quest: plan dev shells from user settings instead of repo hooks

quest/m0/README.md

@@ -104,8 +104,7 @@ ownership-manifest design:
 - [Gate bench](/quest/m0/gate-bench.md) - measure headroom admission on an idle host
 - [Gate load wording](/quest/m0/gate-load-wording.md) - stop naming Spotlight as the cause of a fresh macOS runner's load
 - [Agent workflow helpers](/quest/m0/helpers/README.md) - CLI commands for CI waits, PR status, claims, quest moves, and serialized merges
-- [Worktree hooks here](/quest/m0/worktree-hooks.md) - `quest` reaches Claude Code's hooks and the merge driver in this repository, then `quest init` installs the hooks
-- [Wrapped worktree hooks](/quest/m0/wrapped-hooks.md) - `quest init` accepts a hook that runs `… worktree hook` through a wrapper such as `nix run`
+- [Agents get the dev shell from user settings](/quest/m0/user-dev-shell/README.md) - maintainers' own Claude and Codex settings load the dev shell; repositories drop dev-shell hooks, Quest worktree hooks, and `nix develop` instructions, and `quest init` writes no hooks
 - [Remote control worktrees](/quest/m0/worktree-routes.md) - live-check `claude remote-control --spawn worktree` against the hooks, from a session spawned on claude.ai
 - [Session checkouts released](/quest/m0/session-checkouts.md) - a Claude session's own checkout is released when the session ends and reclaimed on resume
 - [claude rm upstream report](/quest/m0/claude-rm-upstream.md) - waiting on approval: ask Claude Code to run `WorktreeRemove` instead of advising an unlock

quest/m0/session-checkouts.md

@@ -23,12 +23,24 @@ Evidence, gathered 2026-11-02:
 
 Work:
 
-- Install `SessionEnd` and `SessionStart` hook entries through
-  `quest init`, recognised by `init` and `uninstall` like the others.
+- Add `SessionEnd` and `SessionStart` entries to the opt-in user-level hook
+  setup the docs describe. Amended 2026-11-03 (maintainer, /quest-unblock):
+  `quest init` writes no hook entries anymore
+  ([quest init writes no hooks](/quest/m0/user-dev-shell/init-hooks.md)),
+  so this quest no longer installs them through `init`.
 - `SessionEnd` releases (never removes) the session's checkout;
   `SessionStart` claims it again on resume. `/clear` also fires
   `SessionEnd`, so release is idempotent and the next `SessionStart`
   reclaims it. An idle background session never ends; leave it held.
+- Remote control's `bridge-*` checkouts are session checkouts too, and the
+  release covers them if `SessionEnd` fires in a bridge-spawned session.
+  [Remote control's live check](/quest/m0/worktree-routes.md) confirms
+  whether it does. Release only prunes the declared scratch: it keeps every
+  other file and Git's worktree lock (`release` and `prune` in
+  `src/cli/worktree.ts`; only `remove` unlocks). The bridge still decides at
+  session end, as that quest describes: it removes a clean checkout through
+  `WorktreeRemove` and keeps one with changes. Verify that order in the same
+  live check.
 - Claude Code bounds `SessionEnd` hooks by a short budget
   (`CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS`) and does not wait past it.
   Measure a release with scratch to delete, and set the entry's timeout to

