Repeatability
Medium
The mechanical scanning and AST transformation steps are highly repeatable, but the naming convention decisions (e.g., which of three conflicting patterns becomes canonical) require one-time judgment that varies per codebase. The agent can't reliably make those calls the same way a human would each time.
Ambiguity Tolerance
Low
Success criteria are underspecified: 'inconsistent prop naming' could mean camelCase vs snake_case, abbreviated vs full names, or domain-specific aliases — each requiring different resolution logic. 'Auto-generate from actual usage' is also ambiguous when usage is sparse, contradictory, or polymorphic.
Data & Tool Availability
High
The agent needs read access to 40 source files, which is straightforward to provide. AST parsing tools (jscodeshift, ts-morph, Babel) are well-documented and accessible, and the agent can write and test the script in a sandboxed environment.
Error Cost
High
A bad codemod applied across 40 files can silently break prop contracts, introduce type errors that only surface at runtime, or rename props in ways that break downstream consumers not in the repo. Reversible with version control, but debugging a botched mass-rename across a large codebase is expensive.
Human Judgment Required
High
Choosing the canonical naming convention requires understanding team conventions, API contracts, and downstream consumers — context the agent cannot infer from the files alone. TypeScript interface generation from usage patterns also requires semantic judgment about optionality, union types, and intent that static analysis frequently gets wrong.