Run 9 research report: coding agents beyond Claude Code, and Jev versions and aliases (6 Oct 2026)

This dated report records the evidence behind release regular-2026-10-06-coding-agents-versions of 6 October 2026. It covers 18 Jev integrations for coding agents (Codex, OpenCode, Cursor, OpenClaw and Hermes Agent), the Jev model IDs and aliases on 6 routes, a new provider for Jev on Vercel, and the failure checklist, which grows from 51 to 58 rows. The page titles it feeds use phrases that search suggestions offered on 6 Oct 2026, 05:19 UTC (a demand signal, not a volume). The pool ran nothing: no agent, hook, plugin or MCP server was installed or run, and no account, key or Jev call was used.

Run: run:6c40bcd4-3b75-4bfc-bc54-b254713e0f95 (scheduled regular research run 9, on the v0.2 harness) · Pool: pool_jev_catalog · Runner: one runner plus two helper agents (A: the Hermes Agent catalog and OpenClaw; B: Codex, OpenCode and Cursor, then the LiteLLM diff). GitHub search calls: 5 of 10 · Checks: 6 Oct 2026; runner started 05:13:45 UTC; home status written 05:18:34 UTC (minute 5); research notes finished about 05:36 UTC (minute 23).

The run's research notes and its source ledger (80 records, IDs R9-S01 to R9-S120) are held in the pool's private record. Key sources are linked below and on the pages each finding feeds.

Rules kept: nothing was installed or run. No agent, hook, plugin or MCP server was installed or run, and no account, key or inference request was used, so no Jev call was made. Failure cells come only from source at pinned commits. Nothing in this run is Tested. Every finding below is Documented, Reported (a named third party's own claim) or Unverified (a lead only).

Previous report: browser and computer use, framework integrations, Jev on Bedrock or Azure (5 Oct). All dated reports: Research.

1. Summary answer, as of 6 October 2026

1. Access is unchanged (checked 6 Oct 2026, 05:15 UTC). Signups are open with no new-user credit since 27 Sep, 22:33:25 UTC. The API has been operational since 29 Sep, 22:05 UTC (90-day uptime 99.826%). Limits are 100K tokens per second and 80 requests per second; the price is $0.042 per million input tokens. jev-latest and jev-preview both point to jev-1.13.0. Documented
2. One route change: Vercel AI Gateway now lists TypeSafe itself (zero data retention) as a second provider for typesafe-ai/jev, beside DigitalOcean (no zero data retention). Vercel picks between them by uptime and latency unless the caller sets the order. Documented
3. build/coding-agents ships. 18 integrations were read at pinned commits. 6/18 act on Jev's answer, covering Codex, Cursor, OpenCode and Hermes Agent. When Jev errors or there is no key, 4/6 let the tool call run and 2/6 hold it for the user. 1/18 is vendor-published: OpenClaw's own @openclaw/typesafe, which only returns Jev's answer to the agent. 0/18 come from TypeSafe. Documented
4. where-to-use/versions-and-aliases ships. 6/6 routes were read with dates. 4/6 accept a versioned ID: TypeSafe, Netlify and Opper take jev-1.13.0; OpenRouter takes typesafe/jev-1.13. Vercel and Cloudflare accept only an unversioned ID. Documented
5. Guard rows: both are still at their run-7 pins. jev-permission-gate counts on when-jev-fails. jev-llm-guard does not count: its command-line tool exits 0 whatever the verdict. Documented
6. when-jev-fails: n = 58 in 10 domains (adds a new "Coding agents" domain). 32/58 stop or hold the action on a Jev error; 7/58 let it through unchecked. Documented
7. Products: 92 rows (A 75, B 16, C 1); 0/92 independently verified in production.
8. Drift: the LiteLLM 1.104.0 failure path is unchanged. No new release of the six framework pins. ai 7.0.128 and @ai-sdk/typesafe-ai 3.0.13 are patches (drift note only).

