isolated-setup-install

Walk an adopter through the first-time install of the secure agent setup (sandbox, approval and clean-environment layers) for Claude Code, Codex or Gemini CLI. Interactive throughout; never runs sudo,

SKILL.md

GitHub
<!-- Placeholder convention (see AGENTS.md#placeholder-convention-used-in-skill-files): <project-config> → adopting project's `.apache-magpie/` directory <tracker> → value of `tracker_repo:` in <project-config>/project.md <upstream> → value of `upstream_repo:` in <project-config>/project.md -->

setup-isolated-setup-install

Runtime routing (run before the Claude-specific procedure)

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.

Determine the active harness from the session metadata and executable. When it is Codex, follow the install lifecycle in docs/adapters/codex.md, including the project profile merge, static lint, project trust, and agent-iso codex steps. Do not write Claude settings or report a missing .claude file as a Codex failure. After completing the Codex branch, stop; the remainder of this skill is the Claude Code branch.

When the selected runtime is Gemini CLI, follow docs/adapters/gemini.md: resolve the existing extension, snapshot, or framework checkout; propose the workspace profile and guard merge; then guide the wrapper launch and verification. Do not require or write Claude configuration. After completing the Gemini branch, stop before the Claude-specific procedure below.

When the selected runtime is Copilot CLI, follow docs/adapters/copilot.md: register the guard hook, detect and declare mail sources from the MCP servers already present, and ask nothing beyond the single apply confirmation (PRINCIPLES.md §1, Avoiding prompt fatigue). Do not require or write Claude configuration. After completing the Copilot branch, stop before the Claude-specific procedure below.

When the harness is Claude Code, continue below. If the harness cannot be determined from the session or executable, stop and report that; never guess and never prompt.

This skill is the on-ramp for adopters who do not yet have the secure setup running. It is a thin walkthrough wrapper around the canonical install path documented in docs/setup/secure-agent-setup.md. The authoritative content lives there; this skill exists so an adopter can say "set up the secure agent setup" in a fresh session and land in the right step-by-step flow without first reading the document.

Adopter overrides

<!-- BEGIN MAGPIE BLOCK: adopter-overrides — generated from tools/dev/blocks/adopter-overrides.md -->

Before running its default behaviour, this skill consults setup-isolated-setup-install.md in the personal layer (.apache-magpie-local/ when the project adopted Magpie, falling back to the main checkout's in a linked worktree, or <git-common-dir>/apache-magpie/ when Magpie is only installed; applied first, wins on conflict) and .apache-magpie-overrides/setup-isolated-setup-install.md (committed, project-wide) in the adopter repo, if present, and applies any agent-readable overrides it finds. See docs/setup/agentic-overrides.md for the contract.

Hard rule: agents NEVER modify the snapshot under <adopter-repo>/.apache-magpie/. Local modifications go in the override file; framework changes go via PR to apache/magpie.

<!-- END MAGPIE BLOCK: adopter-overrides -->

Snapshot drift

Also at the top of every run, this skill compares the gitignored .apache-magpie.local.lock (per-machine fetch) against the committed .apache-magpie.lock (the project pin). On mismatch the skill surfaces the gap and proposes setup upgrade. The proposal is non-blocking — the user may defer if they want to run with the local snapshot for now. See docs/setup/install-recipes.md § Subsequent runs and drift detection for the full flow.

Drift severity:

  • method or URL differ → ✗ full re-install needed.
  • ref differs (project bumped tag, or git-branch local is behind upstream tip) → ⚠ sync needed.
  • svn-zip SHA-512 mismatches the committed anchor → ✗ security-flagged; investigate before upgrading.

Golden rules

  • Do not auto-run privilege-elevating commands. Anything that needs sudo (apt / dnf installs, system-wide writes) is printed for the user to copy-paste into their own terminal. The skill never invokes sudo itself.
  • Do not edit shell rc files without approval. ~/.bashrc / ~/.zshrc modifications (sourcing agent-iso.sh, the optional alias claude='claude-iso') are surfaced as the exact line to add; the user pastes it themselves. The skill confirms the rc file path with the user first; it does not assume.
  • Do not overwrite an existing settings file silently. If the user already has a project .claude/settings.json or a user-scope ~/.claude/settings.json, the skill diffs the desired merge against the existing file and asks for explicit approval before writing. The merge carries the framework's permissions.allow set (the read-only reads: gh reads, the vetted-ops read dispatcher, archive / mailbox / roster MCP reads, registry WebFetch hosts) alongside deny and ask — leaving it out makes every skill read prompt. Only read-only entries belong in allow. Re-installs / partial-state recoveries are common — the skill must not blow away an unrelated pre-existing hook or permissions.ask rule. The desired merge includes the agent-guard hooks.PreToolUse entry (matcher Bash, command running the user-scope ~/.claude/scripts/agent-guard.py) — the deterministic guard from tools/agent-guard. Install the script as ~/.claude/scripts/agent-guard.py and populate ~/.claude/scripts/guards.d/ from both the engine's bundled guards.d/*.py and every skill-owned skills/*/guards/*.py (so skill-owned guards like mention / mark-ready are active), alongside the other user-scope scripts (Step P); wire the PreToolUse entry once, and preserve any pre-existing hooks entries.
  • Stop on the first failure. If a step fails (manifest read fails, framework path wrong, an existing file conflicts in a way the user has not yet decided about), stop and report. Do not push past a failure to the next step.

