solution · Revision 1
Recommended action: Use a form that works on both: attach a suffix ('sed -i.bak ... file && rm file.bak'), or avoid -i (write to a temp file and mv), or install GNU sed (gsed) on macOS and call it explicitly.
solution · Revision 1
Recommended action: Fall back to the default HTTP Responses transport; confirm model/account support for the websocket mode; handle the reused-connection case with application-level idempotency before retrying. Option: Switch to the HTTP Responses transport [evidence: official_re
solution · Revision 1
Recommended action: Close other installer windows, wait for AV scan, delete %USERPROFILE%\.claude\downloads, rerun installer. Option: Clear downloads and rerun [evidence: official_recommended_action] Applies when: see problem Steps: 1. Remove-Item -Recurse -Force "$env:USERPROFIL
solution · Revision 1
Recommended action: Increase max_wait_seconds or poll file status yourself; check the file's status/status_details for 'error' before retrying the upload. Option: Extend or replace the wait [evidence: documented_workaround] Steps: 1. client.files.wait_for_processing(id, max_wait_
solution · Revision 1
Recommended action: Force LF for shell scripts via .gitattributes ('*.sh text eol=lf'), run 'git add --renormalize .', commit, and rebuild the image. Option: Pin LF for scripts with .gitattributes and renormalize [evidence: official_recommended_action] Applies when: Any repo whos
solution · Revision 1
Recommended action: Add NSAppleEventsUsageDescription to the launching app's Info.plist so macOS prompts; grant in System Settings > Privacy & Security > Automation; if mapping is broken reinstall the host app properly (tccutil reset AppleEvents alone may not help). Option: Make
solution · Revision 1
Recommended action: Pass a base64 data URL, or an object { id: 'file-...' } for uploaded files; use the Responses model for URL file inputs. Option: Send file as data URL or { id } [evidence: documented_workaround] Steps: 1. Upload
solution · Revision 1
Recommended action: Call lane.get_final_response() (or recv until terminal) before the next create; use separate named lanes for concurrent responses; leave max_response_bytes None unless you need the cap. Option: Serialize per lane or use more lanes [evidence: official_recommend
solution · Revision 1
Recommended action: Identify which package/registry: compare the lockfile 'resolved' URL and 'integrity' with what the configured registry serves; regenerate the lockfile entry against the registry CI actually uses (or align the registry config) and investigate unexpected tarball
solution · Revision 1
Recommended action: Configure the gateway to forward anthropic-beta (feature pass-through); otherwise set CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 before launching. Option: Forward anthropic-beta or disable experimental betas [evidence: official_recommended_action] Applies when:
solution · Revision 1
Recommended action: Either drop strict: false and keep the Zod schema, or convert parameters to a raw JSON Schema object if you need non-strict (fuzzy) arguments. Option: Use raw JSON Schema for non-strict tools [evidence: official_recommended_action] Steps: 1. Replace the
solution · Revision 1
Recommended action: Upgrade to pydantic>=2 (check pip show pydantic / dependency pins), or wrap the type in a BaseModel subclass. Option: Use Pydantic v2 or a BaseModel [evidence: documented_workaround] Steps: 1. pip install -U 'pydantic>=2' 2. or class Out(BaseModel): items:
solution · Revision 1
Recommended action: On servers pass transport: 'websocket' (or an OpenAIRealtimeWebSocket instance); in React Native implement a custom RealtimeTransportLayer backed by react-native-webrtc; on Cloudflare Workers use CloudflareRealtimeTransportLayer. Option: Select the WebSocket t
solution · Revision 1
Recommended action: Treat as policy, not a bug: check repository cache policy for untrusted events and ACTIONS_CACHE_MODE; populate caches from trusted runs (e.g. default-branch pushes) so untrusted runs can restore. Option: Seed caches from trusted runs [evidence: documented_wor
solution · Revision 1
Recommended action: Pass an explicit timeout sized to the agent session (and idle_timeout if desired) at Sandbox.create; cancel individual commands by killing the process inside the sandbox rather than terminating the sandbox; check liveness before reuse and recreate/snapshot as
solution · Revision 1
Recommended action: Build tools with openai.pydantic_function_tool(Model) (which sets strict) or add 'strict': True to each function schema (and satisfy strict-schema rules); use chat.completions.create() if you need non-strict or custom tools. Option: Use strict tools [evidence:
solution · Revision 1
Recommended action: If you need to refresh the cache, use a key unique per run (e.g. include github.run_id) with restore-keys for the nearest match; otherwise ignore the warning when the key legitimately exists. Keep in mind the 10GB per-repo limit
solution · Revision 1
Recommended action: Pass timeout=0 (or larger) to commands.run for long commands, or run in background and poll; raise request_timeout for slow API calls; inspect TimeoutException type to tell which fired. Option: Raise command timeout [evidence: official_recommended_action] Appl
solution · Revision 1
Recommended action: Pass api_version='<supported version>' and azure_endpoint='https://<resource>.openai.azure.com' explicitly, or export OPENAI_API_VERSION and AZURE_OPENAI_ENDPOINT. For the v1 GA endpoint use the plain OpenAI client with base_url '<endpoint>/openai/v1/' instead
solution · Revision 1
Recommended action: npm install ws (and @types/ws for TS); ensure bundlers keep it as an external/runtime dependency; in browsers use the native WebSocket variants. Option: Install ws [evidence: official_recommended_action] Applies when: OpenAI Node/TypeScript SDK (openai) / Resp