Releases: FU-max-boop/agentrunproof
Release list
AgentRunProof 0.2.0
AgentRunProof 0.2.0 adds packaged-wheel coverage for OpenAI Agents 0.20.0 and 0.21.0, plus opt-in provider-free generation spans for observability integration tests.
The canonical five-run upstream comparison remains independently checkable and is rebound to the v0.2.0 wheel.
Bundle ID: sha256:da23130794081eaaaa0321953e596a62c5675a451711f552a1f31abb3f5a6349
Verify with:
agentrunproof check-upstream-bundle bundle.json
AgentRunProof remains a community-maintained external project; this release does not imply an official OpenAI dependency or endorsement.
AgentRunProof 0.1.2
AgentRunProof 0.1.2 adds deterministic recursive Agent.as_tool approval-routing checks across live and RunState JSON-restored execution.
The included canonical upstream-comparison bundle records a reproducible five-run causal ladder: openai-agents 0.20.0 fails top-level sibling isolation; the #4413 baseline fixes that case but still fails two-edge recursive routing; the official #4414 merge (50d65f65) passes both live and serialized recursive scenarios.
Bundle ID: sha256:707936202e73f0a969fd6bb6717605ff98e9ecc8342528afafd0cfda02dc92b9
Verify after installing the wheel:
agentrunproof check-upstream-bundle bundle.json
AgentRunProof remains a community-maintained external project; this release does not imply an official OpenAI dependency or endorsement.
AgentRunProof 0.1.1
Adds a public no-network sibling-RunState approval-isolation probe and the state_fork_isolation invariant.
The attached content-addressed certificate records the released openai-agents==0.20.0 counterexample: deciding one sibling state mutates an untouched sibling and permits one protected tool execution in both streamed and non-streamed variants.
The wheel and sdist are reproducibly built from the tagged source; SHA256SUMS covers both distributions. The certificate and bundle are independently checked by CI.
AgentRunProof v0.1.0
First public release of AgentRunProof, a deterministic runtime-conformance harness for openai-agents 0.20.x.
Highlights:
- executes real Runner.run and Runner.run_streamed workflows without model API calls
- checks terminal streaming parity, tool-call linkage, exactly-once effects, selected RunState interruption/resume semantics, and multi-phase session replay
- emits strict content-addressed public or redacted verification certificates
- ships canonical evidence reproducing the fixed boundaries for upstream issues #4322, #4244, and #4125 across six pinned SDK/version cells
- supports Python 3.10 through 3.14
The attached wheel, sdist, checksums, matrix, and bundle are the exact artifacts validated by CI. Streaming scope in v0.1 is terminal-event parity; token/delta timing, backpressure, cancellation, and arbitrary Session implementations remain out of scope.