Token Optimization Agent

Find where your LLM project is using tokens inefficiently and fix it without damaging your product.

Who it's for:

• Teams running Python agents or apps on the Anthropic or OpenAI SDK who
  want to cut their token bill without risking regressions
• Engineers who know their LLM spend is too high but can't safely hand prompt/model edits to a tool
  they don't trust.

What it does:

• Maps every LLM call site in your project (static scan).
• Measures real token usage by running your own scenarios — wrapper telemetry,
  zero code changes.
• Ranks concrete waste (prompt caching, model tiering) with estimated $/run and
  projected $/month savings.
• Applies fixes in isolated git worktrees and verifies them against a golden
  baseline.
• Keeps a durable, independently-measured savings ledger across runs.

Why it's safe:

Every applied edit must pass a do-no-harm gate — your tests pass, your outputs
stay equivalent, and the savings actually materialize — or it is reverted.
Estimated savings never justify keeping an edit; only measured savings do. It runs
read-only by default: it audits and judges, and changes nothing until you
raise the autonomy level.

What you get:

• A read-only audit (meta_report.md) ranking opportunities with dollar figures.
• On opt-in, verified patches that provably preserve behavior.
• A savings ledger and portfolio view across runs.

Requirements:

• A Python project that calls the Anthropic or OpenAI SDK.
• A scenarios.yaml describing what the project must keep doing (the skill helps
  you write one).
• For applying fixes: a git repo. No API key required (a compatible API key only
  sharpens the optional judge).