Coding agents on pi and omp, auto-mode and jcode, and Nace.AI's two sets of terms (21 Oct 2026)
This dated report records the evidence behind the update of 21 October 2026. It adds ohmyjev's pi and omp adapters to coding agents as a counted row, lists auto-mode and jcode as read but not settled, classes 13 new coding-agent leads, names pi and omp on ohmyjev's row of the failure map (still 80 implementations), and quotes Nace.AI's website terms next to its Console terms on alternatives. It also records the day's access check: the two Jev 1.14 notices came off the home page after seven days.
Coding agents: of 27 coding-agent integrations read in source, 14 act on Jev's answer, 10 do not and 3 are unsettled. On a Jev error those 14 are 3 fail-closed, 9 fallback and 2 fail-open.
ohmyjev on pi and omp: on every Jev failure the gate returns without blocking, so the host runs the tool (fail-open with no key, on a timeout, an HTTP error or a malformed answer; acts-anyway below its thresholds). These are the same five terms as its Claude Code module.
Failure map: ohmyjev's row now names its pi and omp adapters; the map still counts 80 implementations and no count on it changed.
Drex: Nace.AI's own legal pages disagree. Its website Terms and Privacy Policy (9 Oct 2026) allow training on customer content by default, with an opt-out, and keep it up to 30 days; its Console Service Terms and Data Processing Addendum (19 Oct 2026) say no training unless you opt in, and deletion within 48 hours (paid) or 7 days (free). Neither page says which controls.
Access: unchanged at 21 Oct 2026, 05:11 UTC. jev-latest still points to jev-1.13.0 a week after TypeSafe said it would move "in a week"; the home page now shows two notices.
The 14 coding-agent integrations that act on Jev's answer, when the Jev call fails (21 Oct 2026)
- Jev error (timeout or HTTP error), in map termsn = 14
- 3/14 hold the call (fail-closed)
- 9/14 a named substitute decides (fallback)
- 2/14 let the call run (fail-open: jev-judge in enforce mode; ohmyjev on pi and omp)
- Does the tool call still run?n = 14
- 5/14 yes (2 fail-open, 3 fallback)
- 2/14 conditionally (both fallback)
- 3/14 no (fail-closed)
- 4/14 routing rows: the turn runs on another model (fallback)
What was read, and at which commits
| Project | Pin | Files read | Result |
|---|---|---|---|
| Novacon/ohmyjev, pi and omp adapters | 78bb82c (HEAD since 15 Oct; unchanged 21 Oct) | adapters/pi.ts, adapters/omp.ts, adapters/shared.ts, hooks/policy.ts, hooks/jev.ts, package.json, README; host docs: pi docs/extensions.md at 16f5f0a, oh-my-pi docs/extensions.md at d5cd750 | Counted on coding agents; the same five terms as the Claude Code module (5 of 5 on each host) |
| zgeoff/auto-mode (Claude Code mod) | read at 2.0.2 (458c41c); cited files re-read at 3.0.0 (c0b9378) | mod register.ts, parse-decision.ts, cli-limits.ts; config.ts, presets.ts, classify-with-jev.ts, classify-action.ts and the outcome and verdict helpers; README and the Jev stage doc | Read, not settled |
| 1jehuang/jcode, PR #1890 | merge f5c4922 (19 Oct) | config types; memory.rs, browser_jev.rs, voice_intent.rs | Read, not settled |
| 13 new leads (19–20 Oct scans) | one commit or package version each | one file each, with the Jev call | Classed; none counted |
Nothing was installed or run and no Jev call was made; jcode's source was read at its pinned commit and not built.
pi and omp: what happens when the Jev call fails
ohmyjev ships one Claude Code module and two adapters, for pi and for omp (oh-my-pi). Both adapters register a tool_call handler that sends Bash, write/edit and exfiltration-type calls to Jev through shared code, with a 1,500 ms timeout. When the call fails for any reason, the shared code returns nothing and the handler returns without blocking. pi's and oh-my-pi's extension docs say a tool_call handler "may block"; one that does not block lets the host run the tool. Documented
- No API key
- fail-open: the call throws and the gate lets the tool run; both adapters log "no Jev key: gates and screens let every call through".
