User repository on GitHub
Reusable GitHub Action: agent fix loop gated by checks, an AI reviewer, and TypeSafe Jev
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Agent Gate Loop
A GitHub Action that turns a labelled issue into a pull request, with a four-part gate before a human sees it: path and test guards, your own check commands, a read-only Claude reviewer, and Jev checking each finding plus whether the change does what the issue asked.
- Stars
- 0
- Language
- TypeScript
- Published by
- Community
- Author
- Ripwords
- Last verified
- Oct 2026
About this project
A GitHub Action that turns an issue into a pull request. An AI agent writes the change, and a gate checks it before a human sees it. The gate has four parts: Guards. No edits to protected paths, no deleted tests, a size limit, and at least one change. Your checks. Your own typecheck, lint and test commands. Reviewer. A separate, read-only Claude session that lists findings.
Quoted from the project's README. Read the full README.
- Languages
- TypeScript
Highlights
- Cheap gates first: guards, your checks and one Jev pre-check, which can skip the AI reviewer entirely.
- Four outcomes, not pass/fail: passed, escalated, failed and needs-info, each mapping to a PR shape.
- review_shortlist greps added lines - hard rules fail the round, judgement rules go to the reviewer.
- The action never merges; a failed gate returns one merged feedback note and the agent retries.
Watch out
The README calls this a prototype and names two known gaps before unattended use: the agent can still change .git/config through its shell, and your check commands can read the job's GitHub token and TypeSafe key from the parent process. It advises using it only on issues written by people you trust. Cost controls are Claude-side only - max_cost_usd (default 2) and max_rounds (default 2) stop the loop, and the README is explicit that max_cost_usd counts Claude spend only, with TypeSafe Jev calls billed separately and not included. Because checks run before each commit and the agent commits every non-ignored file in the working tree, anything missing from .gitignore lands in the diff and the PR. Security-wise the agent's shell shares the process holding the credential, a subscription token cannot be spend-limited and a leak exposes your whole Claude plan, and a PR opened with the default GITHUB_TOKEN will not trigger your other workflows unless you pass AGENT_GITHUB_TOKEN. On private repos without draft PR support, non-passing runs open a normal PR marked with the agent:* label. No accuracy data, no false-positive rate and no comparison against running the agent without the gate are published. 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.
The action never merges. Trigger it by labelling an issue agent:run or commenting /agent, which only people with write access can do, and you must enable Allow GitHub Actions to create and approve pull requests under Settings → Actions → General or the default token cannot open the PR at all. The result decides the PR shape: passed means ready for review unless the issue is high risk, escalated means a draft because Jev was unsure or the reviewer failed, failed means a draft on repeated problems or exhausted rounds, and needs-info means no PR, just a comment on a too-vague issue.
Only Claude spend. max_cost_usd defaults to 2, max_rounds defaults to 2, and the remaining budget is passed into each agent session as its spend limit, but the README states that TypeSafe Jev calls are billed separately and are not included in that cap. The other cost lever is model sizing: the fixer uses claude-sonnet-5 and the reviewer, at reviewer_model: auto, uses claude-haiku-4-5 for changes at or under small_diff_files and small_diff_lines and sonnet-5 otherwise.
The README's own prototype warning is the answer: the agent can still change .git/config through its shell, and check commands can read the job's GitHub token and TypeSafe key from the parent process, so it advises using it only on issues written by people you trust. The mitigations it does list: the agent itself never receives the TypeSafe key or the GitHub token, check commands receive no secret-looking environment variables, and you should use a spend-limited Anthropic key because a subscription token cannot be capped and a leak exposes your whole Claude plan.
Add them to .gitignore before the first run. Checks run before each commit and the agent commits every non-ignored file in the working tree, so anything not ignored ends up in the diff and therefore in the pull request. The README also notes a PR opened with the default GITHUB_TOKEN does not trigger your other workflows, so pass AGENT_GITHUB_TOKEN if you want CI to run on agent PRs.
Copy examples/consumer.yml to .github/workflows/agent.yml, add TYPESAFE_API_KEY plus ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN as secrets, enable Allow GitHub Actions to create and approve pull requests, then label an issue agent:run or comment /agent. Only people with write access can trigger runs.
Guards prevent protected-path edits, deleted tests, oversized diffs and empty changes; your own typecheck, lint and test commands run next; a read-only Claude reviewer lists findings; then Jev checks each finding and asks whether the change does what the issue asked, touches unrelated code and has good tests.
Yes. review_shortlist points at a JSON file of grep rules run on added lines, where hard rules fail the round with the rule's fix text and judgement rules go to the reviewer, and review_rubric points at a markdown rubric the reviewer follows; both are read from the base commit.
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.