Initial commit: workspace setup with skills, memory, config
This commit is contained in:
root
195
memory/2026-02-05.md
Normal file
195
memory/2026-02-05.md
Normal file
@@ -0,0 +1,195 @@
|
||||
# 2026-02-05 — Session Log
|
||||
|
||||
## Major Accomplishments
|
||||
|
||||
### 1. Knowledge Base System Created
|
||||
- **Collection**: `knowledge_base` in Qdrant (768-dim vectors, cosine distance)
|
||||
- **Purpose**: Personal knowledge repository organized by topic/domain
|
||||
- **Schema**: domain, path (hierarchy), subjects, category, content_type, title, checksum, source_url, date_scraped
|
||||
- **Content stored**:
|
||||
- docs.openclaw.ai (3 chunks)
|
||||
- ollama.com/library (25 chunks)
|
||||
- www.w3schools.com/python/ (7 chunks)
|
||||
- Multiple list comprehension resources (3 entries)
|
||||
|
||||
### 2. Smart Search Workflow Implemented
|
||||
- **Process**: Search KB first → Web search second → Synthesize → Store new findings
|
||||
- **Storage rules**: Only substantial content (>500 chars), unique (checksum), full attribution
|
||||
- **Auto-tagging**: date_scraped, source_url, domain detection
|
||||
- **Scripts**: `smart_search.py`, `kb_store.py`, `kb_review.py`, `scrape_to_kb.py`
|
||||
|
||||
### 3. Monitoring System Established
|
||||
- **OpenClaw GitHub Repo Monitor**
|
||||
- Schedule: Daily 11:00 AM
|
||||
- Tracks: README, releases (5), issues (5)
|
||||
- Relevance filter: Keywords affecting our setup (ollama, telegram, skills, memory, etc.)
|
||||
- Notification: Only when significant changes detected (score ≥3 or high-priority areas)
|
||||
- Initial finding: 24 high-priority areas affected
|
||||
|
||||
- **Ollama Model Monitor**
|
||||
- Schedule: Daily 11:50 AM
|
||||
- Criteria: 100B+ parameter models only (to compete with gpt-oss:120b)
|
||||
- Current large models: gpt-oss (120B), mixtral (8x22B = 176B effective)
|
||||
- Notification: Only when NEW large models appear
|
||||
|
||||
### 4. ACTIVE.md Syntax Library Created
|
||||
- **Purpose**: Pre-flight checklist to reduce tool usage errors
|
||||
- **Sections**: Per-tool validation (read, edit, write, exec, browser)
|
||||
- **Includes**: Parameter names, common mistakes, correct/wrong examples
|
||||
- **Updated**: AGENTS.md to require ACTIVE.md check before tool use
|
||||
|
||||
## Key Lessons & Policy Changes
|
||||
|
||||
### User Preferences Established
|
||||
1. **Always discuss before acting** — Never create/build without confirmation
|
||||
2. **100B+ models only** for Ollama monitoring (not smaller CPU-friendly models)
|
||||
3. **Silent operation** — Monitors only output when there's something significant to report
|
||||
4. **Exit code 0 always** for cron scripts (prevents "exec failed" logs)
|
||||
|
||||
### Technical Lessons
|
||||
- `edit` tool requires `old_string` + `new_string` (not `newText`)
|
||||
- After 2-3 failed edit attempts, use `write` instead
|
||||
- Cron scripts must always `sys.exit(0)` — use output presence for signaling
|
||||
- `read` uses `file_path`, never `path`
|
||||
|
||||
### Error Handling Policy
|
||||
- **Search-first strategy**: Check KB, then web search before fixing
|
||||
- **Exception**: Simple syntax errors (wrong param names, typos) — fix immediately
|
||||
|
||||
## Infrastructure Updates
|
||||
|
||||
### Qdrant Memory System
|
||||
