What lmnr-ai/lmnr shipped
Written by FoxPlug from public releases; not affiliated with Laminar. An automatic summary of the public release, pull request and commit data of github.com/lmnr-ai/lmnr. Laminar did not write it and does not use or endorse FoxPlug. Every line links to the public change it describes.
Get a weekly update like this for your product, free
Week of September 21, 2026
What shipped
- Fixed Responses API LLM profiles to use non-strict function tools and set
store: falseon response requests. Pull request #2422 - LLM profiles now reject unused fields instead of silently dropping them, and surface malformed AEAD_SECRET_KEY errors during setup. Pull request #2424
- Fixed shared trace loading with responsive layout improvements and corrected trace page initialization. Pull request #2399
- Refactored span message processing into separate stage functions, breaking down a 720-line function into manageable pieces. Pull request #2390
- Added Redis batch insert with TTL support to the cache layer for more efficient bulk key writes. Pull request #2442
- Search snippets are now withheld under masking policy on dual-mode projects to prevent PII leaks. Pull request #2393
- API keys and CLI tokens now read as masked when accessed on dual-mode projects. Pull request #2392
- Removed the CLUSTERING_ENABLED frontend feature flag since clustering is on by default in EE and OSS produces no clusters. Pull request #2420
- Added a Redis cache key constant for temporary signal throttling per project. Pull request #2439
Why it matters
Custom gateway support through Responses API expands LLM profile flexibility for self-hosted deployments. PII masking improvements strengthen security for dual-mode projects. Refactoring span processing and fixing trace loading improve code maintainability and user experience.
Changelog entry
- Fix Responses API to use non-strict function tools and set
store: falseon response requests Pull request #2422 - Reject unused fields in LLM profile configuration and surface malformed AEAD_SECRET_KEY errors Pull request #2424
- Fix shared trace loading and implement responsive trace layout on shared trace pages Pull request #2399
- Refactor span message processing into separate stage functions for improved maintainability Pull request #2390
- Add
batch_insert_with_ttlto cache layer for efficient bulk Redis key operations Pull request #2442 - Withhold search snippets under masking policy on dual-mode projects Pull request #2393
- Apply masking policy to API key and CLI token reads on dual-mode projects Pull request #2392
- Remove CLUSTERING_ENABLED frontend feature flag Pull request #2420
- Add Redis cache key for temporary signal throttling per project Pull request #2439
Responses API now works with custom gateways. Fixed PII masking in dual-mode projects. Improved trace loading and responsive layouts. Better error handling for LLM profile setup.
This week we expanded LLM profile support with Responses API for custom gateways, strengthened PII protection through masking policies on dual-mode projects, and improved the trace viewing experience with responsive layouts and fixed loading behavior. We also refactored span processing into stage functions and added batch Redis operations for better performance.
Week of September 14, 2026
What shipped
- Dual mode for PII handling now stores canonical text alongside byte-range masks, letting different project members read appropriate versions of the same trace. Pull request #2361
- Fixed realtime SSE to respect dual-mode access policies so masked readers don't receive unmasked agent input over the channel. Pull request #2389
- Refactored PII redactor internals to use SpanVerdict, unify index space, and properly fail batches when project mode lookup fails. Pull request #2384
- Access policies derived from actor roles are now baked into span and trace reads through the query engine. Pull request #2360
- Added dataset CRUD API with create, read, update, delete, and row operations under both project and CLI auth paths. Pull request #2340
- Span detail view now skips reasoning blocks when extracting heuristic previews from reasoning model outputs. Pull request #1865
- Declared the docs sitemap in robots.txt to improve indexing of the 139 documentation URLs. Pull request #2374
- Added static SVG brand assets (icon and wordmark) under frontend/public/brand/ for frontend use. Pull request #2375
- Reverted force-commit-on-batch change to prevent unbounded memory growth in Quickwit at high traffic. Pull request #2362
Why it matters
The PII mask series (events 0–6) completes trace-level access control so projects can store raw text once with masks applied per reader. The dataset API (event 11) adds programmatic management of datasets. Smaller fixes improve span rendering and operational stability.
Changelog entry
- Dual mode for PII now available: stores canonical text and masks, with read paths respecting access policies per actor Pull request #2361
- Fixed realtime
agent_inputSSE to respect dual-mode masks for project subscribers Pull request #2389 - Refactored PII redactor: SpanVerdict type, unified index space, batch-level mode lookup failures Pull request #2384
- Access policies derived from actor roles now applied to span and trace reads Pull request #2360
- Dataset CRUD API: POST/GET/PATCH/DELETE datasets and rows via /v1 (project auth) and /v1/cli (user auth) Pull request #2340
- Span detail view now skips thinking blocks when extracting previews from reasoning model outputs Pull request #1865
- Reverted force-commit batching to prevent unbounded Quickwit memory growth at high traffic Pull request #2362
Shipped dual-mode PII handling: store text once, apply masks per reader. Added dataset CRUD API. Fixed realtime agent_input masking, span preview extraction, and Quickwit memory pressure.
This week we shipped the final piece of trace-level access control: dual-mode PII handling stores canonical text with byte-range masks, letting different project members read appropriate versions of the same trace. We also added a dataset CRUD API for programmatic dataset management, fixed realtime masking for dual-mode projects, improved span preview extraction for reasoning models, and resolved a Quickwit memory issue at high traffic.