Benchmark-Narrowed Tool Recommendations
Agent Analyzer no longer treats a public tool list as proof of token savings. Recommendations now come from the repeated benchmark matrix and only apply when your report shows the matching waste signal. Private MCP names, private skills, raw paths, prompts, and transcript text are not used as recommendation labels.
Default Pack
- Agent Analyzer workflow guidance: 3/3 quality, 23.986% lower published API-rate cost.
- Output-budgeted commands for tool-output bloat.
- Scoped retrieval hygiene for repeated file reads.
- Session hygiene and retry breakers for context pivots and repeated failures.
Conditional Reducers
These reduced cost in the noisy benchmark, but only target specific token categories. The generated plugin should recommend them only when the report evidence matches.
- Semble: path-limited semantic retrieval, 41.5% API-rate savings here.
- context-mode: tool-output/input-context batching, 20.4% API-rate savings here.
- RTK: explicit shell-output compression, 18.2% API-rate savings here; global hooks require separate approval.
- grepai: path-constrained compact retrieval, 14.5% API-rate savings here.
Removed
- Squeez: do not recommend; it conflicts with Spec Kitty workflows despite a positive old shell-output benchmark result.
Measurement, Not Reduction
- ccusage: useful independent accounting; not a direct token reducer.
- ccstatusline: useful awareness outside the prompt path; not a direct token reducer.
Removed From Default Recommendations
claude-context, Probe, Caveman for Claude Code, claude-rlm, claude-token-efficient, and broad ecosystem lists are not default token-saving recommendations because the repeated benchmark was negative, too small, incomplete, or not a reducer.