- Hybrid approach: File-based + vector-based
|
||||
- Enhanced metadata: confidence, source, expiration, verification
|
||||
- Auto-storage triggers defined
|
||||
- Monthly review scheduled (cleanup of outdated entries)
|
||||
|
||||
### Task Queue Repurposed
|
||||
- No longer for GPT delegation
|
||||
- Now for Kimi's own background tasks
|
||||
- GPT workloads moving to separate "Max" VM (future)
|
||||
|
||||
## Active Cron Jobs
|
||||
| Time | Task | Channel |
|
||||
|------|------|---------|
|
||||
| 11:00 AM | OpenClaw repo check | Telegram (if significant) |
|
||||
| 11:50 AM | Ollama 100B+ models | Telegram (if new) |
|
||||
| 1st of month 3:00 AM | KB review (cleanup) | Silent |
|
||||
|
||||
## Enforcement Milestone — 10:34 CST
|
||||
|
||||
**Problem**: Despite updating AGENTS.md, TOOLS.md, and MEMORY.md with ACTIVE.md enforcement rules, I continued making the same errors:
|
||||
- Used `path` instead of `file_path` for `read`
|
||||
- Failed to provide `new_string` for `edit` (4+ consecutive failures)
|
||||
|
||||
**Root Cause**: Documentation ≠ Behavior change. I wrote the rules but didn't follow them.
|
||||
|
||||
**User Directive**: "Please enforce" — meaning actual behavioral change, not just file updates.
|
||||
|
||||
**Demonstrated Recovery**:
|
||||
1. ✅ Used `read` with `file_path` correctly
|
||||
2. ❌ Failed `edit` 4 times (missing `new_string`)
|
||||
3. ✅ Switched to `write` per ACTIVE.md recovery protocol
|
||||
4. ✅ Successfully wrote complete file
|
||||
|
||||
**Moving Forward**:
|
||||
- Pre-flight check BEFORE every tool call
|
||||
- Verify parameter names from ACTIVE.md
|
||||
- After 2 edit failures → use `write`
|
||||
- Quality over speed — no more rushing
|
||||
|
||||
## Core Instruction Files Updated — 10:36 CST
|
||||
|
||||
Updated all core .md files with enforced, actionable pre-flight steps:
|
||||
|
||||
### TOOLS.md Changes:
|
||||
- Added numbered step-by-step pre-flight protocol
|
||||
- Added explicit instruction to read ACTIVE.md section for specific tool
|
||||
- Added parameter verification table with correct vs wrong parameters
|
||||
- Added emergency recovery rules table (edit fails → use write)
|
||||
- Added 5 critical reminders (file_path, old_string/new_string, etc.)
|
||||
|
||||
### AGENTS.md Changes:
|
||||
- Added TOOLS.md to startup protocol (Step 3)
|
||||
- Added numbered steps for "Before Using Tools" section
|
||||
- Added explicit parameter verification table
|
||||
- Added emergency recovery section
|
||||
- Referenced TOOLS.md as primary enforcement location
|
||||
|
||||
### Key Enforcement Chain:
|
||||
```
|
||||
AGENTS.md (startup) → TOOLS.md (pre-flight steps) → ACTIVE.md (tool-specific syntax)
|
||||
```
|
||||
|
||||
## Knowledge Base Additions — Research Session
|
||||
|
||||
**Stored to knowledge_base:** `ai/llm-agents/tool-calling/patterns`
|
||||
- **Title**: Industry Patterns for LLM Tool Usage Error Handling
|
||||
- **Content**: Research findings from LangChain, OpenAI, and academic papers on tool calling validation
|
||||
- **Key findings**:
|
||||
- LangChain: handle_parsing_errors, retry mechanisms, circuit breakers
|
||||
- OpenAI: strict=True, Structured Outputs API, Pydantic validation
|
||||
- Multi-layer defense architecture (prompt → validation → retry → execution)
|
||||
