Move work into work/gallant-ellis-7cd14d#48
Open
fahmiwol wants to merge 141 commits intoclaude/gallant-ellis-7cd14dfrom
Open
Move work into work/gallant-ellis-7cd14d#48fahmiwol wants to merge 141 commits intoclaude/gallant-ellis-7cd14dfrom
fahmiwol wants to merge 141 commits intoclaude/gallant-ellis-7cd14dfrom
Conversation
Root causes of 80% goldset (5 fails):
- Q1/Q4 (current events): synthesizer had no current date/president → fell back
to 2023 corpus data. Fix: inject date + Prabowo fact as PRIORITY GROUNDING
in system prompt so any corpus conflict is overridden.
- Q9 (sanad): answer said "chain of citation" (correct!) but validator only
matched Indonesian keywords. Fix: add "chain", "rangkaian", "jalur" to list.
- Q18 (IHOS): synthesizer had no brand canon → hallucinated "HIDOS". Fix: inject
IHOS definition (Islamic Holistic Ontological System = engineering framework)
into system prompt grounding block.
- Q25 (attention mechanism): answer conceptually correct but lacked vocab
"query/key/value/softmax". Fix: loosen lambda to accept "perhatian", "fokus",
"berfokus", "token", "konteks" as equivalent signals.
- Unicode crash: "Saved → {out}" used U+2192 → UnicodeEncodeError on cp1252.
Fix: replace with ASCII "->".
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…hlight active mode with .mode-active gold ring
…ore modeHolisticBtn.click() indirection
…detection ke root repo file (bukan worktree subfolder)
- mojeek_search.py: Mojeek scraper (no VPS blocking) - lite_browser.py: Playwright headless + trafilatura page fetch - multi_source_orchestrator.py: _src_web_search uses Mojeek primary + deep fetch - omnyx_direction.py: web_search tool uses new pipeline - requirements.txt: +selectolax +trafilatura +playwright
fahmiwol
force-pushed
the
work/gallant-ellis-7cd14d
branch
from
2b3a29b to
9b0a527
Compare
- mojeek_search.py: multiple CSS selectors fallback, debug mode - omnyx_direction.py: add OMNYXDirector.run() wrapper for VPS compat - agent_serve.py: add POST /agent/chat_holistic endpoint (OMNYX primary + legacy fallback)
1. agent_serve.py — UnboundLocalError: omnyx_err referenced after except
block scope ends (Python deletes `as var` after except). Fix: capture
to omnyx_err_str inside the except block, use that in the raise.
2. mojeek_search.py — Mojeek returning 403 from VPS IP ("automated queries").
Fix: add DuckDuckGo HTML scraper as automatic fallback when Mojeek
returns non-200 or 0 hits. Both share MojeekHit interface, callers
unchanged.
3. cognitive_synthesizer.py — new file missing from repo despite being
imported by agent_serve.py and omnyx_direction.py. Add:
- SynthesisResult compat fields (citations, confidence_score, answer_type)
- _strip_yaml_frontmatter() helper imported by omnyx_direction.py
4. knowledge_accumulator.py — new file also missing from repo, imported
by omnyx_direction.py for autonomous knowledge storage.
