Shaduf.Research preview
Jev: Use Cases, Alternatives & Products/Claude Code and MCP guardrails
Build6 integrations read in pinned source · checked , 07:09–07:14 UTC · versions rechecked (oh-my-claudecode checked at v5.6.0; jkudish at 0.13.0 on 2 Oct) · v5.6.1 release noted · 2 guard rows added and rechecked · 2 new releases noted, not read, · oh-my-claudecode v5.6.2 and jev-mcp 0.14.1 checked : failure paths unchanged · jev-effort-router note updated · jkudish/jev-mcp bypass paths read at the audited pin and 0.14.3 release noted · versions rechecked (failure paths unchanged), title updated and 32 leads classified (05:20–05:33 UTC) · 3 guardrails added (read at pinned commits 16 Oct; bypass entry points read 18 Oct) and short answer updated, checked (05:17–05:36 UTC) · ohmyjev's pi and omp adapters read in full, its bypass cell recorded and ohmyjev counted on the failure map; jkudish/jev-mcp read and check commits shown separately · ohmyjev's pi and omp gates rechecked at the pinned commit · ohmyjev 0.13.2 policy file compared · ohmyjev re-read at 0.13.2 (d307c39), lockdown mode named · ohmyjev 0.14.0 and langchain-typesafe 0.0.2 noted · ohmyjev re-read at 0.14.0 (fbc4ea4) and langchain-typesafe read at 0.0.2

Jev Claude Code and MCP guardrails (TypeSafe AI): which hook, plugin or server fails open?

Claude Code hooks, plugins and MCP servers exist that ask Jev whether an agent's next step is safe. This page compares what eleven of them do when Jev has no key, times out, returns an error, returns something malformed or is unsure. It also shows what each one sends to Jev and which mode it uses after installation. Every cell comes from the integration's source code at a stated commit.

Short answer

We read ten Jev guardrails in their source code. Two hold or stop when Jev fails: LangChain AutoModeMiddleware stops the agent, and claude-jev-advisor's Windows delete guard keeps Claude Code's question for you. One, jev-permission-gate, falls back to Claude Code's built-in classifier. Two ship in a shadow or log-only mode, one is advisory by design, and three fail open: danna-zhou/jev-mcp, claude-model-guard and ohmyjev (in lockdown mode without shadow, both off by default). One, gulbaki/jev-llm-guard, only returns a verdict: it is not counted as a decision. The official TypeSafe plugin makes no Jev calls at all. Before you rely on any of them, check three things: its default mode, what it sends to Jev, and its timeout. Documented source at pinned commits, checked 18 Oct 2026; the seven earlier guardrails read 29 Sep to 6 Oct 2026 and versions rechecked 16 Oct 2026, 05:33 UTC, with no failure-path change; claude-jev-advisor, claude-model-guard and ohmyjev read at pinned commits 16 Oct 2026, bypass entry points read 18 Oct 2026; ohmyjev re-read at 0.14.0 (fbc4ea4), 8 Nov 2026

claude-model-guard, claude-jev-advisor and ohmyjev are also counted on when Jev fails (80 implementations; ohmyjev from 19 Oct 2026, after its pi and omp adapters were read in full: bypass table).

jev-kit (Codex and Claude Code), NiazMorshed2007/jev-review and pedroknigge/mcp_jev also target Claude Code; they are compared on coding agents.

Version checks, 16, 18 and 30 Oct and 2 and 8 Nov 2026

Checked at ohmyjev 0.14.0 (8 Nov 2026): Re-read 8 Nov 2026 at 0.14.0 (fbc4ea4): failure cells unchanged; bypass cell adds the shadow setting and the in-repo move skip. The new shadow setting is off by default and lockdown is still off by default; one more key source (ohmyjev key). adapters/codex.ts was not read; the new bin/ohmyjev.mjs is an install and key command, not a gate path. Documented R28-S20 to R28-S23

langchain-typesafe 0.0.2 (8 Nov 2026): read at 0.0.2 (wheel, 8 Nov 2026); failure path unchanged; adds an optional classifier= argument on both experimental middlewares. The AutoModeMiddleware row in the failure matrix keeps its terms; its file references stay as read at 4af7ab8 (0.0.1a3). Documented R28-S211 to R28-S214

Noted 7 Nov 2026 (registries): ohmyjev 0.14.0 (3 Nov 2026) and langchain-typesafe 0.0.2 (5 Nov 2026) were newer than the versions read (0.13.2 at d307c39; 0.0.1a3 at 4af7ab8); both were read on 8 Nov 2026 (above). Documented (npm and PyPI registry records, 7 Nov 2026, 15:29–15:36 UTC; R27-S25)

Checked 2 Nov 2026 (ohmyjev): the row was re-read at d307c39 (ohmyjev 0.13.2, the npm package's commit). The gates run only with lockdown on; lockdown defaults to off (policy.ts L12; .claude-plugin/plugin.json L29–34), and with it off gate returns before any Jev call (ohmyjev.ts L272). In lockdown mode the four failure terms are unchanged; the bypass cell changed in content (auto-approve removed; GitHub push reminder added). Earlier: 30 Oct 2026 policy-file comparison (R24-S38, below). Documented R26-S20 to R26-S23

Checked 30 Oct 2026, 05:20 UTC (ohmyjev): ohmyjev 0.13.2 (29 Oct 2026) removes the auto-approve thresholds from its policy file; the failure-map row was read at an earlier version (re-read at 0.13.2 on 2 Nov 2026: above). The policy files of npm 0.12.0 and 0.13.2 were compared: autoApprove, approveConfidence 0.9, approveRiskMax 0.2 and approveBash are no longer in hooks/policy.ts; a new lockdown setting, off by default, holds the blocking gates; the 1,500 ms timeout is unchanged. The row's entry file hooks/ohmyjev.ts is not in the npm package, so the row's cells rest on the pinned read at 78bb82c; its terms and colours are unchanged. Documented for the files compared R24-S38

Checked 18 Oct 2026: for the three guardrails added to this page, HEAD equals the pin read on 16 Oct: Novacon/ohmyjev 78bb82c (15 Oct), delt96/claude-jev-advisor f2c1791 (14 Oct; npm @delt/claude-jev-advisor 0.5.0) and mcowdery/claude-model-guard bced870 (13 Oct). @jkudish/jev-mcp 0.14.4 was still not on npm (latest 0.14.3). The other integrations were not rechecked on 18 Oct; their 16 Oct check follows. Documented commit feeds, 18 Oct 2026, 05:20 UTC (R18-S200 to R18-S202); npm registry, 05:18 UTC (R18-S230)

Checked 16 Oct 2026: official TypeSafe plugin v0.5.7 (65a39f3, makes no Jev call); danna-zhou/jev-mcp 436469a; jev-claude 7d7e876; jev-mcp 0.14.3 on npm (cells read at 0.11.0); langchain-typesafe 0.0.1a3 (4af7ab8); oh-my-claudecode 5.7.3 (Jev hook files compared with 5.6.2: unchanged apart from one comment); jev-permission-gate v0.2.0 (7349bc9); jev-llm-guard 7c7b1b5. Failure paths read at the pins shown; 1 release compared (oh-my-claudecode 5.7.3); 1 unreleased patch noted, not re-audited (jev-mcp 0.14.4 commits, 15 Oct). Documented commit feeds, release feeds and registries, 16 Oct 2026, 05:20–05:33 UTC; R17-S200 to R17-S215

  • oh-my-claudecode: v5.7.1 and v5.7.2 (13 Oct), v5.7.3 (14 Oct, 03:07 UTC; npm oh-my-claude-sisyphus 5.7.3 published 14 Oct, 03:24 UTC). The Jev hook files in src/hooks/jev/ are byte-identical to v5.6.2 except a comment (types.ts L4). Checked at 5.7.3: failure path unchanged. Documented R17-S209 to R17-S213 Patch releases v5.7.4 (17 Oct) and v5.7.5 (18 Oct 2026) noted, not re-read; the cells rest on the read at 5.7.3. R19-S50
  • jkudish/jev-mcp: HEAD 75cf447 (15 Oct) "chore: release 0.14.4", after "chore: update TypeSafe SDK to 0.7.0"; 0.14.4 was not on npm when checked (16 Oct 2026, 05:20–05:29 UTC). Noted, not re-audited. Documented R17-S203 to R17-S206
  • Others: HEAD equals the pin shown above; no new release. R17-S200 to R17-S202, R17-S207, R17-S208, R17-S214, R17-S215

Which Jev model they ask for. oh-my-claudecode (client.ts L65, at tag v5.7.3) and jkudish/jev-mcp (src/index.ts L86, at the audited 0.11.0 pin) send jev-latest by default, as do claude-model-guard (jev.mjs L132) and claude-jev-advisor (jev.ts L63; below). So on TypeSafe's API these four have received jev-1.14.0 since the move first seen on 24 Oct 2026, 05:12 UTC, without a code change (per TypeSafe's models page; no request was sent); ohmyjev pins jev-1.13.0 on TypeSafe (jev.ts L13–16) and does not move. On Opper, typesafe/jev-latest still mapped to jev-1.13.0 at 05:15 UTC (R23-S05, R23-S27). The integrations' code was not re-read on 24 Oct 2026; the pins are those published here. Before the move: TypeSafe said on 14 Oct 2026 that jev-latest would move to jev-1.14.0 "in a week"; on 23 Oct 2026, 05:45 UTC, it still pointed to jev-1.13.0, and TypeSafe had published no date for the move (R22-S05, R22-S09). TypeSafe published no dated announcement of the move; jev-1.13.0 is still listed and callable by its ID, with no deprecation date (what changed with Jev 1.14). The danna-zhou hooks default to kev-latest (hooks/lib.mjs L12). The model field of jev-claude, LangChain and jev-permission-gate was not read. Documented source at the pins shown, read 16 Oct 2026, 05:33 UTC; R17-S213, R17-S246 to R17-S248

Change, 24 Oct 2026: the paragraph above said that when the alias moves, oh-my-claudecode and jkudish/jev-mcp "will ask 1.14 without a code change"; it now gives the move first seen on 24 Oct 2026, 05:12 UTC, names the four integrations that send jev-latest and ohmyjev's pin, and adds Opper's alias (still on jev-1.13.0 at 05:15 UTC). The checklist's model-ID item is updated the same way. The integrations' code was not re-read (R23-S05, R23-S27).

