2. Following a search

A search runs in the background, so following it means opening a view onto it. Every view reads the same records the engine writes, so you can open and close them as often as you like without touching the search.

While it runs

hillclimb watch is the one to keep open: every search and every coding agent, what each agent is doing right now, and each candidate's score as it lands. hillclimb chart draws the climb, the best score so far against the attempts made, and hillclimb tree draws the search as a tree of which candidate was built on which. The walkthrough shows them on a real search, and Visualisations has each one.

For a quick look without a full-screen view, hillclimb status prints a search's candidates and their scores, and hillclimb show <search> <candidate-id> prints everything about one of them: its score, what the agent said it did, and how its code differs from its parent's.

What it found

The search keeps the best solution so far in its best/ folder, and updates it whenever a better one lands:

runs/<run-id>/searches/<search-id>/
├── best/           # the best solution.py, and what it wrote
├── candidates/     # one folder per attempt: the prompt, the agent's log, the code
└── journal.jsonl   # every candidate and its score, in order

So the answer is there before the search ends. hillclimb summit copies the best solution found across every run of a problem into your working folder. Data storage has the whole layout.