- Timeout or HTTP error
- fail-open: 1,500 ms; same return.
- Malformed answer
- fail-open: the reply parser throws; same return.
- Below threshold
- acts-anyway: Bash is denied only at irreversible ≥ 0.6, irreversible with destructive intent ≥ 0.7 or exfiltration ≥ 0.7; writes only at secret ≥ 0.7 (secrets file ≥ 0.8).
- Bypass
- recorded: an off switch, read-only commands that skip Jev, each gate's own switch, and tools that are never sent to Jev.
- Other handlers
- Model routing, the done-check and (on omp) the subagent hook set no route, keep the model or do not block when Jev fails.
Because the Claude Code module was already one of the 80 on the failure map with the same terms, the map's ohmyjev row now names pi and omp instead of gaining a new row. One file reference in that row's bypass cell was corrected (runBrowse is in hooks/ohmyjev.ts L557–568, not in adapters/shared.ts).
Read but not settled: auto-mode and jcode
- auto-mode 3.0.0 (Claude Code mod, published 20 Oct): its code replaces Claude Code's question with allow or deny; when Jev fails it keeps the question for you (as read: held with no key, on a 5,000 ms timeout, an HTTP error or a malformed answer; acts-anyway below a 0.2 block threshold). Not counted, because how the mod host handles its answer was not read in the host's own source, and the stages before Jev were not read line by line.
- jcode (PR #1890, merged 19 Oct): makes three Jev models configurable. Jev chooses injected memories (none on an error), a browser action and a voice intent; no tool-call gate or model routing was found in the files read. Other callers were not read.
The 13 new leads
5 act on Jev's answer in a host plugin or agent and are candidates for a later full read (jev-plan-review, model-selector, Covre81/jev-router, Keen Code, pi-jev-auto-mode); 1 acts in its own test setup (recoveriq); 1 acts only on context compaction; 2 return the answer to the agent; 1 is a client library; 2 call another model, not Jev; 1 is a feature request with no code (13 in all). No Kilo Code Jev integration was found on GitHub repository search or npm search (21 Oct 2026, 05:12 UTC). The full list is in lead triage.
Nace.AI's two sets of terms
On 20 Oct the Drex row rested on Nace.AI's Console Service Terms. Its website Terms of Service and Privacy Policy and its Console Data Processing Addendum were read on 21 Oct. Both sets of terms say they cover the Console and the APIs; which one controls is not stated on the pages read, and this report does not infer it. A reading of the text, not legal advice. Documented
| Term | Website Terms and Privacy Policy (9 Oct 2026) | Console Service Terms and Data Processing Addendum (19 Oct 2026) |
|---|---|---|
| Training | "We may use Customer Content to develop, train, evaluate, and improve our Models and Services"; opt-out in the Console settings or by email | "No training by default … unless you separately and expressly authorize that use"; the addendum: Nace "will not use Customer Content to train" |
| Retention | deleted content removed "within 30 days"; inputs and outputs "up to 30 days"; API logs "up to 13 months" | "within 48 hours after processing completes for Paid Accounts, and within 7 days" for Free Tier Accounts |
| Zero retention | not found in the text | optional; deleted "within 1 hour after results are delivered", not written to logs or backups |
| Sub-processors and region | the linked sub-processor page returned HTTP 403 and was not read | a schedule that still says processing locations are "to be confirmed before publication" |
drex-v1.5 price: "$0.04" per 1,000 decisions on the Drex product page and "$0.05" per 1M input tokens in the pricing docs (both read 21 Oct 2026). The two use different units; they are not converted into each other.
