PiJ
PiJev: Pi-based coding agent where Jev shortlists skills, ranks code excerpts and triages tool failures while your model writes code.
Overview
PiJev is a terminal coding agent built on Pi with Jev in the loop. Your coding model handles reasoning, edits and tool use, while Jev evaluates small typed questions: shortlisting skills and checking their instructions for relevance at task start, ranking actual file excerpts retrieved by pijev_search, and classifying failed tool calls to suggest what to check next. Suggestions stay advisory; missing credentials or failures fall back to normal coding and lexical search, and Jev scores never replace compilation, tests or acceptance.
The README reports five controlled experiments, 20 retrieval sweeps and 232 agent runs on SWE-bench Verified django tasks and unfamiliar repositories, paired per task against plain Pi with the same model and budget: PiJev made fewer tool calls and reached the right file sooner at the same resolve rate. Jev cost about $0.011-0.014 per agent task. It is early alpha, accesses Jev through Vercel Gateway, and is MIT licensed with a Chinese README.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Score | How relevant is this skill to the task?(inferred) | — |
| Score | How relevant is this code excerpt?(inferred) | — |
| Choice | What kind of failure is this tool error?(inferred) | — |
How it works
- At task start, Jev shortlists relevant skills.
- pijev_search retrieves excerpts and Jev ranks them.
- On tool failures, Jev classifies and suggests next checks.
- Coding model makes all edits; tests decide acceptance.
Strengths
- Fewer tool calls and faster file finding in paired experiments.
- Low Jev cost per task.
- Advisory only; graceful fallback.
Caveats
- Early alpha.
- Evidence limited to the reported task sets.
At a glance
- Type
- Project
- Status
- prototype
- Language
- TypeScript
- Runtime
- CLI coding agent on Pi
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
A Pi-based terminal coding Agent whose main model handles reasoning, edits and tools while Jev provides advice.
logicrw/awesome-jev-projects
Frequently asked questions
Does Jev write code?
No, the coding model writes code; Jev only answers typed questions.
Did it help on SWE-bench?
The README reports fewer tool calls and faster file finding at the same resolve rate.
How is Jev accessed?
Through Vercel AI Gateway.
Related projects
Speed-focused Pi extension combining a local coding model with Jev judgments for file retrieval and cached requests.
ProjectbetaTypeScriptCLI and MCP server giving coding agents Jev-powered semantic search, focused reads and list picking in about a second for about a cent.
BenchmarkbetaPythonDeterministic 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.
ProjectbetaPipe grep or file lists into Jev to get every candidate scored, sorted and cut at a calibrated threshold, so agents read only what matters.
ProjectbetaTypeScript