OpenShell/docs/sandboxes/policy-advisor.mdx

---
# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
title: "Use Policy Advisor"
sidebar-title: "Policy Advisor"
description: "Let sandboxed agents propose narrow policy changes through policy.local while keeping developer approval in the loop."
keywords: "Generative AI, Cybersecurity, Policy Advisor, Policy, Sandbox, policy.local, Agent Policy"
position: 7
---

Policy advisor lets a running sandboxed agent ask for a narrow network policy change after OpenShell denies a request. The agent submits a draft through `policy.local`, a developer approves or rejects it from outside the sandbox, and approved network policy hot-reloads into the same sandbox.

Policy advisor preserves OpenShell's default-deny posture. The structured rule is the approval contract, and the agent's rationale is supporting context. By default every accepted proposal lands in the draft inbox for human review. Opt-in [auto mode](#approval-modes) approves a proposal without a reviewer only when its [prover delta](#what-auto-approval-checks) is empty and the current draft rule produces no security notes. A prover finding or security note keeps the proposal pending for human review.

## Enable Policy Advisor

Policy advisor is disabled by default. Enable it globally when you want every sandbox on the selected gateway to expose the agent proposal surface:

```shell
openshell settings set --global \
  --key agent_policy_proposals_enabled \
  --value true \
  --yes
```

You can also enable it for one sandbox, unless the key is managed globally:

```shell
openshell settings set <sandbox-name> \
  --key agent_policy_proposals_enabled \
  --value true
```

Check the effective setting for a sandbox:

```shell
openshell settings get <sandbox-name>
```

The output shows whether `agent_policy_proposals_enabled` is `global`, `sandbox`, or `unset`. A global value overrides sandbox-scoped values. To return control to sandbox-scoped settings, delete the global key:

```shell
openshell settings delete --global \
  --key agent_policy_proposals_enabled \
  --yes
```

Set the value before creating a sandbox when you want the first denied request to include policy advisor guidance. Running sandboxes poll settings and can enable the surface after startup, but startup enablement gives the agent the clearest first-denial path.

## Approval Modes

Every proposal, mechanistic or agent-authored, is routed through the [policy prover](#what-auto-approval-checks). The gateway also recalculates security notes from the current draft rule before auto-approval. The `proposal_approval_mode` setting decides whether proposals that pass both checks require human review.

| Mode | When unset / `manual` | `auto` |
|---|---|---|
| Empty prover delta and no security notes | Lands in the draft inbox for human review. | Approved automatically. The sandbox hot-reloads the new rule and the agent retries. |
| Any prover finding or security note | Lands in the draft inbox. | Remains pending for human review. |

`manual` is the default. Auto mode is an explicit opt-in; OpenShell's default-deny posture is preserved unless you choose otherwise.

Enable auto mode at gateway scope when you want every sandbox on this gateway to auto-approve eligible proposals:

```shell
openshell settings set --global \
  --key proposal_approval_mode \
  --value auto \
  --yes
```

Enable it for one sandbox when no global value is set:

```shell
openshell settings set <sandbox-name> \
  --key proposal_approval_mode \
  --value auto
```

The shorthand at create time writes the sandbox-scoped setting for you:

```shell
openshell sandbox create --approval-mode auto <name>
```

Only `manual` and `auto` are accepted; typos like `autom` are rejected at configure time. Stale or unknown values found in storage are still treated as `manual` at runtime as a defense-in-depth measure.

**Precedence.** Gateway scope wins over sandbox scope. A reviewer can pin `manual` for a fleet by setting it globally; per-sandbox overrides only apply when no global value is set.

**Audit trail.** Every auto-approval emits a `CONFIG:APPROVED` event with `auto=true`, `source=<mechanistic|agent_authored>`, `prover_delta=empty`, and `resolved_from=<gateway|sandbox|default>` so operators can reconstruct why a given approval ran without human review.

## How It Works

When policy advisor is enabled, the sandbox supervisor turns on three agent-facing surfaces:

- It installs `/etc/openshell/skills/policy_advisor.md` inside the sandbox.
- It also installs `/etc/openshell/skills/policy-advisor/SKILL.md` as a short Codex/generic-agent pointer, and writes a root `/AGENTS.md` pointer only when the image does not already provide one.
- It serves `http://policy.local` from inside the sandbox.
- It adds `agent_guidance` and `next_steps` to L7 `policy_denied` response bodies so the agent can find the skill and local API.

