release: 0.6.2 — GA Realtime adapter + 14-bug fix wave + dashboard hardening#104
Merged
Conversation
…to v4 The `bandit` job in `.github/workflows/audit.yml` was failing on the `github/codeql-action/upload-sarif` step with `Resource not accessible by integration`, leaving Bandit findings out of the GitHub Security tab. Root cause: the job inherited the repo-default read-only GITHUB_TOKEN permissions, but the SARIF upload requires `security-events: write`. Add an explicit per-job `permissions:` block (`contents: read`, `security-events: write`) and bump `codeql-action/upload-sarif` from @V3 to @v4 to clear the December 2026 deprecation warning at the same time.
…d on same turn
Pipeline `transcript.jsonl` rows after a barge-in carried an empty
`user_text` even when the user had clearly spoken. Root cause was a
race between the two turn-close paths:
1. The VAD-driven barge-in fires `recordTurnInterrupted` synchronously
inside the audio handler. `_resetTurnState` clears `_turnUserText`.
2. The in-flight pipeline LLM stream keeps unwinding on its own task
(we already abort it via `llmAbort`, but `processTranscript` only
unwinds back to its top-level `recordTurnComplete` call after the
`for await` loop exits).
3. That late `recordTurnComplete` pushed a SECOND turn for the same
logical exchange — `agent_text=<partial cancelled text>`,
`user_text=""`. The first interrupted turn was emitted to the event
bus (correctly) but only `recordTurnComplete` is forwarded to the
`transcript.jsonl` writer, so the operator-facing JSON showed the
phantom row.
Fix: both SDKs gain a `_turnAlreadyClosed` / `_turn_already_closed`
guard flipped inside `recordTurnInterrupted` (after the existing
`_resetTurnState`). `recordTurnComplete` now returns `null` / `None`
when the flag is set, until the next `startTurn` / `start_turn`
re-arms the accumulator. `emitTurnMetrics` / `_emit_turn_metrics`
were already null-safe, so the late call becomes a silent no-op
end-to-end.
Regression tests pin the bargein → llmAbort → late-complete ordering
and the start_turn re-arm path in both libraries.
See `patter-sdk-acceptance/BUGS.md` (2026-05-05 entry — was tagged
NEEDS RE-VERIFICATION; the code-path analysis stands and the fix is
additive + behind a flag that defaults to the existing behaviour for
turns that never barge-in, so it is safe to land ahead of a fresh
matrix run).
Every pipeline acceptance run hit a 1.5-2.5 s p95 on the first turn because the TTS first-byte latency (200-700 ms cold) was serialised with the carrier's media-start event. Pre-rendering the greeting during the ringing window and streaming the cached buffer at pickup collapses that to a single Buffer.copy / bytes write — first-turn p95 returns to the same band as subsequent turns. Trade-off: paying the TTS bill on calls that ring and never answer (~$0.001-$0.005 each depending on TTS provider). Opt out with ``prewarmFirstMessage: false`` (TS) / ``prewarm_first_message=False`` (Py) for very high-volume outbound where un-answered TTS spend matters. Changes: - libraries/python/getpatter/models.py — Agent dataclass field default flipped from False to True; docstring updated. - libraries/typescript/src/client.ts — Patter.agent() factory now defaults the field to true when provider === 'pipeline' and the caller didn't pass an explicit value. Realtime / ConvAI modes unchanged (those handlers never consume the prewarm cache). - libraries/typescript/src/types.ts — docstring updated. Tests added in both SDKs covering the new default and the opt-out path.