2. Publication

This release shipped both planned pages and the follow-on edits:

• How to use Jev with Codex, OpenCode, Cursor and Hermes Agent (new page).
• Jev 1.13 and jev-latest: model IDs and aliases by route (new page).
• Refreshed: Claude Code and MCP guardrails (the two guard rows); When Jev fails (51 → 58 rows, new "Coding agents" domain); Products (92 rows).
• Refreshed: where to use, Vercel two-provider note; data privacy; model routing (LiteLLM 1.104.0 check); frameworks (recheck line); TypeScript (drift note).
• Status refresh on Home and access status.
• Release 3: 9 of 9 (closed). Release 4: 1 of 4.
• Failure colours aligned across pages (see Research for the list).

3. Corrections to plan facts and pool pages

1. Run 8 spare-time table: jev-curator is in its own repository (anpicasso/hermes-jev-curator, 4e8626c), not in anpicasso/hermes-jev-approvals. Documented
2. Hermes catalog pins: three catalog entries install a commit other than the run-8 HEAD pin: jev (28a4ea3), typesafe-skill-router (e6cdac2) and jev-approvals (28be98a). The catalog says "Every entry installs exactly the commit above". The page cites the catalog-installed commits. Documented
3. Vercel row on where-to-use: "DigitalOcean only (5 Oct)" is no longer current (see section 4).

4. Access status and home notices

• Checked 6 Oct 2026, 05:15:34 UTC. Signups open_with_conditions; service operational. The headline is unchanged: "Available to new users: signups open, but new accounts get no free credit. API operational." Changed since the last run: no. Documented
• New field in the home status record: the alias read. jev-latest and jev-preview both point to jev-1.13.0; TypeSafe's models page is undated (read 05:15:34 UTC).
• Notices shown, unchanged: new-user-credit-disabled (rank 1); lookalike-resellers (rank 2; all three sites return HTTP 200, and the Eye Security report is online); rate-limits-changed-2026-10-03 (rank 3); subprocessors-added-2026-10-02 (rank 4).
• New notice, hidden at rank 5 by manager decision: vercel-typesafe-provider-added-2026-10-06 (route change, Documented). It is not shown on Home. It retires on 13 Oct 2026, 05:15 UTC, or earlier if superseded.
• typesafe-workflow-evals-code stays hidden, now at rank 6. It retires at the 7 Oct run. Retired: none. Timeline additions: none.
• Vercel note for where-to-use and data-privacy: "Vercel lists two providers for Jev (checked 6 Oct 2026, 05:15 UTC): TypeSafe AI (zero data retention) and DigitalOcean (no zero data retention). Vercel picks a provider by uptime and latency unless you set the provider order." Vercel's docs say that by default it "dynamically chooses the default providers … based on a combination of recent uptime and latency". So zero data retention on one provider does not say where a given request goes.
• The typesafe-ai GitHub organisation has 12 repositories. The new one, typesafe-public-examples, holds only a README, so it is not a release.
• TypeSafe's docs file legal.md is 6 bytes larger; the legal "Last updated" dates are unchanged. Not compared line by line; recheck next run.

5. Scan (S1) highlights

• Window: 5 Oct 2026, 05:14 UTC to 6 Oct 2026, 05:14 UTC. Gap 0, overlap 0.
• Window counts: Hacker News "jev" 11 stories and 33 comments; 192 new GitHub "jev" repositories (9 with the topic); npm 14 "jev" and 7 "typesafe" packages published; 22 Hugging Face models modified; 1 DEV.to article. These are search-index counts, not verified integrations or adoption.
• Not readable: Reddit (HTTP 403) and X (login wall). Both are unknown, not zero.
• "jev vision" stays unexplained. It is suggested on all three endpoints. There is no Hacker News hit, no npm package and no TypeSafe vision model (TypeSafe docs say "Text only"). Bing and DuckDuckGo also suggest "jev vlm" and "jev vla".
• Coding-agent leads found and read this run: opencode-tool-gate (created 6 Oct, 03:14 UTC) and jev-kit (npm releases 5 Oct, 19:02 UTC).
• Leads not read (Unverified): abaljeu/jev-mcp ("for thejevai.com"; not a TypeSafe domain, so a possible lookalike watch item); Dotpals (Hacker News); jev-browser-wingman (Hacker News; browser page); n8n-nodes-system-one (n8n); kfchow-llm-value-router (routing); "Pseudo-Relevances with Jev" (reranking).
• No candidate home notice besides the Vercel change.