Up-front confirmations

Before walking any install step, confirm with the user:

  1. OS / distro. macOS, Ubuntu / Debian (apt), Fedora / RHEL (dnf), or Arch / NixOS / other. macOS skips bubblewrap + socat (Seatbelt is built-in); Linux installs both per the distro shortcut.
  2. Framework checkout path. The path to the user's local magpie clone. Required to read tools/agent-isolation/pinned-versions.toml, .claude/settings.json, and the tools/agent-isolation/*.sh scripts. If the user does not have a clone, walk them through git clone first.
  3. Fresh install or re-install. For a re-install on a partial existing state, the skill must enumerate the existing wiring (project settings.json, user settings.json, hooks dir, the agent-guard hooks.PreToolUse entry + ~/.claude/scripts/agent-guard.py, shell rc) before any merge so the user knows what is being preserved vs replaced.
  4. Sync repo (optional). Whether the user maintains a private dotfile-style ~/.claude-config repo per Syncing user-scope config across machines. If yes, the skill installs user-scope scripts as symlinks into ~/.claude-config/scripts/ rather than cp-ing into ~/.claude/scripts/ — the symlink approach is what makes sync push the upgrades to other machines automatically.

Walk-through

Follow the canonical step list at docs/setup/secure-agent-setup.md → Adopter setup → Via a Claude Code prompt. Each step in that list maps 1:1 to a step in this skill. Do not re-write the list here — read the doc, follow it, and surface each sub-step with the user. The doc names are the source of truth; the skill is the runner.

For the verification step at the end, hand off to the setup-isolated-setup-verify skill rather than re-walking the checklist inline.

Agent harness — install @latest, enforce the floor

claude-code is not pinned to an exact version. Install it with npm install -g --no-save @anthropic-ai/claude-code@latest (the command in the doc's install list) — latest always carries the newest permission-rule / sandbox / prompt-injection fixes. The manifest's [tools.claude-code] table in pinned-versions.toml declares a min_version floor, not a pin. Because this install is driven from Claude Code, apply the same hard gate setup-isolated-setup-verify check 5 applies: read the running version (command claude --version, which skips an alias claude=claude-iso) and, if it is below min_version, hard-fail — stop the install, tell the operator to upgrade to @latest, and have them re-run. The secure setup must not be stood up on a below-floor runtime.

Step P — Project-root coverage in the sandbox allowlists

The harness drops the literal . when it pre-resolves sandbox.filesystem.allowRead: ["."], so a session in a fresh clone can write to the project root but not read from it. The fix is to add the root as an absolute path to the project-local settings.local.json.

This step asks one question first — per-project or whole-user scope — and the whole-user answer overrides git's hook lookup for every repository on the machine, so it carries a disclosure the operator has to acknowledge.

Full procedure, both scopes and both whole-user flavours: step-p-sandbox-allowlists.md.

Step V — The vetted-ops split and exclusion

Only applies when the adopter routes forge operations through the vetted-ops dispatcher — i.e. the repo has .apache-magpie-overrides/tools/vetted-ops/config.toml, or the operator asks for the dispatcher to be wired now. If neither is true, skip this step and say so; do not create a policy the project has not asked for.

There are three console scripts over one catalogue, and which of them gets the allow is the entire security question:

Entry point Can write? Permission
vetted-op-read never — refused before policy or --caller is consulted allow
vetted-op-tracker only the tracker rollup / body-field procedures — everything else refused before policy ask
vetted-op yes ask

Never propose an allow on vetted-op. It is tempting, because the policy declares a read-only caller and the invocation names it. That reasoning is wrong: --caller is an argv string chosen by whoever runs the command, so an allow on the write dispatcher grants every operation in the catalogue no matter which caller the example names. The read dispatcher needs no such trust — it refuses writes structurally.

