v26.7.15

Bug Fixes

Sudo Command Refused by LLM

• Root Cause: The [SUDO & PACKAGE INSTALL STRATEGY] section in promptBuilder.ts was written before run_terminal_command_pty existed. It told the LLM "this tool runs in a NON-INTERACTIVE shell" and instructed it to tell the user to run sudo manually — causing the LLM to refuse all sudo commands.
• Fix (promptBuilder.ts): Rewrote the SUDO section. LLM now knows it has two terminal tools (run_terminal_command for non-interactive, run_terminal_command_pty for sudo/interactive), is instructed to ALWAYS use PTY for sudo, and is explicitly told it is "FULLY CAPABLE of running sudo commands." Password injection from config.yaml → security.sudo_password is handled automatically by the PTY tool.

LLM Output Repetition Loop

• Root Cause: When the LLM was corrected by the user (e.g., after giving wrong sports results), it entered an uncertain state and looped — echoing the same phrase in multiple rephrasings within a single response (e.g., "gue kabarin hasilnya kalo mau, gue kabarin hasilnya begitu selesai, gue kabarin aja kabarin aja...").
• Fix #1 (llmProvider.ts): Added frequency_penalty: 0.6 and presence_penalty: 0.3 to all OpenAIAdapter.chat() and .stream() calls. These parameters suppress token-level repetition at the API level, effective for all OpenAI-compatible providers (Groq, OpenRouter, xAI, Mistral, DeepSeek, etc.).
• Fix #2 (promptBuilder.ts): Added [ANTI-REPETITION] rule to the universal discipline prompt: LLM is instructed to never repeat the same phrase, clause, or sentence more than once per response — if it catches itself echoing, it must STOP immediately.

Search Inaccuracy / Hallucination on Factual Queries

• Root Cause (5 bugs identified):
  1. Temporal keyword "tadi" (and others: kemarin, besok, pagi ini, etc.) not detected → date not injected → search engine returned mixed results from multiple years.
  2. LLM frequently ignored depth=2 instruction in tool description → only snippets fetched → scores/data in article body never reached LLM context.
  3. No isFactualQuery detection → sport/news/journal queries not auto-upgraded to depth=2.
  4. No rule forbidding LLM from filling data gaps using training memory after an ambiguous search result.
  5. No source citation enforcement → LLM gave hallucinated facts without accountability.
• Fix #1 (searchWeb.ts): Expanded isTimeSensitive detection with 15+ new Indonesian/English temporal keywords (tadi, kemarin, besok, malam ini, pagi ini, sore ini, minggu ini, bulan ini, baru saja, habis, sudah selesai, yesterday, tomorrow, this week, recent, breaking). Added proper date arithmetic for kemarin/yesterday (now−1 day) and besok/tomorrow (now+1 day).
• Fix #2 (searchWeb.ts): Added isFactualQuery detector covering sports (skor, hasil, pertandingan, piala, liga, klasemen), news (berita, kejadian), journals (jurnal, penelitian, paper, studi), and finance (harga, saham, inflasi). Factual queries auto-upgrade to effectiveDepth = Math.max(depth, 2) regardless of LLM's depth argument.
• Fix #3 (searchWeb.ts): Added [SEARCH_CONFIDENCE: HIGH/MEDIUM/LOW] signal to all search outputs. Confidence is derived from how many top-3 articles were successfully scraped. If LOW on a factual/temporal query, an explicit warning is injected into the tool result instructing the LLM to admit data unavailability instead of guessing.
• Fix #4 (promptBuilder.ts): Replaced <web_search_accuracy> section with stronger [GROUNDED ANSWERS ONLY] rule: after calling search_web, answers must be based strictly on search results. If [SEARCH_CONFIDENCE: LOW] or the specific fact is absent, LLM must say "Gue belum nemu data yang spesifik dari search" — never fill gaps from training memory for 2024–2026 events.
• Fix #5 (promptBuilder.ts): Added [SOURCE CITATION] enforcement: when stating any specific fact (score, result, date, statistic), LLM must include the source URL.