v26.8.12

x402 Agentic Payment Protocol

• x402Handler.ts: Implemented the core interceptor for the x402 payment protocol (Coinbase Ventures). Detects HTTP 402 responses, extracts payment parameters (receiver, amount, currency, chain), and routes the transaction directly to the Policy Engine (/request-tx) without creating any bypass paths. Fully respects policy.yaml rules (require_approval, max_usd_per_tx, whitelist_only).
• fetch_with_x402 Skill (x402Fetch.ts): New cognitive skill allowing the LLM to make HTTP requests to paid agentic services. If an endpoint returns 402, the skill autonomously negotiates the payment, executes the transaction, and retries the request with the proof of payment (X-Payment-Receipt header) in a single seamless tool call.
• Web3MarketPlugin: Registered fetch_with_x402 and injected the <x402_payment_rule> into the Web3 agent's system prompt so the LLM understands when and how to utilize this autonomous payment capability.

Bug Fix: Gemini Repetition Loop

• frequencyPenalty & presencePenalty in Gemini Provider (llmProvider.ts): Added frequencyPenalty: 0.5 and presencePenalty: 0.3 to the generationConfig for both chat() and stream() methods in the Gemini provider. Previously, these parameters were omitted, causing Gemini to occasionally fall into a logit loop at the end of its response and repeat sentences/phrases endlessly.
• Post-Processing Deduplication (deduplicateRepetitions): Added the deduplicateRepetitions() function as a final safety net. Before streaming the response to the user, this function detects and truncates sentence repetitions (Pass 1) as well as progressive clause repetitions (Pass 2) that might escape the model-level penalty. Prevents outputs like "...PC is healthy. PC is healthy. healthy. healthy." from reaching the UI.

Bug Fix: Gemini Internal Monologue Leak

• Stream and Non-Stream Fallbacks (osAgent.ts, web3Agent.ts): Removed a flawed fallback mechanism that inadvertently leaked the AI's internal reasoning_content (thoughts) directly to the user when the LLM produced a "silent stop" (thinking block with no tool calls or text). Implemented a proper Nudge and Prefill system to force the LLM to recover gracefully without exposing its internal monologue.
• Bracket Label Stripping (thinkScrubber.ts): Added a post-processing regex pass to stripThinkBlocks to remove bracket-style internal planning labels (e.g. [Self-Correction], [Action], [User provided details]). Gemini occasionally emits these meta-labels directly into its visible text output instead of wrapping them in XML <think> tags, which caused severe UI clutter and logit loops.

AI Memory & Auto-Correction Engine

• Memory Honesty Guardrail (promptBuilder.ts): Injected a strict language-agnostic MEMORY HONESTY rule to prevent the AI from hallucinating or fabricating memories about the user. The AI is now explicitly forced to invoke the search_memory tool before answering any questions regarding past learnings, preferences, or stored facts.
• Deep Memory Search (searchMemory.ts): Upgraded the search_memory cognitive skill to seamlessly query both the long-term episodic.db (permanent facts and user persona traits) and the FTS5 chat history. Previously, it only searched chat logs, causing the AI to miss actual stored knowledge.
• Auto-Correction Capture (autoCorrectionCapture.ts): Engineered a brand new continuous-learning module. The agent loop now actively monitors for failed tool calls (e.g. malformed bash commands or bad JSON args). If the AI successfully retries and fixes the error on a subsequent turn, the system autonomously extracts the exact difference (the "lesson learned") and permanently writes it to episodic.db as a system_correction. This guarantees the AI never repeats the same operational mistake twice across any tool or command.