Change, 23 Oct 2026: the alias sentence in the paragraph above now gives the 23 Oct check (jev-latest still on jev-1.13.0; no date published for the move), and the 1.14 consequence is conditional on the move; it said TypeSafe "says" the alias moves "in a week". The integrations' code was not re-read.

The three added guardrails (18 Oct 2026). ohmyjev pins jev-1.13.0 on TypeSafe (jev.ts L13–16; ~typesafe/jev-latest when only an OpenRouter key is set, L21–25; jev.ts byte-identical at d307c39, read 2 Nov 2026; at fbc4ea4, 8 Nov 2026, jev.ts changed only in its key sources as compared, its model line not checked separately), so it will not move with jev-latest. claude-model-guard (jev.mjs L132) and claude-jev-advisor (jev.ts L63) send jev-latest. Documented source at the pins shown, read 16 Oct 2026; R17-S236, R17-S239, R17-S243

Earlier version checks (newest first)

14 Oct 2026: @jkudish/jev-mcp 0.14.3 was released on 13 Oct 2026, 18:57 UTC (patch). We have not re-audited it; the cells on this page were read at the version shown. Documented npm registry, read 14 Oct 2026, 16:51 UTC (R16-S64)

13 Oct 2026, 05:22 UTC: @jkudish/jev-mcp 0.14.2 released 13 Oct 2026, 00:02 UTC, and @jkudish/jev-agent-tools 0.2.1 released 12 Oct 2026, 23:48 UTC (patches); not re-audited. Documented npm registry, 13 Oct 2026, 05:22 UTC (R15-S52)

Rechecked 10 Oct 2026: oh-my-claudecode v5.6.2, @jkudish/jev-mcp 0.14.1 and jev-agent-tools 0.2.0 are still the latest releases. Documented registry and release listings, 10 Oct 2026, 05:18:47 UTC; R12-S48, R12-S49

oh-my-claudecode v5.6.2 (8 Oct). Checked at v5.6.2 (GitHub release, 6 Oct 2026; commit 454bae0) on 8 Oct 2026: failure path unchanged (advisory as installed; fallback-heuristic when active; 2,000 ms). Between v5.6.0 and v5.6.2 only points.ts and resolver.ts changed. The Jev call (client.ts L34–54) and config.ts are the same files as in v5.6.0 (DEFAULT_TIMEOUT_MS = 2000, config.ts L36). New in v5.6.2: a separate branch for non-blocking points in active mode, with the same outcome: the heuristic twin answers and the call is logged as degraded (resolver.ts L195–218; "The resolver never throws", L10–13). Documented release and tag commit (R11-S100), source at 454bae0 (R11-S101, R11-S102), read 8 Oct 2026, 05:13 UTC; tag commit confirmed with git ls-remote, 05:21 UTC

jkudish/jev-mcp 0.14.1 (8 Oct). jkudish/jev-mcp: audited 0.11.0; checked 0.14.1 on 8 Oct 2026: failure path unchanged (error returned to the agent; 60 s). 0.14.1 was published on npm on 6 Oct 2026 at 21:41 UTC (tag v0.14.1 = 86eae78, the npm gitHead). Compared with 0.13.0 (5e0ca5c), only dist/provider.js, the README, the changelog and package.json changed; dist/server.js is byte-identical. In dist/provider.js the timeout is still REQUEST_TIMEOUT_MS 60,000 ms (L25) and the Jev call is the same (L263). Errors still go back to the agent: OpenRouter error messages now add an allow-listed code (L304), and an invalid model in the response now raises an error (L315–317). Documented npm publish times (R11-S103) and the 0.14.1 tarball, read and not run (R11-S104), 8 Oct 2026, 05:13 UTC

@jkudish/jev-agent-tools 0.2.0, not audited. jev-mcp 0.14.1 now depends on @jkudish/jev-agent-tools ^0.2.0 (was ^0.1.4). This package carries the TypeSafe-direct and Vercel transports. 0.2.0 was published on npm on 6 Oct 2026 at 20:21 UTC (tag v0.2.0 = 8a237b9; MIT). Its README calls it a "Multi-provider Jev transport with fail-closed validation" Reported. Only the README was read; its code and its timeout were not (R11-S105).

oh-my-claudecode tag v5.6.2 (6 Oct). A git tag v5.6.2 exists with no GitHub release; the latest release is still v5.6.1 (checked 6 Oct 2026, 05:20 UTC). Its failure path was not re-read then; it was checked on 8 Oct (above). Documented (tag and release listing)

oh-my-claudecode v5.6.1 (4 Oct). A patch release, v5.6.1, appeared on 3 Oct 2026 at 06:44 UTC. Its source was not re-read: the rows below give the failure path checked at v5.6.0. No other integration on this page had a new release when checked on 4 Oct 2026. Documented (release listing)

oh-my-claudecode v5.6.0 (3 Oct). oh-my-claudecode's latest release is v5.6.0 (1 Oct 2026). Checked at v5.6.0 (commit e74b22c) on 3 Oct 2026: failure path unchanged; default timeout now 2,000 ms. On a timeout, HTTP error or invalid response the heuristic twin still decides and "the resolver never throws" (resolver.ts L10–12); a point still stops calling Jev after 3 failures in a row; nothing is sent unless OMC_JEV lists points; excerpts are still cut to 200 characters; points still run in shadow by default. New in v5.6.0: OMC_JEV=<point>:active or OMC_JEV=all:active makes Jev decide (a warning is printed), and a script-side judgment channel follows the same degrade-never-block rule. The 250 ms default that made every call fall back (round trips of 465–605 ms reported by the maintainers, #4091, Reported) was raised to 2,000 ms in v5.6.0 (config.ts L36). If you run v5.5.0 or earlier, set OMC_JEV_TIMEOUT_MS=2000. Documented (source at commit e74b22cd, read 3 Oct 2026, 05:21 UTC).

jkudish/jev-mcp (2 Oct).

@jkudish/jev-mcp 0.12.0 (1 Oct) and 0.13.0 (2 Oct, 03:50 UTC) were compared with the audited 0.11.0 at their tagged commits: failure path unchanged. The server still only advises; malformed answers still come back as "review", and from 0.12.0 rate-limit and outage results come back as tool errors. 0.13.0 adds a repository example hook (not in the npm package) that denies a tool call at a block probability of 0.85 or more and, if Jev fails, defers to Claude Code's normal permission prompt. Documented (source diff), checked 2 Oct 2026, 05:22 UTC.

The other four are unchanged: langchain-typesafe is still 0.0.1a3; danna-zhou/jev-mcp was last pushed 23 Sep, brunopivetta88/jev-claude 25 Sep and typesafe-ai/skills 12 Sep (checked 2 Oct, 05:14 UTC). Sources: 0.11.0 to 0.12.0, 0.12.0 to 0.13.0, example hook. Plugins that pick which model answers, rather than allow or deny a tool call, are compared on Jev Router or your own model router?

What you get by default

Installing a Jev guardrail can mean anything from no runtime effect to a guard that halts the agent whenever Jev is unavailable. The figure places each audited integration by what it does as installed. Dashed entries are modes you have to switch on yourself.

Effect on the agent as installed, from none (top) to strongest (bottom)

  1. No Jev callText instructions for your agent only
    • Official typesafe@typesafe-ai plugin
  2. Records onlyShadow or display mode; nothing is blocked
    • jev-claude (default shadow)
    • oh-my-claudecode Jev points (shadow; its built-in heuristic decides)
    • danna-zhou Stop hook
  3. Advises the agentReturns a recommendation; the agent may ignore it
    • jkudish/jev-mcp
    • danna-zhou MCP tools
  4. Gates, but lets the call through if Jev failsCan deny or ask; fails open
    • danna-zhou PreToolUse hook
    • claude-model-guard (blocks a prompt; a model-tier gate)
    • ohmyjev (Bash, Write/Edit and exfiltration gates in lockdown mode without shadow)
    • jev-claude enforce (default failOpen)
  5. Gates, and stops or asks if Jev failsFails closed
    • LangChain AutoModeMiddleware (stops the run)
    • claude-jev-advisor rm guard (keeps its own ask; Windows only)
    • jev-claude enforce with JEV_GUARD_FAIL_OPEN=false (asks a person)
Built from the failure matrix below. Documented source read at the pinned commits, 29 Sep 2026; claude-model-guard, ohmyjev and claude-jev-advisor read 16 Oct 2026, added 18 Oct 2026. It shows behaviour, not quality: none of these integrations was run against the live Jev API.

Failure matrix: what happens to the tool call

Each cell says what happens to the agent's action in one failure class. "Normal permission flow" means the hook emits no decision, so Claude Code's own permission rules apply. It is not an approval. Use the filter to compare default modes with the opt-in ones.

