coffer-cost-review

A Claude Code / Codex skill that audits AI codebases for LLM cost waste — with semantic understanding, not regex.

npm install -g coffer-cost-review

Then in Claude Code or Codex ask:

review my LLM costs

That's it. The skill is installed to:

• ~/.claude/skills/coffer-cost-review/
• ${CODEX_HOME:-$HOME/.codex}/skills/coffer-cost-review/

Why semantic > regex

The first version of this tool was a Python static scanner. We dogfooded it on five well-known agent projects (Aider, smolagents, crewAI, OpenInterpreter, MetaGPT). It produced ~64 "findings" of which most were false positives:

• Aider applies cache_control in a separate file at runtime — a single-file regex can't see that. Findings about Aider's "uncached" system prompts were wrong.
• smolagents caps memory growth via max_steps=20 in the agent loop, not in the append call. Findings about "unbounded history" were wrong.
• crewAI's instructions = f"{instructions}\n\n{output_schema}" looks like a cache-breaking interpolation but output_schema is a fixed Pydantic-model description — static at init. Findings were wrong.
• stagehand_tool's instruction = f"Navigate to {url}" matched the same regex but isn't an LLM system prompt at all — it's a browser navigation command.

Static regex inherently can't see:

• cache_control applied at runtime in a different file
• Caps enforced one level up in an agent loop
• Init-time concatenation that resolves to a static string
• Variable names that suggest "system prompt" but go elsewhere

These all require reading the code with full repo context. That's what Claude does well; that's what this skill leans into.

What it catches

By cost lever (every detector answers "yes" to: does fixing this reduce the bill the provider sends?):

| Lever | Pattern | |-------|---------| | Input tokens | uncached large system prompts (Anthropic / OpenAI aware); dynamic-before-static cache breaks; unbounded conversation history (after checking external caps) | | Output tokens | reasoning_effort="high" default; missing max_tokens on streaming | | Number of calls | unbounded agent loops; temperature > 0 next to a cache layer; LLM doing regex's job | | Provider tier | frontier model for trivial tasks; reasoning model for non-reasoning tasks | | Architecture | retry loops without backoff; public endpoints without rate limit; streaming without abort |

What it deliberately doesn't flag

Reliability, observability, and metering issues are real production problems but they don't change the LLM bill:

• SDK init without timeout= → worker exhaustion, not over-payment
• Missing response.usage capture → metering, not over-payment
• Logging full prompts → Datadog bill, not OpenAI bill
• Missing idempotency_key → correctness, not over-payment

A separate "production-readiness review" skill is the right home for those.

Install paths

npm (recommended, also runs on macOS / Linux / Windows)

npm install -g coffer-cost-review

Or one-shot without global install:

npx coffer-cost-review

Install only one agent target:

npx coffer-cost-review --only=codex
npx coffer-cost-review --only=claude

curl (no Node / npm needed)

curl -fsSL https://cofferwise.com/install-skill.sh | sh

Manual

Clone this repo, copy skill/ to ~/.claude/skills/coffer-cost-review/ or ${CODEX_HOME:-$HOME/.codex}/skills/coffer-cost-review/.

Uninstall

npm uninstall -g coffer-cost-review     # also removes the skill files
# or manually:
rm -rf ~/.claude/skills/coffer-cost-review
rm -rf "${CODEX_HOME:-$HOME/.codex}/skills/coffer-cost-review"

Live runtime tracking

Static review tells you which patterns might be wasting money. It can't tell you which feature, which user, or which prompt is actually burning your bill right now.

For runtime cost attribution per feature / user / prompt: Cofferwise — built by the same team.

Contributing

PRs welcome. New patterns must answer "yes" to "fixing this reduces the bill from OpenAI / Anthropic / a paid LLM provider." Reliability / metering / observability improvements belong in a separate skill.

License

Apache 2.0.