Weekly development log 2026‑W37: Preserving memory choices and exposing failures after connection · partial week

01

What changed

Following W36's code navigation and memory work, this week closed gaps between saved preferences and actual execution: silent export failures, a default model lost after login, and a settings lookup failure that could re-enable memory. OpenDART and public-data connectors also joined the catalog. This is an interim account of 27 main commits through September 8, 23:15 KST.

• Memory export queried removed columns and returned an empty array on failure. It now uses the current schema and records incomplete queries in _meta.partial and failedCategories. That visibility exposed two more stale-column queries in skill and custom-agent exports, prompting a follow-up fix. Automatic extraction and backfill gained audit records, with tests for user isolation at the prompt-injection boundary (#783, #785).
• If the server cannot read memoryLearning, the turn skips memory injection and automatic storage while chat continues. Falling back to a client flag could previously re-enable a feature the user had disabled. Release 1.52.1 closes that path; it does not delete stored memories (#802).
• Automatic grep_code and repo_map discovery excludes credential files and reports how many were skipped. Under the high-risk approval policy, sensitive writes also require approval. Explicit path reads and shell access retain their own gates, so this is not a blanket secret-access block. Local tool definitions gained strict:true so vLLM enforces argument schemas during decoding (#777, #779, #781).
• Saved default models now survive login, and a saved model absent from the current list is no longer shown as automatic. Follow-ups corrected the B.AI probe token floor, displayed saved role assignments outside the catalog, and fixed selector colors. API-key settings gained provider-specific key issuance and authentication links (#789, #791, #794, #796, #798).
• OpenDART and four data.go.kr MCP servers joined the installation catalog, alongside environment-variable defaults and duplicate-start prevention. The read-only ops_metrics tool is administrator-only and exposed for operational questions. Nightly evaluations now retain the beginning of failed responses for diagnosis. Catalog availability is not evidence that every external integration has been exercised successfully (#793, #800, #786, #787).

02

How this week was reconstructed

This week is still in progress. Counts and claims stop at the GitHub main snapshot of September 8, 2026, 23:15 KST (5321ea66). Counts include release commits and use KST week boundaries; changelog release dates may use a different timezone.

03

Team identity and project roles

OpenMake Team is the community identity; openmake_llm is the software project; OpenMake is the product name. Non-developer maker riskpw leads openmake_llm through vibe coding, and professional developer rocky supports its development.