Rules that know what the agent read.
Permissions say what an agent may touch. Decision Horizon's policy also asks what it has seen. A write to a public folder is fine, until the agent has read a secret.
The gap it closes
Identity and access controls answer "is this agent allowed to write here?". They can't answer "should it, given what it just read?". That's where prompt injection does its damage.
Capabilities
Rules in dh.toml or Cedar
Write rules in the config, or in a real Cedar policy file. Hard rules block; soft rules pause for a human.
Horizon-aware
Conditions can use what the agent has read, not just the target, so a secret in view changes what it may do next.
Taint by value
Content that reaches a call's arguments is detected from the bytes, across sessions too, not guessed from file names.
Human checkpoints
Send a soft-deny to a webhook: Slack bot, ticket system or your own UI. The record keeps the hash of exactly what the approver was shown.
Safe defaults
Anything explicitly not granted is blocked, expired authority is blocked, and unknown tools escalate to a human instead of running.
No silent approvals
An unreachable approver is recorded as no human consulted, and the action is blocked.
[[policy.rules]]
id = "public_write_with_secret_in_horizon"; tier = "soft"
action = ["write_file", "edit_file", "move_file", "copy_file"]
when = { targets = "*pub*", horizon = "*secret*" }What it doesn't do
- An approval endpoint is your component. The record writes "the approval endpoint reported", never "a human decided".
See it on one of your own workflows.
We'll put Decision Horizon in front of a real agent workflow of yours and replay the record with you.