OpenHands Agent Canvas Code Review Guidelines

This guide supplements the public code-review skill with rules specific to OpenHands/OpenHands, the Agent Canvas frontend. Read AGENTS.md first; it is the detailed source of truth for current architecture and test conventions.

Be direct and constructive. Review correctness and architecture, not formatting that lint or the compiler already checks.

Review Decision

• Submit exactly one review: APPROVE or COMMENT. Never use REQUEST_CHANGES.
• Default to APPROVE when there are no important findings. Nitpicks and optional cleanup are not reasons to withhold approval.
• Use COMMENT for correctness, security, architecture, missing evidence, or unmet acceptance criteria. Let a human maintainer make the blocking decision.
• Do not approve changes that can affect agent or benchmark behavior—prompts, tool selection, conversation payloads, terminal behavior, planning, memory, or evaluation paths—without human review and appropriate lightweight evals.
• Read the linked issue and include a compact checklist covering each acceptance criterion. Meeting the checklist is necessary but does not replace review for regressions, security, or maintainability.

Repository Ownership

Put behavior in the repository that owns it:

Repository	Owns
OpenHands/OpenHands	Agent Canvas UI, frontend state, backend selection, frontend service integration, and local-stack orchestration
OpenHands/software-agent-sdk	Agent Server, agents, tools, conversations, events, workspaces, and the canonical server API
OpenHands/typescript-client	Browser-compatible typed access to the Agent Server API
OpenHands/extensions	Reusable skills, plugins, and integrations
OpenHands/automation	Scheduling, webhooks, run history, and automation dispatch

The normal dependency direction is Agent Server contract → TypeScript client → Canvas. Flag raw endpoint reimplementations, Canvas-local copies of server contracts, and changes opened in the wrong repository.

Architecture That Guides Agents

Agents tend to copy the nearest pattern and choose the shortest compiling path. Review the codebase as part of the product surface that guides those choices:

1. Make the conventional path cheapest. New work should naturally reuse a named hook, service, store, or feature module instead of adding another branch to a shared root.
2. Fail forbidden dependencies mechanically. Repeated review guidance should become a lint rule, compiler boundary, or architecture test. Do not grow this document when a small executable guard would be clearer.
3. Give durable state one obvious writer. A backend setting, consent value, conversation cache entry, or persisted browser value should have one named owner. Flag second writers and component-local mirrors of authoritative state.
4. Prefer owned feature files over shared switches. Product work should usually extend a feature-owned module. Shared registries and root conditionals need a concrete reason.
5. Keep exceptions narrow and visible. Exceptions belong in a small allowlist next to the guard that enforces the rule and should be reviewed as architecture changes.

Treat “deep module” as a design heuristic, not a line-count target. A good module has a narrow, stable interface and hides cohesive complexity. Do not split a file merely because it is long, and do not create layers that only rename or forward arguments. Prefer a small pure seam when it removes duplicated decisions, makes ownership explicit, or enables focused tests.

React effects

useEffect is for synchronizing React with an external system. Flag effects used to:

• derive render data from props or state;
• respond to a user action that can run in the event handler;
• initialize a value that belongs in a lazy state initializer;
• mirror one store or cache into another component state value; or
• repair ordering created by competing writers.

An effect is not automatically wrong. Subscription, browser API, timer, and network synchronization still belong in effects when cleanup and dependency semantics are explicit.

Blocking Architecture Checkpoints

Agent Server and Cloud API access

src/api/no-direct-agent-server-calls.test.ts is the executable source of truth. Do not approve new raw fetch, axios, shared openHands, or low-level HTTP client access to Agent Server endpoints. Use @openhands/typescript-client with the options from src/api/agent-server-client-options.ts.

Cloud and runtime-sandbox requests must go through callCloudProxy; runtime requests must provide the correct hostOverride and authentication mode. Review changes to the guard's allowlist as architecture changes. Do not copy its current entries into this guide—the test should remain the one authoritative list.

Event wire contracts

The SDK event model is the wire authority, the TypeScript client mirrors it, and Canvas consumes the published client type. Do not approve Canvas-local redeclarations, partial intersections, module augmentation, or presentation fields added to wire-event interfaces.

A contract change should land in this order:

1. SDK model/schema and serialization coverage.
2. TypeScript-client mirror derived from the SDK payload.
3. Published client release.
4. Canvas consumption and rendering/telemetry coverage.

Canvas-only presentation state belongs in a separate view model keyed by event identity.

Telemetry and durable frontend state

• src/services/telemetry.ts is the only owner of the Canvas PostHog client.
• React events go through typed functions in src/hooks/use-tracking.ts; components must not call PostHog directly.
• Consent rendering uses the telemetry consent external store, not mirrored local state. setTelemetryConsent remains the single consent controller.
• A business milestone has one canonical capture. Flag duplicate conditional captures.
• For other durable values, prefer the existing named service/store/hook and flag new storage writes from arbitrary components.

Dependencies and Releases

• Direct dependencies are exact-pinned. Keep package.json and package-lock.json synchronized through npm; do not hand-edit one side only.
• Treat changes to dependency exemptions, git pins, and security overrides as reviewable policy changes. __tests__/package-library.test.ts is the executable source of truth for allowed specs.
• Scrutinize newly published third-party dependency versions for supply-chain risk. First-party OpenHands packages are exempt from a waiting period but not from contract and release-order review.
• Package version changes belong in explicit release PRs and must match the release workflow expectations.

Testing and Evidence

• Require evidence proportional to the behavior changed. For UI behavior, use a screenshot or video from the real app. For CLI, API, or scripts, require the exact runtime command and observed result. Unit tests alone are not end-to-end evidence.
• Prefer tests that exercise real logic and observable state. Do not reward mocks that only prove another mock was called.
• Keep tests focused: one meaningful assertion path per behavior, no duplicated coverage of library behavior, and no brittle presentation-only snapshots.
• Follow the test routing in AGENTS.md. If a change crosses a full-stack flow and lacks suitable coverage, recommend mock-LLM E2E and add the e2e-tests label when appropriate.
• Never broaden live E2E triggers or secret exposure for convenience.

What Not to Comment On

Do not leave review comments for:

• formatting or minor style that tooling handles;
• optional “nice to have” refactors unrelated to the change;
• praise-only observations—approve instead;
• extra tests for straightforward data/config changes when existing checks cover the risk; or
• temporary .pr/ artifacts, which are cleaned up by repository automation.

When raising a finding, trace the relevant call or data flow far enough to show the concrete failure mode. Prefer one high-signal comment over several symptoms of the same ownership problem.

Communication Style

• Be concise, specific, and friendly.
• Explain the user-visible or architectural consequence.
• Suggest the smallest viable correction.
• Use GitHub suggestion syntax for local fixes.
• If the PR is sound, approve it without manufacturing feedback.