Make cc-audit shippable: tests, redaction proof, and a home

Goal

Get cc-audit to a state where it can be merged to master as part of this repo, rather than sitting on wip/audit.

It was deliberately held back from the open-source release: the harness-customization docs are stable, this isn't. This issue tracks closing that gap.

What's on the branch today

audit/ — a dependency-free Python CLI (~150KB, stdlib only) that reads local Claude Code transcripts from ~/.claude/projects/**/*.jsonl and writes a ranked markdown report of what to change to reduce spend.

Verified on a fresh clone: all modules compile under Python 3.14, and --help runs clean.

It is further along than "prototype" — it already handles per-user dated report storage, repo exclusions, worktree-aware agent resolution, a --trend view across runs, and an optional --challenge pass that sends a ~4K-token config-only packet to an LLM to attack its own findings, guarded by a regex that aborts if a path, filename, URL, or shell command would be sent.

Why it isn't merged

1. No tests. This is the blocker. There is no test file anywhere in audit/. For a tool whose entire output is numbers people will act on — and which asserts things like "this would save $X/month" — untested arithmetic is not shippable. Highest-value coverage first:

• pricing.py — cost math, cache multipliers (1.25x / 2.0x write, 0.10x read), family mapping from raw model ids
• parse.py — transcript parsing against fixture .jsonl, including malformed and truncated lines
• store.py — the JSON sanitization boundary, asserting no paths/repo names/session IDs survive
• checks.py — at least the checks that produce a savings figure

2. The redaction boundary is claimed but unproven. The README makes a strong promise: the JSON carries no file paths, no shell commands, no repo names, and the --challenge packet is config-only by construction. Both are load-bearing — the whole point is that reports can be pooled across a team and that the critic pass can't leak source. Neither has a test proving it. This needs adversarial tests with deliberately hostile fixtures (repo named like a path, a prompt containing an API key, unicode tricks).

3. Pricing is a hardcoded table that will silently rot. pricing.py currently has Opus $5/$25, Sonnet $3/$15, Haiku $1/$5, Fable $10/$50, with sanity-range assertions. That's correct today. When prices or families change, every figure in every report becomes quietly wrong with no signal. Decide the policy: a dated as-of stamp printed in the report, a staleness warning after N months, or accept it and document the review cadence.

4. Single-machine assumptions. Paths like ~/.claude/projects and ~/.claude/cc-audit-reports and the <os-username>-<4-char machine hash> user id are reasonable but untested beyond one laptop. Confirm behavior on Linux, and when ~/.claude is absent or empty.

5. Python version floor unverified. README claims 3.9+; it's only been exercised on 3.11 and 3.14. Either test the floor in CI or raise the claim to what's actually verified.

Also worth deciding

• Does it belong in this repo at all? This repo is now scoped to harness customization and is docs-only — no build, no tests, no dependencies, as CLAUDE.md states. Adding a Python CLI with a test suite changes that character, and means CI where there is currently none. The alternative is its own repo (cc-audit), linked from here. Worth settling before investing in the above, since it changes where the CI goes.
• Does the --challenge pass stay? It's the most interesting feature and the biggest review surface — it's the only part that sends anything off the machine. Shipping without it is a smaller, safer v1.
• codex-safe.sh lives in audit/ but is documented in review-loops.md as a .claude/hooks/ script. It's unrelated to the audit tool and probably belongs elsewhere regardless of what happens to cc-audit.

Definition of done

• Home for the tool decided (this repo vs. its own)
• Tests covering pricing math, parsing, and the sanitization boundary
• Adversarial tests proving the JSON and --challenge redaction claims
• Pricing staleness handled (stamp, warning, or documented cadence)
• Verified on a second OS and against an absent/empty ~/.claude
• Python floor either tested or the README claim corrected
• CI running the tests, wherever it lands
• --challenge kept or cut, explicitly

Start by reading audit/README.md and the entry points in audit/cc_audit.py, pricing.py, parse.py, store.py, and checks.py; run the existing compile and --help checks. Map the unresolved home, challenge, pricing, platform, and Python-version decisions before adding coverage. Done means the stated arithmetic, parsing, sanitization, redaction, platform, version, and CI requirements are verified and the tool's final home is explicit.