Agent Smith Plugin — Development skill for Claude Code

agent-smith

a claude code skill that offloads the bulky half of a task to cheaper models, then makes claude verify the result before it counts.

claude scopes, reviews, integrates and delivers. gemini, a paid glm plan, or a local ollama fleet do the drafting, digesting and research. every run is logged, every reviewed run gets a verdict, and the verdicts decide which model gets which kind of work next time. trust is measured in a gym, never asserted.

why

claude tokens are the expensive, rate-limited resource. most of what burns them is not judgment, it is volume: reading a 200 page pdf, drafting boilerplate from a clear spec, running a web search and summarizing ten sources, classifying a thousand records. those jobs are checkable. a cheaper model can draft them and claude can spot-check the output for less than it costs to do the work itself.

the skill exists to make that the default rather than something you remember to do.

what it does

• gemini.py sends a prompt (plus optional files, images, web grounding, json schema) to one of several backends and prints the answer on stdout, metadata on stderr
• smith_agent.py runs a sandboxed tool loop for multi-step scratch builds: fix a bug, add a feature, build a small app in a throwaway directory
• transcribe.py transcribes audio locally with mlx-whisper (apple silicon)
• embed.py does embeddings and reranking via cloudflare workers ai
• a usage ledger records every run with its task shape, model, purpose and archived output
• verdict.py records good/bad after review; the ledger turns verdicts into per-shape routing weights
• gap_report.py joins the ledger against your claude transcripts to show where claude did work a trusted route could have done
• fleet_check.py catches an ollama pull silently replacing a model you had earned trust in
• two non-blocking PreToolUse hooks nudge claude toward the fleet at the moment it reaches for WebSearch or a read-only Agent fan-out

install