User repository on GitHub
TypeSafe AI (Jev) adversarial reviewer and typesafe_ask tool for the omp coding agent
- Forks
- 1
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Omp TypeSafe
omp plugin that runs Jev batteries over tool results, assistant messages and turn ends, gathers deterministic git evidence before each judgment, and can score stale tool output out of later model requests.
- Stars
- 1
- Language
- TypeScript
- License
- MIT
- Published by
- Community
- Author
- siddicky
- Last verified
- Oct 2026
About this project
A reviewer for the omp coding agent, powered by TypeSafe AI's System One model (Jev). It follows omp's advisor pattern, a second set of eyes that watches the session as it unfolds, in one of two roles, set by role in config (default adversarial): adversarial (default): argues the other side, that the last action or claim is wrong, unverified, or incomplete.
Quoted from the project's README. Read the full README.
- Languages
- TypeScript
Highlights
- Three triggers, tool_result, message_end and turn_end, each with its own question battery ending in a shared severity and a defect class.
- Evidence is collected before the judgment: git status, a diff against HEAD and a git grep for definitions the diff removed, each command on a 1.5s budget.
- The role setting swaps the question battery, not the delivery; adversarial argues the other side, advisory mirrors omp's helpful-reviewer framing.
- A note's evidence attribute carries counts, commands run and identifier to path:line, never repo text, so a tracked file cannot put words in front of the agent.
Watch out
Everything the reviewer judges goes to TypeSafe's API, including parts of your transcript and diffs, so read the data-leaving section before enabling it on a sensitive repo. Notes are advice and never block a tool call; only the ambiguity gate, the stop gate and two local pipeline features can change anything. Aborted-work detection is best effort and matches on omp's wording: an error phrased differently is read as a real failure and reviewed, and a review already in flight when you press Esc still finishes and may deliver its note. Reviewing failed commands depends on the result details omp 18.4 sends. A plain bun install pulls roughly 1 GB of host and tooling packages the extension never loads; bun install --production installs about 230 KB. 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.
No. Review notes are advice and never block a tool call, and the note itself is marked to weigh rather than obey blindly. Only the ambiguity gate can refuse a plan submission in plan mode until you have been asked, the stop gate can ask the agent to keep going, and two of the local omp-skills pipeline features can block a specific kind of call without making any Jev call.
Everything it judges, including parts of your transcript and diffs, is sent to TypeSafe's API. The note is narrower than what is sent: its evidence attribute carries counts, an incomplete flag, the commands run and identifier to path:line references, never repo text, so a tracked file cannot put words in front of the agent. The README has a dedicated section on this.
Usually not, but the check is best-effort wording matching against omp's own marks. Stopped commands, aborted turns and tool batches you interrupted are skipped, so a late note cannot restart work you killed. An error worded some other way is treated as a real failure and reviewed, and a command that ran and failed is judged with an error exit status, where a concern or blocker needs a side-effect question to have fired first.
With bun install --production only the runtime SDK installs, about 230 KB. A plain bun install also pulls the omp host package and TypeScript tooling, about 1 GB the extension never loads. You need omp and a TYPESAFE_API_KEY; without the key the extension loads, warns, and stays inactive rather than wedging the agent. Newly added modules need a full omp restart, since /reload-plugins is not enough.
Run omp plugin install github:siddicky/omp-typesafe then export TYPESAFE_API_KEY, or clone the repo, run bun install --production and omp plugin link. Without the key the extension loads, logs a warning, shows a notice at session start and stays inactive; it never wedges the agent.
Set role in config, default adversarial, which argues that the last action or claim is wrong, unverified or incomplete. Advisory mirrors omp's native advisor framing: worth-checking steps, simpler alternatives, missed related updates, ambiguities and unconsidered requirements. Delivery is identical; only the questions and note wording differ.
tool_result reviews edit, write, apply_patch, ast_edit, bash, eval, notebook, debug and task calls; message_end reviews assistant messages of at least 200 characters by default; turn_end reviews the transcript delta since the last review. Each battery ends with a severity score and a defect_class choice.
More Jev resources like this
22GitHub stars
Runnable agent harness that moves pick, score and yes-no loop choices to Jev or Laya in milliseconds while reserving the language model for writing, compiling DECISION.md into executable questions.
Jev Agents & MCP#community#jevbest#catalog-import
Community
zjunlp/JevLoopTypeScript
zjunlp/JevLoopTypeScript15GitHub 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/patdownTypeScript22GitHub stars
Command-line toolkit that wraps Jev judgments into scriptable verify, screen, classify, extract, find, rerank, match, route, ask, compact, and batch commands.
Jev Agents & MCP#community#jevbest#catalog-import
Community
Nasrallah-AL/jev-cliTypeScript
Nasrallah-AL/jev-cliTypeScript24GitHub stars
Wires Jev into agents as seven MCP tools, an embeddable JevDecisionModel library with run functions, and Claude Code hooks at tool and turn boundaries.
Jev Agents & MCP#community#jevbest#catalog-import
Community
Brainwires/jevwireTypeScript
Brainwires/jevwireTypeScript12GitHub stars
Pi extension that puts Jev behind coding-agent hooks: JSON packs ask closed questions on redacted state and turn probabilities into block, confirm, annotate, or steer actions.
Jev Agents & MCP#community#jevbest#catalog-import
Community
HyunjunJeon/pi-quiet-askTypeScript
HyunjunJeon/pi-quiet-askTypeScript7GitHub stars
MCP server exposing Jev Choice, Score, Noul, batched ask, file-reading triage, and model listing as typed tools with full probability distributions.
Jev Agents & MCP#community#jevbest#catalog-import
Community
rashedInt32/jev-mcpTypeScript
rashedInt32/jev-mcpTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.