ai 7.0.165 noted · @ai-sdk/typesafe-ai 3.0.35 and ai 7.0.165 and 7.0.166 compared with the previous versions, registries rechecked Jev LangChain and other framework integrations (TypeSafe AI)
When the pool checked search suggestions on , Google, Bing and DuckDuckGo all offered "jev langchain". "jev pydantic", "jev spring ai" and "jev llamaindex" had no suggestions on any of the three. A suggestion shows that people search for a phrase, not how many do. This page reads seven framework packages for Jev in source: LangChain for Python and for JavaScript, Pydantic AI, Effect, TanStack AI, Spring AI and LlamaIndex. It shows who publishes each one, how young it is, and what it does to your code when the Jev call fails.
Of 7 Jev framework packages read in source, 4 are pre-1.0 and none comes from TypeSafe; 5 raise Jev errors to your code, while the community Spring AI filter passes passages through unscreened and the community LlamaIndex reranker falls back to the retriever's order.
Picked from framework packages builders asked about ("jev langchain") and others chosen for this page; not a survey. TanStack AI was added on 31 Oct 2026, after its 1.0.0 release; on 5 Oct it was listed by name only. The pool installed none of them, ran nothing and made no Jev call. The 5 that raise errors are the LangChain Python and JavaScript classifiers, Pydantic AI, Effect and TanStack AI. The 7 are the packages read, not every framework integration that exists.
Versions checked, 9 Nov 2026 (@ai-sdk/typesafe-ai, ai), 8 Nov 2026 (langchain-typesafe, pydantic-ai-slim) and 31 Oct 2026 (the others); registry rechecked 9 Nov 2026
Read at langchain-typesafe 0.0.2 (wheel, 8 Nov 2026): failure path unchanged; adds an optional classifier= argument on both experimental middlewares. langchain-typesafe 0.0.2 (5 Nov 2026, 22:56 UTC) compared with the 0.0.1a3 read: the classifier and client files are byte-identical, and the handling of errors, timeouts, malformed answers, a missing key and the 0.5 threshold is unchanged. AutoModeMiddleware and ModelRouterMiddleware now accept a TypeSafeClassifier you build yourself (for example one with a custom base_url); without it they build the default one, as before (auto_mode.py L151, L172; model_router.py L144, L156). The public methods are the same set at both versions, so the bypass cells are unchanged. No cell changed. Line numbers cited below move at 0.0.2: auto_mode.py L147 → L150, L164 → L169, L167 → L172, L232 → L237; model_router.py L152 → L156, L197 → L201. Documented (PyPI wheels 0.0.1a3 and 0.0.2 read as files, 8 Nov 2026, 05:19 UTC, nothing installed; the 0.0.2 release page links commit 71881c7, whose source tree was not read; R28-S211 to R28-S214, spot-checked R28-S50)
pydantic-ai-slim 2.66.0 (4 Nov 2026): TypeSafe files identical to 2.64.0 by file hash, 8 Nov 2026. models/typesafe.py and providers/typesafe.py have the same sha256 at both versions, so the 2.64.0 description below still applies. Documented (wheels read as files, 8 Nov 2026, 05:20 UTC; R28-S215)
Compared 9 Nov 2026: @ai-sdk/typesafe-ai 3.0.35 (7 Nov 2026), compared with 3.0.34 on 9 Nov 2026: it now rejects empty Choice or Score criteria and empty Score levels before sending, sends primitive JSON state as text, and returns Score legends in provider metadata; the image and error handling lines are not in that diff (R29-S18). ai 7.0.165 (7 Nov) and 7.0.166 (9 Nov 2026): the decide functions are unchanged from 7.0.164 (compared 9 Nov 2026); the changes are telemetry and documentation (R29-S19). Neither changes a cell on this page. No release in the window for @effect/ai-typesafe (4.1.2), @langchain/typesafe (0.0.2), langchain-typesafe (0.0.2), pydantic-ai-slim (2.66.0), llama-index-postprocessor-jev (0.1.1), @tanstack/ai-typesafe (1.0.0) or LiteLLM (1.108.3) (registries checked 9 Nov 2026). The LiteLLM cells stay as read at 1.106.0 to 1.106.2. Documented (npm and PyPI registry records and package tarballs, 9 Nov 2026, 11:10–11:12 UTC; R29-S18 to R29-S21)
Checked at pydantic-ai-slim 2.64.0 (31 Oct 2026): failure path unchanged, still raises to your code. New in 2.64.0: an error status that a retrying HTTP transport gives up on now raises ModelHTTPError (models/typesafe.py L252–254). Documented (2.64.0 released 31 Oct 2026, 00:22 UTC; tag v2.64.0 = cde1d10; models/typesafe.py and providers/typesafe.py at 2.60.0 to 2.64.0 read as files, 31 Oct 2026, 05:14–05:15 UTC, nothing installed; the provider file is byte-identical, and the model file changed only at 2.64.0; R25-S36, R25-S38, R25-S39)
Checked at @effect/ai-typesafe 4.1.1 (31 Oct 2026): failure path unchanged, still raises AiError; no retry or timeout added. New: image and Locate (Box) decisions, the jev-1.14.0 model ID; a malformed response is now InvalidResponseError (src/internal/errors.ts L42; it was InvalidOutputError at 4.0.1). The no-key cells are unchanged. Documented (4.1.1 released 30 Oct 2026, 21:30 UTC; the 4.0.1 and 4.1.1 package sources compared as files, pinned by npm version, 31 Oct 2026, 05:14–05:15 UTC, nothing installed; R25-S36, R25-S40) Newer release noted, not re-read: 4.1.2 released 2 Nov 2026, 04:03 UTC (npm, checked 2 Nov 2026; R26-S36); this block describes 4.1.1.
