AI SDK Parity Review

SkillCommunication

Review any requested ai-sdk scope against the registered upstream Vercel AI SDK parity baseline. Use for PRs, the current git diff, a branch range, a package or directory such as providers/anthropic, a reported bug, a new feature, stream chunks, SSE framing, provider messages, provider request conversion, orchestration, tools, output, provider options, frontend interop, or conformance fixture changes.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the AI SDK Parity Review skill

What this skill tells your AI

The instructions your AI receives, as published by grafana/ai-sdk in .agents/skills/ai-sdk-parity-review/SKILL.md and read by ahel’s review.

Use this skill to review a user-defined scope for problematic upstream parity issues against the baseline declared in test/conformance/upstream.yaml.

The scope comes from the user request. It can be a PR, the current git diff, a branch range, a package, a directory, a provider, a bug report, or a feature area. If the user does not specify a scope, review the current git diff.

Workflow

  1. Read test/conformance/upstream.yaml and test/conformance/PARITY.md.
  2. Resolve the review scope:
    • PR: inspect the PR diff and relevant changed files.
    • Current git diff: inspect staged and unstaged changes.
    • Branch/range: inspect the diff against the requested base.
    • Package/directory/provider: inspect the implementation and matching tests in that tree.
    • Bug or feature: inspect the affected code path and any existing fixtures.
  3. Identify the parity-sensitive layer and capability from the coverage map: core ai-sdk, provider contract, provider implementation, frontend interop, or conformance harness.
  4. Read the matching upstream TypeScript implementation and tests for the registered baseline when available. Prefer sources in this order:
    • a local upstream checkout at the registered version or matching tag
    • the conformance tool package installed from the registered baseline
    • raw GitHub source for the registered package version Record which source was used when reporting findings.
  5. Check conformance evidence appropriate for the layer: expected.jsonl for core stream/UI behavior, expected-requests.jsonl for provider behavior, expected-object.json for structured output, or a documented manual/gap entry.
  6. For reported bugs or new upstream-visible behavior, check whether conformance-first TDD is possible. If so, prefer adding or updating the fixture before or alongside implementation.
  7. Classify each mismatch using the divergence handling table below.
  8. Report only problematic findings: missing features, behavioral deviations, possible bugs, missing or weak conformance coverage, and undocumented intentional deviations. Do not enumerate parity-preserving matches unless they explain why a suspected issue is not a problem.
  9. For broad scopes, include a compact matrix with the reviewed package/area, upstream files or tests checked, conformance artifacts checked, mismatch classification, action taken, and residual risk.
  10. For PR reviews, put accepted gaps, intentional deviations, and residual parity risk in the PR description or review summary. Do not add broad one-off findings to test/conformance/PARITY.md or test/conformance/upstream.yaml unless the user explicitly asks or the finding changes a long-lived project policy, coverage map, or baseline contract.
  11. Include commands run and residual parity risk in the review summary.

Divergence Handling

ClassificationDefault action
Implementation bugAdd or update the relevant fixture when practical, fix the Go implementation, and rerun focused checks.
Upstream behavior changeTrace the change to upstream source or tests, regenerate the matching conformance artifact, adapt Go behavior, and rerun parity checks.
Intentional Go deviationKeep only with a rationale. Document in the PR description or review summary unless it changes a durable baseline contract.
Coverage gapAdd a fixture or test when cheap and relevant. Otherwise document the residual risk in the PR description or review summary.
Large design gapStop and ask before deferring when the gap affects public API shape, wire format, provider contracts, or broad harness behavior.

Review Criteria

  • Wire shape, chunk type names, JSON tags, ordering, and SSE framing must match upstream unless a documented deviation exists.
  • Provider request bodies and behavior-affecting headers must match upstream request snapshots unless a documented normalization rule applies.
  • Go implementation can use Go idioms, but upstream semantics must be preserved.
  • New gaps must be fixed when practical. Accepted gaps and intentional deviations must be visible in the PR description or review summary; reserve test/conformance/PARITY.md and test/conformance/upstream.yaml updates for durable coverage-map, baseline-contract, or user-requested documentation.
  • If conformance is not updated for a wire/provider-boundary change, the review must explain why existing coverage is sufficient.
  • Provider implementation changes usually need expected-requests.jsonl coverage; core stream changes usually need expected.jsonl coverage.
  • If no upstream equivalent exists, state that and classify the compatibility impact.

Signals

GitHub stars
252
Forks
12
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
ai-sdk-parity-review
Source
github.com/grafana/ai-sdk