Show
Failure classes (a)–(e) for 11 integrations, 16 rows including modes. "Fail-open switched off" means JEV_GUARD_FAIL_OPEN=false. Documented from source at the pinned commits in the next table, checked 29 Sep 2026 (jev-permission-gate and jev-llm-guard: read 4 Oct, rechecked 6 Oct 2026; claude-model-guard, claude-jev-advisor and ohmyjev: read at the pinned commit 16 Oct 2026, bypass entry points read 18 Oct 2026), unless a cell says Tested.
Integration and mode after install(a) No key(b) Timeout (default)(c) HTTP 4xx / 5xx(d) Malformed response(e) Low confidence
Official TypeSafe plugin (typesafe@typesafe-ai)Mode: Instructions onlyn/aNo runtime pathn/an/an/an/a
danna-zhou/jev-mcp: PreToolUse hookMode: Enforcing (deny or ask) once added to settings.jsonProceeds anyway (fail-open)Sends Bearer local; any failure → normal permission flow Tested offlineProceeds anyway (fail-open)8,000 msProceeds anyway (fail-open)Proceeds anyway (fail-open)Holds for a person (held)"dangerous" below 0.6 confidence, or "moderate" → ask. "safe" proceeds unless needs_human ≥ 0.6
danna-zhou/jev-mcp: Stop hookMode: Display onlyFlags only (advisory)The turn ends normallyFlags only (advisory)8,000 msFlags only (advisory)Flags only (advisory)Flags only (advisory)Scores are displayed only
danna-zhou/jev-mcp: MCP toolsMode: AdvisoryFlags only (advisory)Error returned to the agentNo timeout in codeBehaviour on a hang not testedFlags only (advisory)Error to the agentFlags only (advisory)JSON parse error to the agentFlags only (advisory)Raw answer returned; the agent decides
jev-claude ("jev-guard" 0.1.0), as installedMode: Shadow (log and note)Flags only (advisory)Flags only (advisory)Flags only (advisory)Flags only (advisory)Flags only (advisory)Every decision becomes "allow" plus a note
jev-claude, enforce mode with the default fail-openMode: Opt-inProceeds anyway (fail-open)Its own deterministic rules still applyProceeds anyway (fail-open)1,500 ms; hard deadline 8,000 ms → allowProceeds anyway (fail-open)Rules still applyProceeds anyway (fail-open)A 200 without usable answers counts as a failure, not an all-clearAllow, warn, ask or block (conditional)Per-signal thresholds, e.g. destructive: block 0.85, ask 0.6, warn 0.35
jev-claude, enforce mode with fail-open switched offMode: Opt-inHolds for a person (fail-closed)Holds for a person (fail-closed)Holds for a person (fail-closed)Holds for a person (fail-closed)Allow, warn, ask or block (conditional)As above
jkudish/jev-mcp (12 tools), cells read at 0.11.0 (53fe575, 29 Sep 2026)Mode: Advisory: "The server never blocks on its own". Checked at 0.13.0 (5e0ca5c, 2 Oct 2026) and 0.14.1 (86eae78, 8 Oct 2026): failure path unchangedFlags only (advisory)The tool call returns an errorFlags only (advisory)60 s over all attempts (OpenRouter, Cloudflare, compatible endpoints). TypeSafe direct and Vercel: unknownFlags only (advisory)3 attempts on 408/409/429/5xx (fetch transports)Flags only (advisory)invalid_response with a "review" recommendation, not "pass"Flags only (advisory)pass / review / block / skip advice; jev_screen blocks at ≥ 0.75, reviews at ≥ 0.25
LangChain AutoModeMiddleware (langchain-typesafe 0.0.1a3)Mode: Enforcing for listed tools only; other tools bypass itStops (fail-closed)The middleware cannot be constructed (ValueError), so the agent is not builtStops (fail-closed)30 s; the exception ends the agent run and the tool is not executedStops (fail-closed)The run raises; tests assert the tool did not runStops (fail-closed)ExceptionTool runs below 0.5 (acts-anyway)Noul "is_risky" below 0.5 → the tool runs; at 0.5 or above → error message to the agent, which continues. Threshold fixed at 0.5
oh-my-claudecode Jev judgment points, as installed (default shadow) (audited v5.5.0 at 9fd35ec; rechecked at v5.6.0, 3 Oct, and v5.6.2 (454bae0), 8 Oct: failure path unchanged)Mode: Shadow by default: Jev never decides, so the points are advisory; the built-in heuristic decides whether or not Jev answers. A key alone sends nothing: no point runs unless OMC_JEV lists it. After 3 failures in a row a point stops calling for the processadvisoryHeuristic decides; the point is off and nothing is sentadvisoryHeuristic decides; 2,000 ms default from v5.6.0; 250 ms up to v5.5.0advisoryHeuristic decides; Jev's answer is only logged in shadow modeadvisoryHeuristic decides; the response is validatedadvisoryNo gate; Jev never decides at default settings
oh-my-claudecode, active modeMode: Opt-in from v5.6.0: OMC_JEV=<point>:active or all:active makes Jev decide (a warning is printed). On any Jev failure the heuristic decides. Checked at v5.6.2, 8 Oct 2026: unchanged; checked at 5.7.3, 16 Oct 2026: failure path unchanged (version check)Heuristic decides (fallback-heuristic)The point is off; nothing is sentHeuristic decides (fallback-heuristic)2,000 ms default from v5.6.0Heuristic decides (fallback-heuristic)Heuristic decides (fallback-heuristic)The response is validatedNo confidence gateJev's first answer is used
madisonrickert/jev-permission-gate (7349bc9)Mode: Enforcing by default (gate_mode "enforce"); acts only when Claude Code's own engine says "ask" in auto mode. Bypassed when the engine already decided or not in auto mode, for tools other than Bash, WebFetch and WebSearch, for input with secrets or too long, in measure mode and on cache hitsBuilt-in classifier decides (fallback)Claude Code's built-in classifier decides (register.ts L333–341)Built-in classifier decides (fallback)1,500 ms (policy.ts L37; register.ts L360–365, L377–381)Built-in classifier decides (fallback)Error caught (register.ts L377–381)Built-in classifier decides (fallback)hooks/typesafe.ts L54–74; register.ts L198, L377–381Built-in classifier decides (fallback)Defers unless allow (serves_request ≥ 0.5, every risk ≤ 0.1) or deny (a risk ≥ 0.8, serves_request ≤ 0.3) (policy.ts L27–42, L423–445)
mcowdery/claude-model-guard (bced870; package 1.2.0, not on a registry)Mode: Enforcing: a UserPromptSubmit hook blocks a prompt when Jev confidently says a bigger model tier is needed (≥ 0.65; ≥ 0.9 within 3 prompts of a model switch). A model-tier gate, not a safety gate. Documented, read at the pinned commit, 16 Oct 2026; bypass entry points read 18 Oct 2026. Counted on when Jev failsProceeds anyway (fail-open)Throws; exits 0 with no output (jev.mjs L124–127; promptAdvisor.mjs L155)Proceeds anyway (fail-open)2,500 ms (promptAdvisor.mjs L25, L82; jev.mjs L133)Proceeds anyway (fail-open)Error thrown and caught (jev.mjs L135–137; promptAdvisor.mjs L155)Proceeds anyway (fail-open)Missing answers throw, caught at L155 (promptAdvisor.mjs L86)Note to Claude (advisory)Below 0.65 (0.9 after a model switch) a mismatch becomes a note to Claude, not a block (promptAdvisor.mjs L146–152)
delt96/claude-jev-advisor rm hook (f2c1791 = npm @delt/claude-jev-advisor 0.5.0; Windows only)Mode: Enforcing (ask): PreToolUse on Bash|PowerShell. Its own guard asks before a deletion; Jev can only lift that ask (every target ≥ 0.8). Documented, read at the pinned commit, 16 Oct 2026; bypass entry points read 18 Oct 2026. Counted on when Jev failsHolds for a person (held)The guard's ask stands (run.ts L91–92)Holds for a person (held)8,000 ms (jev.ts L4) within a 12,000 ms budget; no call under 1,000 ms left (run.ts L29–30, L54–55)Holds for a person (held)The ask is not lifted (jev.ts L72; run.ts L66, L123)Holds for a person (held)Parse error caught; no score (jev.ts L73–79; run.ts L66–67)Holds for a person (held)The ask is lifted only when every target is ≥ 0.8 (judge.ts L75–77; run.ts L102)
Novacon/ohmyjev (fbc4ea4, 0.14.0)Mode: Enforcing (deny) in lockdown mode without shadow (both off by default): Claude Code module with Bash, Write/Edit and exfiltration gates that run only with lockdown on; with it off, gate returns before any Jev call (ohmyjev.ts L292; policy.ts L12); a new shadow setting, also off by default, runs the lockdown checks but blocks nothing (policy.ts L13, L33–39); no auto-approve path. Documented, read at the pinned commit, 8 Nov 2026 (earlier d307c39: read 2 Nov 2026; 78bb82c: read 16 Oct 2026; every entry-point file read for its Jev paths 18–19 Oct 2026, built Codex hook checked at its gate lines, its source read in full). Counted on when Jev fails from 19 Oct 2026 (bypass). Change, 21 Oct 2026: ohmyjev's pi and omp gates, rechecked at 78bb82c, use the same five failure terms as this Claude Code module; the failure checklist row now names those hosts (still 80 implementations) and build/coding-agents counts them. Change, 30 Oct 2026: ohmyjev 0.13.2 (29 Oct 2026) removes the auto-approve thresholds from its policy file (version check, R24-S38). Change, 2 Nov 2026: re-read at 0.13.2 (version check, R26-S20 to R26-S23); four failure terms unchanged in lockdown mode; bypass changed in content. Change, 8 Nov 2026: re-read at 0.14.0 (version check, R28-S20 to R28-S23); failure cells unchanged; bypass cell adds the shadow setting and the in-repo move skip.Proceeds anyway (fail-open)askJev throws and the call goes to the normal permission flow (ohmyjev.ts L82, L206–208, L302). Keys are taken from the config, $TYPESAFE_API_KEY, a key stored with ohmyjev key, then $OPENROUTER_API_KEY (jev.ts L33–39)Proceeds anyway (fail-open)1,500 ms (policy.ts L95; race at ohmyjev.ts L92–101)Proceeds anyway (fail-open)A Jev error throws, decide returns nothing (ohmyjev.ts L205–215) and the gate returns without blocking (L302, L325, L334)Proceeds anyway (fail-open)Non-JSON or a missing answer throws and the call passes (jev.ts L46–62, L70–81)Tool runs (acts-anyway)Bash is denied only when Jev rates it irreversible at ≥ 0.6, irreversible with destructive intent at ≥ 0.7, or exfiltrating at or above the exfil setting; otherwise the call goes on (policy.ts L434–448; L68–69)
gulbaki/jev-llm-guard (7c7b1b5)Mode: Returns verdict; not a decision. The jev-guard CLI exits 0 for allow, review and block; the demo server enforces nothing. Not a Claude Code hook; not countedReturns verdict; not a decision (no-decision)Throws "TYPESAFE_API_KEY is required" (guard.js L50); the CLI exits 1Returns verdict; not a decision (no-decision)30,000 ms (guard.js L51); throws, the CLI exits 1Returns verdict; not a decision (no-decision)Rethrown (guard.js L61–63); the CLI exits 1Returns verdict; not a decision (no-decision)Strict validators throw (guard.js L27–40, L67–71); the CLI exits 1Returns verdict; not a decision (no-decision)Signal below 0.4 → allow; 0.4 or above → review; 0.8 or above → block, except LLM07 (policy.js L32, L42–45). Nothing enforces it: the CLI exits 0 for all three
  • fail-closed: stops the call or run, or holds it for a person (held)
  • fail-open or acts-anyway: the tool call goes ahead
  • advisory: flags only, Jev never decides; conditional: thresholds you set decide
  • fallback: a built-in heuristic or classifier decides
  • no-decision: an error, and no verdict
  • Not applicable, not set or not recorded (the cell says which)

