47 lines
1.5 KiB
Markdown
47 lines
1.5 KiB
Markdown
# Replace OpenClaw with Hybrid Bedrock Approach
|
||
|
||
**Date:** 2026-06-10
|
||
**Status:** Accepted
|
||
**Author:** Sam
|
||
|
||
## Context
|
||
|
||
OpenClaw has become unreliable for daily use:
|
||
- OAuth token refresh failures (openai-codex provider)
|
||
- Plugin system fragility (codex-supervisor disabled randomly)
|
||
- Model fallback chain breaks when any provider auth expires
|
||
- Requires constant supervision and config debugging
|
||
- Complex TypeScript codebase makes troubleshooting opaque
|
||
|
||
## Decision
|
||
|
||
Replace OpenClaw with a hybrid approach:
|
||
- **Intelligence:** AWS Bedrock (Claude Haiku 4.5) — IAM auth, no OAuth
|
||
- **Orchestration:** Cron + bash scripts + Ansible
|
||
- **Chat interface:** Custom Telegram bot (minimal Python or bash + curl)
|
||
- **Hosting:** racknerd3 (existing VPS) calling Bedrock API
|
||
- **Knowledge base:** Local (ChromaDB or SQLite FTS)
|
||
|
||
## Consequences
|
||
|
||
**Positive:**
|
||
- $2–15/month vs $200/month (ChatGPT Pro subscription)
|
||
- No plugin system to break
|
||
- No OAuth refresh loops
|
||
- Every component is inspectable bash/curl
|
||
- Agents defined by prompt files, not framework config
|
||
|
||
**Negative:**
|
||
- No pre-built skill marketplace
|
||
- Must build Telegram bot manually
|
||
- No self-improving loop (unless we build one)
|
||
- Initial setup effort
|
||
|
||
## Migration Plan
|
||
|
||
1. Build Scribe agent first (note enhancement workflow)
|
||
2. Validate Bedrock connectivity from racknerd3
|
||
3. Wire Telegram bot for conversational access
|
||
4. Port cron jobs (inbox-triage, nightwatch) to bash + Bedrock
|
||
5. Decommission OpenClaw once all workflows migrated
|
||
|