Coding Agent CLI

The adk-rust CLI ships three native coding commands. All run a real agent in a confined workspace, default to a Gemini 3 model, and resolve the API key non-interactively from --api-key or the environment (GEMINI_API_KEY/GOOGLE_API_KEY, OPENAI_API_KEY, …) — so they never block on a setup prompt.

code — one-shot task

adk-rust code "make the failing test pass"
adk-rust code --dir ./project "add a /health route"
adk-rust code --read-only "explain how auth works"   # no writes / no shell

The agent explores, edits, runs commands, and streams its work (tool calls, results, text) plus the final plan.

goal — autonomous goal mode (durable, resumable)

The Codex/Hermes /goal pattern: give a goal and a verifiable success condition (--until is a shell command that must exit 0). The agent loops plan → act → verify, self-correcting from the check's output, until it passes or the iteration budget is reached.

adk-rust goal "make all tests pass" --until "cargo test" --max-iters 8
adk-rust goal "port utils.py to type hints" --until "mypy utils.py" --dir ./svc

Durable & resumable. After every iteration the goal state is atomically checkpointed to <dir>/.adk/goal.json (override with --state <path>). If a run is interrupted, continue it:

adk-rust goal "…" --until "cargo test" --resume

--resume recognizes a completed goal as a no-op, and continues a running or budget-exhausted run from where it left off.

The success condition is what makes goal mode safe and effective — pick something verifiable (a test suite, a type-checker, a build) so the agent has a clear, machine-checkable definition of "done."

ultracode — parallel ultra-review

The Claude Code ultracode/ultrareview pattern: implement the task, then fan out to parallel specialist reviewers (correctness, edge-cases, style), synthesize their verdicts, and revise until they approve — built on adk-graph.

adk-rust ultracode "add input validation to the parser"
adk-rust ultracode --dir ./project --max-rounds 3 "implement a retry wrapper"

You'll see each reviewer's verdict, the synthesized decision per round, and the revise steps, ending in finalize.

Picking a model

All three commands accept the global --model / --provider / --api-key flags. The default is gemini-3.1-flash-lite (fast, reliable multi-step tool use). For OpenAI: --provider openai --model gpt-5-mini (with OPENAI_API_KEY).

Next: Workflows →