Kaeso 0.4.2 lets organization admins add their own providers and rotate keys in one step. Hard budgets and token rates now refuse a single request that could cross them, so operators may see more refusals up front. Streams that finished are no longer logged as client disconnects, and request logs, the console and a Server-Timing header show the time Kaeso adds to each request. Codex and the ChatGPT desktop app can use Claude, Grok and Mistral models, and many documentation and console layouts were reworked for phones.
Contributors: @dvnold
Full changes: v0.4.1...v0.4.2 (öffnet in einem neuen Tab)
Hinzugefügt
- Organization admins can add custom providers, such as a self-hosted vLLM, Ollama or an internal proxy, on any API format Kaeso speaks. Managing them needs the new
providers.custom.manageright, and thecustom_providers.enabledandcustom_providers.allowed_networkssettings control whether they may be used and which networks they may reach. (#427 (öffnet in einem neuen Tab)) - API keys and platform keys can be rotated in one step, from the console or the Platform API. The new key keeps the old key's name, owner, access, models and key limits, and the old key keeps working for a grace period of 0, 1, 24 or 168 hours that you choose. (#422 (öffnet in einem neuen Tab))
- Request logs, the console and a
Server-Timingheader show the time Kaeso itself spends on each model request, apart from the provider's time. The log gainsoverhead_msandstream_processing_mscolumns, and thekaeso_gateway_overhead_secondsmetric records the same per provider. (#423 (öffnet in einem neuen Tab)) - Codex and the ChatGPT desktop app can use Claude, Grok and Mistral models. Namespaced tools become plain function tools, hosted web search maps to Claude's web search tool, and client metadata stays out of other formats. Claude always searches live, so a Codex setting of cached search is refused for Claude with a hint to set live search or turn search off, and the refusal for Mistral gives the same hint. (#396 (öffnet in einem neuen Tab))
Documentation
- The new
kaeso-benchcrate measures the time and memory Kaeso adds under thousands of concurrent streams, client disconnects and slow clients, end to end over real sockets. Its report gives the Kaeso version, commit, CPU and every parameter with each run. (#425 (öffnet in einem neuen Tab))
Geändert
- Requirements for the end-to-end suites are now installed in CI with hash checking, so every package is pinned with its hashes. Two advisory fixes update
undicito 8.12.3 andcrypto-bigintto 0.7.6. (#384 (öffnet in einem neuen Tab)) - Diagrams in the documentation sit centred in a framed well with a caption. Kaeso's own parts are drawn in ink, outside services are dashed, and edge labels no longer have lines running through them, with the frame provided by a new
Figurecomponent in the UI library. (#388 (öffnet in einem neuen Tab)) - On phones, long lines in code blocks wrap with a hanging indent instead of running off the edge, and Copy still copies the code as written. (#402 (öffnet in einem neuen Tab))
- The documentation search no longer shows a count beside each group, while the console's pickers and palette keep theirs. (#403 (öffnet in einem neuen Tab))
- Table headers can wrap onto several lines with the
wrapoption ofTable.HeaderCell, and the documentation wraps headers longer than 24 characters, so the sign-in comparison fits its column. (#405 (öffnet in einem neuen Tab)) - The
densityoption ofTableset to compact puts 8px between columns, and the documentation uses it for tables of five or more columns. Wide Configuration and Chinese KPP tables now fit desktop widths, and long type names may break between words. (#408 (öffnet in einem neuen Tab)) - Development dependencies in the console, docs and UI library move
happy-domto 21.0.4, which fixes a code injection vulnerability in its ESM compiler. (#414 (öffnet in einem neuen Tab)) - On phones, a wide diagram in the documentation opens on the part it is about, the nodes marked in ink, instead of its middle. The
Figurecomponent takes afocusoption for the share of its width to open on. (#418 (öffnet in einem neuen Tab)) - Wrapped code lines keep the indent of the line they continue, so a continued command option sits under its own option. The Platform API error example, the last code block that scrolled on desktop, now wraps at every width. (#430 (öffnet in einem neuen Tab))
Documentation
- The diagrams on the introduction, concept, self-hosting and development pages are redrawn to show the install's boundary, each check's statuses and the path of a request. The Upgrades page gains a diagram that links each update result to the step where it happens. (#389 (öffnet in einem neuen Tab), #390 (öffnet in einem neuen Tab), #391 (öffnet in einem neuen Tab))