TanStack AI @tanstack/ai-typesafe 1.0.0 (TanStack; commit 5651e07): raises to your code on an API error, a malformed answer or a missing key; sets no timeout of its own (a 30 s default is documented but not used). Not a decision. Now a row in the maturity matrix. Documented (read at the pinned commit, 31 Oct 2026, 05:13–05:18 UTC; the error path read twice; nothing installed or run; R25-S45, R25-S227 to R25-S237)
Newer release noted, not re-read: @ai-sdk/typesafe-ai 3.0.31 (30 Oct 2026, 15:39 UTC; patch; newer version exists, not re-read; see TypeScript). Unchanged on 31 Oct 2026: langchain-typesafe 0.0.1a3, @langchain/typesafe 0.0.2, llama-index-postprocessor-jev 0.1.1. typesafe-spring-ai (Maven) was not rechecked. The other cells stay as read at the pinned versions. Documented (npm and PyPI registry records, 31 Oct 2026, 05:14 UTC; R25-S36) Rechecked 2 Nov 2026, 10:55 UTC: @effect/ai-typesafe 4.1.2 released 2 Nov 2026, 04:03 UTC, not re-read; pydantic-ai-slim still 2.64.0, @ai-sdk/typesafe-ai still 3.0.31, @tanstack/ai-typesafe still 1.0.0, and the three packages above unchanged (npm and PyPI registry records; R26-S36).
Change, 9 Nov 2026: @ai-sdk/typesafe-ai 3.0.35 compared with 3.0.34 and ai 7.0.165 and 7.0.166 with 7.0.164; registries rechecked; the 8 Nov registry note moved to the earlier checks. No cell or count changed.
Change, 8 Nov 2026: langchain-typesafe read at 0.0.2 (failure path unchanged; optional classifier= argument on both experimental middlewares; the two LangChain Python rows now name 0.0.2); pydantic-ai-slim 2.66.0 TypeSafe files identical to 2.64.0; @ai-sdk/typesafe-ai 3.0.35, ai 7.0.165 and LiteLLM 1.108.3 noted; the 7 Nov registry note moved to the earlier checks. No cell or count changed.
Change, 7 Nov 2026: newer releases of langchain-typesafe (0.0.2) and pydantic-ai-slim (2.66.0) noted (both read on 8 Nov 2026), and @ai-sdk/typesafe-ai (3.0.34) noted, not re-read. No cell or count changed.
Change, 31 Oct 2026: TanStack AI row added (raises to your code; not a decision); pydantic-ai-slim checked at 2.64.0 and @effect/ai-typesafe at 4.1.1, failure paths unchanged; the counts now cover 7 packages (they covered 6): 4 pre-1.0, 0 published by TypeSafe (framework vendor 5, community 2), 5 raise Jev errors to your code (was 4), 2 decide by themselves. Earlier version notes moved into one list below, newest first. No failure term changed.
Which integration is mature enough: maturity matrix
One row per package, plus the agent middlewares that ship inside the two LangChain packages. The "On a Jev error" and "No key" cells use the pool's shared failure vocabulary: fail-closed (stops or refuses), fail-open (acts anyway), fallback-<what> (a stated substitute), no-decision (nothing returned; your code must handle it). "Raises" means the error reaches the code you write. Documented source at the pins below, read 5 Oct 2026, 05:12–05:27 UTC; 26 key lines spot-checked a second time (all matched); the ModelRouter, JS autoMode and Spring filter decision paths read twice; TanStack AI read at its pin on 31 Oct 2026, 05:13–05:18 UTC, its error path (evaluate.ts L88–121) read twice (all matched)
| Integration (pin) | Publisher | Package and version | Released | Pre-1.0? | On a Jev error | No key | Confidence to your code? | Does the framework code decide? |
|---|---|---|---|---|---|---|---|---|
LangChain TypeSafeClassifier (Python)Read at 0.0.2 (PyPI wheel, 8 Nov 2026; what changed); first read at tag langchain-typesafe==0.0.1a3 = 4af7ab8 | Framework vendor (langchain-ai), not TypeSafe; MIT | langchain-typesafe 0.0.2 (marked @beta(); classifier and client files unchanged from 0.0.1a3, 3 releases when read on 5 Oct) | 0.0.2: . 0.0.1a3: Correction: not 4 Oct. Still the latest version on 4 Oct | Yes | no-decision: raisesTyped error to your code; no retry in the integration; 30 s timeout | fail-closedError at construction (env TYPESAFE_API_KEY) | Yes, on Choice and Score answers | No: your code decides |
LangChain AutoModeMiddleware and ModelRouterMiddleware (Python)Same package and pin; module experimental/middleware/, [experimental] extra | Framework vendor (langchain-ai); MIT | Inside langchain-typesafe 0.0.2, experimental; since 0.0.2 both take an optional classifier= argument Correction: AutoModeMiddleware is not a separate package | 5 Nov 2026 (same package; 0.0.1a3 on 20 Sep 2026) | Yes | fail-closedThe error propagates and the tool handler is not called (AutoMode); the run ends (ModelRouter) | fail-closedError when the middleware is constructed (at 0.0.2, if you pass your own classifier=, the error comes when you build that classifier) | AutoMode: no (it appears only in the block message). ModelRouter: uses the chosen model, no cut-off | Yes. AutoMode blocks risky tool calls; ModelRouter picks the model |
