CoPilot

Features

• Gate write tools behind an in-code approval flow in supervised mode: the agent pauses, shows the exact changes, and only executes after the user approves
• Truncate oversized tool results before they enter the model context (maxToolResultTokens)
• Add support for Opus 5 to Anthropic provider.

Fixed

• Unify the streaming and non-streaming agent loops (empty-stream retries now work on every iteration, stream errors are persisted)
• Add a wall-clock budget (maxAgentSeconds) and an optional token budget (maxTokensPerRequest) per request
• Debug exports now record the provider that was actually used for a turn
• Enforce the changeExecutionMode, changeProvider and changeModel permissions server-side instead of only hiding the UI controls
• Require the editOtherUsersChats permission to append to or compact another user's conversation
• Validate tool arguments against their JSON schema before execution
• The harness now owns the internal _siteHandle argument so the model can no longer redirect tool calls to a different site

• Add support for Opus 4.8, Sonnet 5 and Fable 5 to Anthropic provider.
• Add support for GPT 5.6 Sol and GPT 5.5 to OpenAi provider.
• Add support for Gemini 3.5 Flash to Gemini provider.

• Prevent errors in entries when no api key is configured

• Add support for opus 4.7
• Add support for enabled field in LLMify settings

• Only show default provider options that are actually configured
• Harden up provider init event
• Rename plugin name from coPilot to Copilot

• Change gitHub url from craft-coPilot to craft-co-pilot

• Initial release