- On phones, documentation tables of three or more columns show each row as a list with its labels instead of scrolling sideways. (#394 (öffnet in einem neuen Tab))
- Console, guide, API, self-hosting and contributor pages are split into shorter sentences and paragraphs, with lists where they listed things. No fact, setting or command changed. (#395 (öffnet in einem neuen Tab), #397 (öffnet in einem neuen Tab), #398 (öffnet in einem neuen Tab))
- The KPP reference pages read in shorter sentences, with lists for what each codec refuses and drops. The provider profiles and the Chat Completions providers move to a new page, and links to it are updated. (#399 (öffnet in einem neuen Tab))
- Documentation search shows pages whose title or description match first, then matching sections, then pages that only mention the words. (#400 (öffnet in einem neuen Tab))
- The changelog is no longer built into the documentation. The sidebar's Changelog link goes to the changelog on the Kaeso website, in the page's language, and old changelog addresses redirect there. (#401 (öffnet in einem neuen Tab))
- Long code in documentation tables breaks after separators such as slashes, dots and equals signs, so reference tables such as the Platform API routes fit the desktop column. (#406 (öffnet in einem neuen Tab))
- The two KPP allowlist tables on the cross-format page now fit the desktop width, with scope and origin dialect in one column. (#407 (öffnet in einem neuen Tab))
- Shell commands and code samples are reformatted to fit the page on desktop, with long commands continuing on the next line after a backslash. The upgrade script names the release address once and checks the hash in two steps. (#409 (öffnet in einem neuen Tab))
- Markdown, text and plain-text examples wrap like prose at every width instead of scrolling sideways. (#410 (öffnet in einem neuen Tab))
- On phones, a table column with mostly short values sits beside its labels even when one value is long. (#411 (öffnet in einem neuen Tab))
- The diagram on the benchmark page runs top to bottom and fits a phone without scrolling. The stream counts in its options table are written as 100, 1,000 and 5,000. (#428 (öffnet in einem neuen Tab))
- The key rotation example in the Platform API guide names the key's address once, so its command fits the desktop column. (#429 (öffnet in einem neuen Tab))
- The German and Chinese custom provider examples send the English prompt, so their code matches the other languages. (#431 (öffnet in einem neuen Tab))
- The Configuration settings tables show each variable and TOML key in one column, so most names stay on one line on desktop and phones, and a table cell can now hold a line break. (#432 (öffnet in einem neuen Tab))
Behoben
- A non-streaming Responses request through a ChatGPT plan now returns the reasoning, web search call and message items with their types, as OpenAI does. Before, the items came back without a type. (#416 (öffnet in einem neuen Tab))
- A stream whose last event reached the app is logged as a success, even when the app closes the connection right after it. Before, such requests could be logged as client disconnects with status 499, and a provider that closed its stream after the last event could fail the request. (#421 (öffnet in einem neuen Tab))
- A stream that ends because the app stopped reading it is logged as
client_stalledinstead ofclient_disconnected, and the console and metrics show the new kind. Real disconnects stayclient_disconnected. (#424 (öffnet in einem neuen Tab))
UI library
- On this page marks the section you are reading when you scroll back up, and marks the last section at the bottom of a page. (#393 (öffnet in einem neuen Tab))
- Tables no longer scroll sideways by two pixels when a link ends a text column at the table's edge. (#404 (öffnet in einem neuen Tab))
Documentation
- Long page titles in the documentation sidebar no longer end in an ellipsis, with short navigation labels in English, German and Chinese. (#392 (öffnet in einem neuen Tab))
- A diagram group's title moves along its top edge when an arrow enters the group at its top-left corner, so the arrow is no longer hidden behind the title. (#415 (öffnet in einem neuen Tab))
Sicherheit
- Hard spend budgets and token rates refuse a request whose output bound could cross them, even when it is the only request in flight. Password reset is refused where the organization requires SSO, and the Azure wire server address is blocked as a metadata address. (#385 (öffnet in einem neuen Tab))