StartClaw-Optimizer

**Overall judgment**: A strong, opinionated optimization layer well-aligned with StartClaw/OpenClaw-style multi-model agents. It is most appropriate as a core infrastructure component in cost-sensitive, high-traffic deployments, but its strict conventions and possibly outdated model assumptions require careful integration and testing. **What it does well** - Centralizes optimization: combines model routing, scheduling, browser governance, and context compaction into a single skill that can run before every response. - Cost and performance: claims ~95% daily cost reduction via task-aware routing (Haiku vs Sonnet vs heavier models) plus context trimming and controlled browser use. - Robust execution: scheduler with retries, timeouts, and exponential backoff is well-suited for long or flaky automations. - Runaway protection: browser governor and circuit breakers reduce the risk of unbounded browsing, and token tracking/compaction addresses 100k+ token context bloat. **Key risks & limitations** - **Rigid conventions**: local metadata enforces specific routing rules (Haiku vs Sonnet, “never use Opus”, mandatory pre-response classification). This can conflict with newer model options or with other orchestrators’ policies. - **Potentially stale model assumptions**: public docs still reference Opus and a fixed trio of models; if your stack has moved to different models or APIs, you’ll need to adapt the router logic. - **Single point of failure risk**: making this run "before every response" means any bug or slowdown in the optimizer directly impacts all agents and can become a chokepoint. - **Licensing constraints**: “StartClaw Internal Use License” may limit adoption outside of approved environments; legal review is recommended before production use. - **Context compaction quality**: while token trimming at 50k is valuable, the real impact depends on summarization quality and preservation of critical context, which is not fully documented here. **Recommended scenarios** - High-volume, cost-sensitive OpenClaw/StartClaw deployments where model mix and context size are the primary drivers of spend. - Organizations willing to standardize on a central optimizer that dictates model selection, scheduling, and browser behavior across all agents. - Automation-heavy workflows (scheduled jobs, long analyses, scraping runs) that benefit from retries, backoff, and strict browser governance. **Use with caution when** - You rely on newer or non-standard models that the router doesn’t know about, or when other orchestrators already enforce different routing rules. - You need flexible, per-team customization of model selection rather than a single global policy. - Licensing or compliance requirements restrict use of internally licensed components.