Vishnu Work Done - Ian Ferguson Project

Feb 3, 2026 — 4:37 PM 🔥 Context Overflow — Root Cause & Fix FIX

Problem

Vishnu was experiencing recurring context overflow errors and API rate limit failures (HTTP 429). After investigation, we discovered the root cause: Ian had shared his entire conversation history from ChatGPT and Claude with Vishnu, asking it to save everything permanently.

Root Cause

A transcribed voice message confirmed Ian asked Vishnu to "take all this information and save it and make sure that it's in a place you can reference forever." This caused:

Memory Database Bloat: main.sqlite grew to 5.1 MB with 89 embedding entries (vs 4 in a clean database)
Context Inflation: Every interaction searched the bloated index, pulling in years of historical context
Concurrent Session Overload: Multiple sessions each carrying inflated context triggered cascading rate limits

Pre-Fix State — Vishnu's Mac Mini

                main.sqlite: 16 MB (48 chunks, 48 embeddings, 23 memory files from a single day)

                Total text: 59 KB actual content buried under 1.9 MB of vector embeddings

                Result: Context overflow errors, Claude usage burned through daily limits

Fix Applied — Vishnu's Mac Mini (6:13 PM MST)

                ssh vishnu@100.106.206.127

                cp ~/.clawdbot/memory/main.sqlite ~/.clawdbot/memory/main.sqlite.bak  # backup

                rm ~/.clawdbot/memory/main.sqlite  # remove 16 MB bloated DB

                # System auto-rebuilds fresh, minimal database on next use

Post-Fix Validation

✅ Vishnu database: 16 MB → 0 (clean slate, verified via SSH)
✅ Backup preserved as main.sqlite.bak on Vishnu's machine
✅ System rebuilds clean database from current workspace files only
✅ No more imported ChatGPT/Claude history in the index

Lessons Learned

Never bulk-import conversation history from other AI platforms — summarize key points instead
Embedding databases grow silently (5.1 MB DB had only 8 KB of readable text)
Concurrent sessions multiply memory bloat into rate limit cascades

Feb 3, 2026 — 12:25 PM API Rate Limit Optimization CONFIG

Problem

Vishnu was hitting Anthropic API rate limits (HTTP 429), falling back to Gemini despite having Tier 2 limits (450k tokens/min, 1k requests/min).

Solution

Reordered Fallback Chain: Opus → Gemini → Sonnet → GPT-4o (Gemini moved up as faster fallback)
Preserved Concurrency: Kept all 4 concurrent agents
Smart Retry Logic: 3-minute cooldowns to prevent retry loops

Commands

                cp ~/.clawdbot/clawdbot.json ~/.clawdbot/clawdbot.json.backup.$(date +%s)

                # Updated fallbacks in config: ["gemini-3-pro-preview", "claude-sonnet-4", "gpt-4o"]

                clawdbot gateway restart

Result

When Claude hits rate limits: tries Gemini first (fast), then Sonnet, then GPT-4o. 3-min cooldown prevents retry storms.

Feb 1–2, 2026 Initial Setup & Configuration SETUP

Work Completed

undici installed for ClawdHub on Node v25
Browser set to use clawd profile
Cooldowns reduced from 1hr → 3min
Stuck cooldown state cleared
$60 loaded → Tier 2 rate limits
Local embeddings for memory search
Full explainer website created

Configuration

Primary Model: Claude Opus 4.5
Fallback Chain: Opus 4.5 → Sonnet 4 → Gemini 3 Pro → GPT-4o
API Tier: Tier 2 ($60) — 450k tokens/min, 1k requests/min
Features: Web search, browser, heartbeat, nightshift, memory, multi-provider failover

Vishnu Work Done Log

Problem

Root Cause

Pre-Fix State — Vishnu's Mac Mini

Fix Applied — Vishnu's Mac Mini (6:13 PM MST)

Post-Fix Validation

Lessons Learned

Problem

Solution

Commands

Result

Work Completed

Configuration

About Vishnu