Skip to content
GitHubBuy Me A Coffee

Vault Operator 3.8 ​

3.8.3: community fixes, Gemini OAuth and Italian support ​

  • Failed contextual-enrichment jobs stop repeated provider calls and leave unfinished chunks available for a later retry. Local providers use a circuit breaker without normal cloud rate limiting.
  • Database autosaves leave at least 30 seconds between completed background snapshots. Explicit saves and orderly shutdown still wait for persistence. Tag checks avoid repeated comparisons of identical tag spellings, and implicit-connection discovery yields more often during long calculations.
  • Failed tasks retain their recovery snapshot and checkpoint markers. Active-note context includes a bounded excerpt of the note captured when you send, including unsaved editor text, inside the existing untrusted-content boundary.
  • Complete JSON requests use byte-accurate Content-Length for compatibility with local gateways. GitHub Copilot setup supports Enterprise Cloud accounts on GHE.com and follows the API endpoint returned by authentication.
  • Optional Gemini API OAuth uses your own Google Cloud Desktop client and quota project. Native streaming, tools, images, discovery and embeddings are supported; existing API-key connections are unchanged.
  • Vault search continues beyond the first 500 files. Selecting a custom prompt inserts editable text. The skill file-write wrapper preserves normal note writes and restores the original adapter on unload.
  • Incoming-link blocks count unique existing source paths and use explicit targets for nested notes. Optional row limits prioritize navigation notes; archive exclusions and visible source counts keep long lists readable. Selected notes can be refreshed with scope approval, complete checkpoints and cancellation. Existing settings retain the full graph-producing list.
  • Optional Ollama Qwen reranking uses validated yes/no token probabilities through the Generate API, with a capability check and fallback to the existing local reranker. The local default stays unchanged.
  • Italian is available, based on ivseb's community translation, with the additional current settings translated and checked for key and placeholder parity.
  • Updated vulnerable dependencies within the existing version ranges. Corrected settings guidance, sandbox return-value examples, and the documented boundaries of citation provenance and provider authentication.

The template compatibility fix and Enterprise routing have automated regression coverage; they have not been verified against every third-party template or a maintainer-owned Enterprise tenant. Live Google consent and inference require your own Cloud project and were not verified with maintainer credentials. Ollama server/model compatibility can be checked under Providers > Embeddings > Local reranking; no live 4B-model benchmark is claimed.

3.8.2: harness improvements ​

Released September 15, 2026.

  • A smaller task-focused prompt and tool catalog reduce recurring context overhead. Large tool results remain available in full, including during fast-path execution.
  • Independent read operations run with bounded concurrency. Context compression stops retrying when it does not save space.
  • A work journal preserves hypotheses, artifacts and checks across context compression and task recovery. Local verification and isolated experiments help the agent test its answers without an extra model call for each check.
  • Compatible provider adapters preserve native reasoning continuation. The shared harness continues to support all providers and models with the existing model choices and reasoning budgets.
  • Deep ingest receives the configured attachment and note folders as filtered context. Sending waits for attachment processing, so a file being added is included in the request.
  • Meeting notes accept paged Plaud transcripts.
  • Updated fast-uri and sharp to address dependency security advisories.

These changes do not establish ARC-AGI parity or a guaranteed cost reduction for every task and model. See quality and cost for the verification behavior.