blink
Codebase search where an ensemble of walkers descends the file tree, with Jev ranking which files and folders look relevant at each level.
Overview
blink searches a codebase using Jev and an ensemble of walkers. Given a natural-language query and a directory, Jev scores file and folder names at each level; more likely paths get more walkers, which keep moving until they reach a file. Each result shows the percentage of starting walkers that ended there, so a query like "where is authentication handled?" returns login.ts with most of the mass.
It runs on Bun, takes options for recursive search and the number of walkers, and reports search time, completed API calls and estimated cost. Metadata records the query, settings, shared Jev decisions, token usage, cost and full results, updated after every step even when a search fails. It only uses names, not file contents.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which file or folder at this level is most likely relevant to the query? | entries in the current directory |
How it works
- Start N walkers at the root
- At each directory, Jev scores children by name
- Distribute walkers by probability
- Walkers stop at files; report the share per file
Strengths
- Never reads file contents
- Probabilistic results show confidence
- Cost and token reporting
Caveats
- Relies on file and folder names only
- Many walkers mean many calls
Getting started
bun install, export TYPESAFE_API_KEY, then ./blink "query" "directory" --recursive --n_walkers 100At a glance
- Type
- Project
- Status
- prototype
- Language
- TypeScript
- Runtime
- Bun CLI
- Has code
- yes
Use cases and integrations
How curated lists describe it
Codebase search: an ensemble of walkers asks Jev which file answers a natural-language query
AnotiaWang/awesome-decision-modelsFinds files from a natural-language query using multiple walkers through the directory tree.
logicrw/awesome-jev-projectsNavigates file and directory names with Jev-guided walkers to find codebase paths for a natural-language query.
cobanov/awesome-jevCodebase search powered by Jev from @typesafe-ai.
hellogumbo/awesome-jevSemantic codebase search. At each directory level Jev ranks which files and folders are most likely relevant and sends more walkers there.
MrJev/awesome-jev
Frequently asked questions
Does it read file contents?
No, Jev scores file and folder names.
What do the percentages mean?
The share of starting walkers that ended at each file.
What does it need?
Bun 1.3.14+ and a TypeSafe key.
Related projects
Semantic grep that asks Jev a yes/no about every function for questions no regex can express, with cost reporting.
ProjectbetaDeterministic code-graph context for coding agents on Postgres, AGE and pgvector, with optional Jev or local re-selection of symbols.
MCP serverbetaPythonClaude Code plugin that adds Jev as an independent judge for review findings, debugging hypotheses, design options and search hits.
PluginbetaRetrieval layer for Codex that finds evidence locally, optionally ranks sanitized excerpts with Jev, and returns bounded, cited evidence.
ProjectbetaDeepSeek Harness jev_search tool that runs ripgrep, scores every hit line with Jev, and returns only the top rows above a cut.
PluginbetaJavaScriptCLI that asks one yes/no question of every function in a codebase with Jev and ranks them by probability: grep whose pattern is a question.
CLIbetaPython