Policy

Policy controls what Murph does with a drafted response when a session is running. The selected policy profile is the source of truth for Murph's default autonomy.

For the full policy profile format, runtime order, and metadata reference, use Extending Policy.

Murph Agent is the preferred way to create, preview, and select a custom profile:

murph agent

Ask it to inspect your current policy, create a new profile, preview the result, and select it. The selected profile name is stored in ~/.murph/config.yaml; the profile file itself owns mode.

Shipped profiles

Murph ships role-oriented profiles:

• default
• engineering
• product
• investor
• yolo

The role profiles are conservative and keep auto-send off with mode: manual_review. yolo is the maximum-autonomy preset for trusted local runs after you have verified setup and behavior. Runtime grounding still checks that required read/context work happens before answering.

Custom profile files

Shipped profiles live in the app's policies/ directory. Custom profiles live in ~/.murph/policies/*.md and use a metadata header plus body notes. Set mode in the profile; do not add a separate policy.mode config value.

name: custom
description: My custom policy.
blockedTopics: payroll details, legal advice
alwaysQueueTopics: pricing, customer commitments
mode: manual_review
preferAskWhenUncertain: yes
notes: keep replies concise, avoid promises
---
Extra instructions for Murph when this profile is active.

blockedTopics are hard stops that make Murph abstain. alwaysQueueTopics are guarded topics: Murph can draft, but the result must go to operator review. Grounding for factual replies is always required by the runtime, so it is not a policy field.

Use the CLI to inspect, preview, and select profiles:

murph policy profiles
murph policy preview --profile custom
murph policy preview --profile custom --session-mode dry_run
murph policy set --profile custom

How policy runs

Message arrives
  |
  v
Runtime hard stops
(no session, out of scope, expired, missing owner)
  |
  v
Agent drafts action
(reply, ask, redirect, defer, remind, abstain)
  |
  v
Grounding check
(required read/context tools attempted)
  |
  v
Policy classifier
(request + policy + grounding + draft)
  |
  v
Deterministic final gate
(profile mode, blocked topics/actions, risk, confidence)
  |
  v
Send | Queue | Abstain

Hard stops happen before Murph drafts.

The deterministic final gate is authoritative. Grounding checks whether required retrieval was attempted, but it does not prove factual correctness.

Profile mode is the default

Each policy profile declares its own execution mode. That profile mode is the default autonomy posture for new sessions:

• manual_review queues drafted actions for review.
• auto_send_low_risk can send low-risk actions automatically and queue the rest.

Select the durable default profile from Admin or the CLI:

murph policy set --profile engineering
murph policy set --profile yolo

Session mode is a temporary override. Starting a session without a mode inherits the selected profile mode. A dry run records decisions without side effects, and a manual-review session queues everything for that run. Session overrides cannot increase autonomy beyond the profile mode; a manual-review profile still queues even if a caller asks for auto_send_low_risk.

Runtime hard stops still apply for empty context, out-of-scope threads, high-risk skill context, unsupported action types, and messages Murph cannot safely send.

Runtime Hosts

In a shared messenger channel runtime host, policy resolves at the runtime level. New config-bound sessions snapshot the selected policy profile.

Session overrides cannot raise autonomy beyond the selected profile mode. A manual_review profile still prevents a session request from raising execution to auto_send_low_risk.

YOLO profile

Use yolo when you intentionally want the least restrictive action profile:

murph policy set --profile yolo

yolo sets profile mode to low-risk auto-send and disables the uncertainty preference, but it does not disable runtime grounding. Murph should use materially relevant read-only retrieval or context tools before answering factual questions. It is explicit by design; fresh installs do not select it automatically.