AppWorld execute recovery and verification
Failures across Spotify and phone/venmo tasks share recovery and tool-use patterns distinct from preflight planning.
Retry after partial execute failure
- When a script dies after side effects but before the planned filtered call, rerun the exact intended call with all original task filters.
- Do not drop filters such as
min_follower_count; broader retrieval changes the target set and under-covers bulk mutations. - Fix only the aborting error (e.g. failed import); keep the original query.
Re-fetch after mutation
- A mutating call such as
follow_artistcan change the metric used in the task condition or final answer. - Keep the pre-mutation snapshot for eligibility, but re-fetch the record after mutation and report the post-action value.
- Do not present a stale pre-action count as current without labeling it.
Execute sandbox constraints
- Do not use
open()or other disallowed built-ins in an AppWorld execute block to persist tokens/state; the whole block aborts and prior output is lost. - Keep tokens in variables in the same block, or re-fetch them later.
App-specific login identifiers
- Do not assume the supervisor email works for every app.
- Check each app's login identifier; phone may expect a phone number, venmo with an email can return 422 account-not-found.
- On 422 account not found or 401 invalid credentials, stop cycling variants and consult app docs/provisioning.
Verification
Before answering, confirm: retries kept all original filters, mutated fields were re-read after action, no disallowed built-in was used, and each app used its documented login identifier.
This complements appworld-api-preflight-verification (pre-bulk checks); this
skill covers post-execution recovery and verification.