Skip to content

v1.15.0

Choose a tag to compare

View all tags
@stephanj stephanj released this 28 Oct 11:32
· 6 commits to master since this release
v1.15.0
74e855a

A community release: most of this version came from contributors outside the core team, including two features a Codex agent wrote live on stage during the Devoxx Belgium 2026 opening keynote. The headlines are favorite models, per-project Scan & Copy settings, a fetch_page that returns real Markdown, Claude 5 model support, and the move to the langchain4j 1.20+ MCP client.

Added

  • feat(chat): favorite models and less dropdown clutter. A star next to the model dropdown marks the selected model as a favorite, and a new Favorites entry lists the favorites from every provider and switches to the one you choose. Availability is checked in the background, and a favorite whose provider no longer offers that model stays visible but disabled. Favorites are stored as provider + model-name pairs, survive a restart, and can be removed from the LLM settings page (with Apply/Reset). The provider and model dropdowns are also hidden while they hold only one entry and reappear when there is a choice. Written live by a Codex agent during the Devoxx Belgium 2026 keynote (#1077, #1322, #1327)
  • feat(settings): Scan & Copy settings per project. Included file extensions, excluded folders and files, .gitignore handling and the RAG directory exclusions used to be shared by every project. They are now stored per project in the project's workspace file. A project without its own settings starts from a copy of the current global values, so nothing changes until you edit a project. The settings page names the project it edits, and project scanning, Copy project, Add project to prompt, the token count and RAG indexing all use the settings of the project they run in. Scanner state is now kept per operation, and project-less scanners read the global settings on every call instead of a snapshot taken at construction. Also written live during the keynote (#287, #1329)
  • feat(agent): fetch_page returns Markdown and limits what it fetches. It used to return Jsoup.parse(body).text(), which flattened a Docusaurus page into a single line, dropped line breaks from code samples, turned tables into a run of words and lost every link URL. HTML is now converted to Markdown (HtmlToMarkdown): headings, fenced code blocks with their language, tables, lists and absolute links are kept, while scripts, styles, nav/aside/footer and hidden elements are dropped, and main is used when the page has one. Text, JSON, XML and JS pass through using the declared charset, and binary content is described rather than decoded. The fetch is also bounded: a 10 MB / 60 s cap; same-host redirects only (a cross-host redirect is handed back to the model as a new URL); URLs that are too long, contain credentials, have dot-less hosts, or resolve to loopback/link-local addresses (e.g. 169.254.169.254) are rejected; and pages are cached for 15 minutes (#1338, #1339)
  • feat(models): Claude 5 family: Claude Fable 5.1, Opus 5.5, Opus 5, Sonnet 5 and Sonnet 5.5 added to models.json for both Anthropic and Bedrock. The Anthropic provider no longer sends temperature to Claude models that reject it, using a new shared ClaudeModels rule that Bedrock also uses (#1309, #1310, #1312, #1313)

Fixed

  • fix(agent): the user's task is no longer evicted from memory during a long tool loop. MessageWindowChatMemory evicts the oldest non-system message first, and every tool call adds two messages, so a long agent run could push the original prompt out of the window and keep calling the LLM without it. Sub-agents hit this after only 5 tool calls. A new TaskPreservingMessageWindowChatMemory evicts the oldest finished tool call and its results instead while the loop is still running, then returns to normal eviction once the final answer arrives. It is used by both ChatMemoryService and SubAgentRunner (#1307, #1308)
  • fix(ollama): Ollama models show up even when the /api/show probe fails. Three bugs combined to leave the model list empty and report "LLM provider is not running" while /api/tags was succeeding. First, "provider running" depended on at least one per-model context probe succeeding. Second, a single model answering /api/show with HTTP 200 and an empty body (e.g. gemma3n:e4b) threw an NPE that aborted the whole fetch. Third, the probe used the shared retrying HTTP client, so it could take around 90 s per model. Now a successful listing call is enough to mark the provider as running, a failed probe falls back to a 4096-token context, and per-model failures skip only that model. The probe uses a fail-fast client (3 s connect / 5 s read, no retries), and a provider previously found down is re-probed after 15 s instead of staying "not running" for the rest of the session (#1331, #1324)
  • fix(bedrock): Bearer Token auth works, and Claude Opus 4.7+ no longer returns HTTP 400. In BEARER_TOKEN mode the AWS SDK still picked SigV4, because it comes first in Bedrock's auth-scheme list, and failed with "Unable to load credentials from any of the providers in the chain". httpBearerAuth is now forced on the runtime, async-runtime and control-plane clients. Separately, temperature is no longer sent to Opus 4.7/4.8, Opus 5/5.5, Sonnet 5, Fable and Mythos, which reject it (#1305, #1306)
  • fix(ui): answer bubbles no longer keep empty space after a resize (#1332, #1333)
  • fix(cli/acp): attached images reach CLI and ACP runners, and follow-up prompts keep working afterwards. Images were silently dropped because runners only receive plain text. History replay also called UserMessage.singleText(), which throws on the multimodal message an image turn leaves in memory, so every prompt after an image failed (#1280, #1282, #1284)