Colours follow the shared legend used on when Jev fails and framework integrations. Changed 6 Oct 2026: until then this page coloured "Stops" red and "Proceeds anyway" green. The cells themselves did not change, except that the AutoModeMiddleware low-confidence cell now leads with what happens below 0.5 (the tool runs), as on when Jev fails. Change, 18 Oct 2026: rows for claude-model-guard, claude-jev-advisor and ohmyjev added; the oh-my-claudecode shadow-mode cells now lead with the term advisory, with "heuristic decides" as detail (same outcome as before).

Tool calls that never reach Jev: jkudish/jev-mcp and the three added guardrails

Read at the audited commit on 14 Oct 2026. jev_extract sends no field to Jev when its regex finds nothing, and makes no call at all if no field matches; jev_review and jev_gate input checks return errors without a verdict. Same cell as on when Jev fails. Documented

claude-model-guard, claude-jev-advisor and ohmyjev: entry points from one listing at the pin read 18 Oct 2026. A bypass cell counts as recorded only when every entry point from that listing was read. That holds for claude-model-guard and claude-jev-advisor, which are counted on when Jev fails. For ohmyjev the pi and omp adapters were read in full on 19 Oct 2026, so its cell is now recorded and it is counted on when Jev fails. Documented R18-S203 to R18-S226; R19-S200 to R19-S207

Integration (commit)Code paths that never ask JevWhere (file: lines)Source
jkudish/jev-mcp 53fe575 (0.11.0; read 14 Oct 2026)recordedjev_extract sends no field to Jev when its regex finds nothing (no call at all if no field matches); jev_review and jev_gate input checks return errors without a verdictsrc/index.ts L1168, L1182–1183, L1771–1795, L1929–1975R16-S133, R16-S134
mcowdery/claude-model-guard bced870
counted on the map
recordedEmpty prompt; skip phrase (--skip-guard, [[skip-guard]]) or the CLAUDE_SKIP_MODEL_GUARD env var; continuation words, slash and ! commands; a short prompt with no earlier assistant textpromptAdvisor.mjs L31, L64–65, L70, L77R17-S245, R18-S224 to R18-S226
delt96/claude-jev-advisor f2c1791
counted on the map
recordedrm guard off; tool not Bash or PowerShell, or no command; no deletion found; deletion denied by code; rm.jev off; no key; early returns in judgeTargets (for example more than 5 targets, no transcript, under 1,000 ms left). Each one leaves the guard's own ask or deny, or no decision; none lets an asked deletion throughrun.ts L77–92; L42–55R18-S214 to R18-S223
Novacon/ohmyjev fbc4ea4 (0.14.0), in lockdown mode without shadow (both off by default)recordedClaude Code module: lockdown off, the default, skips every gate before any Jev call; the shadow setting (off by default) runs the gates and Jev's checks but only logs what they would block, so every call runs; with lockdown off or shadow on, a git push or GitHub write is denied once by a code reminder (no Jev call) and runs when repeated; off switch; read-only Bash and in-repo mkdir, touch, cp and mv skip Jev and return without approval unless policies are set (the auto-approve path of earlier versions is removed); writes outside allowed paths denied by code; each gate can be switched off; tools outside HOOKED are never seen; its own exceptions pass the call; under the pi bridge (OHMYJEV_PI=1) the module switches itself off; its own browse tool (registered only with browse on) runs without the exfiltration gate when ohmyjev or lockdown is off (as read at d307c39, 2 Nov 2026). Same shape in its pi and omp adapters (gates need lockdown; same GitHub reminder and the same shadow handling), plus tools outside their gated lists that are never sent to Jev. Read scope at fbc4ea4 (8 Nov 2026): hooks/ohmyjev.ts, hooks/jev.ts, hooks/policy.ts, hooks/hooks.json, .claude-plugin/plugin.json, adapters/pi.ts, adapters/omp.ts and adapters/shared.ts, compared with d307c39; adapters/codex.ts not read; the new bin/ohmyjev.mjs (install and key command) read only for its usage and key check, not a gate path. Earlier at d307c39 (2 Nov 2026): hooks/ohmyjev.ts (full diff against 78bb82c), hooks/jev.ts, hooks/policy.ts, hooks/hooks.json, .claude-plugin/plugin.json, and the adapters at their gate lines. Earlier at 78bb82c: 78 files listed; every entry-point file read for its Jev paths (built Codex hook checked at its gate lines; its source read in full)ohmyjev.ts L292; shadow policy.ts L33–39, L424–425, ohmyjev.ts L313–316, L854; L819–822; L818; L298 (policy.ts L370, L414); L309–320; L24; L746; adapters/pi.ts L508, L522, L538 (reminder L504; shadow L602–606); adapters/omp.ts L535, L563, L590 (reminder L531; shadow L678–682); adapters/shared.ts denyByCode L173–182. As read at d307c39: per-gate switches L276, L286, L304; own exceptions L800; browse L725, L785–790 (adapters/shared.ts runBrowse L552–563); pi off switch L499, read-only L509, writes L525; omp off switch L525, read-only L534, writes L571. Fall-through lines at 78bb82c (not re-read at d307c39 or fbc4ea4): adapters/pi.ts L546–547; adapters/omp.ts L19–29, L149–151, L596R28-S20 to R28-S23; earlier R26-S20 to R26-S24; R18-S203 to R18-S212; R19-S200 to R19-S206

Change, 19 Oct 2026: ohmyjev's pi and omp adapters read in full at the pinned commit; its bypass cell is recorded (two more skip paths in the Claude Code module), and ohmyjev is now counted on when Jev fails (n 80). jkudish/jev-mcp rows now name the commit the cells were read at (53fe575, 0.11.0) separately from the later checks. No other cell changed. Change, 18 Oct 2026: bypass rows added for claude-model-guard and claude-jev-advisor (recorded) and ohmyjev (partial read; not counted as recorded). Change, 14 Oct 2026: jkudish/jev-mcp bypass cell changed from not recorded to recorded. The (a) to (e) cells above are unchanged. We read source only; nothing was run.

