prompt-scan

Every Bash call meets tierGate.ts before Claude Code's allow rules, and a runner step gets only what the gate frees. So the Bash allowlist is the tier table and its classifiers, and .claude/settings.json holds what the gate never sees.

Steps

1. Run npm run -s prompt:scan (-- --sessions <n> --top <n>; 50 and 20 by default). It reads every stop in this repo's recent transcripts and prints one table per kind of stop, each ranked by count. The classifier's calls prompt outside auto mode. Each Bash row says what this checkout's gate does with the command now.

2. Give every row one fix. A row the developer or the classifier answered gets it off its "The gate now" column:

   • free: the gate frees it since; no fix.
   • unclassified: a tier-row proposal, its tier set by what undoing the command costs. A read, or a write git or Google's version history undoes, gets a free row; a command that can't be undone or reaches someone else keeps Claude Code's prompt. An npm run script is freed in freeScripts, per subcommand when its subcommands cost different amounts to undo.
   • asks or refuses: the gate means it. Propose a gate fix only when the why shows it misreading a free call, such as a write into /tmp after a cd it can't follow. A row whose why names a free way through, a heredoc beside it included, is an agent habit: propose the doc or skill line that teaches the free way.
   • not a Bash call: a permissions.allow entry by exact tool name, for a read: an MCP get, list, read or search tool, or WebSearch. Writes stay with their gates.
   • An Edit or Write of .claude/settings.json or another protected path: Claude Code's own prompt, fixed like the removal check below.

   A row of the other tables gets it by its kind of stop:

   • A hook refusal: the asks or refuses fix above, read off its Why column. A read-first refusal is the hook working; none.
   • A sandbox denial or a Hand-back: a reshape of the measure that stopped a routine command: a sandbox setting, an excludedCommands entry, or a repo script that does the step.
   • The removal check: none, counted. When it recurs, a doc line teaching the habit that avoids it.

   Done when every row has a fix or a stated reason for none.

3. Show the developer one table: the row, the proposed fix (a tier row and its tier, a settings entry, a doc line, or none) and the undo cost behind it. The developer picks; the pick is the request a tier-rule change needs.

4. Land the picks in a workspace. A gate fix is a tier row, its classifier under .claude/hooks/lib/tiers/ and a case in tiers.test.ts, in one commit. A settings entry goes beside the others. Rerun the scan: each picked Bash row reads free. Then npm run check and wrap-up.