LangChain JS classifierTag @langchain/typesafe@0.0.2 = 417ddcf | Framework vendor (LangChain org npm accounts, published by GitHub Actions); MIT | @langchain/typesafe 0.0.2 (2 releases) | Yes | no-decision: raises2 retries on 429, 5xx, 408, connection errors and timeouts, then raises; 30 s per attempt | fail-closedTypeSafeError at construction | Yes, on answers | No: your code decides | |
LangChain JS autoModeMiddlewareSame package and pin; src/middleware/autoMode.ts | Framework vendor; MIT | Inside @langchain/typesafe 0.0.2 | 1 Oct 2026 (same package) | Yes | fail-closedThe error is not caught, so the tool does not run (30 s × 3 attempts) | fail-closedThe classifier is built when the middleware is created | Not returned; risk is cut at a fixed 0.5 | Yes. Blocks listed tools at risk ≥ 0.5 |
Pydantic AI TypeSafeModelTag v2.54.0 = 6695132 | Framework vendor (pydantic/pydantic-ai); licence not recorded | In pydantic-ai-slim 2.54.0 (framework version; the TypeSafe model is newer than the framework) | No (ships in a 2.x framework) | no-decision: raisesModelHTTPError, ModelAPIError or UserError (read at 2.54.0); connection failures raise ModelConnectionError since 2.59.0 (whether it subclasses ModelAPIError was not checked); since 2.64.0 an error status that a retrying HTTP transport gives up on also raises ModelHTTPError (L252–254, checked 31 Oct 2026); no retry in the integration; timeout from model_settings, else the SDK default (not read) | fail-closedMissing-key error at provider construction | Yes, on Choice and Score; booleans thresholded at 0.5 | No: it returns typed answers to the agent you write | |
Effect @effect/ai-typesafeTag @effect/ai-typesafe@4.0.1 = 460272d | Framework vendor (Effect-TS; npm maintainers include effect-bot); MIT | @effect/ai-typesafe 4.0.1 (7 versions, including 0.0.x and rc), marked @stability unstable Checked at 4.1.1 on 31 Oct 2026: failure path unchanged (what is new); 4.1.2 released 2 Nov 2026, not re-read | Correction: 4.0.1, not 4.0.0; this page audits 4.0.1 | No (4.0.1), but marked unstable | no-decision: raisesAiError; no retry and no timeout set (uses your HttpClient) | conditionallayerConfig(): ConfigError at build. layer(): no bearer header, so the call fails with 401 and raises | Yes, on Classify and Rate | No: your code decides |
TanStack AI typesafeDecider and createTypesafeDeciderTag @tanstack/ai-typesafe@1.0.0 = 5651e07 (tagged 22 Oct 2026, 15:17 UTC) | Framework vendor (TanStack), not TypeSafe; licence not recorded | @tanstack/ai-typesafe 1.0.0 (the npm src/ files are byte-identical to the tag's) Listed by name only at 0.1.7 on 5 Oct 2026 | (npm) | No (1.0.0) | no-decision: raisesA non-OK response raises an Error with the status (src/adapters/evaluate.ts L100–115); a network failure is rethrown (L92–98); a malformed answer raises "TypeSafe evaluate response had an unexpected shape" (L117–120). No retry in the package. Timeout: none set by the package; a 30 s default is documented in src/utils/client.ts (L30–31) but not used; only your abortSignal is passed on (L90) | conditionaltypesafeDecider(): throws at creation when TYPESAFE_API_KEY is missing (evaluate.ts L176–181; client.ts L60–62). createTypesafeDecider(model, apiKey) does not check the key, so the request fails and raises (L100–115) | Not recorded | No: your code decides. No bypass or default answer found; TanStack AI's core decide() rethrows (packages/ai/src/activities/evaluate/index.ts L556–572, L608–631) |
Spring AI TypeSafe JevDocumentFilter (RAG passage filter)Tag v0.4.0 = 51993bbCompared with other rerankers and passage filters | Community: spring-ai-community organisation, not spring-projects; Apache-2.0 Correction: not part of spring-projects/spring-ai | org.springaicommunity:typesafe-spring-ai 0.4.0 (4 releases) | (Maven Central metadata; the 2 Oct Spring blog cites 0.3.0) | Yes | fail-openPassages that could not be screened are passed through, with a WARN log. SDK: 10 s per attempt, 2 retries, 30 s budget | fail-closedFor a blank key. Blank value: exception at startup. Property absent: no client bean is created | No; the classification is stored in passage metadata | Yes. It drops passages before they reach the prompt |
LlamaIndex JevRerank (community)Repository 421afda (HEAD, 25 Sep)Compared with other rerankers and passage filters | Community (personal repository); MIT. Nothing official in run-llama/llama_index | llama-index-postprocessor-jev and llama-index-selectors-jev 0.1.1 (2 releases) | (PyPI 0.1.1; the repository read is a week newer) | Yes | fallback-retriever orderDefault raise_on_error=False, which the package calls "fail open": returns the retriever's nodes cut to top_n; 2.5 s timeout | fail-closedValueError at construction | Stored in node metadata | Yes. It reorders and cuts the retrieved nodes |
Change, 18 Oct 2026: in 1 failure cell of this table the bold term now shows only the shared failure term (as on When Jev fails); the words that followed it moved into the detail text. No outcome changed.