quest/m0/user-dev-shell/README.md

@@ -0,0 +1,77 @@
+# Agents get the dev shell from user settings, not repository config
+
+## Goal
+
+Agents in moq-dev's repositories run tools without caring about Nix: each
+maintainer's own Claude Code and Codex settings load the repository's dev
+shell into every session's commands, so no repository forces direnv, Nix, or
+`quest` on its contributors through checked-in hooks. No repository carries a
+dev-shell `SessionStart` hook, a Quest worktree hook, or AGENTS.md
+instructions to wrap commands in `nix develop`, and `quest init` writes no
+hook entries.
+
+## Plan
+
+Decided 2026-11-03 (maintainer, /quest-unblock of
+[moq-dev/quest#423](https://github.com/moq-dev/quest/pull/423)'s wait to
+install `quest` on Claude's `PATH`):
+
+- Install `quest` on the `PATH` Claude Code starts with? ✅ No (rejected:
+  `nix profile add github:moq-dev/quest` on every host).
+- How should sessions reach a pinned `quest` and the rest of the dev shell?
+  ✅ A dev-shell hook in the maintainer's own Claude settings (rejected: hook
+  entries that `nix run` the repository's pinned input, as moq-dev/moq does;
+  dropping the hooks with no replacement). Reason: "We don't want to force
+  direnv+nix for other users"; repository config applies to every
+  contributor.
+- Repository-level direnv `SessionStart` hooks in quest, moq, moq.pro, rind,
+  and rust2? ✅ Remove them from all five once the user-level hook loads the
+  dev shell (rejected: keep them).
+- Quest's worktree hooks? ✅ Drop them from repositories, and `quest init`
+  stops writing hook entries; the hooks become an opt-in user-level setup
+  (rejected: this repository and moq only; this repository only; moving them
+  into the maintainer's settings).
+  [moq-dev/quest#423](https://github.com/moq-dev/quest/pull/423) is closed.
+- Codex? ✅ A user-level Codex `SessionStart` hook does the same capture, and
+  the repositories' `.codex/` dev-shell hooks go (rejected: Claude only).
+- A fresh worktree needs `direnv allow`: direnv keys each approval by the
+  SHA-256 of the `.envrc`'s absolute path and contents
+  (`~/.local/share/direnv/allow/`, 1,706 entries on the maintainer's host on
+  2026-11-03), so nothing carries over from the main checkout. ✅ The
+  user-level hook allows a checkout itself when its `.envrc`, `flake.nix`,
+  and `flake.lock` match the remote's default branch (rejected: whitelisting
+  `~/work` and `~/worktrees` in `direnv.toml`, which would also evaluate a
+  fork PR's dev shell; keeping manual allows). A branch that changes them
+  still needs a manual allow. Amended 2026-11-03 (review of
+  [moq-dev/quest#458](https://github.com/moq-dev/quest/pull/458)): those
+  three files alone don't make a checkout reviewed. Any repository's fresh
+  clone matches its own default branch. A fork PR head can change another
+  file the dev shell runs, such as moq.pro's `dev/mbx-shim.sh`. Both were
+  auto-allowed in a reproduction against the merged hook.
+  Decided 2026-11-03 (maintainer, /quest-complete), after weighing the
+  risk as narrow: a session rooted at an outside contributor's fork head,
+  which the guide already says never to run, since agents check untrusted
+  heads out under `.tmp/` rather than as a session root, and an agent
+  session in any checkout runs its code anyway. ✅ One cheap guard: the
+  hook also requires `HEAD` to be on some `origin` branch
+  (`git for-each-ref --contains HEAD refs/remotes/origin/`), so a fork PR
+  head is never auto-allowed (rejected: an owner allowlist, a clean tree,
+  and a tracked `.envrc` on top; matching the whole tree to the merge base;
+  an owner allowlist alone; leaving it as is). Shipped in
+  [kixelated/dotfiles#35](https://github.com/kixelated/dotfiles/pull/35)
+  and live on `desktop` and `nixos2`, where a fresh worktree of moq-dev/moq
+  still loads and one on an unpushed commit is refused.
+- AGENTS.md instructions to run tools through `nix develop`? ✅ Remove them in
+  all five repositories once the user-level hooks are on every host the
+  maintainer runs agents on (rejected: this repository only; keeping them as
+  a fallback).
+
+The user-level hooks live in [kixelated/dotfiles](https://github.com/kixelated/dotfiles),
+installed into each Claude profile's `settings.json` and Codex's user
+config, and generalize this repository's `.claude/hooks/direnv.sh` and
+moq-dev/moq's `.codex/hooks/direnv.sh`.
+
+## Required
+
+- [quest init writes no hooks](/quest/m0/user-dev-shell/init-hooks.md) - `quest init` stops writing hook entries and removes the ones it or a wrapper wrote; the docs describe the hooks as user-level opt-in
+- [Repositories drop dev-shell config](/quest/m0/user-dev-shell/repo-config.md) - quest, moq, moq.pro, rind, and rust2 drop their dev-shell hooks, Quest worktree hooks, and `nix develop` instructions

