agent

@elizaos/agent

Standalone agent host and HTTP/WebSocket backend around the elizaOS runtime.

Hosts explicitly compose core, assistant, storage, and model plugins. Start from the repository root with bun run start; use bun run dev for the app and API together. Configure providers and connectors through the host configuration; never expose host secrets to ungranted agents.

Import public runtime, service, role, and operation APIs from @elizaos/agent. Agent code imports core APIs and catalog maps through @elizaos/core. RegistryClientPluginInfo and RegistryClientSearchResult expose the underlying registry shapes; RegistryPluginInfo and RegistrySearchResult retain the plugin manager's extensions. The roles plugin is exported as rolesPlugin.

Development

Install dependencies with bun install at the repository root. Run from that root:

bun run --cwd packages/agent build  # build
bun run --cwd packages/agent test   # tests

Retain real end-to-end scenarios exercising host, transport, and persistence. The package test, test:e2e, and test:integration commands share that suite; do not reintroduce removed unit, mock, smoke, or source-inspection tests.

Run the native coding CLI end to end with configured provider credentials:

BENCHMARK_NATIVE_CODING_E2E=1 BENCHMARK_NATIVE_MODEL=<model> \
  bun --conditions=eliza-source node_modules/vitest/vitest.mjs run \
  --config packages/agent/vitest.config.ts packages/agent/test/benchmark-coding-live.test.ts

This uses a real provider and isolated Git fixture, verifies executed WRITE/SHELL actions and Python tests, and requires clean process shutdown. Receipts are saved under test-results/agent-native-coding/. Configure provider-specific model settings to match BENCHMARK_NATIVE_MODEL when overriding the shared defaults.

Benchmark JSON separates turn_completed from the optional planner assessment request_fulfilled, and preserves action_results and effect_receipts, including failed attempts that were later recovered. An explicit unfulfilled assessment fails the CLI. These fields are execution evidence; benchmark graders must still verify the requested outcome independently.

Coding benchmark turns use the focused READ, WRITE, EDIT, and SHELL action profile. Create task branches in the supplied workspace; external graders collect that workspace’s commits and do not follow temporary worktrees.

For long coding runs, set ELIZA_CODING_MAX_PROMPT_TOKENS to a positive integer to choose an explicit cumulative planner prompt-token budget. Unset runs retain the default 1,500,000-token limit; cached prompt tokens count toward the budget. The setting does not affect ordinary conversational turns or override an explicit host-supplied planner budget. Record the value with benchmark results.