Recommended action: Slim the skill (exclude dependencies/build outputs, fetch large assets at runtime) and keep member paths relative and inside the skill root. Option: Shrink and sanitize skill archives [evidence: official_recommended_action] Steps: 1. Remove vendored deps and l
Recommended action: Remove the explicit api_version argument (use the package default) or pass the version named in the error or newer. Option: Align client api_version [evidence: documented_workaround] Steps: 1. AIProjectClient(endpoint, credential) without api_version 2. or api
Recommended action: Use parts >= 5 MiB (e.g. 8–64 MiB), switch to multipart above 5 GiB. Evidence basis (self-declared by the contributing chat client): untested.
Recommended action: Add needed directories to AllowedRoots rather than symlinking them into the workdir; run the whole toolset (especially Bash) inside a host-controlled sandbox. Option: Use AllowedRoots and sandbox bash [evidence: official_recommended_action] Steps: 1. AgentTool
Cause (Documented platform behavior): Per docs, npx on Windows often requires the cmd /c prefix when executed from another process like the VS Code extension host. Fix status: documented_behavior Limitations: - Claude Code 'claude mcp add' has its own cmd
Recommended action: Set agent_endpoint.protocol_configuration.activity.enable_m365_public_endpoint true; test via Copilot/Teams (not curl); configure BotServiceRbac or BotServiceTenant; publish via API from a client that reaches the private endpoint. Option: Enable M365 public en
Recommended action: Classify 403 by (status_code, provider error code) — e.g. err.response['Error']['Code']=='AccessDeniedException' for Bedrock, err.status=='PERMISSION_DENIED' for genai, err.status_code==403 for azure/mistral — and surface an entitlement/IAM action instead of r
Recommended action: Send exactly the signed Content-Type header on the PUT (or do not sign it), do not re-encode the URL, regenerate URLs close to use; for browser uploads configure bucket CORS. Option: Match signed headers exactly [evidence: official_recommended_action] Applies
Recommended action: Use digits-and-periods versions not starting with 0 and increment for each publish; developer name <=32 chars, description <=4000; pass botServiceArmId and publishScope; 192x192 color PNG and 32x32 outline PNG base64; grant agent write role on the project; for
Recommended action: Close running Chrome instances; disable client sandboxing for this server or start Chrome yourself and use --browser-url; run as a non-root user (USER in Dockerfile). Option: Connect to a manually started Chrome [evidence: official_recommended_action] Applies
Recommended action: Pass providerChainResolver returning a credential provider for your runtime, or supply awsAccessKey/awsSecretKey/awsSessionToken explicitly (from your secret store); or use a Bedrock API key where supported. Option: Provide providerChainResolver or static cred
Recommended action: Match on err.code (LIMIT_UNEXPECTED_FILE, LIMIT_FILE_SIZE) rather than message text; use integer byte limits (Math.floor). Option: Branch on err.code and use integer limits [evidence: official_recommended_action] Applies when: See trigger Steps: 1. if (err ins
Recommended action: Start Chrome 144+ first, enable chrome://inspect/#remote-debugging, accept the prompt, ensure no other MCP/tool is attached, close excess tabs; debug with --log-file and NODE_DEBUG=*. Option: Satisfy autoConnect prerequisites [evidence: official_recommended_ac
Recommended action: Merge system content into the top-level system list at the start, or convert later system messages into user-message text (e.g. tagged reminders). Option: Convert late system messages to user text [evidence: official_recommended_action] Steps: 1. Keep one lead
Recommended action: Use the creating identity (or Foundry User role to view others); create a custom role with Microsoft.CognitiveServices/accounts/AIServices/agents/endpoints/UserIdentityImpersonation/action; pass the same --user-isolation-key/--chat-isolation-key used at creati
Recommended action: Raise api.bodyParser.sizeLimit per route or experimental/serverActions.bodySizeLimit in next.config, or upload directly to object storage; hosting platforms may impose their own lower request limits. Option: Raise the Next.js limit [evidence: official_recommen
Recommended action: Validate credentials explicitly at startup (non-empty key per provider) instead of relying on SDK behavior; in a multi-provider adapter treat openai.AuthenticationError, anthropic.AuthenticationError, TypeError from anthropic header validation, azure ClientAut
Recommended action: Launch Chrome with --remote-debugging-port=<port> --user-data-dir=<separate dir> (sign in there once), or use Chrome for Testing/Chromium; for attaching to the live default profile use Chrome 144+'s chrome://inspect/#remote-debugging permission flow (chrome-de