fix: parity & bug-fix wave on the 0.6.6/0.6.7 features — EvalSession TS port, explicit-kwarg precedence, env-key path, telemetry chain guard

[nicolotognoni]

Summary

• Consolidation pass on the already-shipped 0.6.6/0.6.7 feature wave: close Python⇄TypeScript parity gaps, fix bugs, and harden integration — no new feature surface beyond parity.
• Biggest item: EvalSession ported to TypeScript (the last big parity gap) — a real-pipeline eval harness driving the actual StreamHandler, with chainable assertions, scripted LLM provider, and an unstubbed getpatter eval CLI.
• Plus five targeted fixes: explicit voice=/model= precedence over engine markers (Python), OPENAI_API_KEY env handling on the provider-string path (both SDKs — TS had a latent dead-call), a telemetry close/chain race guard (TS), barge_in_mode/barge_in_confirm_ms exposed on the Python agent() factory, and missing regression tests for the 0.6.7 telemetry delivery fix.

Implementation

• EvalSession TS (src/evals/): EvalSession constructs a real pipeline-mode StreamHandler and injects turns through the live-call path; fakes only at the STT/TTS/audio-sender/LLM boundary. expect() assertions, LLMJudge, EvalCase.agent/llmProvider routing, JSON/YAML suite loader, CLI eval run with exit codes 0/1/2. Field names, defaults, and report rows byte-compatible with Python. Reviewed post-port: arguments field parity, bounded await of aborted dispatch before timeout throw, crypto-based call ids, judge response guard.
• Explicit-kwarg precedence (Python): voice/model moved to sentinel None defaults so explicit values — even equal to the documented default — win over engine= marker values (mirrors TS ?? resolution). Telemetry model resolution updated consistently.
• Env-key path (both SDKs): provider="openai_realtime" with no configured key now backfills OPENAI_API_KEY from the environment into the local config so the call path actually uses it. Python previously rejected despite its error message promising env support; TS accepted at validation but dialed with an empty key.
• Telemetry (TS): the post-flush chain now stops once close() has begun (parity with Python's not _closed); added the missing regression tests in both SDKs for the 0.6.7 "event recorded during an in-flight flush" fix, using a gated real HTTP collector.
• Python factory parity: barge_in_mode / barge_in_confirm_ms as agent() keywords with validation (were dataclass-only).
• Hygiene: removed leftover temporary DIAG INFO instrumentation from the TS stream handler; scrubbed external-project references from source comments; dropped a stale DTMF TODO.
• Docs: local_recording/localRecording documented across both SDK doc trees (8 pages); runnable examples added for pause-resume barge-in, preemptive generation, and smart-turn semantic EOU (both languages) + examples index.

Breaking change?

No. Two edge-case behavior corrections, both documented under ## Unreleased → Fixed:

• explicit model="gpt-realtime-mini" + engine with a different model now runs the explicit model in Python (previously the engine's — TS already did this);
• provider="openai_realtime" with only an env key is now accepted in Python (previously rejected) and actually works in TS (previously dead call).

Test plan

• Python: pytest tests/ — 2648 passed, 8 skipped, 2 xfailed
• TypeScript: npm test (2096 passed, 9 skipped) + npm run lint + npm run build
• Cross-SDK parity suite: 10/10 — verified both with and without OPENAI_API_KEY in the env (the env-set run is what caught the env-key divergence)
• New regression tests RED-verified against reverted fixes (telemetry chain, kwarg precedence)
• E2E smoke: not required (no pipeline/handler behavior change beyond the eval harness, which is test-infrastructure)

Docs updates

• docs/python-sdk/{features,reference,local-mode,events}.mdx, docs/typescript-sdk/{features,reference,local-mode,events}.mdx (local recording)
• docs/examples/: pause-resume-barge-in.{py,ts}, preemptive-generation.{py,ts}, smart-turn-detection.{py,ts}, README.md index

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
patter-06b046ce	🟢 Ready	View Preview	Jun 11, 2026, 2:28 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.