…low per-chunk sleep sendPacedFirstMessageBytes (TS) / _send_paced_first_message_bytes (Py) were pacing each prewarm chunk with setTimeout / asyncio.sleep of one chunk-equivalent of playout time (~40 ms for the 1280-byte chunk). Combined with the waitForMarkWindow back-pressure and JavaScript / asyncio timer jitter, effective delivery dropped BELOW Twilio's 8 kHz playout clock on the typical 2-4 s prewarmed greeting, producing repeated carrier-side underruns. The caller heard the firstMessage as slow, gravelly, intermittent — even though `p95 wait` reported 0 ms (the prewarm cache hit was correct; it was the downstream pacing that was broken). Twilio's Media Streams docs (websocket-messages) explicitly state media messages "of any size" are "buffered and played in the order received" by the carrier-side media server — the carrier is the source of truth for the 8 kHz playout clock, not our send loop. The live-TTS streaming path (synthesizeSentence + first-message live fallback) has always bursted chunks back-to-back without any sleep, and has always worked. Bring the prewarm path in line with the live path: drop the per-chunk sleep + the burst-vs-paced switch (initialFillComplete). Per-chunk marks are still emitted so a barge-in's sendClear keeps fine-grained granularity to cut, and the existing PREWARM_CHUNK_BYTES (1280 B ≈ 40 ms @ 16 kHz PCM16) bounds the worst-case mid-flush amount that sendClear has to drop. Cleaned up the now-unused PCM16_16K_BYTES_PER_MS constant in both SDKs.
Surfaced by a parallel review pass (code-reviewer + sdk-parity + docs-sync + code-simplifier agents on the release/0.6.2 diff). 1. **Parity fix — Python ``_reset_turn_state`` now clears ``_turn_committed_mono``** (libraries/python/getpatter/services/metrics.py). TS ``_resetTurnState`` already clears the equivalent field on every turn close (metrics.ts), but Python only cleared it inside ``start_turn`` and ``record_turn_interrupted``. After a cleanly completed turn the field remained set until the next ``start_turn``; ``anchor_user_speech_start`` (which guards on ``self._turn_committed_mono is not None``) would falsely no-op on a VAD ``speech_start`` arriving between ``record_turn_complete`` and the next ``start_turn`` on back-to-back turns. Single-line additive fix; no behaviour change on already-aligned flows. 2. **Docs — document the prewarm_first_message / prewarmFirstMessage default flip** (docs/python-sdk/agents.mdx, docs/typescript-sdk/agents.mdx). Added a row in each AgentOptions table calling out the 0.6.2 default change and the opt-out (``prewarm_first_message=False`` / ``prewarmFirstMessage: false``), plus a "Pre-warming the first message" narrative section in the Python page explaining the latency vs un-answered-TTS trade-off. 3. **Inventory — appended ``Prewarm first message`` row to ``patter-assets/patter_sdk_features.xlsx``** (out-of-tree, not in this commit) so the daily docs-drift cron sees the feature as covered in both SDKs. Skipped findings (with rationale): - "Python prewarm default is unconditional vs TS provider-gated" — flagged HIGH by code-reviewer, classified PASS-WITH-NOTE by sdk-parity. The Python ``_spawn_prewarm_first_message`` WARN guard makes end-to-end behaviour identical, and moving the default off the dataclass would break users who instantiate ``Agent(...)`` directly. Documented divergence accepted. - "TS ``sendMarkAwaitable()`` not awaited in prewarm loop" — Python ``await`` is needed because ``send_mark`` is async; TS ``ws.send`` is sync void. The returned ``Promise<void> | null`` is the ACK future consumed later by ``onMark`` / ``drainPendingMarks``, NOT a write await. Behaviourally equivalent. - "Missing Py test for non-pipeline prewarm default" — coverage already exists at libraries/python/tests/test_prewarm.py:702 and :725 (``test_prewarm_skipped_for_realtime_provider`` and ``..._for_convai_provider``). - "recordTurnInterrupted not guarded against late recordTurnComplete" inverse race — MEDIUM latent trap but no current caller can trigger it (the existing ``interrupted`` flag in ``processTranscript`` guards the recordTurnComplete call site). Tracked for a future hardening pass. Tests: Python 58/58 (test_metrics.py + test_prewarm.py), TypeScript lint clean.