- fail-closed: stops the run, the tool or the build. held (on the reranking and when-Jev-fails pages): the item does not go ahead; a person, the host's prompt or, for two passage filters, the code's own cut decides
- fail-open: acts anyway (unscreened passages reach the prompt)
- fallback-<what>: a stated substitute (here the retriever's order)
- no-decision: raises; your code must handle it
- conditional: depends on how it is set up
- 4/7pre-1.0 (0.0.2 Python, 0.0.2 JS, 0.4.0, 0.1.1). Effect is 4.x but marked unstable (read at 4.0.1, checked at 4.1.1; 4.1.2 released 2 Nov 2026, not re-read); Pydantic AI ships in a 2.x framework (read at 2.54.0, checked at 2.64.0; TypeSafe files identical at 2.66.0, 8 Nov 2026); TanStack AI is 1.0.0
- 0/7published by TypeSafe (framework vendor 5/7, community 2/7)
- 5/7raise Jev errors to your code (LangChain Python and JS classifiers, Pydantic AI, Effect, TanStack AI)
- 2/7decide by themselves and do not stop on a Jev error (Spring fail-open, LlamaIndex fallback-retriever order)
Change, 31 Oct 2026: counts across 7 packages (TanStack AI added); they were 4/6 pre-1.0, 0/6 published by TypeSafe (framework vendor 4/6, community 2/6), 4/6 raise and 2/6 decide by themselves.
The two LangChain packages also contain middlewares that decide by themselves; all three fail closed (Python AutoMode and ModelRouter, JS autoMode). They are counted inside their package, not as extra packages.
Corrections to earlier research facts
langchain-typesafe0.0.1a3 was uploaded on 20 Sep 2026, 18:54:39 UTC, not 4 Oct (PyPI).@effect/ai-typesafeis now 4.0.1 (4 Oct 2026, 21:51 UTC), not 4.0.0.- The Spring AI integration is community-published (
org.springaicommunity:typesafe-spring-ai0.4.0). It is not part of spring-projects/spring-ai. - AutoModeMiddleware is inside
langchain-typesafe(experimental module). A separate JS version ships in@langchain/typesafe0.0.2. - LlamaIndex has a community integration (
llama-index-postprocessor-jevand-selectors-jev0.1.1). Nothing official was found (where we looked).
Thresholds and known issues that change behaviour
How to pick your own threshold is on how to choose a confidence threshold. Status codes and retries are on errors and rate limits. Documented source at the pins; issue counts from GitHub search, 5 Oct 2026, 05:17 UTC
- LangChain Python middlewares. AutoMode blocks at risk ≥ 0.5; the value is hard-coded. ModelRouter always takes the chosen model, with no confidence cut. Source observation, not filed as an issue: AutoMode passes
criteria=Noneexplicitly, so its default criteria are never applied, and no test covers this. Open issues include #40694 (classifies calls that inner middleware can rewrite) and #40700 (ModelRouter overrides the ModelFallbackMiddleware model). The failure path is tested. - LangChain Python classifier. 13 open issues with "typesafe" in the title, including #40706 (empty questions sent) and #40867 (no retry or rate-limit handling).
- LangChain JS. autoMode cuts at a fixed 0.5. Tools you do not list are never sent to Jev. 0 open issues.
- Pydantic AI. Booleans are thresholded at 0.5. Routes have a threshold that defaults to 0.0; when you set one, it raises
UnsureRoute. Open issues include #9640 (state always sent as text) and #9641 (LLMJudge drops the probability). - Effect. Threshold left to you. A missing answer key is silently skipped. 0 open issues.
- Spring AI filter. Default policy thresholds: injection 0.70, contradiction 0.70, relevance 0.45, evidence 0.55. A passage whose injection score is below 0.70 is kept. 4 open issues (#3, #7, #21, #27); #27 affects advisors, not RAG.
- LlamaIndex reranker. The optional
confidence_threshold(default none) only sets a metadata flag and never drops nodes. Issues not checked.
Rows added to the "when Jev fails" checklist
A row counts on when Jev fails only if the integration acts on Jev's answer itself. Four rows from this page are added there, in a new "Frameworks" domain. Documented source at the pins, 5 Oct 2026
| Implementation (pin) | Jev error or timeout | Malformed answer | Below threshold | No key | Bypass |
|---|---|---|---|---|---|
Spring AI TypeSafe JevDocumentFilter 0.4.0 (51993bb) | fail-openUnscreened passages kept; 10 s per attempt, 2 retries, 30 s budget | fail-openPassage passes through | acts-anywayA passage with injection risk at or below 0.70 is kept on that test; it can still be dropped by the filter's other rules (relevance below 0.45, evidence at or below 0.55). JevDocumentFilter.java L219, L225–229 | fail-closed for a blank key: startup exception. With the api-key property absent, no client is created and the application starts; the filter is not auto-created (auto-configuration L50, L70–71) | recordedRead at the pinned commit (the filter class and its auto-configuration): only an empty passage list is returned without asking Jev; every non-empty list is screened. JevDocumentFilter.java L147–149 (R17-S68, R17-S124, R17-S125) |
LlamaIndex JevRerank, community (421afda) | fallback-retriever orderCut to top_n; 2.5 s | fallback-retriever order | no-thresholdOptional flag only; never drops | fail-closedValueError at construction | recordedRead at the pinned commit (the postprocessor package): only an empty node list is returned without asking Jev. (Read scope: the postprocessor in full; the Cloudflare and Vercel provider clients at their branch, return and raise lines.) R17-S129 to R17-S133 |
LangChain ModelRouterMiddleware Python 0.0.1a3 (4af7ab8; public methods the same set at 0.0.2, 8 Nov 2026) | fail-closedThe run ends; 30 s | fail-closed | no-thresholdAlways the chosen model | fail-closedAt construction | recordedNone found: read at the pinned commit (libs/partners/typesafe/langchain_typesafe/: the middleware, classifier and client modules); each run asks Jev once before the first model call. R17-S134 to R17-S138 |
LangChain autoModeMiddleware JS 0.0.2 (417ddcf) | fail-closedTool not run; 30 s × 3 attempts | fail-closed | acts-anywayRisk < 0.5: the tool runs (fixed cut) | fail-closedAt construction | recordedTools you do not list are never sent to Jev |
On a Jev error, 1 of these 4 rows lets the action go ahead: Spring, where passages the injection screen could not judge enter the prompt. The LlamaIndex fallback returns the retriever's order by design and is not counted as going ahead, following the checklist's convention for routing fallbacks. Below-threshold results are counted separately from errors.
