AgentX v1.0.4

Non-Breaking Changes / Validation Fixes (No Performance Effect)

• Fixed AgentX profiling liveness validation to accept either late TTFT or late inter-token activity at the existing 98% cutoff.
• Healthy long-running streams no longer fail solely because no new request starts near the profiling boundary.
• Genuinely stalled runs still fail when neither signal reaches the cutoff.
• No changes to request payloads, workload pacing, concurrency, latency calculations, routing, or inference-server execution.

Validation

• 103 focused unit tests passed.
• All applicable pre-commit checks passed.
• Regression coverage includes the observed TTFT=97.9504%, ITL=100% false-positive case.

What's Changed

• fix(agentx): accept either latency signal as liveness by @cquil11 in #39

Full Changelog: agentx-v1.0.3...agentx-v1.0.4

AgentX v1.0.3

Non-Breaking Changes / Bug Fixes (No Performance Effect)

• Fixed concurrent mmap-backed conversation reads by replacing shared-cursor seek()/read() operations with position-independent byte-range slicing.
• Added deterministic regression coverage for parallel conversation reads.
• No changes to AgentX request payloads, workload pacing, concurrency, synthetic AL, latency timing, metric calculations, routing, or inference-server execution. This release only prevents corrupted or truncated dataset records when parallel workers race the shared mmap cursor.

Validation

• 94 targeted dataset and mmap tests passed.
• The full unit test suite completed successfully.
• Pre-commit checks passed for the changed files.

What's Changed

• fix(dataset): make mmap conversation reads concurrency-safe by @cquil11 in #37

Full Changelog: agentx-v1.0.2...agentx-v1.0.3

AgentX v1.0.2

Non-Breaking Changes / Bug Fixes (No Performance Effect)

• Added full_decode_duration, measured from the first parsed content response through explicit HTTP response completion.
• Added full_response_inter_token_latency, aligning the full client-observed decode interval with the reported output-token count.
• Added full_response_output_token_throughput_per_user as the inverse of full-response ITL.
• Preserved all existing parsed-content metrics; the new full-response metrics are additive.
• No changes to request payloads, workload pacing, concurrency, synthetic AL, routing, or inference-server execution. This release adds reporting metrics only and does not alter system performance.

Validation

• 50 targeted metric, registry, and JSON exporter tests passed.
• 833 metrics, property, and exporter tests passed.
• Full pre-commit suite passed.

What's Changed

• feat: add full-response ITL and output throughput by @cquil11 in #36

Full Changelog: agentx-v1.0.1...agentx-v1.0.2

AgentX v1.0.1

Non-Breaking Changes / Bug Fixes (No Performance Effect)

• Added AgentX post-run validation requiring TTFT and inter-token-latency observations to cover at least 98% of each valid duration-based profiling phase.
• Runs with incomplete profiling coverage now exit non-zero, retain diagnostic artifacts, and mark the submission invalid with insufficient_profile_metric_coverage.
• Warmup observations and intentionally short --unsafe-override smoke runs are excluded from the coverage check.
• No changes to request payloads, workload pacing, concurrency, ordering, warmup handoff, spawn/join behavior, latency calculations, routing, or inference-server execution. Healthy-run performance results are unchanged.

Validation

• Full local pre-commit suite passed.
• Targeted changed-path suites passed, with up to 170 tests per independent audit.
• Healthy mock profiles passed with 99.8–100% TTFT coverage and 100% ITL coverage.
• Deliberately stopped or suspended mock servers failed with retained artifacts and the expected invalid-submission reason.

What's Changed

• feat(agentx): fail incomplete profiling metric coverage by @cquil11 in #35

Full Changelog: agentx-v1.0.0...agentx-v1.0.1

AgentX v1.0.0

Final set of Breaking Changes to AgentX v1.0 Harness

Thank you @cquil11 and team for the hard work!

What's Changed

• feat(agentic): add deterministic warmup request budgets by @cquil11 in #27
• fix(agentic): make per-lane warmup additive to primers by @cquil11 in #28
• docs(agentic): clarify additive warmup budget by @cquil11 in #30
• fix(records): abort live on failed request threshold by @cquil11 in #32
• fix(errors): clarify server-triggered aborts by @cquil11 in #33
• fix(agentx): preserve trace timing and joins by @cquil11 in #31

Full Changelog: https://github.com/SemiAnalysisAI/aiperf/commits/agentx-v1.0.0