fix: rebuild codebase-review fixes on main without regressions (#127)

Rebuilds the full-codebase-review changes on top of current main. Brings the
genuine fixes while preserving everything that landed on main after the review
branch was cut: completion-aware call(wait=True)/CallResult, built-in
transfer_call/end_call injection in pipeline mode, Plivo carrier, and the
modern is_ws_alive websocket-liveness check. Closes the one-sided Python<->TS
parity gaps the review left open and corrects defects it introduced.

The default Agent.model is aligned across both SDKs to gpt-realtime-mini
(matching the TypeScript adapter default and the Python agent() helper); both
names are aliases for the same model family and share the same pricing. The
realtime-pricing calibration warning now treats both as default-calibrated.
Internal review artifacts (docs/reviews/*, docs/DEVLOG.md) were intentionally
not carried over.

See CHANGELOG.md (## Unreleased) for the user-visible list.

Co-authored-by: nicolotognoni <nicolo.tognoni1@gmail.com>

Author: FrancescoRosciano

CHANGELOG.md

@@ -7,9 +7,71 @@
   live-call, and metrics panels, alongside Twilio and Telnyx
   (`dashboard-app/`). Also added `"plivo"` to the PyPI / npm package keywords
   so the SDK surfaces in Plivo-related searches. (#123)
+- **`PatterConfigError` (both SDKs).** Added to the TypeScript error taxonomy
+  (`errors.ts`) to match Python's `PatterConfigError`, and exported from the
+  Python package root — raised for invalid SDK configuration.
+- **New Python package exports**: `resample_pcm`, `define_tool`, `DTMF_EVENTS`,
+  barge-in strategy helpers, and pricing constants (`PricingUnit`,
+  `PRICING_VERSION`, `PRICING_LAST_UPDATED`) are now importable from `getpatter`
+  for parity with the TypeScript surface.
+
+### Changed
+
+- **Public config collections are now immutable.** `Agent.tools` / `guardrails`
+  / `text_transforms` / `mcp_servers` and `Guardrail.blocked_terms` are tuples
+  (Python, `frozen=True`) / `readonly` arrays (TypeScript). Code comparing these
+  against a literal `list`/array must compare against a tuple / readonly array.
+- **Default `Agent.model` is now `gpt-realtime-mini`** (was
+  `gpt-4o-mini-realtime-preview`), aligning the Python dataclass default with the
+  TypeScript adapter default. Both are aliases for the same OpenAI Realtime model
+  family and share the same pricing row, so cost is unaffected — this removes a
+  Python↔TypeScript default mismatch and an internal Python inconsistency (the
+  `agent()` helper already defaulted to `gpt-realtime-mini`).
 
 ### Fixed
 
+- **DeepFilterNet noise suppression was a silent no-op.** The resampler could
+  never reach the 48 kHz the model requires, the error was swallowed, and raw
+  audio passed through unsuppressed. Now resampled correctly
+  (`providers/deepfilternet_filter.py`, `providers/deepfilternet-filter.ts`).
+- **Parallel tool-calls returned a 400 from Anthropic and Gemini.** Multiple
+  tool results in one turn produced consecutive same-role messages both
+  providers reject; they are now merged into a single turn in both SDKs
+  (`providers/anthropic_llm.py` / `anthropic-llm.ts`, `providers/google_llm.py`
+  / `google-llm.ts`).
+- **Anthropic prompt-cache savings were not billed in TypeScript.** The TS
+  Anthropic provider now emits cache read/write token counts on its usage chunk
+  (parity with Python), so cached-prompt cost reductions are reflected in the
+  metrics (`providers/anthropic-llm.ts`).
+- **MCP client URLs were not SSRF-guarded.** Both SDKs now validate the MCP
+  server URL (blocking link-local, loopback, and private ranges) before opening
+  the transport (`tools/mcp_client.py`, `tools/mcp-client.ts`).
+- **AMD callbacks were clobbered across concurrent outbound calls.** The single
+  answering-machine-detection slot is now a per-call map in both SDKs, keyed by
+  call SID (`server.py` / `server.ts`, `client.ts`).
+- **`FallbackLLMProvider` crashed in pipeline mode (Python).** `stream()` now
+  accepts and forwards `cancel_event`, matching how the pipeline invokes every
+  provider (`services/fallback_provider.py`).
+- **Blocking model/inference calls ran on the event loop.** Krisp and Whisper
+  now offload to `asyncio.to_thread` so audio is not dropped
+  (`providers/krisp_filter.py`, `providers/whisper_stt.py`).
+- **WebSocket reads could hang indefinitely.** Added read timeouts and surfaced
+  background-task exceptions across providers and the remote-message path
+  (`services/remote_message.py` / `remote-message.ts`).
+- **ElevenLabs API key was a publicly readable field.** It is now private in
+  both SDKs (`providers/elevenlabs_ws_tts.py`, `providers/elevenlabs-ws-tts.ts`).
+- **Async done-callbacks raised `CancelledError` on normal shutdown.** They now
+  guard `cancelled()` before reading `.exception()` (`observability/event_bus.py`,
+  `providers/cartesia_stt.py`, `dashboard/store.py`).
+- **Pipeline could accumulate orphaned conversation-history turns.** The
+  speculative user turn is popped on the no-handler / vetoed path
+  (`stream_handler.py`).
+- **Py↔TS parity drift fixed**: `DTMF_EVENTS` ordering, the OpenAI TTS default
+  model, and the end-of-utterance metric emit guard now match across SDKs.
+- **Telnyx STT logs now use the `getpatter.*` namespace.**
+  `providers/telnyx_stt.py` logged under the stale `patter.providers.telnyx_stt`
+  namespace; aligned to `getpatter.providers.telnyx_stt` like every other module
+  so `getpatter.*` log-level filters capture it.
 - **Pipeline-mode turns after the first now record per-turn metrics and
   broadcast the live dashboard transcript.** `anchorUserSpeechStart()` re-opened
   a turn (set `_turnStart`) without clearing the `_turnAlreadyClosed` guard, so

