A passing Python run leaves a Go deployment question
PR-AF's maintained installation and our measured checkout need separate evaluation. The README says a bare repository URL installed through AgentField selects the Go node. Our October 9, 2026 sandbox record covers the Python checkout at commit 421fbd23bf1c5a2c3916d7046c97b5273958e1f4. Its successful commands are useful setup evidence, but they do not establish that the maintained deployment works end to end.
PR-AF is intended for teams that want an additional AI review layer and can operate the service themselves. The README describes generating review dimensions, assigning focused agents, grounding findings in repository evidence and challenging them before producing GitHub comments. Those are documented design claims. We did not measure their effectiveness on pull requests.
What our recorded run establishes
The stored run used a sandbox configured with 3 CPUs and 8 GB of memory. Its install tail identifies Python 3.12.12. The dependency-install step exited successfully after 15.464 seconds, with 111 packages and 359 MB recorded in the dependency summary. The build step exited successfully after 0.656 seconds; its empty output tail does not identify a built artifact.
The test step exited successfully after 15.485 seconds. Its pytest tail says "126 passed in 13.35s"; the parser summary records 126 passed, zero failed, zero skipped and zero errors. The step duration and pytest's own duration measure different boundaries. We have not reconciled a complete log to establish exhaustive unique test coverage.
The checkout summary records 328 files, 38,746 source lines and 10.9 MB, plus a Dockerfile, Compose file, tests directory and one CI workflow. Pip-audit recorded zero advisories. Its coverage is unknown, so neither the advisory total nor the dependency count establishes that every dependency was checked.
There is no recorded installed-binary check, control-plane deployment, Go test run, model-backed review, GitHub comment delivery or measured review workload. These results cannot establish review accuracy, production reliability, cost or latency.
Choose the implementation before following setup instructions
According to the README, installing the repository URL through AgentField follows the manifest to the maintained Go package. Python remains available through a local-path install, python -m pr_af.app and the root Docker Compose stack. Select the route deliberately rather than treating our Python results as validation of both implementations.
The documented service needs AgentField, OPENROUTER_API_KEY and GH_TOKEN. Railway deploys the reviewer, control plane and PostgreSQL; its public execution endpoint requires X-API-Key. Operators also need to choose the provider, model, workspace location and execution limits.
The README lists default limits of $2.00 per run and 3,600 seconds. These are configuration defaults, not measured spending or guarantees. It describes reviews as typically taking 35-50 minutes, which we did not independently time. Its benchmark leadership and cost comparisons are also project claims, not findings from our sandbox.
Validate empty results and provider accounting
The supplied extract for issue 55 is open and reports a malformed diff returning a successful empty review. The report concerns pr-af-go v0.1.0 on macOS, not our Python sandbox run. It supports checking that invalid input fails clearly; it does not establish that every current installation has the same behavior.
Issue 56, also supplied as open, reports a valid diff-only review without repo_path. The reporter says verification searched the agent's package directory, dropped a data-race finding and returned a location-related finding after 27 minutes and 20 invocations. This is a reported Go-path failure, not a reproduced result from our run. Validate diff-only behavior separately from repository-backed reviews.
Issue 57 describes the same Go review using the OpenCode provider: 20 invocations over 27 minutes while the reported cost remained zero. That report does not establish failure across all providers. It does justify reconciling application accounting with provider records before relying on an application-level budget ceiling.
For a trial, use a pull request with changes you understand. Check that the service reviewed the intended files, inspect the evidence behind its findings and compare reported spending with provider records. Keep existing automated checks and human approval while evaluating those results.
Alternatives serve different review jobs
PR-Agent's README documents an open-source AI assistant with review, description and improvement commands, CLI use, GitHub Actions and several Git-provider integrations. It identifies the project as community maintained and distinct from Qodo's current offering. Consider it when those commands and integration routes fit your workflow; this review does not establish comparative accuracy or speed.
reviewdog takes output from lint tools, filters findings against a diff and posts comments to code-hosting services. Choose it when you need existing diagnostics delivered into review. It does not serve the same purpose as PR-AF's documented agent-based analysis.
How to read the scores
These scores assess the supplied evidence for adoption: 1 means major unresolved obstacles, 2 means limited evidence with consequential unknowns, 3 means a practical evaluation path with operational work, 4 means detailed support for the assessed area, and 5 would require broad corroboration beyond this record. They are editorial judgments, not benchmark results. Unknown adoption and untested compatibility do not earn a high score.
The next useful check is a supervised review through the implementation and provider you intend to deploy. Our Python command results support proceeding to that evaluation; they do not replace it.