Bypass recorded for all 7 counted framework rows; added 16 Oct: Spring AI JevDocumentFilter and LlamaIndex JevRerank skip Jev only for an empty list; LangChain ModelRouterMiddleware: none found. The other 4 framework rows (JS autoModeMiddleware above; the LiteLLM compaction guardrail, jev-reranker and LanceDB on When Jev fails) were recorded earlier: 3 + 4 = 7. Read at the pinned commits, 16 Oct 2026; nothing was run.
Change, 16 Oct 2026: bypass for Spring AI, LlamaIndex and LangChain ModelRouterMiddleware changed from not recorded to recorded. Spring AI's below-threshold and no-key notes are corrected (terms unchanged): other filter rules can still drop a passage, and with the api-key property absent the application starts without the filter. Same cells as on When Jev fails.
How these rows were counted
- JS autoModeMiddleware is a separate row. It is a different package and implementation from the Python AutoModeMiddleware, which stays where it is already counted, on Claude Code and MCP guards. No existing row changes.
- Pydantic AI is not a decision row.
TypeSafeModelreturns typed answers to the agent you write and raises on a Jev error, so your code decides what happens next. Its failure behaviour is on this page only. - The Python classifier and Effect are not decision rows either: they raise to your code.
- TanStack AI is not a decision row (added 31 Oct 2026):
@tanstack/ai-typesafe1.0.0 raises to the caller on an API error, a malformed answer or a missing key, so your code decides. Under the 5 Oct framework rule it is not counted on when Jev fails.
Version rechecks (not audited on this page)
The versions checked most recently are in the version block near the top. Earlier checks are kept below with their dates and sources.
Earlier version checks (14, newest first)
- Rechecked 8 Nov 2026 (registries 05:25 UTC): newer releases noted.
@ai-sdk/typesafe-ai3.0.35 andai7.0.165 (7 Nov 2026) are newer than the versions this page describes; not re-read. Unchanged on 8 Nov 2026:@effect/ai-typesafe4.1.2,@langchain/typesafe0.0.2,llama-index-postprocessor-jev0.1.1,@tanstack/ai-typesafe1.0.0. LiteLLM 1.108.3 (LTS, 7 Nov 2026): release notes do not mention TypeSafe or Jev (checked 8 Nov 2026); whether it contains the eu/jev provider from PR #46875 was not checked in source (the 7 Nov source check covered stable releases up to 1.108.2; see the 7 Nov check). The LiteLLM cells stay as read at 1.106.0 to 1.106.2. Documented (npm and PyPI registry records, 8 Nov 2026, 05:25 UTC; LiteLLM PyPI record and release feed, 05:25 UTC; R28-S47, R28-S45) (@ai-sdk/typesafe-ai3.0.35 andai7.0.165 were compared on 9 Nov 2026, see the version block.) - Rechecked 7 Nov 2026 (registries 15:29–15:36 UTC): newer releases noted.
langchain-typesafe0.0.2 (5 Nov 2026) was newer than the 0.0.1a3 read (read at 0.0.2 on 8 Nov 2026, see the version block). (0.0.1 was also published, 4 Nov 2026.)pydantic-ai-slim2.66.0 (4 Nov 2026) was newer than the 2.64.0 checked (its TypeSafe files were compared on 8 Nov 2026: identical).@ai-sdk/typesafe-ai3.0.34 (5 Nov 2026) was newer than the version this page describes; not re-read. Unchanged on 7 Nov 2026:@effect/ai-typesafe4.1.2,@langchain/typesafe0.0.2,llama-index-postprocessor-jev0.1.1,@tanstack/ai-typesafe1.0.0. LiteLLM's stable releases up to 1.108.2 (6 Nov 2026) do not include the eu/jev provider from PR #46875; it is in pre-releases v1.109.0-rc.1 and v1.110.0-dev.1 (checked in source, 7 Nov 2026; R27-S24, R27-S46); the LiteLLM cells stay as read at 1.106.0 to 1.106.2. The cells stay as read at the pinned versions. Documented (npm and PyPI registry records, 7 Nov 2026, 15:29–15:36 UTC; R27-S25) - Rechecked 30 Oct 2026 (registries 05:20 UTC): two framework packages have newer minor releases; newer version exists, not re-read.
