isolated-setup-doctor

setup-isolated-setup-doctor

Runtime routing (run before the Claude-specific probes)

Use the operator's explicitly requested runtime when supplied; otherwise use the active session's runtime. An installed executable or configuration directory alone does not select a runtime. For the routing below, treat that selection as the active harness.

When the active harness is Codex, first require the static verification in docs/adapters/codex.md, then run the shared live environment probes inside the active Codex sandbox. Attribute failures separately to native sandbox/network denial, approval policy, or the POSIX agent-iso layer. Do not prescribe a .claude settings change for a Codex failure. Then stop before the Claude-specific branch below.

When the selected runtime is Gemini CLI, follow docs/adapters/gemini.md: verify the profile first, then diagnose the actual tool result in the active Gemini session. Distinguish policy refusal, sandbox expansion, hook or trust failures, and wrapper or authentication problems. Do not require Claude configuration or prescribe Claude settings changes; then stop before the Claude-specific probes below.

When the harness is Claude Code, continue below. If the harness cannot be determined, ask once.

This is the diagnostic layer over the secure agent setup. verify checks statically that the setup is installed right, catching drift and missing pieces. This skill checks whether common workflows are functionally blocked by the sandbox as configured, catching over-restrictive allowlists. (install puts the setup in place; update reports framework drift.)

Run verify first when the install itself is in doubt: fresh machine, recent upgrade, sandbox-state surprise. Run doctor when the install is known good and a workflow fails in a sandbox-shaped way: agent unreachable, socket error, port permission error.

Every probe maps to a numbered entry in docs/setup/sandbox-troubleshooting.md. The doctor identifies which entry applies now; it does not re-explain the remediation. If a fail shows a failure mode not catalogued there, propose appending a new entry per the catalog's Adding a new entry section.

Golden rules

• Read-only. Each probe is a small, deterministic, side-effect-free check. The skill never edits a settings file, never runs a command with dangerouslyDisableSandbox, and never installs anything. If a check fails, surface it and point at the catalog entry; do not auto-fix.
• Run every probe, even on early failure. Do not stop at the first ✗. A user may have one of eight independent restrictions or all eight, and finding them one re-run at a time is annoying.
• Distinguish ✗ (failing) from ⊘ (not applicable). ✗ means the probe ran and the sandbox blocked it. ⊘ means the probe was skipped because a prerequisite is absent (e.g. no docker / podman on PATH → docker probe ⊘, not ✗).
• Surface evidence. Each report line names the probe command, the exit code, and the relevant stderr snippet. "Looks blocked" is not useful; "ssh-add -l → rc=2 → Could not open a connection to your authentication agent" is.
• Map each ✗ to a catalog entry. The fail report links directly to the matching section of docs/setup/sandbox-troubleshooting.md. Do not paraphrase the remediation; the catalog is the single source of truth.

The 8 probes

The probes cover the catalog's failure modes that a sandboxed command can detect on its own; the signing entries that need a live touch or a terminal are verified by setup-isolated-setup-verify check 10 instead. New probes are added when new entries land in the catalog, so the two stay in lock-step.

Probe 1 — SSH agent / Yubikey reachable

Tests whether ssh-agent is reachable from inside the sandbox. Two failure modes: SSH_AUTH_SOCK passes through claude-iso's env whitelist but the socket file is not in sandbox.filesystem.allowRead, so the agent's ssh / git push subprocesses cannot even stat(2) it; or, on macOS, the file is readable but its path is missing from sandbox.network.allowUnixSockets, so connect(2) is denied and the agent looks unreachable while the socket is plainly there. A signed commit hits the second as No private key found for public key.

Command:

bash <skill-dir>/scripts/probe-1-ssh-agent.sh

Interpretation:

Result	Status	Meaning
✓ N identities listed	Pass	ssh-add -l returned the key list.
✓ agent reachable, no identities	Pass	ssh-add -l returned rc=1 (the documented "no keys" exit).
✗ socket not stat-able	Fail	Sandbox blocks stat(2) on the socket file.
✗ agent unreachable	Fail	Sandbox blocks connect(2) to the socket.
⊘ SSH_AUTH_SOCK not set	Skip	Either the user does not run ssh-agent, or claude-iso's env whitelist dropped it (separate bug — verify).

On ✗ → remediation: docs/setup/sandbox-troubleshooting.md — SSH agent / Yubikey appears unreachable from inside the sandbox.

