coding_goal

Autonomous goal loop — /goal with durable checkpointing

Codex/Hermes-style goal mode as an ADK-Rust example: set a goal + a verifiable success condition (a shell command that must exit 0), and the agent loops plan → act → verify, self-correcting from the check's output, until the condition passes or the iteration budget is reached.

It's durable: after every iteration the goal state is atomically checkpointed to disk (.adk/goal.json), so a restart can resume from where it left off — and a completed goal is recognized as done.

Run

cargo run --manifest-path examples/coding_goal/Cargo.toml

Requires GOOGLE_API_KEY (Gemini 3, default) — or CODING_PROVIDER=openai with OPENAI_API_KEY. Override the model with CODING_MODEL.

What the demo does

1. Seeds a workspace with a buggy stats.py (mean forgets to divide) + a failing test_stats.py.
2. Runs the goal loop until python3 test_stats.py exits 0, checkpointing each iteration.
3. Prints the persisted checkpoint.
4. Re-runs with resume = true to show a completed goal is a no-op (simulating a restart).
5. Independently verifies the result.

━━ iteration 1/5 ━━
  🔧 bash        (python3 test_stats.py → fails)
  🔧 read_file   stats.py
  🔧 edit_file   return sum(xs)  →  return sum(xs) / len(xs)
  🔧 bash        (python3 test_stats.py → ok)
  ✅ goal met after 1 iteration(s)

══ durable checkpoint (.adk/goal.json) ══
{ "goal": "...", "until": "python3 test_stats.py", "iteration": 1, "status": "done", ... }

══ resume (simulating a restart) ══
  goal already complete (per checkpoint); nothing to do.

══ verification ══
  ✅ PASS python3 test_stats.py (exit Some(0)) after 1 iteration(s)

CLI equivalent

The same capability ships in the CLI, with --resume:

adk-rust goal "make all tests pass" --until "cargo test" --max-iters 8
adk-rust goal "..."                  --until "cargo test" --resume   # continue after a restart