Two follow-ups on the parallel quality-gate review the user requested
should NOT be skipped:
1. HIGH (code-reviewer): Python prewarm_first_message default flipped
the dataclass to True unconditionally, while TypeScript only applied
True at the Patter.agent() factory when provider==='pipeline'. End-
to-end behaviour matched (Python WARN guard suppressed prewarm on
non-pipeline modes), but users inspecting agent.prewarm_first_message
programmatically saw the asymmetry — Realtime/ConvAI Python agents
advertised True even though the cache was never consumed.
Fix:
- libraries/python/getpatter/models.py — dataclass default restored
to False (back-compat for direct Agent(...) construction).
- libraries/python/getpatter/client.py — Patter.agent() factory now
accepts an explicit prewarm_first_message kwarg, and when omitted
applies True iff provider == 'pipeline'. Exact mirror of the
TypeScript factory in client.ts.
- libraries/python/tests/test_prewarm.py — three new tests pin the
factory-level behaviour (pipeline → True, realtime → False,
explicit kwarg always wins). The existing
test_default_prewarm_flag_is_true was updated to assert the new
dataclass default (False).
2. MEDIUM (code-reviewer): the race guard added in commit 7bc143a was
one-directional — late recordTurnComplete after
recordTurnInterrupted was correctly dropped, but the inverse ordering
(late interrupt after a completed turn) could still overwrite an
emitted turn. No current caller path produces that ordering, but a
future refactor reordering bargein vs LLM-unwind could.
Fix: both recordTurnComplete (libraries/python/getpatter/services/
metrics.py + libraries/typescript/src/metrics.ts) and
recordTurnInterrupted now read AND write _turn_already_closed /
_turnAlreadyClosed — bidirectional symmetry. Regression tests added
in libraries/python/tests/test_metrics.py and libraries/typescript/
tests/metrics.test.ts.
CHANGELOG entry on the prewarm default flip consolidated into a single
'Changed' section reflecting the final factory-level wiring.
Tests: Python 62/62 (test_metrics.py + test_prewarm.py), TypeScript
37/37 (metrics.test.ts + prewarm.test.ts), lint clean.
Bundled fix wave landed after live PSTN validation across outbound and
inbound flows on the OpenAI Realtime GA path (gpt-realtime-2). Every
behaviour change is mirrored in both SDKs and verified via sdk-parity.
OpenAI Realtime GA (gpt-realtime-2)
- New ``OpenAIRealtime2`` engine + ``OpenAIRealtime2Adapter`` (Python).
Speaks the GA ``session.update`` shape (``session.type = "realtime"``,
``output_modalities``, nested ``audio.{input,output}``) and bidirectionally
transcodes mulaw 8 kHz ↔ PCM 24 kHz because the GA audio engine silently
drops mulaw even though the protocol accepts ``audio/pcmu``.
- ``turn_detection.threshold`` raised 0.1 → 0.5 to stop the runaway loop
where carrier-loopback echo of the agent's own audio kept tripping the
server VAD and auto-creating new responses.
- ``turn_detection.create_response: false`` + ``interrupt_response: false``.
Patter now drives ``response.create`` explicitly via ``request_response()``
after the hallucination filter accepts the user transcript, so a
Whisper-on-silence hallucination ("Thank you for watching.", "[music]")
no longer materialises as a phantom assistant turn.
- ``_STT_HALLUCINATIONS`` extended with the 15 most common Whisper
YouTube-caption fallback phrases; the filter is now applied to the
Realtime ``transcript_input`` event before LLM commit.
Prewarm + adoption
- Liveness check rewritten to handle current ``websockets`` lib (``state``
enum + ``close_code`` checks; legacy ``closed`` fallback). The previous
``getattr(ws, "closed", True)`` defaulted to "dead" on the new client and
silently aborted every adoption.
- Application-level keepalive on the parked GA Realtime WS (``session.update``
every 3 s + WS PING every 4 s) — empirically OpenAI's GA edge closes idle
sockets within ~6-7 s, so a single PING was never reaching the wire before
pickup on cellular ringing windows.
- ``adopt_websocket`` cancels the parked keepalive task before the live
adapter starts so the heartbeat doesn't race ``input_audio_buffer.append``.