"permissions": {
  "allow": [
    "Bash(uv run --project ~/.claude/magpie/vetted-ops vetted-op-read *)",
    "Bash(uvx --from ~/.claude/magpie/vetted-ops vetted-op-read *)"
  ],
  "ask": [
    "Bash(uv run --project ~/.claude/magpie/vetted-ops vetted-op *)",
    "Bash(uv run --project ~/.claude/magpie/vetted-ops vetted-op-tracker *)",
    "Bash(uvx --from ~/.claude/magpie/vetted-ops vetted-op-tracker *)"
  ],
  "deny": [
    // `Edit(path)` is the path rule for every file-writing tool — Write and
    // NotebookEdit included. A `Write(path)` rule is NOT matched by the file
    // permission check; do not add one alongside, it reads as a second layer
    // that is not there.
    //
    // the operation catalogue — the read `allow` rests on its shape
    "Edit(~/.claude/plugins/cache/apache-magpie/magpie-vetted-ops/**)",
    // the fixed path the rules name, which points into that catalogue
    "Edit(~/.claude/magpie/**)",
    // the policy naming which caller may run which operation
    "Edit(.apache-magpie-overrides/tools/vetted-ops/**)"
  ]
}

Allow both invocation forms, and add both to sandbox.excludedCommands ("uv run --project ~/.claude/magpie/vetted-ops vetted-op-read *" and "uvx --from ~/.claude/magpie/vetted-ops vetted-op-read *"). Skills document uv run --project; read-only gatherer agents use uvx --from, because uv run in the plugin cache needs to write a venv there. A rule for only one form leaves every call in the other form prompting.

Add both forms of vetted-op-tracker to sandbox.excludedCommands too ("uv run --project ~/.claude/magpie/vetted-ops vetted-op-tracker *" and "uvx --from ~/.claude/magpie/vetted-ops vetted-op-tracker *"): the status-rollup and body-field writes need a gh that can verify TLS, which a sandboxed one cannot. Excluding it is acceptable where excluding vetted-op is not, because its catalogue is closed to the tracker procedures — every other operation is refused before the policy is read — and their runner refuses any gh call outside repos/<tracker>/. It stays in ask, never allow, so each write keeps its confirmation. Never exclude vetted-op itself: that would run the whole write catalogue unsandboxed.

The rules name the fixed path ~/.claude/magpie/vetted-ops, never the versioned plugin-cache directory. The plugin's SessionStart hook points that path at the installed version every session, so an upgrade needs no rule change. Never write a rule with * in place of the plugin version. A * also matches spaces, so it approves — and, in excludedCommands, runs unsandboxed — a command with extra uv options spliced in at that position (--with <any package>, a second --from). If an adopter's settings still carry the versioned * form, replace it with the fixed path.

Tell the operator plainly what this buys and what it does not:

  • The catalogue is protected twice — by these rules, and by the plugin cache sitting outside every sandbox.filesystem.allowWrite root, so sandboxed Bash cannot write there either.
  • The policy TOML is protected once, and that is tolerable only because the read dispatcher ignores policy when it refuses a write. Editing the policy can widen which repo is read; it cannot turn a read into a write.
  • Writes are not unattended. They keep the harness confirmation — the dispatcher bounds their shape, not the operator's decision to make them.

Do not describe per-caller scoping as isolation. It is least-privilege hygiene for a cooperating skill, and it is worth having for that, but it stops nothing that chooses to name a different caller.

Step V.1 — Read-only vetted ops MUST never prompt

This is not optional. No vetted-op-read invocation may ever prompt, in any form: bulk skills run it from background gatherer agents, which cannot surface a prompt, so each prompt is a silent denial and the sync returns partial results nobody can verify. Rationale and the full shape: docs/setup/secure-agent-setup.md → Read-only vetted ops never prompt.

Resolve $HOME, read every settings scope (user, project, project-local, and managed where readable), and propose the exact edit for each gap. The operator applies it; never write a settings file on your own.

