Adk Review

ADK Change Reviewer

Review the working-tree diff against the seven dimensions below, report what is wrong, and stop. Fix only what the user then asks you to fix.

Workflow

1. git status and git diff (add --staged for staged work) to get the exact set of added, modified, and deleted files.
2. Review the diff file by file against the checklist.
3. Emit the report in the format below.
4. Stop. Do not edit any file, and do not offer to. Wait for the user to ask.
5. If the user asks for fixes, apply them, then re-run the affected tests (pytest tests/unittests/{path}) and pre-commit run --files {paths}.

Step 4 is the part that is easy to get wrong: an unrequested fix buries the findings the user asked for and mixes review output with new, unreviewed edits.

Checklist

1. Correctness

• Types: no new mypy errors. CI diffs mypy output against the base branch and fails only on newly introduced errors, so a pre-existing error in a file you touched is not a blocker but a new one is.
• Imports: no circular imports; absolute imports where the module already uses them.
• Exceptions: no bare except:; catch a specific type and log with enough context to identify the caller.
• Type discrimination: check an object's type before reading a type-specific attribute (for example confirm a node is an LlmAgent before inspecting mode), so an unexpected node type raises nothing.
• Boundaries: None, empty collections, zero, and empty strings are handled by validation or a fallback default.
• Preconditions: state invariants are checked before the core logic runs.

2. Design

• Complexity: functions or classes that would be clearer split up.
• Coupling: high cohesion, low coupling; no anti-patterns introduced.
• Performance: redundant computation, repeated I/O in a loop, or allocations that scale with input where they need not.
• Security: inputs validated, sensitive data not logged, no injection, resource exhaustion, or exposure of internal state.

3. Style

Cross-reference the diff against the adk-style skill rather than restating its rules here: visibility and _ prefixes, typing, Pydantic v2 patterns, lazy logging, imports, async, and file organization.

Confirm the changed files pass pre-commit run --files {paths}.

4. Architecture and unintended outcomes

• Public API stability: does the change modify, remove, or narrow a public interface, class, method, argument list, or CLI surface under src/google/adk/? A breaking change needs a deprecation cycle first, because the package is released under Semantic Versioning and users pin minor versions.
• Execution and resumption: changes to workflows, nodes, or state must stay compatible with the event execution lifecycle and with session resumption (human-in-the-loop steps and checkpoints). See the adk-architecture skill.
• Concurrency and lifetime: no race conditions; plugins, exporters, and connections are closed on every path, including the error path.
• Database and runtime DDL safety: runtime startup paths (e.g., DatabaseSessionService.prepare_tables() and _ensure_schema_indexes_exist()) must never execute destructive DDL such as DROP INDEX (which takes ACCESS EXCLUSIVE locks and breaks rolling deployments/rollbacks). Any additive runtime DDL must account for multi-container startup races (process-local asyncio.Lock does not prevent cross-container TOCTOU races on checkfirst=True) by wrapping creation in a SAVEPOINT (connection.begin_nested()) and tolerating concurrent creation.
• Scope: a new public name, parameter or flag the fix did not need; a flag that keeps the buggy behavior as the default; a fix in a shared layer when one caller needed it; an existing test assertion edited to match the new behavior.