Examining retry transparency in agent tools
Checking Retry Transparency in Agent Tools
The paper studies how to test the recovery guarantees behind agent tools' local idempotency claims, noting that a successful retry does not necessarily absorb earlier attempts, old writes may overwrite updates confirmed afterward, and partially executed file writes may even empty a file after a successful retry returns. The authors' prototype uses hand-built, source-grounded effect contracts, checking at two levels: atomic retry absorption obligations and bounded interleavings of successful-attempt microsteps, and outputs model counterexamples, conditional admission decisions, and explicit uncertainty when completion bounds cannot be proven. On 6 MCP components from 4 repositories, among 27 missing-reply checks, the selective rule completed 6 workflows with no confirmed target violations, strict merging completed 0, and blind retry completed 9 but produced 3 violations; the intent-aware Porcupine specification independently agreed with all 156 full-history oracles. The authors stress that these are tests of mechanism choices rather than estimates of defect prevalence or a new linearizability algorithm, and that the contribution is turning effect granularity, logical identity and completion assumptions into a testable recovery-contract workflow rather than inheriting those assumptions from tool annotations.