merge_settings cannot remove a stale prompt-type hook group; duplicates accumulate

[millsmillsymills]

Problem

merge_settings() in scripts/install.sh cannot remove a stale prompt-type hook group, so one accumulates permanently in ~/.claude/settings.json.

The merge decides which groups it owns by testing whether any command in the group references /.claude/hooks/:

[ ($eh[$event] // [])[]
  | select(
      [ .hooks[]?.command // "" ]
      | map(contains("/.claude/hooks/")) | any | not
    )
]
+ ($nh[$event] // [])
| unique

A type: "prompt" hook has no command, so .hooks[]?.command // "" yields "", the any is false, and the group is kept as user-authored. The template's group is then appended.

The trailing unique only collapses groups that are byte-identical. When the template's Stop group is [prompt, command] and the stale group is [prompt] alone, the two are distinct objects and both survive — forever, across every re-merge.

Observed

On this machine hooks.Stop held two groups:

1. [prompt] — orphan, prompt only
2. [prompt, command(stop-check-clean-repo.sh)] — the template's

Both carried identical prompt text, so the end-of-turn review ran twice per turn. ./scripts/install.sh settings reported "already up to date -- skipping" because the merge is a fixed point in this state: re-running it reproduces the same two groups.

The code comment at scripts/install.sh:140-141 anticipates this case — "The final unique collapses identical groups, including the prompt-type Stop hook that carries no command for the ours-test to match" — but that only holds while the template's group is also prompt-only. Once a command hook was added alongside the prompt, the two stopped being identical and the orphan became permanent.

Suggested fix

Extend the ours-test so a group is recognized as ours when it has no command hooks but its prompt text matches a prompt in the template's group for the same event. Roughly: treat a group as ours if any command references /.claude/hooks/ or every hook in it is a prompt whose text appears in the template's groups for that event.

Worth adding a regression test that merges a [prompt]-only existing group against a [prompt, command] template group and asserts one group out, not two.

Workaround

Delete the orphan group by hand:

jq --slurpfile c ~/Desktop/Projects/claude-defaults/settings.json \
  '.hooks.Stop = $c[0].hooks.Stop' ~/.claude/settings.json > /tmp/s.json

Already applied here; hooks.Stop is back to a single group.

Related

Found while fixing #158 (permission-denial exceptions in the Stop review prompt).