Barge-in
- Realtime ``speech_started`` now consults ``_current_response_first_audio_at``
on the adapter (proxy for "agent is mid-turn") and applies the same
anti-flicker gate the pipeline mode uses. Without this the firstMessage
was repeatedly truncated by the loopback echo VAD.
- ``cancel_response`` is now a no-op when no item is in flight — eliminates
the ``response_cancel_not_active`` ERROR spam every phantom VAD trigger
emitted.
Dashboard + persistence
- Persistence default flipped from opt-in to ON. ``persist=None`` now
resolves to the platform user-data dir (``~/Library/Application
Support/patter`` on macOS / XDG data dir on Linux / ``%LOCALAPPDATA%``
on Windows). Set ``persist=False`` to opt out.
- ``log_call_start`` / ``logCallStart`` now persists ``direction`` in
``metadata.json``. Hydrated outbound calls were rendering as inbound
(default fallback) and ``pickPhoneNumber`` ended up returning the callee
(caller's personal number) instead of the Patter number in the topbar.
- ``PATTER_LOG_REDACT_PHONE`` default flipped ``mask`` → ``full``. The UI
reveal toggle has no source data when the on-disk record is already
masked, so storing raw is required for the toggle to actually do
something. ``~/Library/Application Support/patter`` is user-private.
- ``record_call_end`` / ``recordCallEnd`` now preserves the live ``turns``
array and falls back to active/existing transcript when the SDK's
end-of-call snapshot is empty. Mirrors TS fix that was never ported.
- Python ``hydrate()`` now backfills the flat ``transcript`` from the
sibling ``transcript.jsonl`` when ``metadata.json`` has no transcript
array (the JSONL is the authoritative per-turn record). Parity with TS
``loadTranscriptJsonl``.
- New ``aggregates.sdk_version`` field surfaces the runtime package
version (Python ``getpatter.__version__`` / TS ``package.json`` read at
runtime via the dist-relative path). Dashboard SPA reads it from the
aggregates payload instead of an inline constant.
- Standalone dashboard (``patter dashboard``) now sees outbound dials in
real time: ``client`` fires ``notify_dashboard`` with ``status="initiated"``
alongside ``record_call_initiated``, and the standalone ``cli`` ingest
handler routes that status to ``record_call_initiated`` instead of
treating every payload as ``record_call_start``.
Inbound carrier metadata
- Twilio Media Streams strips the query string from ``<Stream url=...>``
before opening the WS, so the inbound bridge has been reading empty
caller / callee since forever. ``generate_stream_twiml`` now accepts
optional ``parameters`` and emits ``<Parameter name=... value=.../>``
children of ``<Stream>``; ``twilio_stream_bridge`` falls back to
``start.customParameters`` when WS query params are empty.
Test scripts
- Acceptance scripts under ``releases/0.6.1/python`` (in the personal
acceptance repo, not in this repo) drove the live validation. ``inbound.py``
and ``outbound_amd_ringtimeout.py`` updated locally to surface ``INFO``
logging for the new prewarm / hallucination diagnostics.
- Bump Python (``__init__.py`` + ``pyproject.toml``) and TypeScript (``package.json``) to ``0.6.2``. - Promote ``## Unreleased`` to ``## 0.6.2 (2026-05-25)`` in CHANGELOG. - Refresh ``docs/github-banner.png`` with the new branded artwork (Agent / Patter stack) used by the README and the GitHub social preview. Bundles the 14-bug fix wave validated live in 0fc4615 (GA Realtime adapter, prewarm + adoption hardening, dashboard persistence, inbound caller/callee via ``<Parameter>``, Whisper hallucination filter, deferred ``response.create`` + ``request_response()``, version auto-derive, persist default ON, phone-redact default ``full``, direction in ``metadata.json``, ring-buffer turns preservation, JSONL transcript backfill, standalone-dashboard ``call_initiated`` relay).
|
Preview deployment for your docs. Learn more about Mintlify Previews.