pydantic-ai-slim2.63.0 was released on 28 Oct 2026, 02:03 UTC (2.62.0 on 27 Oct); this site last read the TypeSafe model file at 2.60.0 (23 Oct check).@effect/ai-typesafe4.1.0 was released on 29 Oct 2026, 03:47 UTC (4.0.7 on 28 Oct); the row was read at 4.0.1 and patches up to 4.0.5 were noted on 16 Oct. Patch releases, not re-read: LiteLLM 1.106.6 (28 Oct 2026; the failure path was compared at 1.106.0 and the image files read at 1.106.1 and 1.106.2). Unchanged:langchain-typesafe0.0.1a3,@langchain/typesafe0.0.2,llama-index-postprocessor-jev0.1.1. The matrix and failure terms stay as read at the pinned versions. Documented (npm and PyPI registry records, 30 Oct 2026; R24-S35, R24-S36) Change, 30 Oct 2026: recheck added (pydantic-ai-slim2.63.0 and@effect/ai-typesafe4.1.0 noted, newer version exists, not re-read; LiteLLM 1.106.6 patch noted). No row or term changed. - Checked at pydantic-ai-slim 2.60.0 (23 Oct 2026): the TypeSafe provider is unchanged and a request with images still raises
UserErrorbefore any call. Since 2.59.0, API errors raise subclasses ofModelHTTPErrorby category and connection failures raiseModelConnectionError; 2.58.0 added an optionalmax_retriessetting. Documented (wheels 2.57.0 to 2.60.0 read as files, 23 Oct 2026, 05:23–05:29 UTC, nothing installed;models/typesafe.pyL222–223, L239–250 at 2.60.0; R22-S110, R22-S111, R22-S112, R22-S118) - Images, read 23 Oct 2026. @langchain/typesafe 0.0.2 and langchain-typesafe 0.0.1a3 have no image type: an image in a message is sent to Jev as JSON text, so the decision is made without the image (
messages.jsL47–58;_state.pyL15–52). LiteLLM 1.106.1 passes PNG or JPEG images to Jev only when you send TypeSafe's opt-in header; without it the call is rejected before it is sent, unlessdrop_paramsis on, in which case the image is dropped and the request still goes out (decisions/call.pyL377–387, L486–495; the TypeSafe files are the same at 1.106.2). Images in Jev are experimental and opt-in, and TypeSafe says they are not yet ready for production; no image was sent. Documented (package files read, nothing installed or run; R22-S113, R22-S114, R22-S115, R22-S119) All integrations read: Jev with images: does your SDK or integration send them? Change, 23 Oct 2026: pydantic-ai-slim checked at 2.60.0 (the row's error classes now nameModelConnectionErrorfor connection failures since 2.59.0; the term is unchanged); image handling of both LangChain packages and LiteLLM added. - Rechecked 20 Oct 2026: checked at LiteLLM 1.106.0 (published 18 Oct, minor; tag commit
955bdaa): the Jev classifier's failure path is unchanged. Any error from the Jev call still reaches the same fallback, and a missing key still stops start-up. The classifier type"jev"is renamed"oss_classifier"(the old names are still accepted), and the same option can now call Laya, Bespoke or Databricks instead of Jev. Failures now record a reason (timeout, invalid response, classifier error, circuit open) in the routing log. The TypeSafe guardrail hook is still present; since 1.105.0 its Jev call timeout is configurable (default 30 s). Documented (files at 1.104.07964577and v1.106.0 compared as text, nothing installed; R20-S204–S212, R20-S53) - Checked at pydantic-ai-slim 2.56.0 and 2.57.0 (16 Oct 2026): TypeSafe provider unchanged; the model adds one check, a request with images raises
UserErrorbefore any call; the failure path is otherwise unchanged. Since 14 Oct 2026 TypeSafe accepts images as an opt-in experiment (not production-ready per TypeSafe), so this check is stricter than the API. @effect/ai-typesafe 4.0.5 (patch) was released after our last check; not re-audited. Documented (registry records 16 Oct 2026, 05:28 UTC, and the package files compared as text, nothing installed; R17-S57, R17-S58) - Rechecked 14 Oct 2026 (registries 16:51 UTC): @effect/ai-typesafe 4.0.4 (13 Oct 2026, patch), pydantic-ai-slim 2.56.0 (13 Oct 2026, minor) and litellm 1.105.1 (14 Oct 2026, patch) were released after our last check; not re-audited. Documented (npm and PyPI registry records, R16-S64)
- Rechecked 12 Oct 2026 (registries 05:17 UTC; LiteLLM diff 05:28 UTC): checked at LiteLLM 1.105.0 (published 11 Oct, minor): the Jev classifier's failure path is unchanged; a new option can send the Jev call through a router deployment with no retries and no fallbacks, and its errors reach the same fallback. pydantic-ai-slim (2.55.0) and @effect/ai-typesafe (4.0.3) are unchanged since 11 Oct; the other packages are unchanged. Documented (
complexity_router.pyat 1.104.07964577and at the v1.105.0 tag compared as text, nothing installed; R14-S53, R14-S54) - Rechecked 11 Oct 2026, 05:22 UTC: pydantic-ai-slim checked at 2.55.0 (9 Oct, minor; the row above was read at 2.54.0). The failure path is unchanged:
providers/typesafe.pyis unchanged and the API error mapping inmodels/typesafe.py(L223–243) is unchanged. New in 2.55.0: a request with image input now raisesUserError("TypeSafe does not support image input.") before any request is sent (L219–220). @effect/ai-typesafe 4.0.3 (10 Oct, patch; row read at 4.0.1) and LiteLLM 1.104.2 (8 Oct, patch; failure path read at 1.104.0) are newer releases, not re-audited; the other packages are unchanged. Documented (package registries and the two PyPI wheels compared as text, nothing installed; R13-S35, R13-S119, R13-S120) - Rechecked 10 Oct 2026, 05:18 UTC: pydantic-ai-slim 2.55.0 (9 Oct, minor; row read at 2.54.0; not yet diffed) and LiteLLM 1.104.2 (8 Oct, patch; failure path read at 1.104.0) were newer releases, not re-audited; @effect/ai-typesafe 4.0.2 as noted 8 Oct; the other packages were unchanged. (package registries, R12-S48, R12-S49)
