[Feature]: agent-mesh tests for attempt counts on lost-response retries #4198

Package

agent-mesh

Problem Statement

When a tool's write completes but its response is lost, the agent retries, and session_to_trust_record records a call count of 2 and an affirming appraisal whether one write landed, two did, or none has been seen. That is right for a record of attempts, but no test pins it, so a later change could read the call count as writes.

How does this feature impact your work and what are you trying to achieve?

I maintain agent-evidence-vectors (Apache-2.0, on PyPI), whose observed-effect corpus holds signed records of these three outcomes plus two a verifier must refuse: a duplicate reported as one write, and a timeout read as no write. Testing AGT's retry trace against those records makes a change on either side fail a test.

Timeline

The agent-evidence-vectors 0.14.0 pin in #4171 clears the seven-day Supply Chain Check on October 5, 2026.

Proposed Solution

#4171 adds a test module beside test_trace_model.py whose five cases on Python 3.13 compare the call count with each record's attempts, and #4196 registers the package with the dependency-confusion scanner.

Alternatives Considered

A test with no dependency pins the attempt count but cannot show three outcomes producing one Trust Record, and copied records would drift from the published corpus. An example describes the property; a test beside the trace model's tests fails when session_to_trust_record changes.

Priority

Important

Contribution

• I would be willing to submit a PR for this feature