Probe 2 — Localhost port bind

Tests whether a process inside the sandbox can bind to a loopback port AND then talk to itself over loopback. The catalog documents the second half failing (egress proxy blocks 127.0.0.1).

Command:

bash <skill-dir>/scripts/probe-2-localhost-bind.sh

Interpretation:

Result	Status	Meaning
✓ bound + loopback GET → HTTP 200	Pass	Both bind and loopback HTTP work.
✗ bind: [Errno 1] Operation not permitted	Fail	The sandbox refuses listening sockets outright; sandbox.network.allowLocalBinding is unset. Fails before any egress rule is consulted, so allowedDomains changes do not help.
✗ bind: ... (other errno)	Fail	The sandbox blocks bind(2) on 127.0.0.1 for another reason. Rare; report the literal error.
✗ bind ok, loopback GET: ...	Fail	Bind works but the sandbox egress proxy refuses 127.0.0.1 as a destination. Common shape.

On ✗ → remediation: docs/setup/sandbox-troubleshooting.md — Test cannot bind to a localhost port.

Probe 3 — Podman / Docker through the container gateway

Tests whether the runtime CLI can talk to the container gateway, not the real daemon socket; the sandbox never gets a route to the daemon itself. Run it for each of podman / docker on PATH; ⊘ each one not installed (an absent prerequisite, not a sandbox failure). The remaining checks narrow down which of the three wiring pieces (env var, running gateway, allowed socket) is missing, in the order a fresh install would hit them.

Command:

bash <skill-dir>/scripts/probe-3-container-gateway.sh

status_json comes from the gateway's own read-only status subcommand. The doctor may call it from inside the sandbox, since it neither binds a socket nor touches the daemon. gw_src picks the adopter's pinned snapshot (.apache-magpie/tools/container-gateway/src) when present, else the framework repo's own tree (tools/container-gateway/src), so the same probe runs in an adopter checkout and in this framework's own worktree. The gw_state check runs before the raw socket-file test. A backend that status does not list under serving never gets a socket file, so testing -S "$sock" first would misreport "gateway not running" for the "running, but this backend's machine/daemon is down" case. The -S test is a defensive fallback for an already-serving backend whose socket vanished mid-probe, not the primary check.

Interpretation:

Result	Status	Meaning
✓ <rt> reaches the container gateway at <sock>	Pass	CLI → gateway → daemon all answer.
✗ … unset — gateway not wired into settings	Fail	The reference env block is missing from project settings.
✗ gateway socket missing	Fail	The SessionStart hook did not start the gateway, or it exited; check <project>/.apache-magpie-local/run/container-gateway.log.
✗ gateway running without a <rt> backend	Fail	status reports the gateway up but serving does not list this CLI's backend — the Podman machine or Docker daemon behind it is not running. Start it from outside the sandbox, then restart the gateway.
✗ connect … denied	Fail	The gateway socket is not in sandbox.network.allowUnixSockets.
✗ gateway up, backend down	Fail	Podman machine / Docker not running on the host; start it from your own terminal.
✗ no response in 15s — <rt> info hung	Fail	_probe_timeout killed a stalled <rt> info call; the backend daemon behind the gateway is likely wedged — restart it from outside the sandbox.
⊘ <rt> not on PATH	Skip	Runtime not installed; not a sandbox restriction.

An empty podman machine list from inside the sandbox is a read denial on the machine's directory, not proof that no machine exists. Decide the machine's real state from outside the sandbox, per the catalog entry below.

On ✗ → remediation: docs/setup/sandbox-troubleshooting.md — Docker / Podman command fails with a socket error.

Probe 4 — Per-project scratch directory (TMPDIR)

Tests whether the session has a writable scratch directory. The sandbox mounts the host /tmp read-only and makes only specific subpaths writable, so a session whose TMPDIR falls back to /tmp has no scratch area at all.

TMPDIR landing on the shared session root rather than a per-project directory is not a finding. Claude Code sets TMPDIR itself when it builds the sandbox, and that wins over env.TMPDIR from any settings file, so the shared root is expected and no configuration changes it. Each session still gets a per-project, per-session scratchpad beneath it.