quest/m0/user-dev-shell/init-hooks.md

@@ -0,0 +1,42 @@
+# [S] quest init writes no hook entries
+
+## Goal
+
+`quest init` no longer writes Claude Code hook entries into a repository's
+`.claude/settings.json`, and removes the `quest worktree hook` entries an
+earlier version wrote, including ones a repository wrapped, such as
+moq-dev/moq's `nix run "${CLAUDE_PROJECT_DIR}#quest" -- worktree hook`.
+`quest uninstall` removes both forms too. The README and SETUP.md describe the
+worktree hooks as an opt-in a user adds to their own Claude settings, with
+the entries to add.
+
+## Plan
+
+Decided 2026-11-03 (maintainer, /quest-unblock): repository config applies
+to every contributor, so Quest's hooks become user-level opt-in; see the
+[epic](/quest/m0/user-dev-shell/README.md).
+
+- Replaces the wrapped-hooks quest (moq-dev/moq, 2026-11-03), which widened
+  `isQuestHook` (`src/core/init.ts`) so `init` would accept wrapped hooks.
+  For removal, a command hook is Quest's only when it runs Quest: the plain
+  `quest worktree hook`, or a wrapper whose command ends in `worktree hook`
+  and names the `quest` executable or package as a word or flake attribute,
+  such as moq's `nix run "${CLAUDE_PROJECT_DIR}#quest" -- worktree hook`.
+  Another tool's command that merely ends in `worktree hook` stays.
+- `HOOK_EVENTS` and the hook plan in `src/core/init.ts` stop adding
+  entries; `init` reports each entry it removes. The enrollment setup PR
+  ([moq-dev/quest#431](https://github.com/moq-dev/quest/pull/431)) uses the
+  same plan, so it stops proposing hooks too.
+- The `quest worktree hook` command stays, for users who opt in.
+- Removing moq's four wrapped entries, whether `init` here or
+  [repo-config](/quest/m0/user-dev-shell/repo-config.md) does it first, ends
+  Quest's reclaim of Claude Code checkouts in moq. Nothing replaces it: the
+  maintainer's user-level setup (kixelated/dotfiles) loads only the dev
+  shell, and the epic's decision rejected moving Quest's hooks into the
+  maintainer's settings. So this quest doesn't wait on that setup, and the
+  same end state results in either order. A user who wants reclaim adds the
+  opt-in entries the docs describe. Any removal through the enrollment setup
+  PR shows in that PR's diff.
+- Test `init` and `uninstall` on a repository with plain and wrapped
+  entries, and with other hooks beside them, which stay, including another
+  tool's command that ends in `worktree hook`.

quest/m0/user-dev-shell/repo-config.md

@@ -0,0 +1,31 @@
+# [S] Repositories drop their dev-shell config
+
+## Goal
+
+moq-dev/quest, moq, moq.pro, rind, and rust2 carry no dev-shell hooks and no
+instructions to wrap commands in Nix: each drops its direnv `SessionStart`
+hook from `.claude/settings.json` and `.codex/hooks.json` with the scripts
+and tests behind them, moq drops its four Quest worktree hook entries, and
+each AGENTS.md drops its `nix develop` instructions. Agents there still run
+the pinned tools, through the maintainer's user-level hooks. One PR per
+repository.
+
+## Plan
+
+Decided 2026-11-03 (maintainer, /quest-unblock); see the
+[epic](/quest/m0/user-dev-shell/README.md).
+
+- Keep every other hook: moq.pro's `submodule.sh`, `worktree.sh` (its own
+  submodule-aware worktrees, not Quest's), and `tofu-secrets.py`.
+- CONTRIBUTING.md's human setup (`nix develop` to match CI) stays; only the
+  agent-facing instructions go. In this repository, AGENTS.md's "`quest`
+  means `bun run --silent quest`" note stays, since it names which `quest`
+  to run, not how to load it.
+- Check each repository from a fresh worktree on a host with the user-level
+  hooks: a new Claude Code session and a new Codex session both resolve the
+  dev shell's tools (`which just`, `quest --version` where pinned) with no
+  `nix develop` and no manual `direnv allow`.
+
+## Waiting
+
+- Maintainer: pull kixelated/dotfiles and run `bin/install-dev-shell.py` (or the full `install`, with `codex` on `PATH`) on `nixos` (`kixtop` has neither direnv nor Nix, so the hook has nothing to load there; it got only the effort defaults on 2026-11-03); it changes the live agent config on each host, which an agent here doesn't touch unasked. The repositories' hooks must not go before the user-level ones arrive. [kixelated/dotfiles#33](https://github.com/kixelated/dotfiles/pull/33) is merged. On 2026-11-03, `desktop` and `nixos2` had the hook in `~/.claude-personal/settings.json`, but `nixos` and `kixtop` didn't. Check on each host with direnv and Nix (`desktop`, `nixos2`, `nixos`) that `~/.claude-personal/settings.json` has the dev-shell `SessionStart` hook and that a fresh worktree's session resolves `just`; on `kixtop` the hook is an expected no-op

quest/m0/worktree-routes.md

@@ -23,7 +23,9 @@ agent can't do, so its README description comes from Claude Code's own code:
 - The bridge never runs `git worktree unlock` on a hook-created checkout,
   since it records no Git root for one, so quest's lock stays in place.
 
-Check it from a scratch repository set up with `quest init`, with
+Check it from a scratch repository with the worktree hooks configured (the
+opt-in user-level setup once
+[quest init writes no hooks](/quest/m0/user-dev-shell/init-hooks.md) lands), with
 `QUEST_WORKTREE_HOME` isolated: start `claude remote-control --spawn worktree`,
 spawn a session from claude.ai, end it once without changes and once with an
 untracked file, and watch `quest worktree status`. Fix the README if it