Verified: /agent/chat_holistic now returns valid JSON on VPS.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…c_stream is 404 /agent/chat_holistic_stream endpoint does not exist on backend yet. askHolisticStream() was getting 404 → onError → user sees "Error: Backend error: 404". Fix: keep streaming attempt first (forward-compat), but if streaming fails or returns empty answer, fall back to askHolistic() (non-streaming). Map sources_used → chip updates manually so progress UI still works. Import cleanup: remove duplicate BRAIN_QA_BASE import. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- askHolistic(): POST /agent/chat_holistic non-streaming - askHolisticStream(): POST /agent/chat_holistic_stream SSE (for future use) - ChatHolisticResponse interface with answer/confidence/n_sources/sources_used - Import cleanup: deduplicate BRAIN_QA_BASE Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… line 488 SourceResult dataclass requires 'source' as first positional arg. Call at line 488 was missing it → TypeError on every chat_holistic request. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ed from VPS IP - Reduced DDG timeout 10s→5s (fast fail since VPS IP times out anyway) - Added _wikipedia_search_async() using Wikipedia API (confirmed reachable from VPS) - Wired Wikipedia as third fallback in mojeek_search_async() after DDG fails - Wikipedia searches id.wikipedia.org first, falls back to en.wikipedia.org VPS IP is blocked by both Mojeek (403) and DuckDuckGo (ConnectTimeout). Wikipedia API is always reachable (0.9s, no IP restrictions). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…E2E verify Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- auto_harvest.py: AutoHarvest class with Wikipedia fallback (DDG blocked on VPS) - Fetch Google Trends RSS Indonesia → fallback to static topics - Wikipedia search (id then en) → fetch article extract (300+ chars) - Generate .md notes with YAML frontmatter (title, date, source, tags) - Save to brain/public/omnyx_knowledge/YYYY-MM-DD/ - Auto-reindex BM25 via POST /corpus/reindex after harvest - Dedup via .harvest_log.jsonl (skip already-harvested articles) - scripts/harvest_cron.py: CLI entry point for cron (--dry-run, --topics, --max-topics) - crontab.example: Cron schedule every 6 hours Knowledge flywheel: auto-harvest -> corpus growth -> more passthrough -> faster answers. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…fault Python-urllib) Wikipedia API returns 403 for Python's default User-Agent. Fixed by using RFC-compliant bot UA string for all Wikipedia requests. Same fix applied to Google Trends RSS fetch. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…n token for reindex - Use paths.workspace_root() to resolve absolute path to brain/public/omnyx_knowledge (fixes notes saving to apps/brain_qa/brain/ instead of /opt/sidix/brain/) - Read BRAIN_QA_ADMIN_TOKEN from env for /corpus/reindex X-Admin-Token header - Update crontab.example to source /opt/sidix/.env before running harvest Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…iteria Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…rking Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…fault Python UA) Wikipedia API returns 403 for Python urllib's default User-Agent. Same fix as auto_harvest.py — use RFC bot UA string for all Wikipedia requests. Also strips accidental UTF-8 BOM from file. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- IntentClassifier.classify_complexity() → (intent, tools, complexity, n_persona, synth_model) - COMPLEXITY_MAP: simple→0 persona+1.5b, creative→1 persona+7b, analytical→3 persona+7b - OmnyxDirector.process(): complexity-aware turns + synthesis - _plan_next_turn(): skip persona_brain for simple queries - _synthesize(): corpus passthrough → web direct → synthesis fallback - OmnyxSession: +complexity +synth_model fields - Response: expose complexity + synth_model for observability
Backend: - /upload/image endpoint — multipart upload, save to workspace/uploads/, 5MB limit - /upload/audio endpoint — multipart upload, save to workspace/uploads/, 10MB limit - ChatRequest: +image_path +audio_path fields - /agent/chat_holistic: multimodal path via process_multimodal when image/audio present Frontend: - attach-btn click handler → hidden file input → uploadImage() → backend - pendingImagePath state → passed to askHolistic() opts - Visual feedback: placeholder shows attached filename Integration: - multimodal_input.py already handles image_path → mock vision caption → ReAct - Full end-to-end: user clicks attach → selects image → uploads → asks question → multimodal processing