Owner, version, data sent and grade

IntegrationOwner and typeSends to JevPinned version or commit (date)Grade and label
typesafe-ai/skills: plugin typesafe@typesafe-aiOfficial (TypeSafe AI). Claude Code plugin with one skill; no hook, no MCP server, no scriptNothing. The skill tells your agent to read docs.typesafe.aiv0.5.7, 65a39f3 (12 Sep)A · Documented all contents read · checked 29 Sep
danna-zhou/jev-mcpCommunity. Two Claude Code hook scripts (PreToolUse on Bash, Stop) added by hand to settings.json, plus an MCP serverPreToolUse: the full Bash command, only if it matches a "high-stakes" pattern (git push, publish, kubectl delete, terraform apply, DROP TABLE…); no redaction. Stop: the agent's full final message436469a (23 Sep); no tagsA · Documented; no-key cell also Tested offline · 29 Sep
brunopivetta88/jev-claude ("jev-guard")Community. Claude Code hook plugin (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop); Codex adapter not auditedUser goal (≤ 1,500 characters), tool name and full tool input (regex-redacted; strings ≤ 4,000 characters), last 8 action summaries, working directory. Read-only tools with a clean rule check are not sent0.1.0, 7d7e876; PR #1 merged 25 SepA · Documented · 29 Sep
jkudish/jev-mcp (@jkudish/jev-mcp)Community. MCP server with 12 judgment tools (verify, screen, noul, find, classify, decide, rerank, compare, extract, audit, review, gate); no hookOnly what the agent passes to a tool (for example jev_screen: the text and an optional purpose)Read at 0.11.0 (53fe575, 29 Sep 2026, 05:30 UTC); checked at 0.13.0 (5e0ca5c, 2 Oct 2026) and 0.14.1 (86eae78, 8 Oct 2026): failure path unchangedA for the tool layer and fetch transports; TypeSafe-direct timeout unknown (dependency not read) · Documented; README "never blocks on its own" Reported · 29 Sep
langchain-typesafe AutoModeMiddlewareLangChain (framework vendor, not TypeSafe). Python agent middleware that guards the tools you list; experimental alphaThe proposed tool call (id, name, full arguments), the tool description and the last 30 messages of the agent state, all roles, full text. No redaction0.0.1a3 (released 20 Sep): tag langchain-typesafe==0.0.1a3 = 4af7ab8; sdist sha256 d77647a7…3272A · Documented released package read · 29 Sep
oh-my-claudecode Jev judgment pointsCommunity. A Jev layer inside the plugin's own hooksNothing unless OMC_JEV lists points ("a key alone sends nothing anywhere"); then per-point text, each string cut to 200 charactersv5.5.0 = main = 9fd35ec (22 Sep); dev differsA · Documented · 29 Sep
madisonrickert/jev-permission-gateCommunity. Claude Code plugin on tool.check (hooks/hooks.json L3; hooks/register.ts L293–382); shadow and measure modes optionalLast 3 user and 3 peer messages (each cut to 1,500 characters head and tail), the Bash command (≤ 2,000 characters) and description (cut to 300), WebFetch URL and prompt, WebSearch query, project directory (policy.ts L348–414). Calls carrying a secret are never sent (L279–320)7349bc9 (3 Oct). HEAD 7349bc9 = the pin read on 4 Oct (checked 6 Oct); failure files unchangedA candidate (code read; not verified in production) · Documented · read 4 Oct, rechecked 6 Oct. README "1 unsafe allow in 3,664 risky calls" Reported
gulbaki/jev-llm-guardCommunity. Node library (evaluateGuardrail), jev-guard CLI and demo web server; not a Claude Code hook. Returns verdict; not a decision (not counted)Untrusted text (≤ 20,000 characters; longer input is rejected, not cut; context.js L2, L15), context, optional trusted task and app context; 13 questions (guard.js L18–25); up to 2 more calls for quotations7c7b1b5 (1 Oct). HEAD 7c7b1b5 = the pin read on 4 Oct (checked 6 Oct); failure files unchangedA candidate (code read; not verified in production) · Documented · read 4 Oct, rechecked 6 Oct
mcowdery/claude-model-guardCommunity. Claude Code UserPromptSubmit hook plus a model-switch recorder (install.mjs L16–20, L26–38); a model-tier gate, not a safety gateThe prompt with notifications and system reminders removed; pasted blocks cut to 500 characters; capped at 4,000 characters, head and tail. A prompt under 5 words goes with an excerpt of Claude's last message. No redaction seenpackage 1.2.0 (not on a registry), bced870 (13 Oct); HEAD = pin on 18 OctA · Documented, read at the pinned commit, 16 Oct 2026; bypass entry points read 18 Oct 2026 · R17-S242 to R17-S245, R18-S224 to R18-S226
delt96/claude-jev-advisor (@delt/claude-jev-advisor)Community. CLI that installs a PreToolUse rm guard (Windows only), a Stop and SessionEnd context advisor and a status line; only the rm guard decides on tool callsDeletion target path and kind; for a folder, the file count and up to 10 names; the command with keys redacted; up to 3 earlier tool calls naming a session filenpm 0.5.0 = f2c1791 (14 Oct); HEAD = pin on 18 OctA · Documented, read at the pinned commit, 16 Oct 2026; bypass entry points read 18 Oct 2026 · R17-S239 to R17-S241, R18-S214 to R18-S223
Novacon/ohmyjevCommunity. Claude Code module (hooks/hooks.json), plus Codex, pi and omp adapters that are not the Claude Code row, and from 0.14.0 an install and key command (bin/ohmyjev.mjs, not a gate path)Bash command and context (up to 16,000 characters), Write/Edit path and content (up to 16,000), WebFetch and MCP input (up to 4,000); secrets masked by default; read-only Bash and in-repo mkdir, touch, cp and mv not sent unless policies are set (0.14.0)fbc4ea4 = npm 0.14.0 gitHead (published 3 Nov 2026; read 8 Nov 2026; R28-S20 to R28-S23). Earlier d307c39 = npm 0.13.2 gitHead (published 29 Oct 2026; read 2 Nov 2026; R26-S20 to R26-S23). Earlier pin 78bb82c (15 Oct; read 16–21 Oct 2026; npm 0.12.0, gitHead 0af973e, not read; HEAD = pin on 18 and 19 Oct); 0.12.0 → 0.13.2 policy files compared 30 Oct 2026, 05:20 UTC (R24-S38)A · Documented, read at the pinned commit, 8 Nov 2026 (earlier d307c39: 2 Nov 2026; 78bb82c: 16 Oct 2026; every entry-point file read for its Jev paths 18–19 Oct 2026) · counted on when Jev fails from 19 Oct 2026, in lockdown mode without shadow (both off by default) · R28-S20 to R28-S23, R26-S20 to R26-S23, R17-S236 to R17-S238, R18-S203 to R18-S213, R19-S200 to R19-S206

Grades use the same scale as Products and projects: A means the pool read the code that calls Jev. It does not mean the integration works well or has users. Star counts are not adoption.

What leaves your machine

Data sent to Jev per guarded action, audited range

  1. Official plugin; oh-my-claudecode unless enabledNothing
  2. oh-my-claudecode with points enabledPer-point text, each string cut to 200 characters
  3. danna-zhou PreToolUse hookOne full shell command, only if it matches the high-stakes pattern; not redacted
  4. jev-claudeGoal, full tool input and 8 action summaries, with regex redaction of keys and tokens
  5. LangChain AutoModeMiddlewareThe last 30 messages in full, including tool outputs, plus the full tool call; not redacted
Bar lengths show order, not measured byte counts. MCP servers (jkudish, danna-zhou tools) send only what the agent passes to them. Documented 29 Sep 2026. What each hosted route then does with that data (training, retention, zero data retention) is on data handling by route.

A timeout shorter than the round trip turns the guard off

If the timeout is shorter than the time Jev takes to answer, every call fails. The integration then takes its failure path on every call, which is "proceed" or "heuristic" for most of the integrations above. oh-my-claudecode defaulted to 250 ms up to v5.5.0, below the 465–605 ms round trips its maintainers measured; v5.6.0 (1 Oct 2026) defaults to 2,000 ms.

Default timeouts against one project's measured round trips (shaded band)

  • oh-my-claudecode up to v5.5.0250 ms
  • Measured Jev round trips (#4091, 22 Sep) Reported465–605 ms
  • jev-claude enforce1,500 ms
  • oh-my-claudecode from v5.6.0 (1 Oct)2,000 ms
Scale 0–2,000 ms. Off the scale: danna-zhou hooks 8,000 ms; LangChain 30 s; jkudish 60 s over all attempts (fetch transports). The danna-zhou MCP server sets no timeout. Added 18 Oct 2026, not drawn: ohmyjev 1,500 ms (policy.ts L95 at fbc4ea4, 8 Nov 2026), claude-model-guard 2,500 ms, claude-jev-advisor 8,000 ms within a 12,000 ms budget (read at the pinned commits, 16 Oct 2026). Your own round trip depends on route and region; measure it and set OMC_JEV_TIMEOUT_MS (or the equivalent) above it. Documented timeouts from source, 29 Sep 2026; oh-my-claudecode v5.6.0 read 3 Oct 2026.

Caveats before you rely on this table

  • "Normal permission flow" means Claude Code's own rules apply. It is not an approval. Documented 29 Sep
  • LangChain AutoModeMiddleware 0.0.1a3 sends no risk criteria unless you pass criteria=NoulCriteria(...). The constructor's criteria=None default replaces the documented default criteria (auto_mode.py L66–71, L147, L164), and a unit test asserts criteria is None. Documented 29 Sep
  • danna-zhou hooks: set JEV_MODEL=jev-1.13.0 when you point them at TypeSafe. The default kev-latest is meant for a local Kev server. Whether TypeSafe rejects kev-latest was not tested. Earlier third-party reports say unknown models return 400, and this hook would treat that as "skip review". Documented defaults; the consequence is the pool's inference 29 Sep
  • oh-my-claudecode: defaulted to 250 ms up to v5.5.0, while the maintainers measured 465–605 ms round trips (Reported, #4091, 22 Sep); v5.6.0 defaults to 2,000 ms. On v5.5.0 or earlier, set OMC_JEV_TIMEOUT_MS above your round trip. From v5.6.0, OMC_JEV=<point>:active lets Jev decide, so check your environment for it. Documented 29 Sep; v5.6.0 on 3 Oct
  • jev-claude's own tests run in enforce mode; the shipped default is shadow. Passing tests do not show that the installed plugin blocks anything. Documented 29 Sep
  • No integration was run against the live Jev API, and star counts are not adoption.