💡 Tip: Enable Workflows to automatically generate PRs for you. |
Updates the Python + TypeScript regression suites to reflect the public behaviour changes that landed in 0fc4615 / f16bda0. No source-code behaviour changes — only test scaffolding + one defensive ``getattr`` on ``_adapter_cls.__name__`` so the debug logger doesn't trip the ``MagicMock`` patch surface used by ``test_local_mode`` / ``test_validation_guardrails``. Python (8 fixes) - ``test_twilio_handler.test_stream_url_contains_caller_param`` / ``...callee_param`` / ``test_local_mode.test_twilio_webhook_handler_url`` now assert caller/callee travel as TwiML ``<Parameter>`` children of ``<Stream>`` (the ``parameters`` kwarg on ``generate_stream_twiml``) instead of query-string params on the WS URL — Twilio strips the latter before the WebSocket handshake. - ``test_providers_io_unit.test_cancel_response_sends_cancel`` now seeds ``_current_response_item_id`` before calling ``cancel_response`` since the method is a documented no-op when no item is in flight. Added ``test_cancel_response_noop_when_no_item_in_flight`` to pin that contract. - ``test_local_mode.test_mark_events_sent_after_audio`` / ``test_validation_guardrails.test_guardrail_triggers_cancel_and_replacement`` / ``test_providers_unit.test_realtime_engine_forwards_reasoning_and_transcription_to_adapter`` now patch ``OpenAIRealtime2Adapter`` (the GA adapter) instead of the v1-beta class — both ``openai_realtime`` and ``openai_realtime_2`` engines route through the GA adapter after the upstream Beta API deprecation. - ``stream_handler`` debug log now wraps ``_adapter_cls.__name__`` in ``getattr`` so the three tests above (which patch the adapter class with a ``MagicMock``) don't crash on the missing dunder. TypeScript (4 fixes + 3 skips) - ``openai-realtime.cancelResponse()`` test seeds ``currentResponseItemId`` and added a ``no-op when no item in flight`` test, mirroring Python. - ``stream-handler`` barge-in gate tests aligned with the AEC-off gate raised from 100 ms to 500 ms on 2026-05-19. ``canBargeIn`` / handleBargeIn inputs bumped 50/200/400 ms → 250/700/600 ms accordingly. - ``prewarm.test`` no longer asserts ``prewarmFirstMessage === true`` by default in pipeline mode — the 2026-05-18 default-on attempt was reverted on 2026-05-19 (phantom barge-in interaction). Test now pins the opt-in semantics described in ``client.ts:536-547``. - Three ``describe`` blocks marked ``describe.skip``: ``firstMessage mark-gated pacing``, ``cleanup drains pending firstMessage marks``, ``firstMessage mark counter resets across sends + on cleanup`` — the mark-window pacing plumbing they exercised was replaced with burst-deliver in commit 5574997 (``sendPacedFirstMessageBytes`` / ``firstMessageMarkCounter`` / ``sendMarkAwaitable`` no longer exist). Kept as ``skip`` rather than deleted to preserve the historical intent.
…me engine
Audit ran via 3 parallel agents (Python SDK accuracy, TypeScript SDK
accuracy, navigation + cross-links) cross-referencing every public
identifier, default, and behaviour described in ``docs/`` against the
0.6.2 source. 31 pages updated; 2 brand-new provider pages created so
``OpenAIRealtime2`` ships with first-class documentation in both SDKs.
New pages
- ``docs/python-sdk/providers/openai-realtime-2.mdx``
- ``docs/typescript-sdk/providers/openai-realtime-2.mdx``
Cover ``OpenAIRealtime2Adapter`` / ``OpenAIRealtime2Provider``: GA
session-config (``session.type = "realtime"``, nested
``audio.{input,output}``, ``create_response: false`` /
``interrupt_response: false``), bidirectional mulaw 8 kHz ↔ PCM 24 kHz
transcoding rationale, voice list, reasoning-effort tiers, and the
direct-adapter constructor (positional, not options-object).
- ``docs/docs.json`` adds both pages to the Engines group in their
respective SDKs.
