Shaduf.Research preview
Jev: Use Cases, Alternatives & Products/What changed for Jev builders, 7 to 8 November 2026
Research reportChecks , 05:17–05:42 UTC; previous checks 7 November 2026, 15:29–15:55 UTC; no check between them

What changed for Jev builders, 7 to 8 November 2026

ohmyjev, a Claude Code, pi and omp guard that asks Jev before risky tool calls, released 0.14.0 with a new shadow setting and fewer calls that reach Jev. Its failure handling is otherwise the same. A pi extension that lets Jev pick each request's model was read in source. SubQ's and Cloudflare's data terms were re-read, and TypeSafe's docs index gained one cookbook. TypeSafe's aliases, model IDs, price, signup and credit state did not change.

Key findings

ohmyjev 0.14.0 keeps its gates off by default and adds a shadow setting that logs what lockdown would block without blocking it. Its four other failure cells are unchanged; the failure map still counts 80 implementations.

DocumentedSource read at commit fbc4ea4 (npm 0.14.0, published 3 Nov 2026) and compared with 0.13.2 (d307c39), 8 Nov 2026, 05:20–05:22 UTC (R28-S20 to R28-S23)

Of 29 coding-agent integrations read in source, 15 act on Jev's answer, 10 do not and 4 are unsettled. pi-jev-router (pi) picks each request's model and uses the standard model when Jev fails.

Documented@caph42/pi-jev-router 2.0.0 read at 75ad5b7 with pi 1.5.0 (f9a2636), 8 Nov 2026, 05:17–05:28 UTC (R28-S200 to R28-S210). Counted on the coding-agents page; not added to the failure map

SubQ's docs and Terms now agree that Decisions API inputs and outputs are deleted within 60 days; until at least 14 Oct 2026 they disagreed.

DocumentedSubQ data-retention docs and Terms (Last Updated October 16, 2026), read 8 Nov 2026, 05:24 UTC (R28-S32, R28-S33)

ohmyjev 0.14.0: a shadow setting, and fewer calls reach Jev

ohmyjev is a community guard for Claude Code, with adapters for pi and omp. In its lockdown setting it asks Jev whether a Bash command, a file write or a network call is irreversible or leaks data, and denies the call when Jev's answer crosses a threshold. Version 0.14.0 was published on npm on 3 Nov 2026. Its source was read at the commit npm names for it (fbc4ea4) and compared file by file with 0.13.2 (d307c39, read 2 Nov 2026).

ohmyjev's five failure cells in lockdown mode, 0.13.2 and 0.14.0

Condition0.13.2 (read 2 Nov 2026)0.14.0 (read 8 Nov 2026)
Jev error or timeoutfail-open: 1,500 ms; the call goes to Claude Code's normal permission flowfail-open: unchanged (1,500 ms)
Malformed answerfail-open: the call passesfail-open: unchanged
Below thresholdacts-anyway: denied only at 0.6 / 0.7 / 0.7 or aboveacts-anyway: unchanged thresholds
No keyfail-open: the gate returns without blockingfail-open: unchanged; a key stored with ohmyjev key is one more key source
Bypassrecorded: lockdown off (the default) skips every gate; read-only Bash skips Jev; each gate can be switched offrecorded, with new content: an opt-in shadow setting (off by default) runs the checks but only logs what they would block; in-repo mkdir, touch, cp and mv now skip Jev unless the policies setting is used

Terms as on the failure checklist. Grey marks an unchanged cell, amber a cell whose content changed. The terms and their colours did not change, so no total on the map moves (still 80 implementations). File and line references are on the checklist row (R28-S21, R28-S22).

What ohmyjev 0.14.0 does with a tool call, by setting (as read in source; not run)
SettingDoes a gate ask Jev?Can Jev's answer block the call?
Lockdown off (the default)No: every gate returns before any Jev callNo
Lockdown on, shadow offYes, except read-only Bash and in-repo mkdir, touch, cp, mv (new in 0.14.0)Yes, above its thresholds; fail-open when Jev errs, times out or has no key
Shadow on (opt-in; it wins over lockdown)YesNo: what would be blocked is logged, and the call runs
  • Unchanged: lockdown is still off by default, so a default install asks Jev nothing. In lockdown mode without shadow the guard still acts on Jev's answer. Documented
  • Not read: the Codex adapter (adapters/codex.ts, present in both versions) and the screens, routing, done-check and compaction handlers. The new bin/ohmyjev.mjs is an install and key tool; its key check sends one Jev request and does not decide on a tool call.

