User repository on GitHub
Jev-guided, recoverable tool-output filtering for Oh-my-pi
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Omp Statify
An Oh-my-pi extension that asks Jev which parts of large tool results are worth keeping before they enter the main model's context, and can also trim earlier plain-text assistant messages, keeping the original text locally for exact recovery.
- Stars
- 1
- Language
- TypeScript
- Published by
- Community
- Author
- d3d0n
- Last verified
- Oct 2026
About this project
An Oh-my-pi extension that uses Jev to select useful parts of large text results from any OMP tool before they enter the main model's context. It can also shorten earlier plain-text assistant messages at OMP's pre-model context hook. User prompts and instruction messages are never rewritten. Original text is kept locally for exact recovery; the main model does not change.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- TypeScript
Highlights
- Applies to text results from any OMP tool, and can also shorten earlier plain-text assistant messages at the pre-model context hook.
- It starts off; turning it on selects replace mode, and low-confidence chunks stay visible rather than being cut.
- Key storage uses 0600 permissions and is not encrypted, so it should be enabled only for data you would send to OpenRouter.
- User prompts and instruction messages are never rewritten, and the main model itself does not change.
Watch out
The README calls this a cost experiment, not a guarantee of cheaper or more accurate coding. The two reported trials are weak evidence for causation: the RepoQA figure of $0.189794 down to $0.148549 including Jev is a controlled replay that places captured tool output in a user message rather than making a real agent tool call, and the pytest repair at $0.099682 down to $0.069380 used an explicit starting-file hint while the two arms took different numbers of turns and cache reads, so neither shows Statify alone caused the difference or that accuracy holds on arbitrary repairs. Cheaper main models can end up losing money to the Jev calls. Behaviour degrades conservatively: network errors, unsupported results and replacements that do not reduce the local token estimate all leave the original result in place, and the original text is kept locally for exact recovery, though whether agents actually use that recovery is still an open item in the README's TODO, along with a model-aware cost gate for inexpensive models. 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.
Not on their own, and the README says why. One trial replays captured tool output placed in a user message, which is not a real agent tool call. The other is a single real pytest repair that was given an explicit starting-file hint, with the two arms taking different numbers of turns and cache reads. Neither result shows Statify alone caused the cost difference, and cheaper main models can lose money to the Jev calls.
The originals are kept locally so an exact recovery is possible, and chunks Jev is not confident about stay visible. The open question is the agent, not the extension: the README's TODO still asks whether an agent actually reaches for archive recovery when a needed span was omitted during a repair, which is exactly the failure mode to watch for in real use.
Turning it on with your own OpenRouter key selects replace mode. From then on text results from any OMP tool, and optionally earlier plain-text assistant messages reached at the pre-model context hook, are sent to OpenRouter for the Jev call that picks which chunks survive; anything Jev is not confident about stays visible. Network errors, unsupported results and replacements that do not reduce the local token estimate leave the original result intact, and the key is entered with /statify key add, stored 0600 and not encrypted.
No. User prompts and instruction messages are never rewritten. The candidates are text results from OMP tools, and optionally earlier plain-text assistant messages reached through the pre-model context hook. The main model and your own text are left alone.
It needs OMP and Bun. omp plugin install github:d3d0n/omp-statify, and extension discovery is automatic, so there is nothing to register by hand. In the TUI, /statify key add gives you a command for entering your OpenRouter key without echo, /statify on enables filtering, and /statify status reports the state; the optional statusline is off by default.
On GPT-6-Sol, 20 RepoQA function lookups went from $0.189794 to $0.148549 with 19 of 20 correct in both arms, and one real pytest bug repair with a failing test went from $0.099682 to $0.069380, with the regression and the full testing/test_mark.py suite passing in both arms. Both figures are totals including the Jev calls.
Its TODO list is the useful part. Nobody has checked whether an agent actually reaches for archive recovery when a needed code span was omitted during a repair, multi-repository bug fixes are unmeasured, and there is no model-aware cost gate for inexpensive main models, where the Jev calls can outweigh what they save.
More Jev resources like this
16GitHub 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-harnessTypeScript15GitHub 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/patdownTypeScript1GitHub 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/stepwardenTypeScript48GitHub stars
Agent harness for work that outlives one chat, with resumable turns, routed state machines, and observational memory across sessions and processes.
Jev Agents & MCP#community#jevbest#catalog-import
Community
dzhng/duet-agentTypeScript
dzhng/duet-agentTypeScript37GitHub stars
Assistant framework that stores specs, memory, and specialist agents as plain markdown files in your repo for Claude Code, Codex, Opencode, and similar runtimes.
Jev Agents & MCP#community#jevbest#catalog-import
36GitHub stars
Pi extension that strips the full Agent Skills catalog from every request and loads only the skills a Jev ranking call judges relevant.
Jev Agents & MCP#community#jevbest#catalog-import
From the Jev community
Posts from builders shipping with Jev right now.