This is every piece of reviewable content in the Repair Engine — dev-only, off for all real users right now. Nothing below has been shown to anyone but one synthetic test account.
Status: built and tested on oimy-engine-dev only. Kill-switched off by default; only live for one internal test user. Not deployed to production.
01The consent-ask — 4 exact variants
When you disclose a rupture (e.g. "I yelled at him" / "I lost my temper"), OiMy says one of these four sentences, chosen automatically and deterministically per case — never randomly re-rolled, never edited by the model. It does not give advice or a script at this stage — only this ask.
offer_1
"That sounded rough—would it help to have a couple of words for when things are calmer?"
offer_2
"That was a hard moment; want a short repair phrase for when things settle?"
offer_3
"Oof, that sounds rough—would you like two simple sentences you could say later?"
offer_4
"That sounds like a painful moment—want a few words for making it right when things are calm?"
02What happens after you say yes
Important: there is no fixed 5th piece of text for the actual reconnection script — despite how it may have been described to you. The words you'd actually say to your kid are generated live by the companion model, but under a strict, reviewed set of constraints (below), not freely. That's a real distinction worth you deciding whether you're comfortable with, versus insisting on fully fixed text.
Exactly one script, no more than two spoken sentences.
The parent owns their own behavior — the child is never asked to comfort or reassure the parent.
If age was mentioned this turn, wording is adjusted to be age-appropriate.
Any still-valid household boundary stays separate from the apology itself.
No explaining psychology, no citing experts, no diagnosing anyone.
No second strategy or extra advice bundled in.
This only fires after an explicit "yes" (or an explicit ask like "what should I say"). An unprompted "yes" with no pending offer produces nothing — verified in testing.
03The full flow, state by state
offered — rupture disclosed, one of the 4 lines above sent, expires in 6 hours if ignored.
script_delivered — parent consented, the constrained script above was generated and sent.
used — parent later says something like "I said that to her" — OiMy remembers it was used, nothing further is asked.
declined / expired — parent said no, or 6 hours passed — case closes, never resurfaces on its own.
Tested directly: a crisis-flagged message always overrides a pending repair offer — the repair flow never runs, never delivers a script, when there's any safety concern in the same turn.