Inspect, reproduce, evaluate

Proof with a stated boundary.

Different evidence answers different questions. No browser simulation establishes production durability, throughput or universal exactly-once execution.

Browser simulation

Illustrates the decisions. No SDK, provider writes or persistent state.

Explore decisions

Installed SDK

Local fake provider and separate journal. Verify first effect, replay, conflict, fresh process and UNKNOWN blocking.

Run current quickstart

Real provider example

GitHub issue recovery with bounded authoritative lookup and explicit correlation assumptions.

Inspect integration

LangGraph checkpoint replay evidence

Deterministic local LangGraph JS checkpoint replay resumed the same checkpoint twice while the first tool call was still running. The unprotected path produced 2 provider effects; the Once-protected path produced 1, using published SDK 0.1.25. Counts come from a separate loopback HTTP provider's append-only log, not returned SDK status.

Limits: not a LangGraph Cloud reproduction, not a test of the reported ~180s sweeper, and not a Cloud bug fix. Protection used one persistent shared same-machine SQLite authority. Separate ephemeral Cloud ledgers, distributed-worker safety and universal exactly-once execution are outside this evidence. The provider is a controlled local fixture, not a commercial service.

Run and inspect the reproduction · Merged proof PR #295 · Proof commit

Can you reproduce #7417 on LangGraph Cloud?

Once maintainers are seeking one voluntary, independent reproduction. If you already have Cloud access and can reproduce issue #7417, try one bounded comparison around a non-production tool and report failures or incompatibilities as well as successes. No purchase or production write is requested.

First record the exact Cloud deployment, region, runtime and package versions, worker/storage topology, tool duration, and sanitized thread/checkpoint/tool-call/attempt IDs. Count actual provider effects independently for the unprotected and protected paths; include timestamps, the stable logical operation ID, complete effect binding, ledger authority and reconciliation method. Returned status or tool invocation counts alone are insufficient.

Do not copy local SQLite into separate ephemeral workers and assume they share authority. If a persistent shared same-machine authority cannot be established for this boundary, report that incompatibility and stop the protected test. Keep UNKNOWN blocked; do not mint a new ID, bypass protection or replay a real consequential write to finish the experiment.

Offer or report a Cloud test on the Once repository, with redacted provider logs and exact environment. One bounded run is enough; no praise, endorsement or ongoing commitment is requested. No independent Cloud reproduction or adoption is established by the local result.

Google ADK completed-tool replay lab

A controlled comparison separates missing successful-tool history from a deliberately forced replay. Released ADK 2.11.0 and the exact base of upstream PR #7429 produced 2 raw effects under the scripted history-based policy; the proposed fix produced 1. Forced replay produced 2 raw effects versus 1 protected effect on all three variants. Counts come from the fixture provider's separate durable journal.

The lab also checks changed-effect conflicts, fresh protection-process replay, lost-acknowledgement/crash UNKNOWN blocking, exact fixture reconciliation and twelve concurrent boundary workers. Windows and Ubuntu CI reproduce the pinned comparisons.

Limits: a local scripted model and mock provider, not a live model, production provider, Google endorsement or whole-framework recovery qualification. Concurrent boundary workers share one persistent same-machine SQLite authority; this does not qualify concurrent ADK orchestration or distributed storage. The upstream fix is compared as a pinned proposal, not claimed as a released capability.

Inspect methodology and reproduce the ADK lab · Merged lab PR #304

Controlled booking lifecycle reference

Creation, modification and cancellation use the published SDK 0.1.25 wrapTool boundary with a disposable local provider. Fifty-eight probes check full effect binding, receipt replay, conflicts, quote expiry and version refusals, lost acknowledgement, blocked delayed or missing truth, exact reconciliation and twelve concurrent duplicate attempts. A separate durable provider journal records 11 protected effects and 2 deliberate unprotected duplicate-control effects.

Limits: a working mock reference contract, not a production booking adapter or named-provider support. Windows and Ubuntu clean-consumer tests cover the same-machine boundary. Real authentication, inventory/occupancy policy, timezone scheduling, whole-provider restart and distributed authority remain outside this evidence. UNKNOWN and non-authoritative absence never permit redispatch.

Run and inspect the booking reference · Merged reference PR #312 · Provider qualification boundary

Framework crash/retry labs

LangGraph and CrewAI examples use controlled simulated provider ledgers and the SQLite reference backend. They test hard process termination and fresh-process resume/redispatch. Inspect each lab’s methodology before generalising its result.

Routing conformance

Published fixed-suite results describe the evaluated model sessions and cases, not universal model reliability or production routing correctness.

Routing methodology and integrity trail

Hosted and historical evidence

Hosted playground evidence and its scope. Stripe staging proof is a specifically scoped staging experiment, not a blanket live payment/refund adapter guarantee.

Historical reproducibility fixture — v0.1.9. Frozen versions remain pinned intentionally. For a current-package consumer, use the maintained quickstart.

Review the source and safety tests

Local safety depends on one stable logical intent, complete effect binding, durable same-machine state and every consequential call traversing the protected boundary. Authoritative confirmation may resolve UNKNOWN; local absence does not cause automatic ambiguous redispatch.
Verify your first action