[codex] upgrade voice runtime for realtime-1.5 and responses ws

[human-bee]

Issue and User Impact

PRESENT voice runtime needed a production-grade upgrade for the February 2026 voice model/transport releases:

• adaptive session-level tiering for gpt-realtime-1.5 vs gpt-realtime-mini
• direct Responses WebSocket transport path with gpt-audio-1.5
• preserved queue/idempotency/tool envelope correctness in tool-heavy chains

Without this, model quality/tiering intent and transport improvements were inconsistent, and WS mode introduced reliability/contract regressions.

Root Cause Analysis

Primary root causes identified during implementation/review:

• Adaptive model selection was bypassed by resolver defaults injecting models.voiceRealtime as an explicit override.
• WS reply lifecycle diverged from existing reply timeout controls and could drop turns on repeated transport failures.
• Responses tool catalog/validation drifted from LiveKit tool schemas.
• Participant-aware dedupe/idempotency logic had ordering gaps in voice and client quick-apply fallbacks.
• Tool event telemetry could misattribute provider path/model under responses_ws.

Why This Fix Solves Root Cause

This change removes the adaptive bypass, aligns WS timeout behavior with existing reply controls, enforces schema-backed WS tool validation, preserves turns via controlled fallback after repeated WS failures, and normalizes participant-aware dedupe/idempotency before fingerprint/hash generation.

What Changed

Control Plane and Settings

• Added/extended model control fields:
  • models.voiceRealtimePrimary
  • models.voiceRealtimeSecondary
  • knobs.voice.realtimeModelStrategy
• Updated resolver defaults to keep models.voiceRealtime explicit-only (no implicit forcing).
• Updated model settings guidance/reference entries and marked transport/responses model knobs as env-only coverage.

Voice Runtime

• Added transport abstraction + new ResponsesWebSocketTransport (wss://api.openai.com/v1/responses).
• Added adaptive realtime selection plumbing and explicit STT model wiring for multi-participant transcription.
• Added transport-aware model/provider-path telemetry.
• Added WS tool catalog generation from the same tool context schemas.
• Added strict WS tool argument validation via runtime schemas.
• Added WS failure handling with retry and fallback to realtime session after repeated transport errors (no silent turn loss).

Queue/Dedupe/Idempotency Hardening

• Injected participant identity before voice dedupe fingerprinting and dispatch idempotency hashing.
• Updated browser quick-apply fallback idempotency hash to include participant identity.

Tests

• Added resolver.test.ts for explicit-vs-adaptive model default behavior.
• Added responses-ws-transport.test.ts for WS tool loop and continuation behavior.
• Extended existing config tests.

Backward Compatibility Analysis

• External browser tool envelope contract remains unchanged (tool_call shapes and status contract preserved).
• Queue-first orchestration and mutation idempotency semantics are preserved and hardened for participant-aware cases.
• Runtime cutover remains direct to responses_ws by default, but repeated WS failures now degrade to in-process realtime reply generation for turn preservation.
• No DB schema or migration changes.

Validation and Checks

Ran successfully:

• npm run typecheck:agent
• npm run lint (passes; repo has existing warnings)
• npm test -- src/lib/agents/realtime/voice-agent/__tests__ src/lib/agents/control-plane/*.test.ts src/components/tool-dispatcher/hooks/useToolRunner.test.tsx

Reviewer Passes and Resolved Findings

Independent reviewer lanes covered:

• correctness/regression/perf risk
• maintainability/architecture
• compatibility/contracts

Resolved findings included:

• adaptive model bypass
• WS timeout guardrail mismatch
• WS turn-loss on repeated errors
• schema-validation bypass in WS tool execution
• participant-aware dedupe/idempotency ordering gaps
• provider-path attribution drift in dedupe telemetry
• client quick-apply participant-agnostic fallback key

Remaining Risks and Follow-ups

• Direct Responses WS cutover remains operationally sensitive to env entitlement/network conditions.
• Follow-up recommended: production soak benchmark for 20+ tool-call chains and reconnect contention windows.

UI Evidence

• UI-visible change is limited to model settings/reference metadata copy and coverage labels.
• No behavior-heavy frontend interaction changes; no before/after recording attached for this PR.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
present	Building	Preview, Comment	Feb 26, 2026 9:00pm