feat(sdk): TokenJamClient LiteLLM Integration

[anilmurty]

Summary

Adds tokenjam.sdk.TokenJamClient — a thin HTTP client that POSTs a single LiteLLM call as an OTLP JSON span to a running tj serve. Unblocks the upstream BerriAI/litellm named-callback PR (litellm.success_callback = [\"tokenjam\"]), which depends on a public surface that doesn't require running inside a tokenjam-aware app.

Bumps version to 0.2.2.

Public surface

```python
from tokenjam.sdk import TokenJamClient

c = TokenJamClient(
endpoint="http://localhost:7391\", # default
ingest_secret="...", # optional Bearer auth
)
c.emit_litellm_span(
kwargs={"model": "openai/gpt-4o-mini", "metadata": {"tj_agent_id": "my-agent"}},
response_obj=response,
start_time=start,
end_time=end,
success=True,
)
```

What gets captured

• gen_ai.request.model (provider prefix stripped — `openai/gpt-4o-mini` → `gpt-4o-mini`)
• gen_ai.provider.name (from response._hidden_params['custom_llm_provider'], falling back to the model prefix)
• gen_ai.usage.input_tokens / output_tokens / cache_read_tokens / cache_creation_tokens
• tokenjam.cost_usd (from kwargs['response_cost'] or response._hidden_params['response_cost'])
• gen_ai.agent.id / gen_ai.conversation.id (from kwargs['metadata']['tj_agent_id' | 'tj_session_id'])
• Status code (OK / ERROR) and message on failure

Design notes

• Non-blocking: every error path is caught and logged at debug — the method never raises into the caller's request path. Matches the idiom expected by LiteLLM, Langfuse, Helicone, etc.
• No SDK initialization required: the client is a plain HTTP POSTer with no dependency on bootstrap/ensure_initialised() or the OTel TracerProvider, so it can be safely instantiated from inside LiteLLM without side effects on the host process.
• Endpoint normalization: accepts either the server base URL or the full /api/v1/spans path.
• For in-process tokenjam users, patch_litellm() remains the preferred path — it produces equivalent spans via the OTel pipeline.

Files

• tokenjam/sdk/client.py (new, 230 lines) — TokenJamClient + pure-function payload builders
• tokenjam/sdk/__init__.py — re-export TokenJamClient
• tests/unit/test_litellm_client.py (new, 10 tests) — covers payload shape, provider/model fallbacks, dict vs pydantic usage, error status, auth header, swallowed connection / build errors
• pyproject.toml, sdk-ts/package.json, sdk-ts/package-lock.json — bump to 0.2.2

Test plan

• `pytest tests/unit/ tests/synthetic/` — 316 passed
• `ruff check` clean on new files
• `mypy tokenjam/sdk/client.py` clean
• CI green
• After merge: create v0.2.2 GitHub release to trigger PyPI publish
• Then proceed with upstream BerriAI/litellm PR

🤖 Generated with Claude Code