Worked example: the danna-zhou hook with no key

Tested (offline, no Jev call) Node v24.20.0, 29 Sep 2026, about 07:12 UTC. The hook (86 lines) and its helper (62 lines) use only Node built-ins and were read in full before running. They were fetched at commit 436469a2 into a temporary directory, then deleted. Nothing was installed.

The pool piped four PreToolUse events into the hook with an empty environment (no key). The endpoint pointed at ports on 127.0.0.1 that were confirmed closed.

CaseCommand in the eventEndpointExitOutputSent to Jev?
Agit push origin maindefault 127.0.0.1:8009 (closed)0 (159 ms)"jev guardrail unreachable, skipping review: fetch failed"Attempted; failed → proceeds
Bgit push origin main127.0.0.1:9 (closed)0 (152 ms)Same messageAttempted; failed → proceeds
Cls -la127.0.0.1:9 (closed)0 (83 ms)NoneNo: not in the high-stakes pattern
Drm -rf / --no-preserve-root127.0.0.1:9 (closed)0 (73 ms)NoneNo: not in the high-stakes pattern

Reading. Exit 0 with no permissionDecision means Claude Code's normal permission flow applies, so the Jev layer fails open. Case D shows that the pattern list, not Jev, decides what this guard checks: a destructive command outside the list is never sent. The run does not show the behaviour against the real API with a missing or invalid key (the 401/403 path). The code sends that path to the same fail-open helper.

Runner script and raw output
# bash, pool-authored; run in a temporary directory with the two hook files
run(){ echo "--- case: $1"; shift; start=$(date +%s%N)
  out=$(echo "$PAYLOAD" | env -i PATH=$PATH "$@" node hooks/pretooluse-guardrail.mjs 2>err.txt); code=$?
  end=$(date +%s%N); echo "exit=$code ms=$(( (end-start)/1000000 ))"; echo "stdout=$out"; echo "stderr=$(cat err.txt)"; }
PAYLOAD='{"hook_event_name":"PreToolUse","tool_name":"Bash","tool_input":{"command":"git push origin main"}}'
run "A no key, no env (default base 127.0.0.1:8009, closed)"
run "B no key, JEV_BASE_URL=http://127.0.0.1:9 (closed)" JEV_BASE_URL=http://127.0.0.1:9
PAYLOAD='{"hook_event_name":"PreToolUse","tool_name":"Bash","tool_input":{"command":"ls -la"}}'
run "C non-high-stakes command, closed port" JEV_BASE_URL=http://127.0.0.1:9
PAYLOAD='{"hook_event_name":"PreToolUse","tool_name":"Bash","tool_input":{"command":"rm -rf / --no-preserve-root"}}'
run "D rm -rf / (not in high-stakes regex), closed port" JEV_BASE_URL=http://127.0.0.1:9
--- case: A no key, no env (default base 127.0.0.1:8009, closed)
exit=0 ms=159
stdout={"hookSpecificOutput":{"hookEventName":"PreToolUse","additionalContext":"jev guardrail: could not reach http://127.0.0.1:8009 (fetch failed); skipping review."},"systemMessage":"jev guardrail unreachable, skipping review: fetch failed"}
--- case: B no key, JEV_BASE_URL=http://127.0.0.1:9 (closed)
exit=0 ms=152
stdout={"hookSpecificOutput":{"hookEventName":"PreToolUse","additionalContext":"jev guardrail: could not reach http://127.0.0.1:9 (fetch failed); skipping review."},"systemMessage":"jev guardrail unreachable, skipping review: fetch failed"}
--- case: C non-high-stakes command, closed port
exit=0 ms=83
--- case: D rm -rf / (not in high-stakes regex), closed port
exit=0 ms=73

File hashes: lib.mjs sha256 2064e808…089e; pretooluse-guardrail.mjs sha256 d6890148…d079. Empty stderr lines are omitted from the output above.

Before you install: eight checks

  1. Owner. Official means github.com/typesafe-ai only. Every other server, hook or plugin is community code. (The github.com/TypeSafeAI organisation is an unofficial community organisation.)
  2. What it installs. A skill is instructions only. An MCP server is advisory, because the agent can ignore it. Only hooks can deny or ask. Look for hooks/hooks.json, .claude-plugin/plugin.json or settings.json snippets in the README.
  3. Default mode. Two of the ten audited guardrails ship in a shadow or log-only mode (jev-claude shadow; oh-my-claudecode shadow points). Installing them gates nothing until you change the mode.
  4. What leaves your machine. Find where the state is built. The audited range runs from nothing to the last 30 messages in full (see the figure above).
  5. Failure branch. Find the catch, the timeout constant and the no-key check (see the next section). Decide whether that behaviour is acceptable for the action being guarded. The general choice between failing closed, holding for review and failing open is on Errors and rate limits.
  6. Timeout against round trip. A guard whose timeout is shorter than the real round trip fails on every call. oh-my-claudecode defaulted to 250 ms up to v5.5.0 and to 2,000 ms from v5.6.0.
  7. Model ID. Check the model string the integration sends on your route: jev-1.13.0 direct, typesafe/jev-1.13 on OpenRouter, typesafe-ai/jev on Vercel (route matrix). Jev 1.14 (released as a preview on 14 Oct 2026) is jev-1.14.0 direct, typesafe/jev-1.14 on OpenRouter and typesafe-ai/jev-1.14.0 on Vercel (listed 16 Oct 2026, 05:21–05:25 UTC). An integration that sends jev-latest to TypeSafe's API has received 1.14 since 24 Oct 2026 (which model they ask for). A local-model default such as kev-latest sent to TypeSafe is a likely silent failure.
  8. Key scope. Prefer a plugin's sensitive userConfig field or an environment variable over a key in a project settings.json that may be committed.

Documented method derived from the six audits, 29 Sep 2026; it makes no claim about live behaviour

How to check an integration yourself

  1. Pin a commit. Open the repository and press y on a file view, or use the tags page, to get a permanent URL.
  2. Search the source for fetch(, systemone, evaluate( and system_one( to find the one call site.
  3. From the call site, read four things: the timeout constant (timeout, AbortController, timeoutMs), the catch block, the check for a missing key, and what the code does when the response lacks the expected field.
  4. Map each branch to one of four outcomes: stops, asks a person, proceeds, or only flags.
  5. Search the tests for "fail", "timeout", "unreachable" and "invalid". Note whether they use a fake server and which mode they set.
  6. If the hook is a small standard-library script, run it with no key and the endpoint pointed at a closed local port, as in the worked example, and read the exit code and output. Testing without a key covers mocks for your own code.

