools() signature differences across versions)
• a small bug in response dict access (output.get[...] vs output.get(...))

Expected

• Be able to run 1 task end-to-end without external API keys (at least for validating logging + control flow).
• Robust tool parsing fallback if tool parser is unavailable or signature differs.

Proposed changes

1. Offline user sim

• Allow user_model_provider="stub" (or provider="stub") so env step doesn’t call external LLM
• Returns a canned user message like "(stub user) OK."
• Cost = 0

2. Tool parsing compatibility
   Add a wrapper in openai_tool_adapter.py that:

• detects parse_tools signature (2-arg vs 3-arg)
• supports different return schemas (e.g. missing normal_text)
• falls back to “no tool calls” when parser unavailable

3. Bugfix
   Use:

raw_response = output.get("text", "")

(currently it is output.get["text"])

Repro

• Start local sglang http endpoint (or stub server) on 127.0.0.1:30000
• Environment: macOS, CPU-only, no triton installed
• Sanity check:

curl -s -X POST http://127.0.0.1:30000/generate \
  -H "Content-Type: application/json" \
  -d '{"text":"Say ONLY: OK123","sampling_params":{"temperature":0,"max_new_tokens":8}}'

• Run the tau-bench example with user_model_provider="stub"

Result after patch

• episode.jsonl is generated and env_step proceeds with observation (stub user) OK.
• No external API calls are needed (no Gemini/OpenAI keys), so no 429.