Two tools with minimal descriptions (both under 100 chars) and incomplete parameter documentation. insert_prefix and lookup_prompt lack descriptions of what values are expected, valid ranges, or error conditions. No output schema documented. Parameter descriptions are absent entirely, 'prompt' and 'kv_cache_id' have no guidance on format, length, or constraints. No error handling guidance. Tool names are verb-based but descriptions do not explain WHEN to use each tool or what distinguishes them. No mention of idempotency, side effects, or recovery paths.

• Expand tool descriptions to 100 - 200 chars. For insert_prefix: 'Cache a system prompt or template prefix in the radix tree to enable KV-state reuse across requests. Reduces TTFT latency for repeated prompts. Idempotent: calling with the same kv_cache_id overwrites the previous entry.'
• Add parameter descriptions: prompt → 'The text to cache, tokenized by whitespace. Max 10,000 tokens. Required.' kv_cache_id → 'Unique identifier for this cached prefix (e.g., sys_001, tool_schema_v2). Used to track cache hits. Required.'
• Document output schema in tool metadata. E.g., insert_prefix returns {status: string, cache_entries: int, estimated_savings_percent: float}. lookup_prompt returns {total_tokens: int, cached_tokens: int, hit_ratio: float, estimated_ttft_reduction: string, is_full_cache_hit: bool}.
• Add error handling guidance: 'If kv_cache_id is empty, returns error "kv_cache_id required". If prompt is null, returns error "prompt required". Retry with valid inputs.'
• Declare insert_prefix as a write/destructive operation in tool metadata (destructiveHint: true) so LLMs know it modifies state and should not be retried blindly.