Version checks and route lines
- Cloudflare: its AI model catalog lists
typesafe/jev-1.14.0as a third-party model served through AI Gateway, with a 32,000-token context window and images and Box marked beta (checked 21 Oct 2026, 05:13 UTC; first seen by this catalogue on 21 Oct; it may have been listed earlier). Its Jev overview page still shows 1.13 examples. No request with that ID was tested (versions and aliases). Documented - TypeSafe's n8n node
@typesafe-ai/n8n-nodes-typesafe-ai0.10.0 was published on 20 Oct; its release note says it is for image support. The change itself was not read. Documented - Patch releases noted, not re-read:
@ai-sdk/typesafe-ai3.0.24,ai7.0.148,@jkudish/jev-mcp0.14.4,@delt/claude-jev-advisor0.5.1,litellm1.106.1.pydantic-ai-slim2.58.0 and 2.59.0 were not checked for changes to its TypeSafe files. - LiteLLM pull request #46875 (eu/jev) is still open (21 Oct, 05:16 UTC), so no release ships it.
- Clairic (Cloudflare, post dated 19 Oct): a preview of an extraction model on Workers AI that returns typed values, not Jev-style decisions; no price in the post. A different layer, not listed as a substitute. Documented
- "nano jev": a likely source is the Hugging Face model
C-Tianyu/NanoJev("A nano replica of Jev"); Jev-named, not published by TypeSafe, and not shown to be the source of the search suggestion. Reported
Access and notices
Checked 21 Oct 2026, 05:11 UTC: signups open, no free credit for new accounts, API operational (90-day uptime 99.826%; last event resolved 14 Oct). Price $0.042 per million input tokens for jev-1.14.0 and jev-1.13.0, output free. TypeSafe's models page and quickstart changed size since 20 Oct; the changed lines were not identified, and their access, price, alias and limit text reads the same. The home page's two Jev 1.14 notices (the preview, and the experimental image, Box and long-context features) came off after their seven-day shelf life; both facts stay on versions and aliases and limitations. The home page shows two notices: no credit for new users, and lookalike reseller sites. Documented
Dates the scan covered
This update's scan covered 20 Oct 2026, 05:13 UTC to 21 Oct 2026, 05:10 UTC (23 h 57 min), starting at the previous scan's cutoff, so there is no gap and no overlap; the 72-hour look-back for hosted decision models started 18 Oct, 05:10 UTC; no new launch was found on the surfaces searched (21 Oct 2026, 05:14–05:16 UTC). Reddit and DEV.to search returned HTTP 403, and X, YouTube and TikTok were not searched, so their coverage is unknown.
What was not verified
- No Jev call, account or bill; no Drex call; nothing installed, built or run.
- pi's and omp's own code after a handler returns without blocking (the row rests on their extension docs and ohmyjev's README).
- auto-mode: the mod host's handling of its answer and the stages before Jev. jcode: callers outside memory, browser and voice.
- The 13 new leads rest on one file each; the 5 that act on Jev's answer need a full read before they can be counted.
- Nace.AI's sub-processor page (HTTP 403) and the Fireworks AI terms; which Nace.AI terms document controls.
- The lines that changed on TypeSafe's models page and quickstart (earlier copies were not kept).
Method and sources
Cells were read at the pinned commits and set in the failure map's shared terms; counts on the coding-agents page, the build hub and in this report were recomputed by script from the page's own table rows. Sources read 21 Oct 2026: access and alias checks R21-S01 to R21-S13; reseller rechecks R21-S14 to R21-S18; route checks R21-S22 to R21-S31; scan and lead checks R21-S34 to R21-S47; ohmyjev R21-S100 to R21-S112; auto-mode R21-S115 to R21-S138; leads R21-S200 to R21-S229; jcode R21-S230, R21-S231; Nace.AI and DeepInfra terms R21-S232 to R21-S245.
- All research reports
- 20 Oct research report (the self-check and the first Drex row)
- 19 Oct research report (ohmyjev's Claude Code module joins the map)