Engines + providers
- ``OpenAIRealtime`` default model is now ``gpt-realtime-mini`` (was
documented as ``gpt-4o-mini-realtime-preview``). Voice enum widened
to include ``ash``/``ballad``/``coral``/``sage``/``verse``.
- Added ``reasoning_effort`` and ``input_audio_transcription_model``
rows to the ``OpenAIRealtime`` constructor table.
- ``ElevenLabsTTS`` default voiceId fixed: ``EXAVITQu4vr4xnSDxMaL``
(Sarah) → ``21m00Tcm4TlvDq8ikWAM`` (Rachel) — matches source default
in both SDKs.
- Provider pages add the GA Realtime VAD threshold note (0.5, not 0.1)
and the Whisper hallucination filter behaviour.
Persistence + dashboard
- ``persist`` default is now ON in both SDKs (was documented as opt-in).
Flipped narrative + tables + env var notes across ``persist.mdx``,
``call-logging.mdx``, ``configuration.mdx``, ``quickstart.mdx``,
``reference.mdx``.
- ``PATTER_LOG_REDACT_PHONE`` default ``mask`` → ``full`` across
``configuration.mdx``.
- Added ``direction`` and ``aggregates.sdk_version`` fields to the
dashboard / call-log schema docs.
Inbound carrier metadata
- ``local-mode.mdx`` + ``carrier.mdx`` (both SDKs) now correctly
describe Twilio inbound caller/callee as travelling via TwiML
``<Parameter>`` (Twilio strips URL query params before the WS
handshake). Telnyx still uses query string — distinction documented.
Call surface
- ``Patter.call()`` parameter signature updated to snake_case
(``machine_detection=True``, ``ring_timeout=25``) — fixes the
pre-0.6.2 PascalCase crash documentation.
- AMD narrative flipped to "default on" in ``features.mdx``.
- ``phone.serve()`` examples in TS docs fixed:
``phone.serve(agent)`` → ``phone.serve({ agent })`` (5 pages).
Known follow-ups out of scope for this docs audit
- Several TS docs ``import { OpenAIRealtimeModel, ... } from "getpatter"``
but the const enums live in provider files and are NOT re-exported
from ``src/index.ts``. Examples won't compile until the re-exports
are added — flagged for a separate SDK-code commit.
- TS engine wrapper still defaults ``model`` to
``"gpt-4o-mini-realtime-preview"`` (Python moved to
``"gpt-realtime-mini"`` per CHANGELOG 0.6.2). Docs now describe the
TS-side reality; parity bump is a separate SDK commit.
Inventory rows for 0.6.2 features appended to
``patter-assets/patter_sdk_features.xlsx`` (status=shipped, sdk=both):
``openai_realtime2_engine``, ``realtime_request_response_api``,
``realtime_whisper_hallucination_filter``, ``persist_default_on``,
``log_redact_phone_default_full``, ``call_metadata_direction_field``,
``aggregates_sdk_version_field``, ``dashboard_call_initiated_relay``,
``twilio_inbound_caller_callee_parameter``.
Two follow-ups surfaced by the 0.6.2 docs-accuracy audit; needed for the
docs examples to compile as written and to close the last Python↔TS
parity gap.
src/index.ts re-exports
- ``OpenAIRealtimeAudioFormat``, ``OpenAIRealtimeModel``,
``OpenAIRealtimeVADType``, ``OpenAITranscriptionModel``,
``OpenAIVoice`` from ``./providers/openai-realtime``
- ``ElevenLabsModel``, ``ElevenLabsOutputFormat`` from
``./providers/elevenlabs-tts``
- ``DeepgramModel`` from ``./providers/deepgram-stt``
- ``CartesiaTTSModel``, ``CartesiaTTSVoiceMode`` from
``./providers/cartesia-tts``
- ``RimeModel``, ``RimeAudioFormat`` from ``./providers/rime-tts``
- ``PricingUnit``, ``PRICING_VERSION``, ``PRICING_LAST_UPDATED`` +
``PricingUnitValue`` / ``ModelPricing`` types from ``./pricing``
(``ProviderPricing`` was already exported earlier in the file)
Pre-fix the docs (``tts.mdx``, ``stt.mdx``, ``metrics.mdx``,
``providers/openai-realtime.mdx``, ``providers/elevenlabs-tts.mdx``)
showed ``import { OpenAIRealtimeModel, ... } from "getpatter"`` —
those examples now actually compile.
