Code Review

Review Semiotic pull requests

Produce a findings-first review grounded in executable behavior. Treat AGENTS.md as the repository authority. Do not restate the diff or praise the change; identify defects that the author can act on.

Establish the review contract

1. Read the pull request description, changed files, and relevant sections of AGENTS.md.
2. Classify the change surface: chart HOC, Stream Frame/runtime, browser interaction, SSR/server, package surface, docs/example, AI/schema, generated artifact, test infrastructure, or custom lint.
3. Identify the user-visible or package-level behavior the change claims to preserve or introduce.
4. Use GitHub MCP when the pull request references an issue, incident, prior pull request, or failing check. Read the referenced acceptance criteria and compare them with the implementation. Do not infer requirements from an identifier without retrieving it.
5. Inspect available CI results through GitHub MCP when they can confirm a suspected regression or reveal an unverified path. Distinguish a product defect from a stale generated baseline or an unstable-rule disagreement.

Trace behavior, not files

Follow changed values across component boundaries. A local implementation can be type-correct while breaking a downstream scene, renderer, or package contract.

For chart HOCs and Stream Frames, check:

• Accessor, grouping, style, tooltip, hover, selection, legend, margin, and frameProps precedence through the final Stream Frame props.
• Line-object versus flat-row normalization and preservation of parent series metadata.
• Static mode versus push mode. Static and serialized paths require real data; React push mode omits data. data={[]} is not push mode.
• Controlled updates, refs, rAF scheduling, cleanup, transition continuity, stable configuration identity, and retained-scene invalidation.
• Coordinate-space assumptions involving margins, legends, scales, responsive dimensions, device pixel ratio, canvas bounds, and pointer type.
• Browser, SSR, hydration, static renderer, and serialized/MCP parity when the shared behavior crosses those paths.

For public API changes, check all affected surfaces together:

• Runtime implementation and exported TypeScript types.
• Family entry points and package exports without accidental graph widening.
• Chart specs, ai/schema.json, behavior contracts, prompts, examples, and reference coverage when agent-visible behavior changes.
• Documentation examples using high-level charts and family subpath imports.
• Generated sections and manifests updated only through their owning generator.

Demand meaningful evidence

Evaluate whether tests prove the changed behavior rather than merely mounting a component or finding a canvas/SVG element.

• Prefer assertions against scene summaries, rendered marks, callbacks, accessible output, or user-visible behavior.
• In canvas interaction tests, account for automatic legends, margins, responsive sizing, and plot coordinates before trusting literal pointer positions.
• Require regression coverage for the exact input form and interaction mode that failed.
• Treat regenerated bundle, package, and measurement baselines as evidence of an intentional contract change only when the implementation explains the delta.
• Recommend focused checks from AGENTS.md; do not claim a command passed unless CI or a tool result shows that it completed successfully.