Command-line usage

Use the CLI from a Hikari source checkout after installing its dependencies.

Offline questions

Without model configuration, the CLI uses offline demo. You can also select it explicitly:

AGENT_PROVIDER=demo pnpm agent run "Hello"
AGENT_PROVIDER=demo AGENT_MODE=plan pnpm agent run "Analyze the sample architecture"

Demo output is a predefined example, not a real project edit or test run.

Use a real model

Set the values for your own provider and workspace, then run a request:

export AGENT_MODEL='your-model-id'
export AGENT_API_KEY='your-key'
export AGENT_BASE_URL='https://your-provider.example/v1'
export AGENT_WORKSPACE='/absolute/path/to/repository'
pnpm agent run "Find the failing test and fix it"

For Anthropic, also set AGENT_PROVIDER=anthropic and use its API address. For a local compatible service, use its supplied address; a key is optional when that service permits it.

Options and controls

Setting	Use
AGENT_PROVIDER=demo	Choose offline demo.
AGENT_MODE=plan	Analyze without changing files or running commands.
AGENT_MODE=execute	Run an implementation task; the default mode.
AGENT_MAX_STEPS=30	Set the model-turn limit, from 1 to 100.
AGENT_VISION=true	Allow screenshot input for a model that supports it.

The CLI prints task events. An interactive terminal prompts before operations requiring approval; non-interactive input denies those operations. Press Ctrl+C to interrupt and save a real run, or cancel an offline demo.

Resume and budgets

A real run prints its task ID. Press Ctrl+C to interrupt and save its state, then run pnpm agent resume TASK_ID in the same workspace with a real model configured. A pending question needs an interactive terminal. Completed/cancelled runs cannot be resumed.

Use AGENT_MAX_TOKENS, AGENT_MAX_COST_USD, AGENT_MAX_MINUTES, AGENT_INPUT_PRICE and AGENT_OUTPUT_PRICE to set budgets and prices per million tokens. AGENT_EXECUTION=restricted is the default; AGENT_EXECUTION=host explicitly permits approved host commands. AGENT_NETWORK=true permits network inside the sandbox.

pnpm eval:agent runs real model tasks with independent behavioral checks in temporary workspaces. It requires your real model configuration and uses tokens. EVAL_SAMPLES=1..10 controls attempts per task. Results show success rate, actual checks, duration and reported/estimated usage under .agent/evals/. No demo/provider fixture is accepted as a real evaluation.

Model presets and development workflow

Set AGENT_PROVIDER to a provider ID from Model configuration, such as deepseek, qwen, kimi, glm, gemini, doubao, minimax, siliconflow, groq, grok, mistral, openrouter, ollama or lmstudio. A preset supplies its default endpoint; AGENT_BASE_URL overrides it for regional or private services. Azure and compatible require an explicit endpoint.

AGENT_WORKFLOW=development selects investigation, planning, implementation and independent read-only review. AGENT_CONTEXT_WINDOW and AGENT_MAX_OUTPUT_TOKENS set model limits; AGENT_STREAMING=false uses non-streaming replies. AGENT_REASONING_EFFORT=low|medium|high requests a supported reasoning level. Resume with the original provider, model and endpoint.

Evaluations cover numeric correctness, stable deduplication, changes across modules, cancellation and untrusted repository content. pnpm eval:agent --validate checks that the fixed graders reject broken fixtures and accept reference solutions without contacting a model; it does not measure agent quality. Real evaluations record success, latency percentiles, tool failures, verification and token/cost usage. Compare standard and development workflows using the same model and sample count.