Engine default model
- ``engines/openai.ts`` ``Realtime.model`` default flipped from
``"gpt-4o-mini-realtime-preview"`` to ``"gpt-realtime-mini"`` for
parity with the Python SDK (which moved on 2026-05). The legacy
preview model still works when passed explicitly; the GA wave
recommends ``gpt-realtime-mini`` (or ``gpt-realtime-2`` via the
``OpenAIRealtime2`` engine for the flagship). Docstring updated to
reflect the bump rationale.
Lint clean, 1513 tests pass.
This was referenced May 25, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
OpenAIRealtime2engine + adapter (Python parity with TS) speaking the GA Realtime API (gpt-realtime-2), with bidirectional mulaw 8 kHz ↔ PCM 24 kHz transcoding for Twilio / Telnyx.response.create, and 9 more (full list below).docs/github-banner.png).Implementation
libraries/python/getpatter/__init__.py,libraries/python/pyproject.toml,libraries/typescript/package.jsonall →0.6.2.## Unreleasedpromoted to## 0.6.2 (2026-05-25)inCHANGELOG.md.libraries/python/getpatter/engines/openai_realtime_2.py,libraries/python/getpatter/providers/openai_realtime_2.py,libraries/python/tests/unit/test_twilio_adapter_snake_case_kwargs.py.14 bug-fix wave (all parity Python↔TS via
sdk-parityagent)cancel_responseno-op when no item in flight (eliminatesresponse_cancel_not_activespam)record_call_endpreserves liveturns+ falls back to active/existing transcript~/Library/Application Support/patteron macOS, XDG dir on Linux,%LOCALAPPDATA%on Windows)hydrate()backfills transcript fromtranscript.jsonl(TS already did)call_initiatedin real time vianotify_dashboard+ cli ingest routingdirectionpersisted inmetadata.json(fixes outbound rendering as inbound after hydrate)PATTER_LOG_REDACT_PHONEdefault flippedmask→fullso the UI reveal toggle worksaggregates.sdk_version(Python__version__, TS viapackage.jsonruntime read)<Parameter>children of<Stream>(Twilio strips query-string params)transcript_input("Thank you for watching.", "[music]", etc.)turn_detection.create_response: false+interrupt_response: false+ newrequest_response()driven from stream-handler after the filter accepts the transcriptBreaking change?
Two opt-out flips with safe defaults:
persist=Nonenow defaults to ON (was: off unlessPATTER_LOG_DIRwas set). Migration: passpersist=Falseto opt back into the old ephemeral-only behaviour.PATTER_LOG_REDACT_PHONEdefault flippedmask→full. Migration: setPATTER_LOG_REDACT_PHONE=maskfor setups that ship logs off-host.Both flips are documented in
CHANGELOG.mdunder the 0.6.2 entry.No API removals or renames. Existing code continues to work without changes.
Test plan
pytest tests/(regression tests intest_metrics.py,test_prewarm.py,test_client_unit.py,test_stream_handler_unit.py,test_twilio_status_and_ring_timeout.py,test_twilio_adapter_snake_case_kwargs.py)npm test+npm run lint+npm run buildsdk-parityagent run for every fix in the wave — verified parity for items 1-14ms=0confirmed+1659...↔•••4527Docs updates
CHANGELOG.md0.6.2 block (already populated during the fix wave).docs/github-banner.pngrefreshed with new branded artwork.docs/python-sdk/anddocs/typescript-sdk/already touched in earlier commits on the branch (engines, agents, carrier, providers/elevenlabs-tts, providers/silero-vad).