spendcaplisted

# Spendcap Most routers guess how hard a task is and pick a tier. They infer the task and **assume the wallet**. Spendcap inverts that: ask the human what their economics actually are, compile the answers into a policy file that persists, then route inside those declared limits. Two modes. Pick one before doing anything. - **Configure** — no policy file exists, or the user's plan, balance, or model catalog changed. Run the interview, compile, emit. Go to *Configure*. - **Enforce** — a policy file exists. Read it, route within it, respect its gates. Go to *Enforce*. Check for an existing policy first: `.spendcap/policy.md` in the project, then `~/.spendcap/policy.md`. If one exists and nothing changed, do not re-interview. ## Configure 1. **Scan the project first.** Run the script — do not reimplement it by hand: ```bash python3 scripts/scan.py <path> # human-readable proposal python3 scripts/scan.py <path> --json # machine-readable python3 scripts/scan.py <path> --explain # every signal, including misses ``` It is deterministic, runs in well under a second, reads structure only (paths, counts, extensions, git metadata, config-file existence) and never opens a file. `references/scan.md` documents the signal-to-inference table it implements; read that only when you need to argue with a conclusion or extend the rules. 2. **Discover the real catalog.** Never hardcode model names; they get renamed and retired. `references/tiers.md` covers