repro: it calls `create_hold` before it has a customer id and passes the string "customer_id"
0.51, agent mode, four tools. the order that has to happen is `lookup_customer` then `list_slots` then `create_hold`. it is written in the rules file in plain english, one line, numbered 1 2 3. what it does about 1 in 4 re-rolls is fire `create_hold` first with customer_id set to "customer_id", or once to "<id>", and one time to an empty string. my endpoint takes all three because the validator only checks the type. so there is a hold sat in the db against nothing and nothing errored anywhere. the part that gets me is it prints the order back correctly when i make it plan first. plan says 1 lookup 2 slots 3 hold. then the first call in the trace is the hold. so it isnt a case of it not knowing the order `apps/api/src/holds.ts:61` is where it lands if anyone wants the shape of it. no repro on a fresh project with the same 4 tools, which is why i havent filed anything with anyone. things i tried: numbering the tools in their own descriptions, saying `never` call hold without a real id, saying it twice in two places. 15 re-rolls or so across all of that. best i got was 1 in 6. separate thing, did tool descriptions get truncated somewhere around the tuesday build or is that just me
Not run against a model yet
1 works · 0 fails
Times copied by users