The loop has seven steps:

1. A sandboxed process attempts a network request that policy denies.
2. For inspected REST traffic, OpenShell returns a structured `403` body with fields such as `layer`, `host`, `port`, `binary`, `method`, `path`, `rule_missing`, `agent_guidance`, and `next_steps`.
3. The agent reads the policy advisor skill, inspects the current policy, and optionally reads recent denial log lines.
4. The agent submits one or more `addRule` proposals to `http://policy.local/v1/proposals`.
5. The gateway stores accepted proposals as pending draft chunks for the sandbox and runs the [policy prover](#what-auto-approval-checks) against the proposed delta.
6. Before auto-approval, the gateway reloads the current stored rule, recalculates its prover verdict, and regenerates its security notes. Under `auto` mode, it approves the proposal only when both the prover delta and security notes are empty. Any finding or security note keeps the proposal pending. Under `manual` mode, every accepted proposal lands in the draft inbox for a developer to approve or reject.
7. The agent waits on `/v1/proposals/{chunk_id}/wait` until a decision is available. Approved proposals hot-reload into the sandbox; rejected proposals return `rejection_reason` and `validation_result` so the agent can revise.

When a proposal is approved, `/wait` reports `policy_reloaded: true` only after the local sandbox policy covers the approved rule. At that point the agent can retry the original denied action once. If a proposal is rejected, `/wait` returns `rejection_reason` and `validation_result` so the agent can revise or stop. `validation_result` carries the categorical prover findings — `link_local_reach`, `l7_bypass_credentialed`, `credential_reach_expansion`, `capability_expansion` — so the agent can narrow the next attempt to the specific concern the prover flagged.

## What Gets Proposed

OpenShell has two proposal paths:

| Path | Source | Typical rule shape |
|---|---|---|
| Mechanistic mapper | Aggregated denial summaries from the sandbox. | Groups by host, port, and binary. If L7 request samples are available, it can draft REST method and path rules. Otherwise it drafts an L4 endpoint. |
| Agent-authored proposal | The in-sandbox agent, using `policy.local`. | Usually a REST `addRule` with exact host, port, binary, method, and path from the structured denial. It can also propose L4 rules for opaque protocols. |

For REST APIs, prefer L7 rules over broad L4 access. A good proposal allows one method and the smallest safe path:

```json
{
  "intent_summary": "Allow gh to update documentation in NVIDIA/OpenShell.",
  "operations": [
    {
      "addRule": {
        "ruleName": "github_contents_docs_write",
        "rule": {
          "name": "github_contents_docs_write",
          "endpoints": [
            {
              "host": "api.github.com",
              "port": 443,
              "protocol": "rest",
              "enforcement": "enforce",
              "rules": [
                {
                  "allow": {
                    "method": "PUT",
                    "path": "/repos/NVIDIA/OpenShell/contents/docs/**"
                  }
                }
              ]
            }
          ],
          "binaries": [
            {
              "path": "/usr/bin/gh"
            }
          ]
        }
      }
    }
  ]
}
```

The current `policy.local` JSON shape covers L4 endpoints and REST method or path rules. Use [Customize Sandbox Policies](/sandboxes/policies) or [Policy Schema Reference](/reference/policy-schema) for policy fields that are not part of the agent-authored proposal surface, such as WebSocket credential rewrite, GraphQL operation matching, endpoint path scoping, and provider-owned policy bundles.

Policy advisor proposals do not add `allowed_ips` automatically. If an advisor-proposed hostname resolves to an internal or private address, OpenShell's SSRF protections still block the connection until a developer explicitly adds the required `allowed_ips` entry. Exact hostname trust for user-declared policy endpoints does not apply to advisor-generated proposal binaries.

Private RFC 1918, CGNAT, IPv6 ULA, and other special-use destinations classified as internal produce advisory security notes when they appear as literal endpoint IPs or in `allowed_ips`. CIDR intersections are included, and hostless `allowed_ips` rules receive an additional warning because they can match any hostname resolving into the configured range.

Always-blocked destinations are not advisory. Loopback, link-local, and unspecified IPs or CIDRs, plus `localhost` and known metadata endpoint hostnames, are excluded from security notes. Submit and edit can store such a draft, but approval fails when merge validation prevents it from entering the active policy. Runtime SSRF protections continue to enforce the same boundary.

## What Auto-Approval Checks

Auto-approval requires all three conditions: the effective mode is `auto`, the prover delta is empty, and recalculating security notes from the current stored rule produces none.

The policy prover runs against mechanistic and agent-authored proposals alike and asks four formal questions about the proposed change. Each "yes" is one categorical finding. Any finding blocks auto-approval. An empty delta is necessary but not sufficient.

| Category | Triggered when |
|---|---|
| `link_local_reach` | A rule reaches `169.254.0.0/16`, `fe80::/10`, or a known metadata hostname. |
| `l7_bypass_credentialed` | A binary using a wire protocol the L7 proxy cannot inspect (`git-remote-https`, `ssh`, `nc`) gains reach to a host where a credential is in scope. |
| `credential_reach_expansion` | A binary gains credentialed reach to a `(host, port)` it could not reach before. |
| `capability_expansion` | On a `(binary, host, port)` that already had credentialed reach, the proposal adds a new HTTP method. The finding cites the specific method. |

Findings are categorical. There is no severity tier. The reviewer reads the category and the structured evidence to decide.

