@j1nn0/agent-tool-policy-pi

What it is

@j1nn0/agent-tool-policy-pi is a minimal Pi adapter that enforces a caller-declared tool policy at Pi's pre-execution tool_call boundary. It provides the human-only /agent-tool-policy command namespace and registers no model-callable tools.

The adapter is safe by default: a session with no policy entry or with a corrupt newest entry blocks tool calls. An explicit clear --yes operation is the only way to select the pass-through state.

Installation

Install the published adapter from the registry:

pi install npm:@j1nn0/agent-tool-policy-pi

Add -l to install it project-locally instead of for your user.

From a local checkout (development)

Build and install the package from a local checkout:

REPO_DIR=/abs/path/to/agent-primitives
PACKAGE_DIR="$REPO_DIR/packages/agent-tool-policy-pi"
cd "$REPO_DIR"
pnpm install
pnpm --filter @j1nn0/agent-tool-policy-pi build
pi install "$PACKAGE_DIR" -l

For a single run without changing Pi's installed extensions, load the built extension directly:

pi -e "$PACKAGE_DIR/dist/extension.js"

Enforcement states

The adapter loads only the newest agent-tool-policy-state entry on the current Pi branch. It never falls back to an older entry or repairs a corrupt entry.

State	Tool-call behavior	How it is selected
unconfigured	Block all calls	No matching session entry exists.
disabled	Pass through all calls	The newest entry is the explicit policy: null marker written by clear --yes.
enforcing	Delegate every call to the frozen core judge	The newest entry contains a valid policy.
corrupt	Block all calls	The newest matching entry has a wrong schema, extra keys, or an invalid policy.

Silent allow happens only in disabled, after an operator explicitly writes the null marker with clear --yes. Missing, malformed, or contradictory configuration is never silently treated as disabled.

Commands

The extension registers the human command /agent-tool-policy. A bare command is the same as status:

/agent-tool-policy
/agent-tool-policy status
/agent-tool-policy set <policy-json>
/agent-tool-policy judge <tool>
/agent-tool-policy clear
/agent-tool-policy clear --yes

For example, configure a deny-by-default policy while allowing search, denying writes, and requiring approval for a deployment tool:

/agent-tool-policy set '{"default":"deny","allow":["mcp__search__query"],"deny":["write"],"requiresApproval":["deploy"]}'

set consumes the remainder of the command as one JSON value. JSON parsing and the frozen core's policy validation must both succeed; otherwise the state is unchanged. A successful set replaces the in-memory policy and persists exactly one full snapshot.

Use judge for a transient verdict without changing state:

/agent-tool-policy judge mcp__search__query

clear is gated and changes nothing until explicitly confirmed:

/agent-tool-policy clear
/agent-tool-policy clear --yes

clear --yes writes an explicit DISABLED marker with policy: null; it does not delete session history. Repeating clear --yes while already disabled reports that the policy is already disabled and writes nothing.

Approval flow

When the core returns requires_approval, the adapter:

• blocks immediately when ctx.hasUI is false;
• calls ctx.ui.confirm when dialog-capable UI is available;
• allows the call only when confirmation resolves to exactly true;
• blocks when confirmation is false, undefined, or throws.

The current ctx.signal is forwarded to the confirmation dialog. Aborting the turn therefore cancels the dialog rather than turning an unresolved approval into an allow. Approval dialogs may include a bounded, allowlisted operand summary for known built-in tools. This summary is a human decision aid only and is never used for policy matching; unknown or custom tools and any preview failure fall back to the name-only prompt.

Persistence

Each successful set or confirmed clear --yes appends one exact envelope to the current Pi session branch:

{
  "schemaVersion": 1,
  "policy": {
    "default": "deny",
    "allow": ["mcp__search__query"],
    "deny": ["write"],
    "requiresApproval": ["deploy"]
  }
}

The disabled marker has the same envelope shape with "policy": null. Verdicts and approval decisions are transient and are never persisted.

What this adapter does not do

• It has no model-callable tools, so the model can never change its own policy.
• It does not support patterns, globs, prefixes, regular expressions, or other matching semantics.
• It does not inspect tool arguments for policy matching; approval previews are limited to allowlisted built-in operands.
• It does not write an audit log or persist verdicts or approvals.
• It does not implement session-once approvals; that is a future candidate.
• It never uses Pi's terminate result.
• It does not discover tools, execute tools, infer a policy, or make provider/model calls.

Relationship to @j1nn0/agent-tool-policy

@j1nn0/agent-tool-policy is the frozen, harness-agnostic core. This adapter validates and persists the caller-declared policy, routes each Pi tool_call through judgeToolPolicy({ tool, policy }), and maps the returned outcome to Pi enforcement. The adapter owns approval UI and fail-safe blocking; the core owns exact-name policy semantics. The adapter does not replace or reinterpret the core judge.

Coexistence with the other Pi adapters

This adapter can be loaded alongside the other seven Pi adapters. Each adapter owns an independent command namespace, tool family, and session entry type:

Adapter	Command	Tools	Session entry
Context Guard	/context-guard	none	agent-context-guard-state
Agent State	/agent-state	agent_state_*	agent-state-state
Agent Progress	/agent-progress	agent_progress_*	agent-progress-state
Agent Retry Guard	/agent-retry	agent_retry_*	agent-retry-state
Agent Evidence	/agent-evidence	agent_evidence_*	agent-evidence-state
Agent Handoff	/agent-handoff	agent_handoff_*	agent-handoff-state
Agent Budget	/agent-budget	agent_budget_*	agent-budget-state
Agent Tool Policy	/agent-tool-policy	none	agent-tool-policy-state

It does not merge or map the other adapters' state, and those adapters do not change this policy.

Limitations

Version 0.1 supports a session-scoped policy only. Matching is exact-name and case-sensitive. A corrupt newest policy requires a manual /agent-tool-policy set <policy-json> before any tool call can pass; the adapter never falls back to an older valid entry. The core's requires_approval result is limited to this adapter's interactive confirmation flow, with no session-once approval memory.