6. Findings per page and ship verdicts

6.1 build/coding-agents (new; closes release 3): ships

• Page: How to use Jev with Codex, OpenCode, Cursor and Hermes Agent (TypeSafe AI). Not claimed: that coding-agent plugins are unsafe, or anything about hosts or plugins that have no row. Short form: "Of 6 Jev coding-agent integrations that act on Jev's answer, 4 let the tool call run when Jev errors and 2 hold it for you."
• Editorial correction: the runner's draft said "12 only advise the agent". Of the 12 not counted, 8 return Jev's answer to the agent, 2 give advisory hints and 2 are not settled.
• Title query source: search suggestions checked 6 Oct 2026, 05:19:23–05:19:30 UTC. Google (client=firefox) suggests "how to use jev with codex" and "how to use jev with opencode" exactly, plus "jev in cursor" and "jev openclaw". "jev hermes" and "jev codex" appear on Google, Bing and DuckDuckGo. Every host named in the title has a counted row.
• Ship rule: 4 or more implementations at pinned commits, covering at least 3 of 5 hosts, with the required columns from source. Met: 18 implementations; the 6 counted rows cover 4 hosts.
• Spot-checks: 4 rows checked against the raw files (the failure cell and the "still runs" cell): bloudhood and rh-guard (helper B), jev-judge and OpenClaw (helper A). All matched.

Not counted (12 of 18):

• Return Jev's answer to the agent (8): @openclaw/typesafe (vendor-published), jev-typesafe, jev-skill-router-mcp, ourines jev v0.1.2, NiazMorshed2007/jev-review, burnigtm/jev-mcp, minhgv/jev-mcp and pedroknigge/mcp_jev.
• Advisory hints (2): typesafe-skill-router and wellkilo/codex-jev-preflight.
• Not settled (2): jev-approvals sends its verdict to Hermes core, which was not read; jev-curator was only partly read, so counting is deferred.

6.2 Guard rows and build/claude-code-mcp

• gulbaki/jev-llm-guard: HEAD 7c7b1b5 on 6 Oct; failure files unchanged. Not counted. Its jev-guard command-line tool returns 0 for allow, review and block (src/cli.js L31–40) and exits 1 only on errors (L41). The demo server only displays the result. Written as "returns verdict; not a decision". Documented
• madisonrickert/jev-permission-gate: HEAD 7349bc9 on 6 Oct; failure files unchanged. Counted in "Claude Code / MCP guards" with fallback-built-in classifier on error, malformed answer, below threshold and no key; bypass recorded. Documented
• Both rows go on Claude Code and MCP guardrails, using the run-7 cells. If the page counts the guardrails it read, the count sentence ("We read five Jev guardrails …") becomes seven.
• A one-line pointer on that page leads to coding agents for jev-kit, jev-review and mcp_jev, which also target Claude Code.

6.3 build/when-jev-fails: recount (n = 58, 10 domains, each with 3 or more rows)

• The base is the published 51-row map, read at 05:25:44 UTC. This run adds 1 guard row and 6 coding-agent rows. No existing row changes. Totals were recomputed by a script that checks every row and that each total sums to n = 58. Below-threshold totals are kept apart from error totals.
• New page meta: "58 Jev implementations in 10 use cases, read in source: what each does on a Jev error, a bad answer, an unsure answer, a missing key or a bypass."