Source notes per integration

Official plugin typesafe@typesafe-ai
  • Install: claude plugin marketplace add typesafe-ai/skills, then claude plugin install typesafe@typesafe-ai. Other agents: npx skills add typesafe-ai/skills --skill typesafe-ai ("project-local by default; add -g to install globally"). Source: docs.typesafe.ai/agent-skill. Documented
  • The tree at the tag has 9 entries: marketplace and plugin manifests, README, licences and one SKILL.md (about 150 lines). The skill says "The live TypeSafe docs are the source of truth" and "Keep API credentials server-side in web apps" (SKILL.md). Any Jev call you see after installing it is made by your agent, not by the plugin. Licence MIT.
  • The skill links a v1 migration guide that returns 404. This is open issue #13, and the pool confirmed the 404 on 29 Sep. Issue #2 asks that the skill stop triggering on "TypeScript" requests. Reported
  • Whether claude plugin install installs for the user or the project is not stated in TypeSafe's docs and was not tested.
danna-zhou/jev-mcp (hooks and MCP server)
  • Fail-open helper: every exception (no key, unreachable, 8,000 ms abort, non-2xx, invalid JSON, missing fields) prints a warning and exits 0 (lib.mjs L46–61; pretooluse-guardrail.mjs L24–85).
  • Mapping: Choice risk (safe / moderate / dangerous) and Noul needs_human. "dangerous" with confidence ≥ 0.6 → deny.
  • Stop hook: SHOULD_BLOCK = false. A local log ~/.claude/jev-quality-log.jsonl stores scores, cwd, session ID and a 200-character preview (stop-guardrail.mjs).
  • Defaults: JEV_BASE_URL=http://127.0.0.1:8009, JEV_API_KEY="local", JEV_MODEL="kev-latest" (index.ts L16–18). No retries, no tests. package.json says MIT, but there is no LICENSE file.
brunopivetta88/jev-claude ("jev-guard")
  • Defaults: mode: 'shadow', timeoutMs: 1500, failOpen: true, redaction on (config.mjs). Shadow mode turns every decision into allow (policy.mjs L81–105).
  • Missing key, timeout, non-2xx or "no usable answers" all count as failure (jev-client.mjs). An 8 s watchdog exits with "allowing" (entry.mjs).
  • Payload built in run.mjs L131–139. Redaction masks private keys, cloud and chat tokens, sk-… keys, JWTs and *SECRET/TOKEN/PASSWORD/API_KEY= values. Results are cached for 60 s under .jev-guard/ in the project directory.
  • A deterministic rule floor runs first, with or without Jev; a test asserts rm -rf / is blocked with no key, in enforce mode. The manifests name a different owner (bruno-ship-it); this was not investigated.
jkudish/jev-mcp
  • Malformed or missing answers become status: "invalid_response" with recommendation: review (index.ts L296–372). Error texts redact the configured key.
  • Transports: OpenRouter, Cloudflare and compatible endpoints use one 60 s deadline, up to 3 attempts, retrying 408/409/429/5xx with 0.5–4 s jittered backoff (provider.ts L28–49). TypeSafe direct and Vercel go through @jkudish/jev-agent-tools, which was not read: timeout and retries unknown.
  • "The server never blocks on its own; enforcement stays with the calling agent" (README L242). Reported
  • Node ≥ 22; depends on @typesafe-ai/sdk ^0.6.0. Extensive mocked tests exist; not run by the pool.
LangChain AutoModeMiddleware
  • Class langchain_typesafe.experimental.middleware.AutoModeMiddleware, installed with the [experimental] extra. It was read from the PyPI source distribution, not installed. The release tag langchain-typesafe==0.0.1a3 is commit 4af7ab8, read for framework integrations on 5 Oct 2026.
  • "Classification failures propagate and the tool handler is not called, so failures are fail-closed." "This middleware blocks risky calls; it does not request human approval." (auto_mode.py L94–98)
  • Timeout 30 s; no retry code (classifier.py L36–38). This matches the single-attempt row on Errors and rate limits. With no key, TypeSafeClassifier() raises "TypeSafe API key is required" (classifier.py L255–265).
  • Payload: messages[-30:], the tool call id, name and full arguments, and the tool description (auto_mode.py L170–182). The trace policy hides this payload from LangChain traces only.
oh-my-claudecode Jev judgment points
  • Configuration: OMC_JEV=off or a list of points; OMC_JEV_TIMEOUT_MS default 250 (2,000 from v5.6.0); OMC_JEV_EXCERPT_CHARS default 200; ACTIVATED_POINTS is empty, so enabled points run in shadow (config.ts L1–31). Model jev-latest is hard-coded.
  • On timeout, HTTP error or invalid response the heuristic "twin" decides and the call is logged as degraded (resolver.ts). The project's decision record rejects fail-closed: "blocking the orchestration on an external judgment API would make OMC hostage to one dependency" (ADR 03665).
  • Timeout history: PR #4092 raised it to 2,000 ms and was merged on 22 Sep, 22:21 UTC, into the dev branch. On 29 Sep, v5.5.0 and main still set 250 ms (L27); only dev set 2,000 ms (L36). The change was released in v5.6.0 on 1 Oct 2026 (config.ts L36 at e74b22c, read 3 Oct). v5.6.0 also adds environment activation (OMC_JEV=<point>:active, commit 3681dcb3, #4093) and fixes the request shape (lowercase types; Score criteria sent as a list). Documented
  • No retries ("the interactive path degrades instead"). Tests use injected fetch stubs; not run by the pool. Which hook events each point runs in (points.ts) was not read.
madisonrickert/jev-permission-gate
  • Pin 7349bc9 (3 Oct). HEAD checked by git ls-remote on 6 Oct 2026, 05:20:19 UTC (R9-S70): HEAD 7349bc9 = the pin read on 4 Oct; failure files unchanged. Cells from the read of the source, 4 Oct 2026 (R7-S85). Documented
  • Default gate_mode "enforce" (register.ts L88, L222). Timeout 1,500 ms (policy.ts L37). No key (register.ts L333–341), timeout (L360–365) and errors (L377–381) fall back to Claude Code's built-in classifier. Malformed answers: typesafe.ts L54–74; register.ts L198. Thresholds and defer: policy.ts L27–42, L423–445.
  • Bypass: engine already decided or not auto mode (register.ts L305–316); tools other than Bash, WebFetch and WebSearch; secrets, too-long or empty input (policy.ts L28, L348–383); measure mode; cache hits reuse a stored verdict. Opt-in logs in enforce mode; a write failure never affects the decision (register.ts L47–70). 4 test files; not read.
gulbaki/jev-llm-guard
  • Pin 7c7b1b5 (1 Oct; Show HN 49919953, 1 Oct). HEAD checked by git ls-remote on 6 Oct 2026, 05:20:19 UTC (R9-S70): HEAD 7c7b1b5 = the pin read on 4 Oct; failure files unchanged. Cells from the read of the source, 4 Oct 2026 (R7-S84). Documented
  • Returns verdict; not a decision: the jev-guard CLI prints Decision: <ACTION> and exits 0 for allow, review and block (cli.js L31–41); it exits 1 only when evaluation throws (L41) or input is over 20,000 characters (L28). The demo server returns the analysis as JSON and enforces nothing (web-server.js L76–100). A CI step that runs jev-guard fails only on an error, not on a "block" verdict.
  • No key, 30,000 ms timeout, HTTP errors and malformed answers all throw (guard.js L50–71). Thresholds 0.4 and 0.8 are set in code (policy.js L32, L42–45). No audit log; tests stubbed, not run.
claude-model-guard, claude-jev-advisor and ohmyjev (added 18 Oct 2026)
  • claude-model-guard. Blocks only when the session's model is known; downgrades only nudge unless CLAUDE_GUARD_BLOCK_DOWNGRADES=1 (promptAdvisor.mjs L89). Thresholds in decision.mjs L6–10. It may also fit model routing, because it gates on model tier rather than safety. Documented R17-S242 to R17-S244
  • claude-jev-advisor. The install adds the rm hook on win32 only, with a 15 s hook limit (features.ts L6–19). When Jev lifts an ask, the hook prints a systemMessage with no permission decision, so Claude Code's normal permission flow applies (src/rm/run.ts L122). A crash outside the Jev step ends with no output (hook.ts L13–14). Documented R17-S240, R17-S241, R18-S216
  • ohmyjev. ohmyjev 0.14.0 (read 8 Nov 2026) keeps its gates off by default and adds a shadow setting that logs what lockdown would block without blocking it. Documented R28-S20 to R28-S23 In lockdown mode, with user policies set, a violates_policy answer at or above the bashDestructive threshold denies first; a deny is dropped when Jev says the user asked for it (policy.ts L379–389 at d307c39; L318–327 at 78bb82c). After a Jev error only the status text changes; every later call still tries Jev. The Codex, pi and omp adapters have the same skip paths in shape; they are not the Claude Code row. Documented R17-S238, R18-S205, R18-S207, R18-S210, R18-S211, R26-S23
  • None of the three was installed or run; no Jev call was made.

Seen, not audited

These are one-line self-descriptions by their authors, recorded on 29 Sep 2026 (Reported). The pool did not read their failure paths, so this page makes no fail-open or fail-closed claim about them.

Routing and context-compaction plugins (Jevgrep, fast-jev-compaction) are not covered here because they do not guard tool calls. jev-effort-router, a Hermes Agent model router, is audited on coding agents: on a Jev error the turn keeps the configured model, and below its 0.5 threshold the code rewrites the request to deepseek-v4.1-flash at medium effort (read at 562dc2a, 10 Oct 2026; R12-S202). Its routing is on in the plugin's own config, but the plugin is off until enabled in Hermes; with Hermes defaults (model "") it is inert, because it acts only when the provider is Ollama:Cloud (R13-S217; added 11 Oct 2026).

Claude Code and MCP leads seen 6–16 Oct 2026: 32 projects, classified, not counted

32 Claude Code and MCP projects seen in scans from 6 to 16 Oct 2026, classified from their README and manifest (not counted): 16 hooks or plugins that allow, block, hold or reroute; 12 that return Jev's answer to the agent; 1 skill only; 3 not Jev. 3 of the 16 were read at a pinned commit (Novacon/ohmyjev, delt96/claude-jev-advisor, mcowdery/claude-model-guard); since 18 Oct 2026 they are in the failure matrix; all three are counted on the failure map (ohmyjev from 19 Oct 2026). The other 13 of the 16 are leads (README and file listing only). Reported READMEs, manifests and file listings at the pins shown, 16 Oct 2026, 05:30–05:33 UTC; R17-S217 to R17-S261

Leads by class: earlier lists (6–10 Oct and 14 Oct scans) and this update's scan (14 Oct, 16:45 UTC to 16 Oct, 05:19 UTC)
ClassEarlier listsThis update's scanAll
Hook or plugin that allows, blocks, holds or reroutes (counts once read at a pin)9716
Returns to the agent; not a decision7512
Skill or instructions only101
Not Jev213
Not readable000
Total191332

Of the 16 in the first class: in the earlier lists, 3 allow, block or hold (ohmyjev, claude-jev-advisor, claude-model-guard), 1 has gates that are off by default (jev-mod) and 5 are model or skill routers (9); in this update's scan, 1 is a PostToolUse guard (u007), 1 is a router with unread safety hooks (guigetelina) and 5 are routers (7). Routers pick the model; they are compared on model routing. Pins are the repository HEAD on the date shown. "yes" in the coding-agents column links to the lead triage on coding agents, where the lead is already named.

