[Agentproto] Re: AgentProto implementation evidence: Cross-domain dialog recovery after mediator failure (B5-2)

[Nikhil D <[e-mail]>]

Hi Nicholas and Lars,


Thank you both for your detailed technical feedback and independent review.

I have now completed native macOS verification of the corrected AgentProto
revision:

Commit: 22ff90d7085cbeb47c852f0ad0a8a9fd7da1cf47

Environment: macOS 26.6.2, Apple M5 (arm64), Python 3.15.0.

Verification results:

   - Full test suite: 379/379 passed.
   - Restart suite: 10 consecutive runs, with 74/74 tests passing in each
   run.
   - Previously affected global-store restart-identity test: 30/30 separate
   runs passed.
   - Total: 1,149 passing test executions, with no failures or skips.
   - No ConnectionRefusedError, Errno 61, or wait_entered timeout was
   observed.

The logs, JUnit results, and run summaries were reviewed for consistency.

These results confirm that the previously reported failures were not
reproduced under this macOS test schedule. They do not conclusively
establish the original root cause or eliminate the possibility of
intermittent failures.


I have attached the reviewed verification report for reference.

Thank you again for helping strengthen the reproducibility and reliability
of this implementation evidence.


Best regards,

Nikhil Reddy Donapati


On Mon, 12 Oct 2026 at 11:10, Nicholas Templeman <[e-mail]> wrote:

> Hi Nikhil and Lars,
>
> Yes. The source and acceptance-test definitions at
> 278fbf573522b6a45366c9e7f33a0235f4f715b8 now represent the SENT-before-ack
> restart case and distinguish a missing acknowledgement from confirmed
> non-execution.
>
> Case C preserves the original operation identity, leaves the replacement
> process IN_DOUBT while the result is unresolved, and accepts the original
> COMMITTED result without a replacement invoke. Case D establishes
> NOT_EXECUTED only after fencing the original attempt; a late copy is
> refused, and an explicit retry keeps the operation identity. Case I
> preserves an unresolved outcome when the downstream system cannot be
> checked. The unsupported no-ticket case also shows why MAPPING_NOT_FOUND
> cannot establish non-execution: the original operation can already have
> executed.
>
> Pinned tests:
>
>
> https://github.com/nikhilred2619/agentproto-privacy-preserving-dialog-recovery/blob/278fbf573522b6a45366c9e7f33a0235f4f715b8/tests/test_process_restart.py
>
> This was a source review, not a new run. The published 74 restart passes
> and 379 suite passes remain reported results; I have not independently
> established execution of that exact final revision. The reasoning depends
> on atomic downstream fencing, one live Agent A per ledger, and a durable
> authoritative B ledger.
>
> On the reproduction, the 96/97 figure compares documentation claims. The
> methodology's 23 is stale against the generated summary and raw result of
> 13. The 270 unsupported no-ticket refusals do not demonstrate detection by
> the later fault checks.
>
> The published Mac log records 1 failed and 73 passed, with
> ConnectionRefusedError before the affected restart assertions completed.
> The earlier timeout remains a separate reported observation. The backlog
> probe does not establish the cause of either failure.
>
> Reproduction summary and linked run receipts:
>
>
> https://github.com/MoltyCel/reproduction-runs/blob/266d1018f8b8baf844895622cd055a6899c1bbc0/agentproto-dialog-recovery-278fbf57/run-summary.md
>
> This answers the narrow acceptance-case question within the prototype's
> stated assumptions; platform reliability remains open.
>
> Best,
>
> Nicholas
> _______________________________________________
> Agentproto mailing list -- [e-mail]
> To unsubscribe send an email to [e-mail]
>