- Common failure modes: parameter hallucination, type mismatches, missing fields
|
||||
- Research paper "Butterfly Effects in Toolchains" (2025): errors cascade through tool chains
|
||||
- **Our unique approach**: Pre-flight documentation checklist vs runtime validation
|
||||
|
||||
---
|
||||
*Session type: Direct 1:1 with Rob*
|
||||
*Key files created/modified: ACTIVE.md, AGENTS.md, TOOLS.md, MEMORY.md, knowledge_base_schema.md, multiple monitoring scripts*
|
||||
*Enforcement activated: 2026-02-05 10:34 CST*
|
||||
*Core files updated: 2026-02-05 10:36 CST*
|
||||
|
||||
## Max Configuration Update — 23:47 CST
|
||||
|
||||
**Max Setup Differences from Initial Design:**
|
||||
- **Model**: minimax-m2.1:cloud (switched from GPT-OSS)
|
||||
- **TTS Skill**: max-tts-custom (not kimi-tts-custom)
|
||||
- **Filename format**: Max-YYYYMMDD-HHMMSS.ogg
|
||||
- **Voice**: af_bella @ Kokoro 10.0.0.228:8880
|
||||
- **Shared Qdrant**: Both Kimi and Max use same Qdrant @ 10.0.0.40:6333
|
||||
- Collections: openclaw_memories, knowledge_base
|
||||
- **TOOLS.md**: Max updated to match comprehensive format with detailed tool examples, search priorities, Qdrant scripts
|
||||
|
||||
**Kimi Sync Options:**
|
||||
- Stay on kimi-k2.5:cloud OR switch to minimax-m2.1:cloud
|
||||
- IDENTITY.md model reference already accurate for kimi-k2.5
|
||||
|
||||
## Evening Session — 19:55-22:45 CST
|
||||
|
||||
### Smart Search Fixed
|
||||
- Changed default `--min-kb-score` from 0.7 to 0.5
|
||||
- Removed server-side `score_threshold` (too aggressive)
|
||||
- Now correctly finds KB matches (test: 5 results for "telegram dmPolicy")
|
||||
- Client-side filtering shows all results then filters
|
||||
|
||||
### User Preferences Reinforced
|
||||
- **Concise chats only** — less context, shorter replies
|
||||
- **Plain text in Telegram** — no markdown formatting, no bullet lists with symbols
|
||||
- **One step at a time** — wait for response before proceeding
|
||||
|
||||
### OpenClaw News Search
|
||||
Searched web for today's OpenClaw articles. Key findings:
|
||||
- Security: CVE-2026-25253 RCE bug patched in v2026.1.29
|
||||
- China issued security warning about improper deployment risks
|
||||
- 341 malicious ClawHub skills found stealing data
|
||||
- Trend: Viral adoption alongside security crisis
|
||||
|
||||
### GUI Installation Started on Deb
|
||||
- Purpose: Enable Chrome extension for OpenClaw browser control
|
||||
- Packages: XFCE4 desktop, Chromium browser, LightDM
|
||||
- Access: Proxmox console (no VNC needed)
|
||||
- Status: Complete — 267 packages installed
|
||||
- Next: Configure display manager, launch desktop, install OpenClaw extension
|
||||
|
||||
### OpenClaw Chrome Extension Installation Method
|
||||
**Discovery**: Extension is NOT downloaded from Chrome Web Store
|
||||
**Method**: Installed via OpenClaw CLI command
|
||||
**Steps**:
|
||||
1. Run `openclaw browser extension install` (installs to ~/.openclaw/browser-extension/)
|
||||
2. Open Chromium → chrome://extensions/
|
||||
3. Enable "Developer mode" (toggle top right)
|
||||
4. Click "Load unpacked"
|
||||
5. Select the extension path shown after install
|
||||
6. Click OpenClaw toolbar button to attach to tab
|
||||
**Alternative**: Clone from GitHub and load browser-extension/ folder directly
|
||||
Reference in New Issue
Block a user