Before auto-approval, the gateway reloads the draft chunk, recomputes its prover verdict, and regenerates security notes from its current rule. These checks apply after edits and deduplicated resubmissions: an edited stored rule controls the decision even when the duplicate incoming proposal has an empty prover delta. The recalculated verdict is used for this decision but is not written back, so a later `validation_result` read can still show the submit-time verdict. Failures leave the chunk pending. Security notes flag concerns such as internal or private destinations and `allowed_ips`, wildcard hosts, hostless `allowed_ips`, ephemeral ports, and well-known database or service ports. Draft reads and bulk approval also regenerate notes from the stored rule. Any prover finding or security note keeps the chunk pending.

The full reasoning model lives in [`crates/openshell-prover/README.md`](https://github.com/NVIDIA/OpenShell/blob/main/crates/openshell-prover/README.md). Provider profiles composed in via [Providers v2](/sandboxes/providers-v2) are part of the effective policy the prover reasons over.

## Review Proposals

Review pending chunks from the host:

```shell
openshell rule get <sandbox-name> --status pending
```

Under `auto` mode, proposals with a prover finding or any recalculated security note remain pending for human review. Proposals that pass both checks are visible under `--status approved` with the auto-approval audit fields described in [Approval Modes](#approval-modes). Under `manual` mode, every accepted proposal shows up as pending regardless of the prover verdict or security notes.

The output shows the chunk ID, status, rationale, binary, and endpoint summary. For L7 proposals, the endpoint summary includes the protocol, method, and path:

```text
Endpoints: api.github.com:443 [L7 rest, allow PUT /repos/NVIDIA/OpenShell/contents/docs/**]
```

Approve only when the structured rule matches the access you intend to grant:

```shell
openshell rule approve <sandbox-name> --chunk-id <chunk-id>
```

Reject with guidance when the rule is too broad or points at the wrong target:

```shell
openshell rule reject <sandbox-name> \
  --chunk-id <chunk-id> \
  --reason "Scope this to docs/ paths only."
```

The rejection reason is returned to the agent through `policy.local`. The agent can use it to draft a narrower proposal.

## Agent API

`policy.local` is available only inside the sandbox and uses plain HTTP:

| Endpoint | Purpose |
|---|---|
| `GET /v1/policy/current` | Returns the current effective sandbox policy as YAML. |
| `GET /v1/denials?last=10` | Returns recent denied OCSF shorthand log lines, newest first. Query strings are redacted before lines are returned to the agent. |
| `POST /v1/proposals` | Submits `addRule` operations. The response includes `accepted_chunk_ids` and `rejection_reasons`. |
| `GET /v1/proposals/{chunk_id}` | Returns one proposal's current `pending`, `approved`, or `rejected` status. |
| `GET /v1/proposals/{chunk_id}/wait?timeout=300` | Holds one HTTP request open until the proposal is approved, rejected, or the timeout expires. |

If policy advisor is disabled, every route returns `404 feature_disabled`, the skill is not installed for new sandboxes, and L7 deny bodies do not advertise `policy.local` routes or include `agent_guidance`.

## What to Expect

Approved network rules hot-reload without restarting the sandbox. HTTP L7 keep-alive connections are closed at the reload boundary so the next parsed request uses the new policy. Raw streams remain connection-scoped, as described in [Customize Sandbox Policies](/sandboxes/policies#policy-structure).

Policy advisor emits audit events into the sandbox log. Use these lines to trace the full loop:

```shell
openshell logs <sandbox-name> --since 10m
```

Look for `HTTP:* DENIED`, `CONFIG:PROPOSED`, `CONFIG:APPROVED` or `CONFIG:REJECTED`, `CONFIG:LOADED`, and the final allowed request if the agent retries successfully. Auto-approved chunks emit `CONFIG:APPROVED` with `auto=true`, `source=<mechanistic|agent_authored>`, `prover_delta=empty`, and `resolved_from=<gateway|sandbox|default>`.

## Next Steps

- Use [Customize Sandbox Policies](/sandboxes/policies) for manual policy updates and L7 rule syntax.
- Use [Policy Schema Reference](/reference/policy-schema) for full YAML field details.
- Use [Logging](/observability/logging) to interpret OCSF shorthand log entries.