airlock studio — paste-a-transcript rehearsal sandbox¶
Landed in v0.6.0. Implementation src/agent_airlock/studio/; CLI
src/agent_airlock/cli/studio.py.
A local HTTP sandbox for rehearsing a policy bundle against an agent transcript. Paste a transcript into the form, and each line is evaluated against the loaded bundle with inline verdicts plus a diff against the previous run — so you can change a policy and see exactly which lines changed verdict.
Stdlib http.server renders it, so the default path needs no extra.
Usage¶
| flag | default |
|---|---|
--host HOST |
127.0.0.1 |
--port PORT |
8010 |
Why it exists¶
Reading a policy and knowing what it will do to a real transcript are different skills. The diff-against-previous-run is the part that matters: tightening a rule usually changes more verdicts than the author expects, and this shows which.
Positioning anchor: Microsoft's Agent Governance Toolkit.
Honest scope¶
- Rehearsal only. Studio evaluates a transcript you paste. It does not sit in a request path, does not gate a live agent, and writes no policy.
- The transcript is text you supply, so fidelity is your responsibility — a hand-written transcript rehearses the policy against what you imagine an agent sends, not against what it sends.
--host 0.0.0.0binds publicly with no authentication, and the form accepts whatever is pasted into it. Keep it on loopback unless you have put something in front of it.- FastAPI is gated behind an extra; the stdlib server is the always-available default and is what these instructions describe.
See also¶
airlock console— the terminal equivalent, over an audit logairlock policy compile / explain— authoring the bundle