Blocked: Active optimization requires OpenClaw hook mutation support

[evansantos]

Context

SlimClaw is currently running in shadow mode only — measuring potential savings but unable to actually optimize requests.

The Problem

OpenClaw's llm_input and llm_output hooks are observation-only (return void). We cannot:

• Modify historyMessages (for windowing)
• Change model (for routing)
• Mutate systemPrompt at the right scope

Upstream Issue

I've commented on the OpenClaw feature request for this:
[Feature]: Make llm_input / llm_output modifying hooks for middleware patterns
openclaw/openclaw#20416 (comment)

The proposed API would give us:

type PluginHookLlmInputResult = {
  prompt?: string;
  historyMessages?: unknown[];  // ← What we need!
  systemPrompt?: string;
  model?: string;               // ← For routing
};

Current Workaround

SlimClaw tracks metrics and reports what could be saved. Users can leverage OpenClaw's built-in alternatives:

• contextPruning config for windowing-like behavior
• compaction config for automatic context management

Next Steps

1. Monitor [Feature]: Make llm_input / llm_output modifying hooks for middleware patterns (PII redaction, guardrails, content filtering) openclaw/openclaw#20416 for progress
2. When mutation hooks land, implement active optimization
3. Consider contributing a PR to OpenClaw if maintainers are open to it

Labels

• blocked
• upstream
• enhancement

[evansantos]

Phase 1 landed in PR #24 — cross-provider model support for 12 models (OpenAI, Google, DeepSeek, Meta, Qwen) via OpenRouter. Tier-based pricing, inference, and downgrade detection all working.

Current state of the blocker:

• Upstream PR openclaw#20426 covers prompt + assistantTexts mutation — still open, not merged
• For active routing we need model mutation in the hook return type, which isn't in scope for that PR yet
• Without it, SlimClaw can classify and recommend but can't actually reroute requests

What we're doing now (Phase 2a):
Building everything that doesn't need hook mutation — provider resolution, tierProviders glob matching, shadow routing with "would have routed to X, saving Y%" recommendations, and preparing the routing decision function for when mutation lands.

When upstream ships model mutation, we flip from shadow to active with minimal changes.