  1. Allow both launchers. Bash(<launcher> ~/.claude/magpie/vetted-ops vetted-op-read *) MUST be allowed for each launcher in {uvx --from, uv run --project}: two rules, in the committed project .claude/settings.json (as in the framework's own) or in user scope. The literal ~ is the only spelling: the agent-guard vetted-ops-spelling guard refuses an expanded home path, so no absolute-path rule is needed. If an earlier run left <resolved $HOME>/.claude/magpie/vetted-ops allow rules or exclusions, propose removing them.
  2. No ask rule. Propose removing every permissions.ask rule, in any scope, that matches a vetted-op-read command in either launcher form — a broad rule such as Bash(uvx *) counts. An ask rule prompts even when an allow rule matches.
  3. Read block off. If permissions.blockReadsOutsideWorkingDirectories is true in any scope, propose removing it (or setting it false). With it on, every call prompts, because the check cannot resolve the literal ~ the commands spell, and no allow rule overrides it. Never offer an additionalDirectories entry or /add-dir for the checkout instead (see item 7). Say plainly that Bash stays sandboxed and the credential Read denies still apply. If managed settings force it on, the imperative cannot be met on this machine: tell the operator, and that bulk skills must not run their gatherer agents there.
  4. Tamper protection. The three Edit deny rules above MUST be present, in the project settings or in user scope; propose any that is missing. They are what makes the unprompted read safe.
  5. Exclusions for both launchers. sandbox.excludedCommands keeps the vetted-op-read and vetted-op-tracker entries in both ~ launcher forms. Without them an allowed call runs inside the sandbox, where its gh cannot verify TLS: a silent failure again.
  6. The spelling guard. The agent-guard plugin MUST be installed, so its vetted-ops-spelling guard turns an expanded spelling into a refusal that names the ~ command, instead of a sandboxed call that fails silently.
  7. The checkout is not a working directory. No permissions.additionalDirectories entry in any scope may cover $HOME/.claude/magpie (that path, its ~ spelling, anything inside it such as ~/.claude/magpie/vetted-ops, or a parent such as $HOME/.claude or $HOME). A working directory is writable to sandboxed Bash, and item 5's exclusions run that checkout outside the sandbox. sandbox-add-project-root.sh removes the ~/.claude/magpie entry it once added from each worktree's settings.local.json; propose the hand edit for any other covering entry, naming the file.

This covers the read dispatcher only. vetted-op-tracker and vetted-op stay in ask, never in allow, and vetted-op is never excluded.

Step R — The adversarial-review exclusion

Only applies when the magpie-adversarial-review plugin is installed. If it is not, skip this step and say so.

The reviewer CLIs it runs (codex, copilot, gemini, grok, claude) need network access and read their own credentials (~/.codex, ~/.copilot, ~/.gemini, ~/.grok, ~/.claude), which this sandbox denies. The tool therefore runs outside it, through one exclusion. It names ~/.claude/magpie/adversarial-review, a link the plugin's SessionStart hook points at the installed version every session, never a * where the version sits: that would also match uv options spliced in at that position, and run them outside the sandbox.

"sandbox": {
  "excludedCommands": [
    "uvx --from ~/.claude/magpie/adversarial-review adversarial-review *"
  ]
},
"permissions": {
  "deny": [
    // the tool runs unsandboxed, so the code it runs must not be editable
    // by the agent that calls it
    "Edit(~/.claude/plugins/cache/apache-magpie/magpie-adversarial-review/**)",
    "Edit(~/.claude/magpie/**)"
  ]
}

Do not add an allow for it. Every run sends the change to other model providers and costs money; the harness prompt is the gate, by design.

Tell the operator what the exclusion covers and what it does not:

  • It matches only the single-line form, spelled with a literal ~ and an unquoted path. A pipe, $(…), &&, a redirection, quotes, or an expanded home directory put the command back in the sandbox, where the reviewer CLIs fail to read their credentials and report unavailable.
  • The * in the pattern can match more than a version directory — a path with .. segments would still match. The prompt that every run keeps is the gate against that; read the path in it before approving.
  • The tool refuses a --body-file or diff: file outside the repository or a temporary directory, so an approved run cannot be pointed at ~/.ssh or a private checkout to send it to a model.
  • Outside the sandbox the tool only runs each reviewer CLI in its own read-only mode, and writes nothing to the repository. codex's read-only mode still reads files anywhere on the machine; see the tool's README for what that means for a machine that also holds a private checkout.

Step W — Working directories for the read block

Where the vetted-ops dispatcher is used, Step V.1 already requires permissions.blockReadsOutsideWorkingDirectories to be off, and these entries cannot stand in for that: they do not stop the vetted-op-read prompt. They remain useful for the agent's own reads of the scratch root where the block is on for other reasons.

Nothing extra to write: the sandbox-add-project-root.sh run in Step P also adds /tmp/claude-<uid>, resolved, to permissions.additionalDirectories of each worktree's project-local settings.local.json. Under permissions.blockReadsOutsideWorkingDirectories the skills read it on every run; without the entry each read prompts, and a bulk sync multiplies that by its gatherer agents. The script also removes $HOME/.claude/magpie where an earlier version added it.

