Blink — ellipsis-dev

No, it never reads file contents. Jev scores file and folder names, more likely paths get more walkers, and each walker keeps moving until it reaches a file, so it answers where something is handled at the path level rather than finding a string. It complements content search. Empty folders come back marked unresolved and the table shows the top 10 with the rest grouped as OTHER.

The distribution is the point: each result shows the percentage of starting walkers that ended there, so 74.0% means roughly three quarters of 100 walkers walked to that file. Splitting walkers across paths is also what makes recursion work, since one walker following a single path would only ever land on one file. Without -r or -n it ranks only immediate subdirectories.

The summary reports elapsed time, how many API calls completed (not counting retries), and estimated cost, and Jev is priced at $0.042 per million input tokens with output free. Only file and folder names go into each call, so a search is cheap, but the repository publishes no measured per-search figure, so the summary line is the only real number you get.

No: bun test runs against fake API responses and needs no key. A real search needs Bun 1.3.14 or newer plus an exported TYPESAFE_API_KEY, and you can exclude names at every depth with an ignoredNodes list in settings.json. One thing to flag before reuse: the repository ships no license file, so the reuse terms are unclear.

The form is ./blink "query" "directory" [--recursive] [--n_walkers 100]. -r follows one walker down the most likely path to a file, -n 100 splits 100 walkers across paths and enables recursion, and -v prints the requests, responses and walker allocations before the final table. Without -r or -n it ranks only the immediate subdirectories.

Recursive searches write output beside the CLI under output/searches/search_abc123/, with metadata.json plus one trace file per walker. Each walker keeps a unique walker_id such as walker_abc123 for its whole path, and its trace records every choice, probability, final destination and status. Metadata holds the query, settings, timestamps, shared Jev decisions, token usage, cost and full results, and files are rewritten after every step, including when the search fails.

The table prints the top ten nodes with the percentage of starting walkers that ended there and groups the remainder as OTHER, with paths shown from the folder you searched. A folder containing no files comes back marked unresolved rather than silently dropped, so pointing at the wrong directory shows up as unresolved entries instead of an empty result that reads like nothing matched.