Jev Bfs

It claims neither. The default is a beam search across BFS levels that keeps the five highest-ranked links at each level, and the README says plainly that link selection can remove the shortest route. It reports observed paths, not a proof of shortestness. If you want breadth-first exhaustiveness, --width 0 retains all candidates in BFS order instead of pruning.

You set the caps. There is no default hop limit and no default page limit, so the search runs until it finds the target, exhausts the selected routes, or you press Ctrl+C, which returns exit 130. --depth bounds path edges and --max-pages bounds fetch attempts. Exit 0 means a path was found, 1 means no path or a runtime error, and 2 means invalid arguments.

No. Each request receives the target query, link titles and URLs, and nothing else. Chromium reads the pages and Python extracts English article links; Jev only ranks them, scoring up to 128 candidates per request with up to five requests in flight, against the jev-latest alias. That keeps the token bill tied to link counts rather than article length.

Three places, checked in order: the process environment, then ./.env, then ~/.config/jev-bfs/.env, with the first one holding a key winning; JEVON_KEY and TYPESAFE_API_KEY are both accepted. For scripts, --json emits only machine-readable output and --save writes the path, scores, timing and API usage, with exit 0 for a path found, 1 for none or a runtime error, 2 for bad arguments and 130 for an interrupt. The live display wants 100 columns by 40 rows and falls back to a compact view below that.

Python 3.11 or later and uv, then `uv tool install git+https://github.com/komikat/jev-bfs.git` followed by `jev-bfs --install-browser` for Playwright's Chromium. Run `uv tool update-shell` if the shell cannot find the command. On Linux, install Playwright's system dependencies if they are missing. The implementation is MIT-licensed and three modules: core.py, cli.py, and terminal.py.

It fetches up to five pages concurrently, extracts English article links from the article content, checks for a direct link to the target, then asks Jev to favor links with the fewest expected remaining hops and keeps the highest-ranked candidates for the next level. Repeat until the target is found, the selected routes are exhausted, or you interrupt it. No ChatGPT controller and no search API are involved at any step.

Do not expect it. The demo carries the caption that it is an example from a recorded run and that results and timings vary. What comes back is an observed path with its link scores, progress events, errors, timing and API usage, not a guaranteed route, and your --width, --depth, and --max-pages settings decide which candidates survive each level. Use --save to keep that trace so you can diff two runs.