After Step P, show the operator the resulting additionalDirectories list and check three things:

  • Never ~/.claude/magpie. If any scope lists $HOME/.claude/magpie or a parent of it, propose removing it. A working directory is writable to sandboxed Bash, and the vetted-ops commands run from that checkout outside the sandbox. /add-dir has the same effect for a session.

  • No glob and no ~. If the operator already has an entry such as /tmp/claude-* in any scope, say it does nothing: the setting matches literal paths only, so the glob is listed but never matched.

  • Nothing per-host in a synced or committed file. If the operator put the paths in a user-scope ~/.claude/settings.json that is synced across machines, or in the committed project settings, suggest moving them: they name this host's uid.

Rationale and the rules in full: docs/setup/secure-agent-setup.md → Working directories under the read-outside-working-directories block.

Step T — Local agent settings

The committed .claude/settings.json is only the shared floor; the rest of a working setup is personal and goes in .claude/settings.local.json. Build and hand over that file exactly as setup upgrade → Reconcile the local agent settings does: the MCP tools for families installed here but not adopted by the project and, only when this project is the Magpie framework checkout, its template (tools/agent-isolation/claude-settings.local.template.json). Never stage that template into another project: it would widen that project's sandbox beyond what its maintainers configured. An agent cannot write that file, so stage it and give the user the copy command.

Steps K, L, M and B — optional extras

None of these is needed for a working install. Walk the one the operator asks for: a hardware security key (K), the container gateway (L), the eval-harness exclusion (M), or the exclusion for the third-party roborev CLI (B) — offer B unprompted when command -v roborev finds it.

Procedures: optional-steps.md.

After the install lands

Tell the operator what to look for in the footer, and what each state means. This is the one piece of the install they will see on every render afterwards, so a wrong reading here persists:

Tag Means
[sandbox] green sandboxed, and still prompting per command
[sandbox-auto] yellow sandboxed, not prompting per command — auto-allow; wider blast radius, which is why it is not green
[NO SANDBOX] bold red not sandboxed; the state this install exists to make impossible to miss

Captures of each are in What a session looks like; point the operator there rather than describing the colours twice.

Two things to say explicitly, because both are counter-intuitive:

  • The tag reads the settings files, not the running process. A CLI flag that disables sandboxing mid-session is not visible here. Pair it with the bypass-warn hook, which fires per call.
  • Yellow is not a warning that something is broken. Auto-allow is a deliberate, common choice; the colour distinguishes two postures rather than flagging a fault.

Suggest two follow-up routines the user can wire later:

  • setup-isolated-setup-verify — re-run after every Claude Code upgrade or settings-file edit, to confirm denials still fire as expected. The "did a denial silently turn into an allow?" signal is exactly what this skill exists for.
  • setup-isolated-setup-update — periodic check for framework updates, pinned-tool upgrade candidates, and drift between the installed user-scope copies and the framework's source-of-truth. The pre-flight of every skill proposes it for them once the install is recorded (below): after an upgrade that changes the secure-setup files, and weekly otherwise.

Record the install so that pre-flight knows the isolated setup is used on this machine and what it was installed against:

PYTHONPATH=".apache-magpie-local:$(git rev-parse --git-common-dir)/../.apache-magpie-local:$(git rev-parse --git-common-dir)/apache-magpie" \
  python3 -m setup_preflight.isolated record-update

It writes the isolated_setup block of the personal layer's reconciled.json (<git-common-dir>/apache-magpie/ when the project has not adopted Magpie, .apache-magpie-local/ when it has) and nothing else. Skip it when the command cannot find setup_preflight — say that /magpie-setup:setup config installs the checker, and that the reminders start once it is there.

Always propose shared-config sync once the install lands. Regardless of whether the operator already maintains the ~/.claude-config sync repo, proactively offer to run setup-shared-config-sync as a follow-up:

  • If the ~/.claude-config sync repo is already in place, the skill commits + pushes the local modifications (the user-scope scripts, hooks, and settings this install just wired up) so the other machines pick them up.
  • If the operator does not yet have ~/.claude-config, the setup-shared-config-sync skill bootstraps it (clones the default private remote if it exists, or creates a fresh private remote and scaffolds the layout). Mention this so a first-time operator knows the follow-up will set sync up from scratch — the point of proposing it is precisely so the just-installed config does not stay machine-local.

Surface it as an offer for the operator to accept, not an auto-run — the sync skill has its own confirmation gates before it commits or pushes anything.