Skip to content

Latest commit

 

History

History
142 lines (96 loc) · 6.6 KB

File metadata and controls

142 lines (96 loc) · 6.6 KB

CLI Reference

The evals executable uses Commander. Run evals --help or evals <command> --help for generated help from the installed version. This page covers the built-in commands in this repository; @hasna/events may append event-related commands at runtime.

Global options

Option Behavior
-V, --version Print the package version
-h, --help Print help

evals run <dataset>

Loads a JSONL/JSON dataset or glob, runs it against one adapter, and exits 1 when any result fails or errors. Runs are only persisted with --save.

Option Default Behavior
--adapter <type> http http, anthropic, openai, mcp, function, or cli
--url <url> HTTP endpoint or OpenAI-compatible base URL
--method <method> POST HTTP GET, POST, PUT, or PATCH
--headers <json> JSON object whose values must be strings
--response-mode <mode> json HTTP response parser: json or text
--input-path <path> Dot path where the HTTP input is written
--output-path <path> auto Dot path read from an HTTP JSON response
--timeout-ms <n> adapter-specific Positive adapter timeout in milliseconds
--model <model> Required for Anthropic and OpenAI adapters
--max-tokens <n> provider default Positive maximum output tokens for model adapters
--api-key <key> provider environment Provider API key override
--api-key-env <name> Read the provider API key from a named environment variable
--base-url <url> OpenAI-compatible base URL
--system <prompt> Anthropic/OpenAI system prompt
--module <path> Required function-adapter module
--export <name> default Function-adapter export
--command <cmd> Required CLI-adapter command; supports {{input}}
--mcp-command <cmd> Required MCP server command, split on spaces
--tool <name> Required MCP tool name
--concurrency <n> 5 Number of cases in each parallel batch
--repeat <n> 1 Run each case repeatedly unless the case defines repeat
--tags <tags> Comma-separated tag filter; matching is OR
--no-judge false Skip configured LLM judges and run assertions only
--output <format> terminal terminal, json, or markdown
--save false Save the run in the local SQLite store
--limit <n> 20 Compact terminal row limit
--verbose false Show every terminal result row
-j, --json false Alias for --output json

evals ci

evals ci run <dataset>

Runs and always saves the dataset, then compares it with a named baseline. Adapter options match evals run except this command has no --concurrency, --repeat, --tags, or --save options.

Option Default Behavior
--baseline <name> main Baseline to compare
--fail-if-regression <pct> 0 Exit 1 when pass-rate drop exceeds this percentage
--output <format> terminal terminal or markdown
--limit <n> 20 Compact run/diff row limit
--verbose false Show all run/diff rows
-j, --json false Print the full run as JSON

It also exits 1 when the new run contains failed cases. Adapter setup and validation options are listed under evals run.

evals ci set-baseline <name>

Sets a named baseline to --run-id <id> or, when omitted, the most recently saved run. It exits 1 when no saved run exists.

Run history

evals runs list

Option Default Behavior
--limit <n> 20 Maximum saved runs returned
--cursor <n> 0 Non-negative pagination offset
--dataset <path> Exact dataset-path filter
-j, --json false Compact JSON summaries and pagination metadata

evals runs show <id> / evals runs inspect <id>

Both commands accept a full run ID or an unambiguous prefix.

Option Default Behavior
--limit <n> 20 Compact result row limit
--verbose false Show all human-readable rows
--markdown false Print a complete Markdown report
-j, --json false Print the full run JSON

Comparison and judging

evals compare <before> <after>

before accepts a run ID/prefix or baseline name. after also accepts latest. The command exits 1 when regressions are present.

Options: -j, --json, --markdown, --limit <n> (default 20), and --verbose.

evals judge

Required options are --input <text>, --output <text>, and --rubric <text>. Optional flags are --expected <text>, --model <model> (default claude-sonnet-4-6), --provider <anthropic|openai> (default anthropic), and -j, --json. A FAIL verdict exits 1; PASS and UNKNOWN exit 0.

evals calibrate <gold>

Loads cases with metadata.gold_verdict, judges each case's input against its expected text, and reports agreement plus Cohen's kappa.

Options: --model <model>, --provider <anthropic|openai>, --limit <n> (default 20), --verbose, and --json.

Dataset utilities

evals estimate <dataset>

Estimates judge calls using 800 tokens per judged case without making provider calls. Options: --model <model>, --no-judge, and -j, --json.

evals generate

Requires --description <text>. It asks Anthropic Claude for JSONL and writes valid cases to --output <path>.

Options: --seeds <path>, --count <n> (default 10), --output <path> (default generated.jsonl), --model <model>, and -j, --json.

evals capture

Requires --app <url> and starts a local reverse proxy. The output file is truncated at startup. JSON request/response samples are appended as staging cases.

Options: --port <n> (default 19441), --rate <n> (default 0.1), and --output <path> (default captured.jsonl). Route client traffic through http://localhost:<port> while capture is running.

Operations

evals doctor

Checks Anthropic/OpenAI keys, local database access, and example-dataset availability. Use -j, --json for machine-readable output. The OpenAI key and example dataset are optional; required-check failures exit 1.

evals completion <shell>

Prints a completion script for bash or zsh; any other shell exits 1.

evals mcp register

Registers evals-mcp for Claude Code, Codex, Gemini, or all three with --claude, --codex, --gemini, or --all. With no flag, it selects Claude Code.

evals mcp start

Starts evals-mcp in its default Streamable HTTP mode. See MCP server for transport options and endpoints.