Updated with this reading: when Jev fails, Claude Code and MCP guardrails, coding agents, products and the build hub. The 2 Nov readings stay on those pages as dated history.

pi-jev-router: Jev picks the model for each pi request

@caph42/pi-jev-router 2.0.0 (published 1 Nov 2026) is a community extension for the pi coding agent. It registers a virtual model, jev/auto; for each request it asks Jev whether the task is complex and sends it to the complex or the standard model it was configured with. pi acts on the choice in code: its runtime resolves the virtual model and sends the request to the model the extension returns. Read at 75ad5b7 (the commit npm names; the published file is byte-identical) with pi 1.5.0 at f9a2636. A different, unscoped project with the same name (r33drichards/pi-jev-router) was not read.

  • On a Jev error or timeout (3,000 ms, no retry): fallback, to the standard model on the first routed turn, otherwise the current model is kept.
  • Malformed answer: the same fallback. No key: the same fallback, per request, with no startup error. Below threshold: the standard model (a routing rule; there is no separate confidence cut).
  • Bypass (recorded): compaction summaries, retries, a reused planning model, continuation turns and, in its fixed mode, every turn after the first are routed without asking Jev.
  • Counting: counted on the coding-agents page (it reroutes a model choice and the host enforces it in code, read at a pinned commit). It is shown under "Read in source; not on the failure map yet"; the failure map's 80 are unchanged. Documented

Counted coding-agent integrations by what happens on a Jev error

7 Nov 2026 (14)3 fail-closed, 9 fallback, 2 fail-open
8 Nov 2026 (15)3 fail-closed, 10 fallback (pi-jev-router added), 2 fail-open

Green: fail-closed; blue: a named fallback decides; red: fail-open. 3 + 10 + 2 = 15 counted of 29 read in source (15 counted, 10 not counted, 4 not settled; 28 community projects and 1 vendor-published, none from TypeSafe). These are the coding-agents page's own counts; the failure map is a separate count of 80 and did not change.

The coding-agents page also lists 19 coding-agent projects seen between 24 Oct and 8 Nov 2026 for Codex, OpenCode, pi and Claude Code, each with one quote, as leads not read in full; further names seen in that period were not sorted. Details: coding agents.

Data terms: SubQ Snap, Cloudflare and OpenAI

Hosted decision APIs from other providers: data terms re-read on 8 Nov 2026, 05:24 UTC
ProviderBeforeAs read on 8 Nov 2026
SubQ Snap (Subquadratic)Two texts that disagreed (14 Oct 2026): Terms "no more than thirty (30) days"; data-retention docs "Retain indefinitely (default)"Both say 60 days: docs "deleted within 60 days of processing"; Terms "no more than sixty (60) days" (Last Updated October 16, 2026). Changed between 14 Oct and 8 Nov 2026 (R28-S32, R28-S33)
Cloudflare Workers AI (Clef family)Page dated Oct 16, 2026 (read 31 Oct): no training; zero data retention for standard inference, with exceptions for storage services and AI Gateway logging or cachingSame date and sentences; the exceptions list also names asynchronous or batch features and file or adapter uploads (not quoted on 31 Oct; whether it was there then was not recorded) (R28-S31)
OpenAI Decisions APINot used for training; 30-day abuse-monitoring retention; ZDR and Private Retention with PSP listed for eligible customers (30 Oct 2026)Unchanged (R28-S30)

These are other providers' models; your data goes to that provider under its terms, not TypeSafe's. Legal text is quoted, not interpreted.

TypeSafe's docs: one new cookbook

TypeSafe's models, Images, quickstart, legal and changelog pages were identical to the 7 Nov copies. Its docs index (llms.txt) added one page, "Verifying LLM answers against their source" (cookbooks/verify_llm_outputs.md), which uses three Noul questions to check whether an LLM's answer follows from the paragraphs it was given and sends doubtful answers to review. The index now lists 6 patterns, 26 cookbooks and 2 demos. The cookbook reports TypeSafe's own figures from jev-1.14.0 on 7 Nov 2026 (1,000 questions answered by gpt-6-luna; the check passed 73% of answers, and 95.7% of those were right) Reported; the pool ran nothing (R28-S15, R28-S16).

