Kimi Code CLI

The README says the CLI can read and edit code, run shell commands, search files, and fetch web pages, giving it broad capabilities. The supplied material does not show fine-grained permission limits or confirmation for risky operations, so least privilege and user confirmation are deducted. The README mentions plugin trust levels and lifecycle hooks, but does not explain data flows, secret protection, confirmation of external effects, or recovery. Those criteria receive only 1. The repository provides a vulnerability reporting channel, but publisher identity is unverified and the supplied code offers limited evidence about provenance and dependency risk, keeping source attribution and dependency security at 1.

The README lists development build, test, typecheck, and lint commands. CI uses frozen dependency installation and includes sharded tests, a build smoke check, and typechecking, supporting ordinary maintenance; self-consistency scores 2. The evidence does not fully establish dependency availability or known failure handling, so that criterion scores 1. Doctor tests cover missing configuration, parse errors, field paths, and aggregated errors, while CLI tests check error guidance, supporting a score of 2 for failure messages. These static files do not mean this review executed or verified those tests.

The README identifies terminal coding tasks and describes video input, subagents, MCP, and IDE ACP scenarios, supporting a score of 2 for audience and scenarios. It lists many tool capabilities but the supplied material does not clearly define safety boundaries or unsupported tasks, so capability boundaries score 1. The project-directory launch, login steps, IDE setup, and doctor command provide clear entry points, scoring 2 for trigger precision. Installation covers macOS, Linux, and Windows, notes the Git Bash requirement, and gives Node/pnpm development requirements, scoring 2 for environment fit.

The README is organized around installation, quick start, features, IDE use, docs, and development, though it relies on external documentation, scoring 2 for information architecture. Install commands, first-login steps, and the Windows prerequisite are concrete, scoring 2. CLI, ACP, and doctor naming is mostly consistent across README, package scripts, and tests, scoring 2. Quick start and configuration/IDE documentation links are present, but the supplied files include no FAQ or broad example set, scoring 1. The supplied files do not systematically document user-facing limitations, scoring 1. The complete MIT terms are present, scoring 3. Package versioning, Changesets scripts, and CI branches are visible, but no changelog for this revision is supplied, scoring 2. SECURITY.md states that only the latest release is supported and provides reporting channels, but does not give response deadlines, so maintenance responsibility scores 2.

The README describes code editing, command execution, and iterative next-step selection; doctor tests also show diagnostic output and field-level errors, supporting a score of 2 for output usability. IDE ACP, subagents, MCP configuration, and video input expand coding workflows, but these are largely README claims, so marginal value scores 2. Single-command installation and single-binary distribution reduce setup steps, while model login/credentials are required and the agent can perform broad actions; the evidence does not quantify operating costs or risk-benefit, so cost-benefit scores 1.

Most README claims about features, speed, and experience are not directly quantified or individually corroborated by the supplied source files, so claim traceability scores 1. README, tests, package scripts, and CI provide some cross-checks for commands, diagnostics, build, and test workflows, scoring 2 for cross-source corroboration. The evidence mixes checkable installation/configuration facts with unquantified promotional claims without explicitly separating facts from inference, scoring 1.