dashboard-app/src/components/CallTable.tsx

@@ -12,53 +12,53 @@ import {
 } from './icons';
 
 export interface CallCost {
-  telco?: number;
-  llm?: number;
-  stt?: number;
-  tts?: number;
+  readonly telco?: number;
+  readonly llm?: number;
+  readonly stt?: number;
+  readonly tts?: number;
   /** @deprecated Sum of stt+tts kept for legacy aggregate-spend callers. */
-  sttTts?: number;
-  cached?: number;
-  total?: number;
+  readonly sttTts?: number;
+  readonly cached?: number;
+  readonly total?: number;
 }
 
 export type CallMode = 'realtime' | 'pipeline' | 'convai' | 'unknown';
 
 export interface Call {
-  id: string;
-  status: 'live' | 'ended' | 'no-answer' | 'queued' | 'fail';
-  direction: 'inbound' | 'outbound';
-  from: string;
-  to: string;
-  carrier: CallCarrier;
+  readonly id: string;
+  readonly status: 'live' | 'ended' | 'no-answer' | 'queued' | 'fail';
+  readonly direction: 'inbound' | 'outbound';
+  readonly from: string;
+  readonly to: string;
+  readonly carrier: CallCarrier;
   /** ms epoch — set for any call we know started, live or ended. */
-  startedAtMs?: number;
-  durationStart?: number;
-  duration?: number;
-  latencyP95?: number;
-  latencyP50?: number;
+  readonly startedAtMs?: number;
+  readonly durationStart?: number;
+  readonly duration?: number;
+  readonly latencyP95?: number;
+  readonly latencyP50?: number;
   /** avg(llm_ms) across this call's turns — for the waterfall llm bar. */
-  llmAvg?: number;
-  sttAvg?: number;
-  ttsAvg?: number;
+  readonly llmAvg?: number;
+  readonly sttAvg?: number;
+  readonly ttsAvg?: number;
   /** Number of completed turns. p50/p95 are statistically meaningful only when this is >= 5. */
-  turnCount?: number;
+  readonly turnCount?: number;
   /** p50 of agent_response_ms (wait time after user stops speaking) — user-perceived latency. */
-  agentResponseP50?: number;
+  readonly agentResponseP50?: number;
   /** p95 of agent_response_ms — user-perceived latency outlier. */
-  agentResponseP95?: number;
-  cost: CallCost;
-  agent?: string;
-  model?: string;
-  mode?: CallMode;
-  sttProvider?: string;
-  ttsProvider?: string;
+  readonly agentResponseP95?: number;
+  readonly cost: CallCost;
+  readonly agent?: string;
+  readonly model?: string;
+  readonly mode?: CallMode;
+  readonly sttProvider?: string;
+  readonly ttsProvider?: string;
   /** Model identifier within the provider, e.g. "ink-whisper", "eleven_flash_v2_5", "gpt-oss-120b". */
-  sttModel?: string;
-  ttsModel?: string;
-  llmModel?: string;
-  transcriptKey?: string;
-  endedAgo?: number;
+  readonly sttModel?: string;
+  readonly ttsModel?: string;
+  readonly llmModel?: string;
+  readonly transcriptKey?: string;
+  readonly endedAgo?: number;
 }
 
 interface CallRowProps {

dashboard-app/src/components/CostPanel.tsx

@@ -18,7 +18,7 @@ function titleCase(s: string): string {
 }
 
 export function CostPanel({ call }: CostPanelProps) {
-  if (!call || !call.cost?.telco) return null;
+  if (!call || call.cost?.telco == null) return null;
 
   const c = call.cost;
   const telco = c.telco ?? 0;

dashboard-app/src/components/LatencyPanel.tsx

@@ -3,6 +3,9 @@ import type { Call } from './CallTable';
 import { fmtDuration, fmtPhone } from './format';
 import { IconForward, IconHangup, IconMic, IconRecord } from './icons';
 import { CARRIERS } from '../lib/mappers';
+import type { TranscriptTurn } from '../lib/mappers';
+
+export type { TranscriptTurn };
 
 interface LiveDurationProps {
   start: number;
@@ -17,14 +20,6 @@ function LiveDuration({ start }: LiveDurationProps) {
   return <>{fmtDuration((Date.now() - start) / 1000)}</>;
 }
 
-export interface TranscriptTurn {
-  who: 'user' | 'bot' | 'tool';
-  txt?: string;
-  args?: Record<string, string | number>;
-  typing?: boolean;
-  lat?: { stt?: number; llm?: number; tts?: number; total?: number };
-}
-
 export interface LiveCallPanelProps {
   call: Call | null;
   transcript: TranscriptTurn[];

dashboard-app/src/components/LiveCallPanel.tsx

@@ -83,14 +83,14 @@ function LatencyView({ call }: { call: Call }) {
           <div className="latbox">
             <div className="l">end-to-end p50</div>
             <div className="v">
-              {showPctRt ? p50 || '—' : '—'}
+              {showPctRt ? p50 ?? '—' : '—'}
               {showPctRt && <span className="u">ms</span>}
             </div>
           </div>
           <div className={'latbox' + (showPctRt && p95 > 600 ? ' warn' : '')}>
             <div className="l">end-to-end p95</div>
             <div className="v">
-              {showPctRt ? p95 || '—' : '—'}
+              {showPctRt ? p95 ?? '—' : '—'}
               {showPctRt && <span className="u">ms</span>}
             </div>
           </div>