What else changed

Each change is on its page with this wording and date. Old values stay on the pages as dated history.
FactBeforeNowChanged between
langchain-typesafe0.0.1a3 read; 0.0.2 (5 Nov) not re-readRead at 0.0.2: failure path unchanged; adds an optional classifier= argument on both experimental middlewares (R28-S211 to R28-S214)Release 5 Nov 2026
pydantic-ai-slim 2.66.0Not comparedIts TypeSafe model and provider files are identical to 2.64.0 by file hash (R28-S215)Release 4 Nov 2026
Integration releases@ai-sdk/typesafe-ai 3.0.34, ai 7.0.164, auto-mode 3.1.6@ai-sdk/typesafe-ai 3.0.35 and ai 7.0.165 (7 Nov); auto-mode 4.0.0 (8 Nov, a major release); none re-read (R28-S47)Release dates as listed
LiteLLM1.108.2 (LTS)1.108.3 (LTS, 7 Nov 2026); its release notes do not mention TypeSafe or Jev (R28-S45)Release 7 Nov 2026
Status-page uptime figures99.941%, 99.841%, 99.882%99.942%, 99.842%, 99.883%; the third is read as the 90-day figure; the windows are not labelled (R28-S04)7 Nov, 15:29 UTC and 8 Nov, 05:17 UTC
OpenRouter's ZDR endpoint list1,119 endpoints1,117 endpoints; both TypeSafe endpoints listed (R28-S46)7 Nov and 8 Nov 2026

Re-read on 8 November and unchanged

  • Access: signups open, no free credit for new accounts, API operational ("All services are online"; the status page's history lists no event after 30 Oct). The console showed the same sign-in page; signup itself was not observed (R28-S01 to R28-S14). Details: access status.
  • Models and price: jev-latest and jev-preview give jev-1.14.0; jev-1.13.0 is listed with no deprecation date; $0.042 per million input tokens, output free; the models page was read again at 05:42 UTC and was still identical (R28-S05, R28-S09, R28-S11).
  • The six routes: TypeSafe, OpenRouter (461 models; no Jev ID ending :free), Vercel, Opper, Netlify and Cloudflare list the same Jev IDs; Opper's typesafe/jev-latest still gives jev-1.13.0 (R28-S40 to R28-S44). Details: versions and aliases.
  • OpenCode's free ID (third-party route; not run by TypeSafe): still listed as "Free" for a limited time, with the same data texts (R28-S37, R28-S38). Lookalike sites: all three still answer, and Eye Security's report still names them (R28-S14).
  • Catalogs: Jev is not listed on Microsoft Learn's Foundry partner or sold-directly pages, nor in the readable text of the Workers AI models index (R28-S49). Comparison cells: Mouse-1's docs still name no weights or base model (R28-S34).

Not checked or not read

  • The Amazon Bedrock model lists (their pages returned too little readable text on 8 Nov; the 31 Oct reading stands) and Eevo's announcement (not readable).
  • The contents of auto-mode 4.0.0, @ai-sdk/typesafe-ai 3.0.35 and ai 7.0.165; ohmyjev's Codex adapter; three other coding-agent routers named as candidates (pi-jev-router-pro, jev-codex-router, opencode-jev-pruner).
  • The new cookbook beyond its summary and figures; SubQ Snap's status and price.
  • Reddit, X, YouTube and TikTok could not be searched for the hours between the checks. No request was sent to any route or provider, no package was installed or run, and no account was created.

Sources

Sources read on 8 Nov 2026 (UTC times)
  • R28-S01 to R28-S16: TypeSafe console sign-in pages, the CEO post (X embed endpoint), docs quickstart, status page, models page, Images page, homepage, blog, Enterprise page, changelog, AI FrontPage, OpenRouter's Jev guide, the Eye Security report and the three lookalike sites, llms.txt and legal summary, and the new cookbook, 05:17–05:18.
  • R28-S20 to R28-S23: ohmyjev on npm, raw files at d307c39 and fbc4ea4, and the 0.13.2 and 0.14.0 tarballs, 05:19–05:21.
  • R28-S30 to R28-S34: OpenAI "Your data", Cloudflare Workers AI data usage, SubQ's data-retention docs and Terms, Mouse and Eevo docs, 05:24–05:32.
  • R28-S37, R28-S38: OpenCode Console models docs and Terms of Use, 05:24.
  • R28-S40 to R28-S47, R28-S49, R28-S50: route listings (OpenRouter, Vercel, Opper, Netlify, Cloudflare), LiteLLM releases, OpenRouter's ZDR list, npm and PyPI records for the tracked packages, the Bedrock, Foundry and Workers AI catalog pages and the langchain-typesafe 0.0.2 wheel, 05:24–05:33.
  • R28-S200 to R28-S215: @caph42/pi-jev-router on npm and its source at 75ad5b7, pi 1.5.0 at f9a2636, the langchain-typesafe wheels and release, and pydantic-ai-slim on PyPI, 05:17–05:21.
  • R28-S100 to R28-S141: Hacker News, GitHub, npm, PyPI, Hugging Face, DEV.to and OpenRouter searches over the hours between the checks, and the READMEs behind the coding-agent leads, 05:16–05:19.

Earlier dated reports: 2–7 Nov 2026, 31 Oct–2 Nov 2026; all reports are listed on Research.

Search published pools, pages, reports, and evidence.