Changed

  • feat(mcp): migrated to the langchain4j 1.20 MCP client. langchain4j 1.19 removed the legacy 2024-11-05 HTTP+SSE transport with no replacement, so the earlier 1.18.1 pin is gone. Instead, HTTP_SSE is now a legacy alias for HTTP (Streamable HTTP with the subsidiary SSE channel), so existing configurations still load and are re-saved as HTTP when edited, and new servers choose STDIO or HTTP. SSE-only MCP servers (#1151) are no longer supported. The protocol version is no longer pinned: the client auto-detects 2026-07-28 and falls back to 2025-11-25, with detection bounded by your configured timeout. Test Connection in settings now goes through the same transport builders as real execution (#1298, #1301)
  • refactor(mcp): stop using the McpLogMessage JsonNode API that langchain4j 1.20 deprecated. As a side effect, object and array log payloads now show as JSON text instead of an empty string (#1350)
  • chore: remove a JVM crash log that was committed by accident and ignore *.log files (#1319)

Tests

  • fix(test): platform fixtures are disposed, stopping a nested temp-dir leak. JUnit 5 and MockitoJUnitRunner never call LightPlatformTestCase.tearDown(), so about 45 platform tests leaked their fixture and nested every later test's temp dir one level deeper, until setUp() failed with "File name too long". This caused three order-dependent failures (e.g. BedrockModelFactoryTest, OpenRouterChatModelFactoryTest). tearDown() is now an EDT-wrapped @AfterEach, skipped when setUp() never ran (#1357)

Documentation

  • docs: add the Groma architecture map (groma/: actors, externals, flows, containers and components) and Groma agent instructions in AGENTS.md/CLAUDE.md (#1335)

Dependencies

  • chore(deps): upgrade langchain4j 1.18.1 → 1.20.0 → 1.22.0, AWS SDK BOM 2.54.16 → 2.55.1, commonmark 0.30.0 → 0.31.0, Netty 4.2.17 → 4.2.19.Final, Logback 1.6.3 → 1.6.5, Mockito 5.23.0 → 5.24.0, mockwebserver 5.5.0 → 5.5.1, jsoup 1.23.1 → 1.23.2, sqlite-jdbc 3.53.2.1 → 3.53.4.0, Lombok 1.18.46 → 1.18.48, Kotlin plugins 2.4.10 → 2.4.21 and Gradle wrapper 9.6.1 → 9.8.2 (#1298, #1302, #1351)
  • chore(deps): bump docs toolchain: body-parser 1.20.6 → 1.20.9, brace-expansion 1.1.18 → 1.1.21, browserslist 4.28.6 → 4.28.9, colord 2.9.3 → 2.10.0, compression 1.8.1 → 1.8.2, fast-uri 3.1.5 → 3.1.9, follow-redirects 1.16.0 → 1.16.1, image-size 2.0.2 → 2.0.4, joi 17.13.3 → 17.13.8, js-yaml 4.3.1 → 4.3.2, launch-editor 2.14.1 → 2.14.2, proxy-addr 2.0.7 → 2.0.8, qs 6.15.2 → 6.15.4, shell-quote 1.10.0 → 1.12.0, source-map-js 1.2.1 → 1.2.2, plus baseline-browser-mapping, http-proxy-middleware, postcss-selector-parser and svgo (#1287, #1290, #1292–#1295, #1297, #1299, #1304, #1336, #1341, #1343–#1349, #1354–#1356)

Contributors

Contributors

stephanj, dliubarskyi, and 3 other contributors
Assets 3
Loading