- Rechecked 8 Oct 2026, 05:18 UTC: @effect/ai-typesafe 4.0.2 (7 Oct, patch; row read at 4.0.1) and LiteLLM 1.104.1 (7 Oct, patch; failure path read at 1.104.0) were newer patch releases, not re-audited; the other four packages were unchanged. (package registries, R11-S48)
- Rechecked 7 Oct 2026, 05:21 UTC: no new release of any of the six packages; LiteLLM still 1.104.0. (package registries, R10-S48)
- Rechecked 6 Oct 2026, 05:20 UTC: no new release of any of the six packages; LiteLLM 1.104.0 failure path unchanged. The Spring AI repository HEAD has moved to
2c30a58, but tagv0.4.0is still51993bb; the LlamaIndex repository HEAD is still421afda. Documented (package registries andgit ls-remote)- LiteLLM 1.104.0 (PyPI, 3 Oct 2026, 22:42 UTC). Checked on 6 Oct 2026 at tag commit
7964577: the Jev classifier's failure path is unchanged from8d28e8d(the caller block is byte-identical; the classifier file differs only in comments). The tag may predate8d28e8d; their order was not verified. Routing cells stay on model routing. The TypeSafe proxy guardrail hook (proxy/guardrails/guardrail_hooks/typesafe/typesafe.py) exists at both commits. Read at the tag in spare time on 6 Oct: it is a context-compaction guardrail, not a safety block (it removes tool results Jev judges no longer relevant, below 0.2); on a Jev error it forwards the request uncompacted by default; with no key it fails at start-up. It is not a row on this page yet. - Vercel
@ai-sdk/typesafe-ai3.0.15 (7 Oct 2026, 01:41 UTC), after 3.0.14 (6 Oct, 22:20 UTC) and 3.0.13 (5 Oct, 18:26 UTC); patch releases, not re-audited (R10-S35). See TypeScript. - Pydantic AI Gateway does not route Jev. Its documented providers are OpenAI, Anthropic, Google Vertex, Groq and Bedrock ("More providers coming soon"), and
providers/gateway.pyat v2.54.0 has no TypeSafe entry (docs and source search, 05:19 UTC). - Mastra: no package found.
@mastra/typesafereturns 404, and no Mastra-published package appeared in the npm search (05:16 UTC).
- LiteLLM 1.104.0 (PyPI, 3 Oct 2026, 22:42 UTC). Checked on 6 Oct 2026 at tag commit
Where we looked for an official LlamaIndex integration
- PyPI
llama-index-llms-typesafe,llama-index-typesafe,-postprocessor-typesafe(-rerank),-tools-typesafeand-llms-jev: 404 (05:12–05:18 UTC). - The PyPI simple index, searched for "typesafe" and "jev": only the two community packages (05:18 UTC).
- npm
@llamaindex/typesafe: 404 (05:13 UTC). - run-llama/llama_index
llama-index-integrations/foldersllms,postprocessorandselectors: no TypeSafe or Jev entry (05:18 UTC).
Other framework packages seen in registry searches (names only, not audited)
Names, versions and publishers from npm and PyPI searches on 5 Oct 2026, 05:16 and 05:18 UTC. Not read in source and not counted above.
- npm:
@composio/typesafe0.1.1;@openclaw/typesafe2026.9.8;@arizeai/openinference-instrumentation-typesafe0.1.0;@effect-uai/typesafe-ai0.17.0;@effect-agent/ai-typesafe0.1.0-beta.100. - PyPI:
agent-framework-typesafe1.0.0a261002 (author listed as Microsoft);composio-typesafe0.25.0;flyteplugins-typesafe-ai2.10.7;openinference-instrumentation-typesafe0.1.4;kitaru-typesafe-evaluator0.1.0;llm-typesafe0.1a0;langgraph-jev0.1.0;ragas-jev0.3.0;mellea-jev-adapter0.1.1.
Change, 31 Oct 2026: @tanstack/ai-typesafe, listed here at 0.1.7 on 5 Oct 2026, moved into the maturity matrix, read at 1.0.0.
What is on other pages
- Calling the Jev SDK directly, without a framework: Python and TypeScript.
- DeepEval and LangSmith Evals: Jev as a judge.
- LiteLLM routing: model routing.
Plugins, hooks and MCP servers for coding agents (Codex, OpenCode, Cursor, OpenClaw, Hermes Agent) are compared on Jev with coding agents (6 Oct 2026); Claude Code guards are on Claude Code and MCP guards.
What was not verified
- Nothing was installed or run, and no API key was used. Every cell is read from code; behaviour at run time may differ.
- SDK defaults not read: the
@langchain/coreretry backoff used by LangChain JS, and which typesafe-sdk version, with which retry and timeout defaults, a Pydantic AI install uses (typesafe-sdk 0.7.4 itself defaults to 10 s per HTTP operation, 2 retries and a 30 s retry budget, read from its wheel on 11 Oct 2026; R13-S300, R13-S301). - How Effect's core DecisionModel handles thresholds.
- TanStack AI (read 31 Oct 2026): what confidence values reach your code, its licence and its tests were not recorded. The 30 s timeout is documented in
client.tsbut not read anywhere in the files read; whether another TanStack package applies it was not checked. - pydantic-ai-slim 2.64.0 and @effect/ai-typesafe 4.1.1 (31 Oct 2026): only the TypeSafe files were compared; the rest of each framework was not read.
- The Spring auto-configuration lines (no-key behaviour) were read once and not spot-checked a second time.
- Test contents for Effect, Spring and LlamaIndex (file names only).
- The LiteLLM TypeSafe guardrail hook's registration, config and tests (the hook file was read at
7964577on 6 Oct; it exists at both commits, 17,570 and 18,224 bytes; the difference was not read). - Pydantic AI's licence field. LlamaIndex community issues.
- The LlamaIndex line numbers come from the repository at
421afda, which is a week newer than PyPI 0.1.1. - How many people use any of these packages.
- Coding-agent plugins (Hermes, Codex, OpenCode, Cursor, OpenClaw) are not on this page; 18 were read in source on 6 Oct 2026 for Jev with coding agents.