- Analisa posisi SIDIX vs visi bos (40+ riset notes + 12 dokumen + codebase audit) - Identifikasi 5 gap fundamental (Sanad=spec, Hafidz=storage, Pattern=isolated, ToolSynth=missing, Persona=prompt-level) - Roadmap baru 4 fase (Organisme Hidup → Creative Agent → Persona Mandiri → Vision) - Sprint plan A+B: Sanad Orchestra + Hafidz Injection (NEXT PRIORITY) - Update dokumen outdated: SIDIX_CAPABILITY_MAP, STATUS_TODAY, LIVING_LOG
Sanad Orchestra (sprint A): - Multi-source consensus validation engine (sanad_orchestra.py) - Claim extraction (LLM + regex fallback) - Claim verification (corpus + web + tools) - Consensus calculation with weighted confidence - Relative thresholds per query type (simple/analytical/creative/tool) - Verdict: golden | pass | retry | fail - Retry synthesis with failure context - Stats endpoint: /agent/sanad/stats - Validate endpoint: /agent/validate Hafidz Injector (sprint B): - Two-Drawer memory: Golden Store (>=threshold) + Lesson Store (<threshold) - Pre-query: retrieve few-shot examples + lesson warnings - Post-query: store to Golden or Lesson based on Sanad score - BM25-based retrieval with persona filtering - Prompt injection: build_hafidz_prompt() - Stats endpoint: /agent/hafidz/stats Integration: - OMNYX Direction wired: pre-query Hafidz retrieve + post-synthesis Sanad validate - ChatResponse extended: sanad_score, sanad_verdict, hafidz_injected, hafidz_stored - CognitiveSynthesizer: _try_corpus_passthrough helper Tests: 34 passed (16 Sanad + 18 Hafidz) E2E: Verified end-to-end flow Query→Hafidz→OMNYX→Tools→Synthesis→Sanad→Store
…lf-Train Fase 1
Agency Kit 1-Click:
- agency_kit.py: 6-layer pipeline DAG (brand → content → copy → campaign → thumbnail → synthesis)
- Background threading, progress 0-100%, job store max 50
- POST /creative/agency_kit (async), GET /creative/agency_kit/{job_id}, GET /creative/agency_kit/list
- Frontend: wizard modal, progress bar, result gallery, export MD
- Debate Ring integration per layer via _run_debate_if_available()
- CQF scoring per output
Debate Ring REAL:
- debate_ring.py: 3-round Creator↔Critic consensus via generate_sidix()
- 5 pre-defined debate pairs + custom persona support
- Heuristic CQF scorer (Relevance/Quality/Creativity/Brand/Actionability)
- POST /creative/debate, GET /creative/debate/personas
- debate_layer_output() helper for Agency Kit integration
- Fail-open: returns best text if any round fails
Self-Train Fase 1:
- curator_agent.py: 6-dimension rule-based scoring (relevance, sanad_tier, maqashid, dedupe, length, structure)
- corpus_to_training.py: weekly cron script → JSONL instruction-tuning format
- dataset_id_sea_collector.py: enhanced to output training-compatible JSONL
- GET /training/stats, POST /training/curate, GET /training/data/latest
- Output: brain/public/training_data/YYYY-MM-DD/corpus_pairs.jsonl
All py_compile PASS. Frontend build PASS. Self-hosted ONLY.
DoRA Adapter:
- dora_adapter.py: PERSONA_ADAPTERS registry + PERSONA_SYSTEM_PROMPTS logical fallback
- adapter_exists(), load_persona_adapter(), unload_persona_adapter()
- generate_with_persona(): physical adapter if exists, else logical (system prompt + temperature)
- local_llm.py: generate_sidix() accepts persona parameter, delegates to dora_adapter
- agent_react.py: run_react() passes persona to generate_sidix()
- agent_serve.py: /agent/generate + /agent/chat_holistic propagate persona
- Frontend: persona selection in test generate button
Voyager Protocol P1:
- voyager_protocol.py: dynamic tool creator from natural language intent
- AST security scan + pattern scan + import whitelist + py_compile validation
- Restricted namespace exec() for generated code
- agent_tools.py: TOOL_REGISTRY supports dynamic generated tools with metadata
- Endpoints: POST /app/voyager/create, GET /app/voyager/tools,
GET /app/voyager/tools/{tool_name}, POST /app/voyager/tools/{tool_name}/delete
- Generated tools persist to agent_workspace/generated_tools/
All py_compile PASS. Frontend build PASS. Self-hosted ONLY.
…tion Voyager Protocol Phase 2 — Skill Library Pattern: - Usage tracking (call_count, success_rate, avg_latency) - Skill discovery (keyword overlap, block create if similar >= 0.6) - Self-refinement (auto-improve tool kalau success_rate < 50%) - Anthropic Agent Skills v1 compatible metadata - 5 endpoint baru: /stats, /discover, /refine, /skills-format Maqashid Auto-Tune Phase 2 — Hybrid Judge + Trace-Aware: - HistoricalJudge: self-hosted judge learns dari thumbs up/down feedback - Trace-aware evaluation: score every reasoning step - /app/maqashid/feedback endpoint untuk judge calibration - Hybrid routing: heuristic fast-path + enhanced judge for borderline Raudah Protocol v0.2 — TaskGraph DAG + /raudah/run: - Dependency-based topological sort (explicit depends_on) - Specialist tool integration: call ReAct tools sebelum LLM - POST /raudah/run endpoint exposed Research: note 317 — AI Landscape 2026 gap benchmark (12 sources, 9 gaps) All py_compile PASS. Smoke test PASS.