Command:

bash <skill-dir>/scripts/probe-4-scratch-dir.sh

Interpretation:

Result	Status	Meaning
✓ per-project + writable	Pass	TMPDIR resolves under this project's path slug and accepts writes.
✓ writable; shared session root	Pass	The expected value on current Claude Code. Every project on the machine shares this directory, so write through the per-session scratchpad beneath it, or use unique filenames — but there is nothing to fix.
✗ TMPDIR not set	Fail	Tooling falls back to /tmp, which is read-only inside the sandbox.
✗ directory missing	Fail	TMPDIR names a path nothing has created yet.
✗ not writable inside sandbox	Fail	TMPDIR points outside sandbox.filesystem.allowWrite.

On ✗ → remediation: docs/setup/sandbox-troubleshooting.md — Temp files fail with "Read-only file system" under /tmp.

Do not propose env.TMPDIR in a settings file as the fix. Claude Code overrides it when it builds the sandbox, so the setting is accepted and silently has no effect. The giveaway is a directory that exists, is named exactly as configured, and stays empty. The catalog entry above covers what is actually actionable.

Probe 5 — Signing key readable (gpg.format=ssh)

Tests whether the public key git hands to ssh-keygen -Y sign can be opened from inside the sandbox. Failure mode: the framework denies ~/.ssh/ wholesale, so with gpg.format=ssh every signed commit fails before the hardware key asks for a touch. The touch overlay, which waits for ssh-keygen to block, never sees it block.

Command:

bash <skill-dir>/scripts/probe-5-signing-key.sh

Interpretation:

Result	Status	Meaning
✓ readable inside sandbox	Pass	ssh-keygen will be able to open the public key.
✓ literal key	Pass	user.signingkey holds the key text itself; no file is involved.
✗ not readable inside sandbox	Fail	The sandbox's ~/.ssh/ read deny covers the public key; add that one file to allowRead.
⊘ gpg.format is not ssh	Skip	Signing goes through gpg (or is off); the agent-socket rules behind Probe 1 are what matter.
⊘ user.signingkey unset	Skip	Misconfigured signing, not a sandbox problem — mention it, do not fail the probe.
signing-program → ✓	Pass	git's signing program (the touch overlay's gpg-touch-wrap-* wrapper, or whatever gpg.ssh.program / gpg.program names) can be started from inside the sandbox. Nothing printed when neither key is set: git uses its default ssh-keygen / gpg from PATH.
signing-program → ✗	Fail	The program git is configured to sign with is read-denied inside the sandbox — for the overlay wrapper, ~/.claude/scripts/ is. Every sandboxed signed commit fails at once with cannot exec; add the wrapper's two files to allowRead.

On ✗ → remediation: docs/setup/sandbox-troubleshooting.md — Signed commit fails before any touch when git signs with ssh for the key, and docs/setup/sandbox-troubleshooting.md — Signed commit fails with "cannot exec" of the touch-overlay wrapper for the program.

Probe 6 — gh runs outside the sandbox

Tests whether gh can reach GitHub from a sandboxed Bash call, and if not, whether the sandbox.excludedCommands: ["gh *"] exclusion the framework reference relies on is in place. On macOS a sandboxed gh cannot verify TLS or read the keychain (x509: OSStatus -26276 / HTTP 401), so only the exclusion makes it work. The exclusion applies only when every segment of a Bash invocation is cd … or gh ….

The probe deliberately runs gh through sh -c so the exclusion cannot apply to it, showing what an un-excluded gh does on this machine.

Command:

bash <skill-dir>/scripts/probe-6-gh-outside-sandbox.sh

Interpretation:

Result	Status	Meaning
✓ gh works inside the sandbox	Pass	Platform lets gh verify TLS and read its token inside the sandbox (typical on Linux).
✓ … "gh *" is in excludedCommands	Pass	The known macOS shape, and the framework's exclusion is present. Calls still fail if they are not cd/gh-only invocations — see the catalog entry.
✗ … NOT found in excludedCommands	Fail	gh cannot work inside the sandbox on this machine and nothing runs it outside.
⚠ gh failed for another reason	Warn	Not the catalogued shape (network down, not logged in, …); inspect the message.
⚠ catch-all "Bash(gh *)" in permissions.ask	Warn	Ask beats allow regardless of specificity, so this rule prompts on every read-only gh call. Replace it with the explicit write-subcommand list from the reference .claude/settings.json. Extra line, printed after the main result.
⊘ gh not on PATH	Skip	gh not installed; not a sandbox restriction.

~/.claude/settings.json is usually unreadable from inside the sandbox, so the exclusion check may see only the project-scope files. If the user keeps the exclusion at user scope, a ✗ here is a false alarm; say so when reporting.