Sources and check times (5 Oct 2026, UTC; later checks dated in each entry)
- R28-S211 to R28-S215, R28-S50 (8 Nov 2026, 05:19–05:33 UTC): PyPI JSON and the 0.0.1a3 and 0.0.2 wheels of
langchain-typesafe(read as files); the langchain-typesafe==0.0.2 release page (links commit71881c7; source tree not read); PyPI JSON and wheels ofpydantic-ai-slim2.64.0 and 2.66.0 (file hashes). R28-S47 (05:25 UTC): npm and PyPI registry records for the 8 Nov version recheck. R28-S45 (05:25 UTC): LiteLLM PyPI record and release feed. - R25-S227 to R25-S237 (31 Oct 2026, 05:13–05:14 UTC): npm record for
@tanstack/ai-typesafe(1.0.0 published 22 Oct 2026, 15:19:42 UTC) and the paired@tanstack/ai1.0.0; the jsDelivr copy of the npm files; tag@tanstack/ai-typesafe@1.0.0at 5651e07. Paths underpackages/ai-typesafe/src/:adapters/evaluate.tsL90, L92–98, L100–115, L117–120, L176–181;utils/client.tsL30–31, L37, L60–62;byok.ts; andpackages/ai/src/activities/evaluate/index.tsL556–572, L608–631. R25-S45:evaluate.tsL88–121 read a second time at the same commit (05:18 UTC). - R25-S38, R25-S39 (31 Oct 2026, 05:14–05:15 UTC):
pydantic-ai-slimmodels/typesafe.pyat v2.60.0 to v2.64.0,providers/typesafe.pyat v2.60.0 and v2.64.0,_model_errors.pyL316 at v2.64.0; release v2.64.0 (commitcde1d10). - R25-S40 (31 Oct 2026, 05:14–05:15 UTC):
@effect/ai-typesafe4.0.1 and 4.1.1 package sources (internal/errors.ts,TypeSafeClient.ts,TypeSafeConfig.ts,TypeSafeDecisionModel.ts) from jsDelivr; npm record for 4.1.1 (no git commit recorded). - R26-S36 (2 Nov 2026, 10:55 UTC): npm and PyPI registry records for the 2 Nov version recheck (
@effect/ai-typesafe4.1.2, released 2 Nov 2026, 04:03 UTC; the others unchanged). - R25-S36 (31 Oct 2026, 05:14 UTC): npm and PyPI registry records for the version checks in the version block.
- R8-S11: search suggestions from Google (
client=firefox), Bing (osjson) and DuckDuckGo (ac), 05:14:54–05:15:03. - R8-S80: PyPI JSON for
langchain-typesafe; taglangchain-typesafe==0.0.1a3at 4af7ab8 (read 05:13). Paths underlibs/partners/typesafe/langchain_typesafe/:classifier.pyL38, L255–265, L366–371;client.pyL333–344;types.pyL241, L269;experimental/middleware/auto_mode.pyL39, L96–97, L147, L164, L167, L232;experimental/middleware/model_router.pyL85–86, L152, L197. Test:test_auto_mode.pyL317–327. - R8-S81: npm record for
@langchain/typesafe; tag@langchain/typesafe@0.0.2at 417ddcf (read 05:13). Paths undersrc/:classifier.tsL26, L128–131, L197;middleware/autoMode.tsL26, L128, L144–146, L154, L157. - R8-S82: PyPI JSON for
pydantic-ai-slim; tagv2.54.0at 6695132 (read 05:14). Paths underpydantic_ai_slim/pydantic_ai/:models/typesafe.pyL222–242, L231–232, L293–306;models/decision.pyL556, L623–624, L798;providers/typesafe.pyL78–83. Test filetests/models/test_typesafe.py. - R8-S83: npm record for
@effect/ai-typesafe; tag@effect/ai-typesafe@4.0.1at 460272d (read 05:13). Paths underpackages/ai/typesafe/src/:TypeSafeClient.tsL76, L145;TypeSafeDecisionModel.tsL72–90;internal/errors.tsL53–111;DecisionModel.tsL69–92. - R8-S84: Maven Central for
typesafe-spring-ai0.4.0; tagv0.4.0at 51993bb (read 05:17).typesafe-spring-ai/src/main/java/org/springaicommunity/typesafe/rag/JevDocumentFilter.javaL118–132, L170–206 (fail-open path; spot-checked a second time);RetryPolicy.javaL34–35;TypeSafeConstants.javaL48; auto-configuration L50, L70–71 (read once, not spot-checked). - R8-S85: Spring blog, 2 Oct 2026, cites 0.3.0 (web-search snippet). Reported
- R8-S86: llama-index-jev at 421afda (HEAD 25 Sep; read 05:24) and PyPI JSON (0.1.1, 18 Sep). Paths under
packages/llama-index-postprocessor-jev/llama_index/postprocessor/jev/:base.pyL69–83, L103, L145–151, L266–272;utils.pyL92–96. - R8-S87: LlamaIndex search boundary: PyPI JSON for the official-style names (404), the PyPI simple index, npm
@llamaindex/typesafe(404), and run-llama/llama_index integration folders (05:12–05:18). Absence within that boundary only. - R8-S88: PyPI JSON for LiteLLM 1.104.0 (read 05:15).
- R8-S89: Pydantic AI Gateway docs (read 05:19).
- R8-S90: GitHub issue search for "typesafe" and "jev" in 8 repositories (05:17:50); title matches counted per repository.
- R8-S91: npm and PyPI searches for other framework packages, including
@mastra/typesafe(404) (05:16:38 and 05:18:15).