pi-agent-config/docs/isolation-baseline.md

Isolation Baseline for a Dedicated Pi Agent

Applies to every scenario in scenarios/. Mechanisms verified against pi 0.84.3 — see pi-runtime-notes.md for the source quotes and evidence/ for reproduction.

A "dedicated agent" is not achieved by narrowing what the model is told. It is achieved by narrowing what the process loads and what it can call. Pi offers four independent layers. Use all four; each one covers a different failure mode.

---

The four layers

┌──────────────────────────────────────────────────────────────┐
│ L1 Loading      what enters the process at all               │
│   --no-extensions -e / --no-skills --skill /                 │
│   --no-prompt-templates / --no-themes / -nc / project trust  │
├──────────────────────────────────────────────────────────────┤
│ L2 Personality  who the agent is                             │
│   .pi/SYSTEM.md (replace) · APPEND_SYSTEM.md · AGENTS.md      │
│   AGENTS.override.md · --system-prompt                        │
├──────────────────────────────────────────────────────────────┤
│ L3 Capability   which tools exist and are active             │
│   --no-builtin-tools · --tools · --exclude-tools              │
│   pi.setActiveTools() · same-name registerTool override       │
├──────────────────────────────────────────────────────────────┤
│ L4 Invocation   per-call authorization                       │
│   pi.on("tool_call") -> { block: true, reason }               │
└──────────────────────────────────────────────────────────────┘

Layers 3 and 4 must be enforced inside the extension, not only on the command line, so that a flag mistake in a systemd unit cannot silently widen the agent's reach.

---

Baseline flag set

pi --mode rpc \
   --session-id   "<scenario>-<subject>" \
   --session-dir  "<dedicated session dir>" \
   --no-builtin-tools \
   --no-extensions        -e     "<abs>/scenario-tools.ts" \
   --no-skills            --skill "<abs>/.pi/skills/<name>" \
   --no-prompt-templates \
   --no-themes \
   --approve \
   --provider "<provider>" --model "<model>" --thinking "<level>" \
   --name "<display name>"

Launched with:

Property	Value	Reason
cwd	the scenario workspace	bounds AGENTS.md and .pi discovery
env	explicit minimal allowlist	keeps backend API keys out of the node process
start_new_session=True	yes	lets the parent kill the whole process group
timeout handling	killpg, not kill	pi is a node CLI and may leave children

Why each flag

Flag	Failure mode it prevents
--no-extensions + -e <abs>	~/.pi/agent/extensions/*.ts (currently herdr-agent-state.ts, pi-memo-trust.ts) being injected into an unrelated agent. The absolute -e path also loads before project trust resolves, so the extension works without trusting the directory.
--no-skills + --skill <abs>	~/.agents/skills/* leaking in. Probed leak: ["find-skills","modsearch","summarize"] — find-skills actively instructs the agent to install more skills.
--no-prompt-templates	~/.pi/agent/prompts/*.md becoming callable slash commands.
--no-themes	pointless discovery I/O and a source of run-to-run variation.
--no-builtin-tools	bash/edit/write/grep/find/ls being active. Chosen over --tools because it imposes no registry-level allowlist, so an extension may register tools dynamically. Built-ins stay registered-but-inactive; the extension governs the active set.
--approve	required to load .pi/SYSTEM.md and .pi/settings.json. Combined with --no-skills it does not re-open skill auto-discovery.
--session-id + --session-dir	sessions colliding with the user's interactive history.

Do not use --no-tools

It disables extension tools as well, which removes read, which removes the skills block from the system prompt (see pi-runtime-notes.md §1). Any SKILL.md then becomes dead weight. This was a real, long-lived defect in the Curator scenario.

---

Required extension responsibilities

A scenario extension must do all five:

const ALLOWED_TOOLS = ["read", /* … scenario tools … */];

export default function scenarioGuard(pi: ExtensionAPI) {
  // 1. path-restricted `read` override — shadows the built-in by name
  pi.registerTool({ name: "read", /* allowlist-checked implementation */ });

  // 2. scenario tools, each with promptSnippet so they appear in the prose list
  pi.registerTool({ name: "…", promptSnippet: "…", promptGuidelines: [ "…" ] });

  // 3. declarative active set, reasserted after resource discovery
  const restrict = () => pi.setActiveTools(ALLOWED_TOOLS);
  pi.on("session_start", async () => restrict());
  pi.on("resources_discover", async () => restrict());

  // 4. per-call gate — defence in depth against layer-3 mistakes
  pi.on("tool_call", async (event) => {
    if (!ALLOWED_TOOLS.includes(event.toolName)) {
      return { block: true, reason: `<scenario> blocks tool: ${event.toolName}` };
    }
  });

  // 5. every tool truncates its own output (50 KB / 2000 line caps)
}

The read override must still permit the skill directory, otherwise skill bodies cannot be loaded (progressive disclosure, pi-runtime-notes.md §2).

---

Measured effect

Same workspace, same extension, only flags differ:

Configuration	system prompt	active tools	skills visible	coding-assistant persona	pi-docs paths
--no-tools (Curator, as found)	1859	(none)	no	yes	yes
--no-builtin-tools, no isolation	3413	own + read	3 foreign	yes	yes
+ --no-skills --skill	2619	own + read	1, correct	yes	yes
+ .pi/SYSTEM.md --approve	960	own + read	1, correct	no	no

Full baseline is 72 % smaller than the unisolated default and strictly more capable, because the skill is finally reachable.

---

Residual risk

1. No kernel boundary. Pi ships no sandbox; extensions run with the process's full permissions. The read override is application-level. A future step is bubblewrap (examples/extensions/sandbox/) or a container.
2. env discipline is the only secret boundary. Without an explicit env= allowlist the node process inherits every *_API_KEY in the unit file.
3. Parent-directory context files. Nothing prevents a future ~/AGENTS.md from layering into every scenario. Mitigate with -nc plus SYSTEM.md, or an AGENTS.override.md in the workspace.
4. Workspace should be read-only to the service. --approve trusts everything project-local, so a writable .pi/ is a code-execution path. Enforce with systemd ReadOnlyPaths=.
5. Prompt injection is not solved by any of this. Untrusted text must be delimited and labelled as data, and any write must be gated by deterministic code outside the model.

---

Per-scenario conformance

Scenario	L1 loading	L2 personality	L3 capability	L4 invocation	env minimal
curator	target	target	target	target	target
memo-inbox	gap	gap	ok (setActiveTools)	ok (tool_call)	gap
pi-grok	not assessed	not assessed	not assessed	not assessed	not assessed

memo-inbox already implements L3 and L4 correctly and is the reference for those layers; its L1/L2/env gaps are tracked as a follow-up and must not be changed in the same pass that migrates its configuration into this repository.

Check conformance with:

scripts/pi-diff.sh <scenario>