Curated twice a day from three sources: TypeSafe AI's #show-and-tell Discord channel (read in full from 2026-07-22), GitHub search, and X posts found through web search. Deliberately wide; use the search and filters.
Jev primitives: Choice (pick one), Noul (yes/no), Score (0 to 1). No temperature. No open-value extraction.
Picked from every entry below by reading each candidate's repository: license, tests, last push, what the code does, and what it sends off the machine. Re-judged weekly against the week's new entries.
Prose linter. Regex rules run locally; judgment rules ask Jev. Custom rules load from .snifftest.yaml.
Sends Nothing for regex rules or --dry-run. Judgment rules ask before sending text.Checked MIT, tests, on npm. 182 ms and $0.013 per 100 paragraphs (author).
Claude Code Stop hook. Blocks a "done" when files changed and no test, build or lint passed afterwards; one Jev call, only on those stops.
Sends On flagged stops only: the turn's prompt and closing message, secret-redacted.Checked MIT, offline test suite. AUROC 0.976 on 100 model-labelled stops (author).Note CI runs on Linux only, and the test-run detector reads Bash tool calls, not PowerShell. Has a shadow mode.
Rewrites a Jev question's wording against your labels, asking you to label only the uncertain rows each round.
Sends The evaluated rows to TypeSafe; reflection prompts to a second model.Checked Apache-2.0, tests, 286 stars.
pytest fixture that asserts what model output says (holds, lacks, choice, score), all claims in one Jev call.
Sends The text under test and the context passed with it.Checked MIT, tests. 110x cheaper than Sonnet 5 as the judge (author).
Trims large tool output in Claude Code: redacts secrets, caps per content type, keeps error lines and test totals, stores the full output on disk.
Sends Nothing. No model calls; telemetry is opt-in.Checked MIT, about 60 test files. 14.5 to 18.9% fewer tokens on tool results (author).
Screens fetched pages and MCP results for injected instructions before the agent reads them. Its other hooks can be switched off.
Sends Each screened result, up to about 128,000 characters.Checked MIT, tests. 210 ms p50 warm (author).
Crawls a site toward a question: Jev scores each link and passage, and the agent receives only the selected passages, copied verbatim with URL, heading path and character offsets.
Sends Link and passage text from crawled pages to TypeSafe for scoring; pages are fetched from the target site.Checked Apache-2.0, CI, on PyPI, 52 stars. On 50 WebWalkerQA tasks: answer page found 54% against 33% with Claude picking links and 11% for breadth-first; 40% against 30% answered correctly; $0.025 against $0.194 per task (author).Note Reads plain HTML only unless the js extra and Chromium are installed. Repository created 2026-10-11; the benchmark is 50 tasks.
Two guards: jev-gate approves clearly safe agent commands, blocks clearly dangerous ones and defers the rest to the normal prompt; review-router tells an AI PR reviewer when to read the whole pull request.
Sends jev-gate: each command, not the conversation. review-router: the pull request diff. Both through OpenRouter.Checked MIT. 0 of 12 dangerous commands ran against 8 of 12 under Claude Code auto mode, median 0.28 s; 66% of 3,622 commands auto-approved with no rm, git push, git reset or sudo approved; 13 of 13 CVE fixes on 561 commits sent to full review (author).Note Ran 9 of 12 safe commands without a prompt. The README states it catches mistakes, not a determined attacker. Starts in watch-only mode.
Nothing matches. Loosen a filter or clear the search.
quiz_handover.md in channel)self-test on Choice, Noul, Score and the cascade patternsstate; it is only as good as the facts you give it.