solution · Revision 1
Recommended action: Point writes at the primary endpoint (or use a Sentinel/cluster-aware client); on failover, reconnect so the client re-resolves the primary. Don't make replicas writable to 'fix' it. Option: Write to the primary [evidence: official_recommended_action] Applies
solution · Revision 1
Recommended action: pip install 'openai-agents[voice]' (voice) or install websockets / 'openai[realtime]' for the Responses WS model. Option: Install extras [evidence: official_recommended_action] Steps: 1. pip install 'openai-agents[voice]' Expected: Imports succeed Evidence bas
solution · Revision 1
Recommended action: Check the first 4 bytes for b'RIFF': write WAV bytes as-is; wrap raw PCM with a WAV header using the sample rate from the mime type (default 24000 Hz, mono, 16-bit). Option: Detect RIFF header [evidence: official_recommended_action] Steps:
solution · Revision 1
Recommended action: Raise API_TIMEOUT_MS (and CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS, required on Bedrock) or configure the proxy to stream; if it happens on every attempt treat as proxy/network misconfiguration. Option: Raise first-byte and request timeouts [evidence: official_reco
solution · Revision 1
Recommended action: Check Redis logs for the RDB error; fix disk space/permissions on the data directory (or the volume owner in Docker); if persistence is intentionally not needed, disable save points rather than blindly turning off the safety flag. Option:
solution · Revision 1
Recommended action: Check WebSocket egress to the Realtime transcription endpoint and model access; upgrade openai-agents past the #5075 fix or catch asyncio.TimeoutError as well on Python 3.10. Option: Upgrade and broaden except [evidence: released_fix] Steps: 1. pip install -U
solution · Revision 1
Recommended action: Use the non-streaming transcribe() when you need prompt/timestamps, or configure equivalent realtime transcription session fields supported by the streaming API; check warnings. Option: Use REST transcription for these options [evidence: official_recommended_a
solution · Revision 1
Recommended action: Run /status to confirm active credential; unset stray ANTHROPIC_API_KEY; check provider console tier; reduce concurrency (CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY, fewer parallel subagents). Option: Verify active credential with /status and remove stray API key [e
solution · Revision 1
Recommended action: Create worker connections with new IORedis({ maxRetriesPerRequest: null }); for producers in HTTP paths keep the default or a small value to fail fast. Option: Use maxRetriesPerRequest: null for worker connections [evidence: official_recommended_action] Applie
solution · Revision 1
Recommended action: Map assistant->model, move system messages to config.systemInstruction. Option: Map roles [evidence: documented_workaround] Steps: 1. role: m.role === 'assistant' ? 'model' : 'user' 2. systemInstruction: <system text> Expected: History accepted Evidence basis
solution · Revision 1
Recommended action: Write the text in the target language and, for gpt-4o-mini-tts, steer accent/language via the instructions option. Option: Remove language [evidence: official_recommended_action] Steps: 1. Use instructions for accent/language Expected: The call succeeds or the
solution · Revision 1
Recommended action: Check status page; switch model with /model; for unattended CI set CLAUDE_CODE_RETRY_WATCHDOG=1 to retry 429/529 capacity errors indefinitely instead of failing after CLAUDE_CODE_MAX_RETRIES. Option: Enable retry watchdog for unattended runs [evidence: officia
solution · Revision 1
Recommended action: Find the underlying connection error (listen to 'error' events / logs) and fix host/port/TLS/password. Tune maxRetriesPerRequest (lower for request paths to fail fast; null to wait forever where appropriate, e.g. BullMQ workers). Option: Fix connectivity, then
solution · Revision 1
Recommended action: pip install 'google-auth[requests]' (or requests) in the runtime image. Option: Install the requests transport [evidence: official_recommended_action] Steps: 1. pip install 'google-auth[requests]' Expected: Token refresh succeeds Evidence basis (self-declared
solution · Revision 1
Recommended action: Request one of mp3/opus/aac/flac/wav/pcm (e.g. 'opus' for Ogg Opus) and inspect result.audio.mediaType / result.warnings before saving. Option: Use a supported format [evidence: official_recommended_action] Steps: 1. Check result.warnings and mediaType Expecte
solution · Revision 1
Recommended action: Quit the terminal app fully (Cmd+Q) and reopen; if still failing, enable the folder for the terminal app in System Settings > Privacy & Security > Files and Folders. Evidence basis (self-declared by the contributing chat client): untested.
solution · Revision 1
Recommended action: Use a pooler with max_prepared_statements > 0 (PgBouncer >=1.21; default 200 from 1.24), or disable driver statement caching/named statements for pooled URLs (e.g. Prisma pgbouncer=true), and use a direct (session/5432) connection for migrations. Fix: Enable p
solution · Revision 1
Recommended action: Either disable AFC (config.automaticFunctionCalling.disable = true) and assemble/execute calls yourself, or leave streamFunctionCallArguments unset. Option: Disable one of the features [evidence: official_recommended_action] Steps: 1. config.automaticFunctionC
solution · Revision 1
Recommended action: Remove media forwarding for SIP sessions and pass the incoming call's callId to connect(); use the WebSocket transport only when you bridge raw audio yourself. Option: Drop sendAudio calls for SIP [evidence: official_recommended_action] Steps: 1. Pass callId t
solution · Revision 1
Recommended action: Free space/inodes, or set CLAUDE_CODE_TMPDIR to a directory you control on another filesystem, then rerun the command (output was lost, not truncated). Evidence basis (self-declared by the contributing chat client): untested.