32 leads: 1–19 from the earlier lists, 20–32 from this update's scan
#LeadPin (HEAD, date)ClassReason (README and manifest)On coding agentsSources
1PeterFujiyu/jev-status-marketplacedd98512, 7 Octreturns to the agent; not a decisionPlugin that shows Jev's done/ship verdict in a status band after each turn. Nothing is allowed or blockedyesR17-S217
2apatlasdev/fast-decide-compaction88e4d85, 7 Octnot JevCompaction uses Mercury Decide, not JevyesR17-S218
3@delt/claude-jev-advisor 0.5.0 (delt96/claude-jev-advisor)f2c1791, 14 Oct (= npm 0.5.0 gitHead)hook or plugin that allows, blocks, holds or reroutesPreToolUse rm guard (Windows only) asks before real deletes; Jev can only lift the ask. Read at a pin; in the failure matrix since 18 Oct 2026yesR17-S224, R17-S239 to R17-S241
4mcowdery/claude-model-guardbced870, 13 Oct (earlier pin 8b6ade2)hook or plugin that allows, blocks, holds or reroutesUserPromptSubmit hook blocks a prompt when Jev confidently wants a bigger model tier. Read at a pin; in the failure matrix since 18 Oct 2026yesR17-S219, R17-S242 to R17-S245
5VictorGambarini/jev-mod3c24002, 12 Oct (earlier pin c39353a)hook or plugin that allows, blocks, holds or reroutesREADME: model routing, a tool-call gate (puts calls to you) and a stop gate, all off by default; screening of fetched text is on. Not read at the pinyesR17-S220
6DiegoSalazar/ask-jev-skille2e313d, 8 Octreturns to the agent; not a decisionSkill plus a jev CLI; Claude reads the answersyesR17-S221
7kanedayo/clef-mcp91679b6, 8 Octnot JevMCP server for Cloudflare Clef on Ollama, not JevyesR17-S222
8Sad1mus/agentic-portfolioa9cfa37, 7 Octreturns to the agent; not a decisionDemo MCP server and skill; its benchmarks are run on mocked fallbacks (README)yesR17-S223
9@muthuishere/jevx 0.12.13npm 0.12.13, 7 Oct (gitHead 4ad05c6); repo HEAD ff51603, 12 Octreturns to the agent; not a decisionSkill plus CLI. Its guard hooks (for example a bash guard) run only after jevx plugin enable: "no hook runs until you enable a plugin" (Reported)yesR17-S225
10Novacon/ohmyjevfbc4ea4 (npm ohmyjev 0.14.0, 3 Nov; read 8 Nov 2026); earlier d307c39 (npm 0.13.2, 29 Oct; read 2 Nov 2026); earlier 78bb82c, 15 Oct (npm 0.12.0, 12 Oct)hook or plugin that allows, blocks, holds or reroutesClaude Code module hook with Bash, Write and exfiltration gates in lockdown mode without shadow (both off by default). Read at a pin; in the failure matrix since 18 Oct 2026 and counted on the failure map from 19 Oct 2026yesR17-S226, R17-S236 to R17-S238, R26-S20, R28-S20
11Oqura-ai/optim-jev212b702, 10 Octhook or plugin that allows, blocks, holds or reroutesPlugin with compaction (on), plus model routing and skill choice (off, per the coding-agents lead row). It reroutes the model; it does not allow or block. A model-routing leadyesR17-S227
12Akurganow/model-router427530b, 15 Oct (v0.11.0)hook or plugin that allows, blocks, holds or reroutesSets model and effort on each Agent tool call. Reroutes; a model-routing leadyesR17-S228
13dominicrico/jev-router39712e3, 12 Octhook or plugin that allows, blocks, holds or reroutesPicks the model and effort per prompt, step and subagent; the lean preset is on by default (README). A model-routing leadyesR17-S229
14cloud-f1/jev-agent-kit242c56a, 12 Octreturns to the agent; not a decisionShortens long Bash output before the model reads it; the full output is kept; observe mode by default. No allow or blockyesR17-S230
15UditAkhourii/quicksilver (copy seen: vyupfqkjey/quicksilver 4d221ac)5d6fe5c, 25 Sepreturns to the agent; not a decisionSkill: Claude hands bulk judgments to Jev and gets a shortlist back. "86% fewer Claude tokens" is the author's claim (Reported)noR17-S231
16GorOhanyan/jev-offload-mcp (now named decisions-mcp)4c74a44, 14 Octreturns to the agent; not a decisionMCP server for OpenRouter decision models, with Jev as the default; results go back to the agentnoR17-S232
17DarkisuSu/Jev-Claude-Router8eacae8, 15 Oct (v0.5.0)hook or plugin that allows, blocks, holds or reroutesOn each prompt, asks Jev which skills fit and which model and effort to use. Reroutes; a model-routing leadnoR17-S233
18sutan13/claude-jev-routerca76882, 14 Octhook or plugin that allows, blocks, holds or reroutesRoutes each prompt to the cheapest model and effort (README in Indonesian). A model-routing leadnoR17-S234
19michelle8961/building-with-typesafe-jev692cd71, 8 Octskill or instructions onlyAn unofficial skill that teaches agents to build with Jev, with links to community projectsno (a copy, Nancy97940/…, is a coding-agents lead)R17-S235
20u007/claude-plugins (plugin jev-guard)0b8e9fc, 15 Octhook or plugin that allows, blocks, holds or reroutesREADME: a PostToolUse guardrail that scans inbound web, MCP and network results with Jev. It acts after the tool has run. Not read at the pinnoR17-S249
21ricky94511/jev0e1a6b7, 16 Octhook or plugin that allows, blocks, holds or reroutesAsks Jev on each prompt and switches the model automatically (haiku at confidence ≥ 0.7, per README). A model-routing leadnoR17-S250
22chrisDeFouRire/jevrulesb0d046b, 16 Octreturns to the agent; not a decisionChecks code against Cursor rule files from Claude Code, Codex, a git hook or CI. A report, not a gate in Claude Code (README)noR17-S251
23trungnt13/claude-system1bf9a074, 15 Octhook or plugin that allows, blocks, holds or reroutesChooses reasoning effort before model requests; on by default; Jev or another decisions API. A model-routing leadnoR17-S252
24davidmclaughlinemail/jev-routerae846b9, 15 Octhook or plugin that allows, blocks, holds or reroutesRoutes every turn to Haiku, Opus or Fable; "Routing is enforced, not suggested" (README); asks you below a threshold. A model-routing leadnoR17-S253
25guigetelina/claude-jev-codex92c1ba2, 14 Octhook or plugin that allows, blocks, holds or reroutesThe Jev router picks the model per message. Its safety hooks (confirm before credentials, block unreviewed push) were not read, so whether they use Jev is unknownnoR17-S254
26audichuang/jev-skill-suggestion23eda7c, 15 Octreturns to the agent; not a decisionJev picks at most one skill and attaches its SKILL.md to the prompt. Nothing is allowed or blockednoR17-S255
27flupkede/fast-compaction70450cd, 15 Octreturns to the agent; not a decisionCompaction: Jev scores old tool calls for removalnoR17-S256
28ThingsLikeClaude/ani-steer-publicfdbf50d, 15 Octreturns to the agent; not a decisionCorrection memory: Jev decides whether a user message is a correction, and matching notes are attached to later requestsnoR17-S257
29nim-f/jev-pilot90cc1f1, 14 Octhook or plugin that allows, blocks, holds or reroutesDemo for a video: Jev runs single read-only steps itself and routes effort. A model-routing leadyes (named)R17-S258
30kevinCodeSavvy/cockpit8743bdb, 15 Octhook or plugin that allows, blocks, holds or reroutesStatus panel plus a "Jev model router" plugin. A model-routing leadnoR17-S259
31thiago-lage/jev-claud30603ca, 16 Octreturns to the agent; not a decisionCuts noise (passing test output, large data reads) before Claude reads it; also on PyPI as jev-claudnoR17-S260
32mdimamhosen/jevea71fe2, 14 Octnot JevIts own model router and endpoint named "JEV" (jev-auto, jevrouter on npm). The README does not describe TypeSafe's JevnoR17-S261

Not triaged: two later copies with the same description, DianeUchida2774/quicksilver (created 15 Oct; the earliest copy found is UditAkhourii/quicksilver, 25 Sep) and Cook58010/building-with-typesafe-jev (created 15 Oct; a likely copy of lead 19); a guide on flaviocopes.com (an article, not code); npm jev-guard 0.3.1 (leepokai/jev-guard, 17 Sep; R17-S216). Change, 16 Oct 2026: this table replaces the "Leads seen, not yet read (6–10 Oct scans)" line, which named 14 leads without a class.

What was not verified

  • No integration was run against the real Jev API. The offline run covered unreachable endpoints only; the 401/403 no-key path of the danna-zhou hook is read from code.
  • jkudish/jev-mcp's timeout and retries for TypeSafe direct and Vercel live in @jkudish/jev-agent-tools (0.2.0 from jev-mcp 0.14.1). Only its README was read on 8 Oct 2026; its code was not audited.
  • The danna-zhou MCP server sets no timeout; what happens on a hanging endpoint depends on Node and the MCP client and was not tested.
  • jev-claude's deterministic rule list and its Codex adapter were not read line by line; oh-my-claudecode's per-point hook events were not read, and its npm package name was not verified.
  • Whether TypeSafe accepts model: "kev-latest" is unknown.
  • 16 Oct 2026: 13 of the 16 hook or plugin leads were classified from their README and file listing only (including jev-mod, u007's jev-guard and guigetelina's safety hooks); jev-mcp 0.14.4 was not published on npm at the check and was not read; the model field of jev-claude, LangChain and jev-permission-gate was not read. Nothing was installed or run.
  • 18 Oct 2026: ohmyjev's pi and omp adapters were read at their tool-call gates only; their other handlers (output screen, done-check, routing, compaction) were listed, not read line by line. claude-jev-advisor's CLI, status line and context run files were checked by a scripted scan, not line by line; none is a tool-call path. The npm 0.5.0 tarball of @delt/claude-jev-advisor was not compared with the pin on 18 Oct. Nothing was installed or run.

Search published pools, pages, reports, and evidence.