Add forgetConversation method to Brain that clears conversationHistory and lastLlmInputSnapshot, expose as forget_conversation() global in JavaScriptPlanner sandbox, update brain-prompt with value-first rule mandating read->return->act pattern for world/query-dependent requests to avoid acting on unresolved variables, add tests verifying forgetConversation only clears conversation state without touching ll
Move brain-prompt system prompt from inline TypeScript string to external brain-prompt.md file, implement template loading with fs.watch hot-reload in development mode, add renderTemplate helper for {{variable}} substitution to inject toolsFormatted, cache loaded template with ensureTemplateLoaded/ensureWatcher helpers to avoid repeated disk reads
Add NO_ACTION_FOLLOWUP_SOURCE_ID constant, implement queueNoActionFollowup method that schedules system_alert event with no_actions reason when planner returns zero actions during observation mode, suppress follow-up if already in follow-up chain to prevent infinite loops, update brain-prompt with feedback loop guard guidance to avoid chat->feedback->chat cycles and clarify feedback:true usage for diagnostic verification only
Add LlmInputSnapshot interface tracking systemPrompt/userMessage/messages/conversationHistory/updatedAt/attempt, store lastLlmInputSnapshot in Brain and inject into RuntimeGlobals as llmInput, expose llmInput/llmMessages/llmSystemPrompt/llmUserMessage/llmConversationHistory as frozen readonly globals in JavaScriptPlanner sandbox, add source:'manual'|'llm' field to DebugReplResult to distinguish user-triggered vs autonomous
Add toCoord/cloneVec3 helpers, return structured telemetry from goToPlayer/goToCoordinate with ok/startPos/endPos/movedDistance/distanceToTargetBefore/distanceToTargetAfter fields, track lastPlannerOutcome summary (actionCount/okCount/errorCount/returnValue/logs) in Brain and inject as [SCRIPT] context with truncation, implement expect/expectMoved/expectNear guardrail tools in JavaScriptPlanner to validate action
Replace async followPlayer skill with sync reflex-based auto-follow that runs during idle mode, add clearFollowTarget action to disable auto-follow before independent movement, add autonomy state to ReflexContext tracking followPlayer/followDistance/followActive/followLastError, implement reconcileAutoFollow in ReflexRuntime using pathfinder GoalFollow that pauses during work/wander/alert modes, add followControl:
Replace string return type with unknown in ActionRegistry/TaskExecutor to support structured responses, change searchForBlock/searchForEntity from navigation to pure search returning structured data with found/block/entity/distance fields, add validation result handling in js-planner to return failed validations as ActionRuntimeResult instead of throwing, use nullish coalescing (??) instead of OR (
Add reconnect option to MineflayerOptions with enabled flag and maxRetries (default: 5), implement tryReconnect method to handle kicked/end events, extract setupBotEventHandlers to reuse on reconnect, track reconnect state with reconnectAttempts/isReconnecting/isStopping flags, add @types/node dependency for setTimeout typing
Replace manual reasoning prefix concatenation with proper reasoning field in Message type, enable reasoning with configurable effort (default: low) in LLM calls, extract reasoningText from response instead of reasoning property
Replace stateful neuri.handle with handleStateless, maintain conversation history manually using Message[] array, append user/assistant messages only after successful parsing to avoid dirty data on retry
- Remove unused ActionLLMHandler and createActionNeuriAgent from adapter.ts
- Replace execution:'parallel'/'sequential' with readonly:true for query tools
- Add 'skip' action for explicit no-op turns
- Remove error suggestion logic (now handled by brain retry)
- Simplify TaskExecutor to use ActionAgent.performAction directly
- Update tests to use createNeuriAgent() without bot parameter
1. Wrap neuri tool invocations to catch ActionError and return as
result strings instead of throwing. This allows the LLM to learn
from tool failures during its reasoning phase.
2. Enhance craftRecipe to use the recipe planner for auto-crafting
intermediate materials. If the bot has raw materials (e.g., logs)
but not direct ingredients (e.g., planks), it will automatically
craft the intermediates first.
3. Add memory note about the dual tool interface discovery (neuri
native tool calls vs JSON actions through TaskExecutor) for future
optimization work.
In Zod 4, _def.defaultValue is the value itself, not a function.
This was causing 'curr._def.defaultValue is not a function' error
when the debug dashboard tried to extract tool definitions.
- Change brain retry behavior to abort on error instead of throwing
- Add JSON parsing utilities with better error messages for malformed LLM responses
- Treat JSON syntax errors as recoverable for retry logic
- Update conscious signal subscription to use 'conscious:signal:*' pattern
- Move RuleEngine and rule utilities from cognitive/os to cognitive/perception/rules
- Update tests to reflect new abort behavior (returns
cleanup comments and rename debug source
fix(minecraft): fix system message regression
Rin