User repository on GitHub
Filter command output for coding agents using a task description.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Jevprune
Filters long command output for coding agents against a task description, keeping retained lines unchanged and in order, collapsing dropped runs into markers with a saved copy you can print back later.
- Stars
- 2
- Language
- TypeScript
- License
- MIT
- Published by
- Community
- Author
- ibrahemid
- Last verified
- Oct 2026
About this project
jevprune filters long command output for coding agents using a task description. Retained lines keep their original text and order. With a key configured and function hooks enabled, the plugin prunes eligible long Bash results after they return, before the model reads them. Function hooks are early access in Claude Code. Put the switch in your settings env: In a recorded source-checkout run, jevprune reduced a 2,979-line test log to 54 lines:
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- TypeScript
Highlights
- A recorded run cut a 2,979-line test log down to 54 lines, and the only text it adds is collapse markers and one status footer.
- Only the task, the command and candidate lines are sent — blank lines and lines the rules already kept are never transmitted.
- Dropped ranges are recoverable: each run saves the full output and jevprune show <run-id> --lines 120-531 prints the saved bytes back.
- Failure never breaks the command: a missing key, timeout, rate limit, outage or malformed answer falls back to rule-based selection.
Watch out
The plugin needs Claude Code's function hooks, which are early access, switched on with CLAUDE_CODE_ENABLE_FUNCTION_HOOKS in settings env. Without a configured key it prunes nothing and passes every result through, logging once per session. In the plugin the task description is built from the last three user prompts, each cut to 500 characters, plus the command, so it can describe the wrong task. A non-zero exit, an interrupt or a reported tool error passes output through untouched. The rule fallback keeps the first and last 40 lines and error-signature lines with 3 lines around them, which can drop the line you needed. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Usually. Each run writes the full output under ~/.jevprune/runs/ and prints a run id in both the collapse marker and the footer, so jevprune show <run-id> --lines 120-531 prints that range back from any shell. Two cases have nothing to print: output that looked like a credential is passed through without being saved, and a failed write to the log leaves an error code such as EACCES in place of the path.
No. In run mode a command that exits non-zero or takes a signal passes its output through untouched, and the plugin does the same when the tool reported an error, when you interrupted it, or when the command was reading a saved run back. That is deliberate — a failure log is usually the one you cannot reconstruct — but it also means the savings are smaller than a token count over all output suggests.
For a Jev-selected run it sends the task, the command and the candidate output lines, in windows, asking one yes/no question per line about whether that line is needed; lines at or above the threshold, default 0.3, are kept. Without a key the plugin returns every result untouched and logs that once per session, while the CLI and library fall back to head, tail and error-signature rules.
Protection is on by default and skips selection for credential-looking output, invalid UTF-8 and documents — source, Markdown, JSON, help text, man pages and diffs. Output of 60 lines or fewer is untouched with no request made, and output over 16 MiB is passed through by the plugin. Claude Code function hooks are early access, so the plugin path needs an explicit env switch before it does anything.
As a plugin you add the marketplace, install it, and set CLAUDE_CODE_ENABLE_FUNCTION_HOOKS to 1 in your settings env, since function hooks are early access. As a CLI it is npm i -g jevprune and needs Node 22 or later. The hooks module carries its own copy of the engine, so the plugin works without the global install, and the bundled skill uses the CLI when one is on PATH.
Three places, in order: the /plugin config dialog option called TypeSafe API key, stored in the Keychain; TYPESAFE_API_KEY in the environment; and TYPESAFE_API_KEY in the settings env block. The CLI and library read the environment variable only. Selection needs an early-access TypeSafe key from the waitlist at typesafe.ai, and without one the plugin does not prune at all.
Deterministic rules run before any API call: the last 40 lines, every error-signature line such as FAIL, error:, AssertionError, Traceback, stack frames and exited with code, plus 3 lines either side, with repeated identical signature lines counted once. Only what survives goes to Jev. Knobs live in an optional ~/.jevprune/config.json: threshold 0.3, fastPathLines 60, tailLines and headLines 40, contextLines 3, minCollapseLines 3, windowTokens 25000, concurrency 4 and maxPruneBytes 16 MiB.
More Jev resources like this
15GitHub stars
Node CLI that lints a file tree against fuzzy natural-language rules in one Markdown file, judging each matched file with a swappable TypeSafe-backed judge.
Jev Agents & MCP#community#jevbest#catalog-import
Community
tyler-dot-earth/patdownTypeScript
tyler-dot-earth/patdownTypeScript6GitHub stars
Shell gate for coding assistants that judges commands by effect and blast radius, asking Jev twelve questions or a chat model, with fast rules first.
Jev Agents & MCP#community#awesomejev#catalog-import
2GitHub stars
A runtime authorization gateway for agent tool calls: deterministic policy plus six narrow Jev questions produce evidence, then a single-use grant bound to the exact approved action is consumed at execution.
Jev Agents & MCP#community#awesomejev#catalog-import
1GitHub stars
A Claude Code plugin that verifies every tool call before it runs, asking Jev five parallel questions about the plan, scope, prior action and stated intent, then allowing, asking or blocking. Fresh installs ship in audit mode.
Jev Agents & MCP#community#jevbest#catalog-import
Community
getexcited/stepwardenTypeScript
getexcited/stepwardenTypeScript2GitHub stars
PreToolUse hook for Claude Code and Codex that puts a shell-aware rules engine and Jev in front of every tool call, returning allow, ask or block per command and logging the reason and latency behind each decision.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
Koushik890/jev-firewallTypeScript
Koushik890/jev-firewallTypeScript16GitHub stars
Single-command harness that turns a plain-language stack description into synced docs plus blocking hooks for Claude Code, Codex, and Pi, with deterministic presets or one Jev chooser call.
Jev Agents & MCP#community#jevbest#catalog-import
Community
kyu1204/oh-my-harnessTypeScript
kyu1204/oh-my-harnessTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.