Per PR DiamondLightSource#337 review: nodejs, gh, and glab are slated to move into the
ubuntu-devcontainer base image once it ships on Ubuntu 26.04, where
all three are available from apt at recent enough versions. Add TODO
comments at each install block so the cleanup is obvious when 26.04
lands.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

VS Code's default port auto-detection forwards the port when the user
clicks the link in sphinx-autobuild output, so the explicit
forwardPorts: [8000] and remote.autoForwardPorts: false override
weren't pulling their weight as defaults. Per PR DiamondLightSource#337 review: this
was sphinx-autobuild-specific config that crept in and shouldn't ship
with every sphinx project.

Removed from both the meta repo's own devcontainer.json (the
add_claude=no baseline) and the templated copy, keeping the drift
test happy.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Previously postCreate.sh only existed when add_claude=yes; the
add_claude=no case used a long inline bash string in
postCreateCommand. Tidier to always have a script and gate the
Claude-specific bits with Jinja {% if add_claude %} inside it, per
PR DiamondLightSource#337 review.

The meta repo now has its own .devcontainer/postCreate.sh that
mirrors the add_claude=no rendering, and its devcontainer.json
calls it the same way. Drift test still passes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

git submodule update --init runs on every postCreate, which on
container rebuild moves any submodule whose HEAD differs from the
parent's pinned SHA into detached HEAD — silently yanking
in-progress branch work in initialized submodules. The data isn't
lost (reachable via reflog and existing branches) but recovery
requires submodule literacy that not every user has.

Filter to only the uninitialized submodules using git submodule
status output (leading '-'), preserving the first-clone build
protection without touching already-checked-out submodules. Applies
to all generated projects, not just the Claude-sandbox case.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Reinstate the README → README-CLAUDE.md table row removed in the
previous commit, using {{repo_url}}/blob/main/... so the link works in
both renderers:

- myst (Sphinx) treats it as an external URL and skips xref resolution,
  so --fail-on-warning no longer trips on myst.xref_missing.
- GitHub renders it as a normal link to the file in the repo.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The previous absolute URL pointed at the example repo's main branch,
which won't have README-CLAUDE.md until the example is regenerated
with add_claude=yes — chicken-and-egg, so linkcheck 404s.

Raw <a href="./README-CLAUDE.md"> works in both renderers without that
dependency:
- myst: passes through as inline HTML, no xref_missing warning.
- sphinx linkcheck: doesn't follow raw HTML hrefs (verified locally),
  so no 404.
- GitHub: renders as a normal clickable link to the file in the repo.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Always install gh and glab when add_claude is on. The base image will
ship them on Ubuntu 26.04, at which point the Dockerfile install blocks
can be dropped wholesale; in the meantime, having them unconditional
removes two questions and the matching jinja gating across the
Dockerfile, devcontainer.json, postStart.sh and justfile.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

If the user opens the devcontainer before running `git init`, uv sync
fails opaquely because setuptools-scm can't compute a version, and
pre-commit can't install hooks into a non-existent .git/hooks. VS
Code's wrapper then hides the real stderr behind a generic
"postCreateCommand failed" exit-1.

postCreate.sh now detects a missing .git directory up front, prints
a clear explanation to stderr, and exits 1. The user is told exactly
what to run on the host to recover. Failing fast beats silently
bootstrapping a half-broken environment.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…derived skill triggers

Three sandbox guards (SSH_AUTH_SOCK, VSCODE_GIT_IPC_HANDLE, GIT_ASKPASS) match
exactly what `just claude` clears, so the failure messages now name that
recipe. The other guards keep their original remediations because `just
claude` would not fix them.

Also extend the copier-derived skill description to fire on .claude/,
CLAUDE.md, and README-CLAUDE.md edits — these are template-managed but
weren't listed as triggers, so the skill was being missed for sandbox edits.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…tings

Pin three Dev Containers settings in devcontainer.json so VS Code never
forwards host git credentials into the container in the first place,
rather than relying on env-var clearing in `just claude`:

  git.terminalAuthentication: false              -> stops VS Code's Git
    extension from setting GIT_ASKPASS / VSCODE_GIT_IPC_HANDLE in
    integrated terminals (source: vscode/extensions/git/src/askpass.ts)
  dev.containers.gitCredentialHelperConfigLocation: "none"
                                                 -> stops Dev Containers
    from writing a credential.helper line into /etc/gitconfig
  dev.containers.copyGitConfig: false            -> stops the host
    ~/.gitconfig (with its url.insteadOf rewrites and per-host helpers)
    from being copied into the container

A compromised model can no longer reach a leaked VS Code IPC socket by
editing the hook or recipe, because the socket is never wired up.

Knock-on cleanup:
- devcontainer.json: drop the now-redundant remoteEnv blanks for
  GIT_ASKPASS / VSCODE_GIT_IPC_HANDLE / VSCODE_GIT_ASKPASS_{MAIN,NODE,
  EXTRA_ARGS}. The new settings prevent VS Code from setting them at
  all; sandbox-check.sh catches any regression. SSH_AUTH_SOCK= stays
  in remoteEnv — there is no VS Code setting alternative for it.
- justfile: drop VSCODE_GIT_IPC_HANDLE= / GIT_ASKPASS= from `just claude`
  for the same reason; SSH_AUTH_SOCK= stays.
- sandbox-check.sh: env-var failures for the credential bridge now
  point at "rebuild the devcontainer" since `just claude` no longer
  fixes them.
- README-CLAUDE.md: rewrite the credential-injection section to lead
  with the settings as the primary defence; postStart.sh and the
  sandbox-check.sh hook are belt-and-braces verifications, not the
  only enforcement.

Requires devcontainer rebuild for the settings to take effect.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Block prompt submission when a VS Code reconnect drops either gitconfig
bind: /root/.gitconfig grep'd for vscode-remote-containers credential
bridges, /etc/gitconfig checked via `git config --system --list` so any
leaked system-scope setting (core.hooksPath, url.insteadof, http.proxy,
credential.helper) trips the hook, not just credential helpers.

verify-sandbox skill: replace the narrow 6c (system credential.helper
only) with 6c (/etc/gitconfig masked) + 6d (system scope is empty).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Jinja's default settings keep the newline after `{% endif %}`, so the
rendered postCreate.sh ended with `bash\n\n` (a trailing blank line).
end-of-file-fixer in pre-commit flagged this, breaking lint/run and
all `test_template_*` jobs that exercise the generated project.

`{% endif -%}` strips the trailing newline; removing the blank source
line above `{% if %}` keeps the spacing tidy when add_claude is false.
Also normalises the meta repo's own .devcontainer/postCreate.sh to
match.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>