diff --git a/CHANGELOG.md b/CHANGELOG.md index 7e73ac36d..3727adbcf 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,11 @@ # Changelog +## 1.0.15-beta.0 — 2026-11-16 + +### Docs + +- **Policies suggested from your agents' code.** A new page under Prevent failures › Get a policy explains how connecting GitHub turns the rules your agents' code states into tested policies you deploy, in observe mode, with one yes (#1006) + ## 1.0.14-beta.0 — 2026-11-16 ### Features diff --git a/docs/docs.json b/docs/docs.json index 729aefafa..8eae679c6 100644 --- a/docs/docs.json +++ b/docs/docs.json @@ -181,6 +181,7 @@ "pages": [ "policies/authority", "policies/editor", + "policies/suggested", "policies/packs" ] }, diff --git a/docs/policies/suggested.mdx b/docs/policies/suggested.mdx new file mode 100644 index 000000000..04d551125 --- /dev/null +++ b/docs/policies/suggested.mdx @@ -0,0 +1,46 @@ +--- +title: "Policies suggested from your code" +description: "Connect GitHub and Failproof AI reads the rules your agents' own code states, writes and tests a policy for each, and lets you deploy it with one yes." +icon: "sparkles" +--- + +Your agents' code already says what they must never do: a prompt that says "never force-push to main", a tool that deletes files, an approval step before an email goes out. Once you connect GitHub, Failproof AI reads those rules, writes a policy for each, tests it, and puts it in front of you. Nothing runs until you say yes. + +## How it works + +1. **Connect GitHub** under **Settings › Integrations › GitHub** and select the repositories your agents live in. +2. Failproof AI links each agent to the folder that runs it and writes the agent's profile: what it is for, what it must and must not do, and which of its tools can't be undone. Each rule points at the line of code it came from. +3. For the rules a policy can enforce — "must not", "needs approval", "limit", and tools that can't be undone — it writes a policy and tests it on example calls it must stop and calls it must allow. When the agent already has traffic, the policy is also replayed against its recent calls. +4. Only policies that pass their tests are suggested, at most three per agent. + +Suggestions are refreshed when the agent's code changes. + +## Where you see them + +- **Home › needs you** shows the best suggestion, one at a time. Choose **review** to open it, or say no in place. +- **Policies › suggested from your code** lists them all. Each card shows the rule, the line of code it came from, how it was tested, and the policy itself. + +Every card says how much testing stands behind it, in plain numbers: for example *tested on 9 examples (4 to stop, 5 to allow) · no traffic yet, tested on examples only*, or *replayed on 1,240 real calls from 7 days: would have acted 3×*. + +## Saying yes + +**Yes, deploy** opens a short sheet: + +- The machines that run this agent's kind of agent (Claude Code, Codex, …) are already selected. Uncheck any you don't want. +- The policy is deployed in **observe** mode: it runs on every call but only records what it would have stopped, and blocks nothing. +- With no machines connected yet, **publish only** keeps it ready to deploy later from its page. + +Once an observed policy has acted on live traffic without stopping a working call, it shows as **ready to enforce** on Home and on the policies list. Switching it to enforce is a separate, deliberate step. See [deploy a policy](/policies/deploy). + +## Saying no + +**No** dismisses the suggestion. It stays dismissed until the code it came from changes. + +## Who can do what + +| Action | Permission | +| --- | --- | +| See suggestions | `policies:read` | +| Say yes or no | `policies:write` | + +Disconnecting GitHub removes every suggestion that quotes your code. Policies you already published stay.