- VISI_TRANSLATION_MATRIX: coverage shift Pencipta 45%->55%, Cognitive 90%->93% - LIVING_LOG: handoff entry with WIP carry-over and anti-menguap checklist - Ready for next session continuation
…sprint batch 2026-05-08 - Research Note 318: Cognitive Expansion synthesis (20 sources, 5 dimensions) - BACKLOG: 6 sprint queued (Input, Orchestration, Methods, Output, Apps, Blueprint) - VISI_MATRIX: target shift ~87% -> ~93% post-batch - LIVING_LOG: session entry with hypothesis + risk matrix
…ts + tools registry - NEW document_parser.py: Word/Excel/CSV/JSON/TXT parser (pure Python) - ADD /upload/audio/transcribe endpoint (STT via audio_capability) - ADD /tts endpoint (TTS via audio_capability) - ADD /upload/document endpoint (auto-detect format, parse structured) - ADD 3 tools to TOOL_REGISTRY: transcribe_audio, synthesize_speech, parse_document - py_compile: 3/3 PASS - smoke test: 2/2 PASS
… 8 tools - NEW vision_analyzer.py: VLM image/video analysis via Ollama (moondream/llava) - NEW coding_agent_enhanced.py: lint, debug trace, test gen, dependency analysis, code review - NEW brand_guidelines.py: color system, typography, spacing, component tokens, voice & tone - NEW mcp_web_fetch_expanded.py: Reddit, YouTube, GitHub, arXiv, HackerNews (no API key) - ADD 8 endpoints to agent_serve.py: /upload/image/analyze, /upload/video/analyze, /code/lint, /code/debug, /code/tests, /code/review, /brand/guidelines, /web/fetch - ADD 8 tools to TOOL_REGISTRY - py_compile: 6/6 PASS - smoke test: 4/4 PASS
- NEW runpod_connector.py: connector ke RunPod serverless (image/3D/TTS/design) - NEW dataset_collector.py: read-only scan Mighan-Web/Mighan-3D assets - ADD 5 endpoints: /generate/image, /generate/3d, /dataset/scan, /dataset/collect, /dataset/sources - ADD 4 tools: generate_image_runpod, generate_3d_runpod, scan_dataset, collect_dataset - py_compile: 4/4 PASS - smoke test: 2/2 PASS
- NEW: dataset_web_collector.py — Unsplash, Pexels, Wikimedia, LAION-5B - ADD: 6 tools (search_unsplash, search_pexels, search_wikimedia, search_dataset_web, analyze_dataset_dna, get_laion_info) - ADD: 7 endpoints (/dataset/web/*, /dataset/dna, /dataset/laion) - RESEARCH: WEB_DATASET_SOURCES_ANALYSIS.md — legal risk analysis, DNA metrics, recommendations - Decision: reject commercial scraping (Shutterstock/Adobe/Getty/Canva/Instagram) due to copyright+ToS risk - Tools: 50 -> 56 - Research Note: 319
- NEW: dataset_drive_collector.py — Google Drive API via pure HTTP (urllib) - ADD: OAuth2 flow (auth_url, exchange_code, refresh_token) - ADD: list_drive_images + collect_drive_dataset + auto-tagging by folder path - ADD: export_drive_dataset_jsonl for training pipeline - ADD: 4 tools (drive_auth_url, drive_exchange_code, drive_list_images, drive_health) - ADD: 4 endpoints (/dataset/drive/auth, /exchange, /list, /health) - No external deps: pure urllib to avoid cffi/cryptography incompatibility with Python 3.14 - Tools: 56 -> 60 - Research Note: 320
- ADD: Multi-account support via env var suffix (_FAHMIWOL, _TIRANYX, etc) - ADD: explore_drive_structure() — recursive folder tree + image count per folder - ADD: get_account_overview() — user info, storage, total images, top folders - ADD: batch_collect_drive_datasets() — collect from multiple accounts at once - ADD: get_account_config_instructions() — step-by-step for 4 accounts - ADD: 4 tools (drive_explore, drive_overview, drive_batch_collect, drive_config) - ADD: 4 endpoints (/dataset/drive/explore, /overview, /batch, /config) - UPDATE: drive_list_images + drive_health support account param - Auto-tag: account name added as tag (fahmiwol, tiranyx, etc) - Tools: 60 -> 64 - Research Note: 320-v2
