Extending Policy

Policy profiles control Murph's autonomy after a draft exists.

Use policy when the question is "may Murph send this?" Use skills when the question is "how should Murph interpret or answer this?"

File location

Policy profiles live in:

~/.murph/policies/*.md

Murph loads the shipped app profiles first, then Markdown files in ~/.murph/policies. A user profile with the same normalized name overrides the shipped profile. README.md is ignored.

File shape

name: engineering
description: Engineering coverage for technical status, incidents, releases, and implementation threads.
blockedTopics: compensation, performance reviews, legal advice
alwaysQueueTopics: production incidents, security vulnerabilities, data loss
blockedActions:
mode: manual_review
allowAutoSend: no
requireGroundingForFacts: yes
preferAskWhenUncertain: yes
notes: cite technical evidence, avoid changing commitments
---
Reply with bounded engineering continuity from thread context, tickets, docs, or code references.
Queue incident ownership, security impact, data changes, release decisions, and access or secrets requests.

The header is simple key: value metadata. The body becomes additional notes for the agent and policy classifier.

Metadata

Field	Purpose
name	Stable profile name. Defaults to filename.
description	Profile summary shown in setup surfaces.
blockedTopics	Comma-separated topics that force abstain.
alwaysQueueTopics	Comma-separated topics that force review.
blockedActions	Comma-separated action types to block.
mode or executionMode	manual_review or auto_send_low_risk.
allowAutoSend	Legacy boolean fallback for mode.
requireGroundingForFacts	Whether factual replies should require grounding attempts.
preferAskWhenUncertain	Whether uncertainty should bias toward asking or queueing.
notes	Comma-separated classifier and agent notes.
scopedRules	Optional JSON for narrower rule overrides.

Runtime order

Murph applies policy after the main agent drafts.

The runtime still has hard stops before drafting, such as no matching session, owner-authored events, expired sessions, unsupported action types, or missing context.

After drafting, policy classification and the deterministic final gate decide whether the action sends, queues, or abstains.

Policy mode

manual_review queues drafted actions for review.

auto_send_low_risk can send low-risk actions automatically and queue the rest.

Session mode can temporarily reduce autonomy, but it cannot increase autonomy beyond the durable policy mode.

Commands

List profiles:

murph policy profiles

Preview a profile:

murph policy preview --profile engineering --mode manual_review

Select a profile:

murph policy set --profile engineering

Set durable policy mode:

murph policy set --mode manual_review
murph policy set --mode auto_send_low_risk

Use yolo only when you intentionally want maximum local autonomy. It does not disable runtime grounding.