Even with the exclusion present, a gh call is excluded only when every part of the Bash invocation is cd … or gh …. A pipe, a $(…) substitution, a loop, or any file redirection (> file, even > /dev/null) puts it back in the sandbox. The redirection case is a Claude Code regression tracked in anthropics/claude-code#95532; the catalog entry shows the gh tofile alias that works around it. When the user reports a gh failure this probe does not reproduce, ask for the exact command line; the shape is usually the answer.

On ✗ → remediation: docs/setup/sandbox-troubleshooting.md — gh fails with TLS OSStatus -26276 or HTTP 401 inside the sandbox.

Probe 7 — prek and uv usable inside the sandbox

Tests whether the dev tools the framework's hooks and Python tools run on are found and can write their cache from a sandboxed Bash call. They live in ~/.local/bin, which the harness reads only when the worktree's .claude/settings.local.json grants it; the committed project-scope entries are not applied. Inside the sandbox a read-denied directory looks exactly like an absent one, so the probe reads settings.local.json to tell a missing grant from a tool that is not installed.

Command:

bash <skill-dir>/scripts/probe-7-dev-tools.sh

Interpretation:

Result	Status	Meaning
✓ (found: …; ~/.cache writable)	Pass	prek / uv run inside the sandbox and can write their caches.
✗ (found: …; ~/.cache not writable …)	Fail	The tools start but cannot write their cache or state.
⚠ (… not found; ~/.local/bin is not in … settings.local.json)	Warn	The grant is missing; the tools are most likely installed but hidden.
⊘ (… not found; ~/.local/bin is granted …)	Skip	The grant is present and neither tool is installed there; not a sandbox restriction.

On ✗ or ⚠ → remediation: docs/setup/sandbox-troubleshooting.md — prek or uv not found, or cannot write its cache, inside the sandbox.

Probe 8 — Global git hook dir readable (whole-user scope)

Tests whether git run inside the sandbox can see the hooks a global core.hooksPath points at. Git treats a hook it cannot see as one that does not exist, so an unreadable hook dir costs every hook — pre-commit and prek included — without an error. Only a look from inside the sandbox shows it, which is where this probe runs.

Command:

bash <skill-dir>/scripts/probe-8-git-hooks.sh

Interpretation:

Result	Status	Meaning
✓ (… readable; hooks visible: …)	Pass	Sandboxed git runs the listed hooks.
✗ (core.hooksPath … not readable inside sandbox …)	Fail	The hook dir is hidden; every sandboxed commit skips every hook.
✗ (… readable, but not the target of: …)	Fail	The dir is visible but a hook's symlink target is not — typically hooks symlinked from a sync repo the sandbox does not grant.
⚠ (… holds none of pre-commit, commit-msg, pre-push, post-checkout)	Warn	core.hooksPath is set to a dir with none of the common hooks; check it is the intended one.
⊘ (no global core.hooksPath …)	Skip	Per-project scope: hooks live in each repo's .git/hooks, inside the project root.

On ✗ → remediation: docs/setup/sandbox-troubleshooting.md — Git hooks silently skipped for commits made inside the sandbox.

After the report

If every probe is ✓ or ⊘:

All eight probes pass (or are not applicable). The sandbox is not currently blocking the known failure modes catalogued in docs/setup/sandbox-troubleshooting.md. If you hit a different sandbox-shaped failure, follow the catalog's Adding a new entry section and (optionally) extend this skill with another probe so future runs catch the same shape automatically.

If any probe is ✗:

1. Surface every fail in one report (do not stop at the first).
2. For each fail, print the troubleshooting-doc anchor link from the probe's On ✗ → remediation line above.
3. Suggest the user read the catalog entry's symptom → root cause → fix, then apply the settings.json widening themselves. Do not propose to apply the widening from this skill; settings.json widenings are sandbox-bypass-adjacent and need an explicit user-driven edit.
4. After the user has applied the widening (in a separate flow), re-run setup-isolated-setup-doctor to confirm the probe now passes.

If a probe surfaces a fail shape not catalogued in docs/setup/sandbox-troubleshooting.md:

1. Report the fail with the literal probe command + exit code + stderr.
2. Suggest the user add a new entry to the catalog per its Adding a new entry section (symptom verbatim, root cause, fix, notes).
3. Once the catalog has the new entry, extend this skill with a matching probe in the same shape so the next doctor run catches it automatically.

Adding a probe

When the troubleshooting catalogue grows an entry, add a matching probe: scripts/README.md.