… sound) - NEW: elevenlabs_connector.py — Pure HTTP ElevenLabs API client - ADD: generate_tts() — TTS dengan voice settings (stability, similarity, style) - ADD: list_voices() — All voices + Guru Trainer recommendations - ADD: clone_voice() — Voice clone dari audio samples - ADD: generate_sound_effect() — Sound effects dari text - ADD: get_user_info() — Quota & usage - ADD: 6 tools (elevenlabs_tts, voices, clone, user, sound, health) - ADD: 6 endpoints (/tts/elevenlabs/*) - UPDATE: .env.sample dengan ElevenLabs + Google Drive + Unsplash + Pexels - API key HANYA via env var — tidak disimpan di repo - Tools: 64 -> 70 - Research Note: 321
- Add RUNPOD_API_KEY, RUNPOD_MEDIA_ENDPOINT_ID, RUNPOD_3D_ENDPOINT_ID placeholders - API key received from bos, set as env var locally (not committed) - No code changes needed — runpod_connector.py already reads from env
…red) - NEW: dataset_spark_curation.py — ethical dataset curation pipeline - ADD: validate_license() — whitelist/blacklist checker (Pinterest BLACKLISTED) - ADD: create_content_credential() — C2PA-like manifest per asset - ADD: audit_bias() — gender/western/professional bias detection - ADD: curate_ethical_dataset() — main pipeline - ADD: 5 tools (spark_curate, spark_validate, spark_bias, spark_pinterest_warn, spark_provenance) - ADD: 5 endpoints (/spark/*) - RESEARCH: SPARK_DATASET_CURATOR.md — Pinterest risk + Muse analysis + Adobe approach - DECISION: Pinterest scraping REJECTED (ToS + DMCA + copyright) - Tools: 70 -> 75 - Research Note: 322
- Add scripts/exchange_drive_tokens.py — CLI helper for Google OAuth2 code→token exchange - Update .env.sample with clearer Google Drive setup instructions - Add .env.drive.tokens template (gitignored, for local use) - Update LIVING_LOG Refs: Google Drive multi-account collector, Spark ethical dataset
- Add drive_admin_manager.py — runtime token store (JSON) with CRUD + refresh
- Add 6 admin endpoints: /admin/drive/accounts, /connect, /exchange, /refresh, /account/{name}
- Add Google Drive tab to admin.html — OAuth wizard, multi-account manager, folder browser
- Update dataset_drive_collector.py — fallback token read from admin store
- py_compile PASS
Bos can now connect Drive accounts via ctrl.sidixlab.com/admin
without touching .env files.
- Fix 40 import paths (dataset_drive_collector, web_collector, spark_curation, elevenlabs_connector, runpod_connector, vision_analyzer, coding_agent_enhanced, audio_capability, document_parser, mcp_web_fetch_expanded, brand_guidelines, dataset_collector) — ganti absolute jadi relative - Fix creative_framework.py syntax error (line number artifacts) - Add output modality auto-detect (_detect_output_modality) + wire ke /agent/chat - Add ChatResponse.attachments field (image/audio/3d) - Add QA audit script (scripts/qa_tool_audit.py) - py_compile PASS
- Add PDF parser to document_parser.py (pymupdf primary, PyPDF2 fallback) - Wire /agent/vision → vision_analyzer.analyze_image() (Ollama VLM) - Wire /agent/audio → audio_capability.transcribe_audio() (Whisper) - Wire /agent/chat multimodal routing: image_path → vision, audio_path → transcription - Fix sensorial_input.py broken import tts_engine (sys.path hack to apps/audio/) - py_compile PASS Beta Sprint 1: vision + audio + document input fully functional
- Mark Sprint 1 (Input Expansion) as DONE in BACKLOG.md - Deliverable: VLM vision, Whisper ASR, Piper TTS, PDF parser, multimodal chat routing - Also mark Sprint 4 (Output Modality Auto-Detect) as DONE - Append final LIVING_LOG entry for 2026-05-08 sprint batch
- Mark Sprint 1 (Input Expansion) and Sprint 4 (Output Modality) as DONE - List 5 commits in batch, production status, known blockers - Define next actions: deploy VPS, Sprint 2/3/6, or fix broken tools
Session end. 6 commits delivered. Resume checklist in STATUS_TODAY.md.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Convert embedded gallant-ellis-7cd14d into tracked files and move accidental work off main. See docs/LIVING_LOG.md and brain/public/research_notes/308_sprint_decision_2026-04-30.md for details.