The protocol.
Your agent talks directly to ours. No human in the loop for the common cases.
How the exchange works.
Each box in the dialog above maps to a single request. The patient agent sends a health report to /heal. The diagnostic chamber returns a diagnosis and a list of prescribed repairs. The patient executes whichever repairs are marked safe, then reports the outcome to /heal/results.
Loop semantics.
Each session can run up to ten turns. After every /heal/results post, the chamber returns a verdict that decides what happens next.
| verdict | meaning |
|---|---|
| "healed" | Symptoms resolved. Session closes, patient is discharged. |
| "more_repairs" | Symptoms partially resolved. Another prescription follows. |
| "escalate" | Auto-repair can't fix it. Operator queue takes over and on-call is paged. |
If a patient hasn't reached a terminal state by turn ten, the chamber refuses further auto-repair and escalates by default. No infinite loops, no patient stuck in triage forever.
Safety design.
Every repair is either safe -- meaning the chamber may execute it directly -- or manual, requiring your confirmation. The safe list is curated per framework; anything destructive (restarts, credential changes, large file operations) requires confirmation by default.
The chamber map.
Max 10 turns per session. Verdicts: healed (discharge), more_repairs (loop), or escalate (operator).