6.4 products: row changes

• 20 new A rows: the 2 guard rows and 18 coding-agent projects.
• The existing B rows "jev-review" (devagrawal09) and "openclaw-typesafe-ai" (Olli0103) are different projects from the new rows, so both stay.
• New total: 92 rows (A 75/92, B 16/92, C 1/92), was 72 rows (A 55/72, B 16/72, C 1/72). Independently verified in production: 0/92.

6.5 where-to-use/versions-and-aliases (new; first release-4 page): ships

• Page: Jev 1.13 and jev-latest (TypeSafe AI): model IDs and aliases by route. Not claimed: that jev-latest is safe, or anything about the next version.
• Title query source: "jev 1.13" and "jev-latest" appear on all three endpoints (checked 05:19:23–05:19:30 UTC); Google now suggests "jev-latest" too. "jev 1.14", "jev versions" and "jev model version" got no supporting suggestion.
• Ship rule: at least 4 routes with IDs, aliases and dates read this run, plus TypeSafe's own alias statement. Met: 6/6 routes read 05:15–05:24 UTC. 4/6 show a date of their own; TypeSafe's page and Cloudflare are marked "undated; read at …".

7. Drift (S5)

• LiteLLM 1.104.0 (tag commit 79645770fedc7ec2627e6468d31062f20f82aecc): the Jev classifier's failure path is unchanged from 8d28e8d. The runner compared the whole caller block by script and found it identical. Line for model routing: "Checked at LiteLLM 1.104.0 (tag commit 7964577, 6 Oct 2026): the Jev classifier's failure path is unchanged from 8d28e8d." Caveat: the tag may predate 8d28e8d; their order was not verified. Documented
• Framework pins: no new release of any of the six. Line for frameworks: "Rechecked 6 Oct 2026, 05:20 UTC: no new release of any of the six packages; LiteLLM 1.104.0 failure path unchanged."
• Drift notes only: ai 7.0.128 and @ai-sdk/typesafe-ai 3.0.13 (TypeScript), and the oh-my-claudecode tag v5.6.2 (no release; Claude Code and MCP guardrails).
• New this run: @osuki-dev/opencode-osuki-agent 0.2.8 is already out (the row is pinned at 0.2.7). Its page note says "0.2.8 not read".

8. Spare time

1. The LiteLLM TypeSafe proxy guardrail at v1.104.0 is a context-compaction guardrail, not a safety block. It blanks tool results that Jev judges no longer relevant (below 0.2). Its default model is jev-latest. On a Jev error it forwards the request uncompacted by default; it can be set to return HTTP 502 instead. With no key it fails at start-up. It is a candidate decision row for the next frameworks or routing refresh. Documented
2. Item 2 (reranking studies) and item 3 (database error behaviour) were not reached.

9. Evidence gaps and uncertainties

• Host behaviour after a hook decision: Codex, Cursor, Hermes core and OpenClaw core were not read.
• OpenCode's default permission configuration.
• rh-guard's default mode and its thresholds.
• osuki 0.2.8, and whether the 0.2.7 package matches its tag.
• How jev-approvals errors are handled in Hermes core.
• jev-curator's plan path.
• The contents of all test files.
• Commit dates for 4 MCP repositories.
• What each route actually serves (no request was sent).
• Whether OpenRouter accepts patch-level IDs.
• Route changelogs: only TypeSafe's was checked, and it returns 404.
• SDK default models beyond TypeSafe's statement.
• The legal.md size change.
• The Internet Archive (connection error).
• Reddit and X (unknown).

10. Not done (for run 10)

• The reranking-rag condition check (spare item 2). It decides run 10's page.
• The databases-and-documents error behaviour (spare item 3).
• Unread coding-agent leads: about 20, listed in the research notes, including the 7 further Hermes catalog jev-* plugins and 0xNatoshi/jev-codex-router.
• The per-SDK default-model check.
• The legal.md comparison.
• Drift-list additions from run 10: @openclaw/typesafe, @osuki-dev/opencode-osuki-agent and the three @jev-kit packages.

11. Timings (UTC, 6 Oct 2026)