Changelog
Release notes and version history synced from the Overseer desktop repository.
v0.1.4
STABLELATESTMay 9, 2026Summary
Overseer v0.1.4 overhauls the core engine with a new Architecture, delivering intent-aware synthesis, multimodal perception, grounding decoupling, and user attachment visibility — alongside the first binary size optimization pass for desktop.
HIGHLIGHTS
- +**Intent-Driven Synthesis**: The final synthesizer now matches response detail to the user's actual question — concise answers for specific queries, brief summaries for general descriptions, and full evidence only when explicitly requested.
- +**3-Tier JSON Robustness**: A multi-stage extraction pipeline prevents internal LLM artifacts from leaking into user-facing responses.
- +**Conversational Intelligence**: Simple greetings bypass the Lattice for low-latency direct chat. Creative tasks preserve narrative framing instead of collapsing into status reports.
- +**Tiered Media Perception**: Multimodal inputs are processed through native Rust parsers, local OCR, or a query-guided Vision Assistant depending on complexity.
- +**Grounding Decoupling: Generic image descriptions are cached permanently while turn-specific perception is transient, eliminating contextual echo.
- +**User Attachment Visibility**: User-sent file attachments now render inline in chat bubbles with download capability and immediate optimistic display.
- +**Release Profile Optimization**: LTO, symbol stripping, and single codegen-unit compilation for significantly smaller release binaries.
OPERATOR NOTES
- !Evidence governance now only surfaces sources when explicitly requested — vision and media queries receive natural conversational answers.
- !Grounding enforcement uses silent delivery for operational tasks, suppressing technical noise in favor of clean success messages.
v0.1.3
STABLEMar 28, 2026Summary
Artifact-native workspace sessions, explicit work items, and more mature selector-first Windows desktop control.
HIGHLIGHTS
- +Artifacts are now first-class outputs that can be inspected, selected, reopened, and reused across follow-up tasks.
- +Work items now act as explicit tracked units of work for humans and agents instead of loose instructions buried in chat.
- +Persistent workspace sessions let artifact and work-item surfaces stay open in the main pane.
- +Windows desktop control now leans further into selector-first UI Automation flows, app hints, blocked diagnostics, and post-action verification.
OPERATOR NOTES
- !Native local tasks like screenshots, app launch, file inspection, and some accessibility-driven actions now survive more model-backend outages through local fast paths.
- !UIA selector misses now fail clearly with retryable diagnostics, candidate hints, and search-scope details.
- !Public packaged release availability is still Windows-first, while macOS and Linux remain source-build paths for now.
v0.1.2
STABLEMar 25, 2026Summary
Native computer fast paths, reliable screenshot-to-chat rendering, and stronger machine-state inspection across the desktop operator flow.
HIGHLIGHTS
- +Added native computer fast paths for screenshots, background-process inspection, file search, browser state, file open, and file read requests.
- +Screenshot capture now completes through the native desktop tool path instead of fragile UI detours like opening Snipping Tool.
- +Assistant responses can now attach and render generated desktop screenshots directly inside chat.
- +Computer Control now emphasizes the live session state, pending approval, and the core safety controls without the noisier secondary fields.
OPERATOR NOTES
- !Machine-state requests now prefer native inspection first, model reasoning second, and UI automation last.
- !Screenshot tasks treat the saved screenshot artifact itself as grounding evidence for successful completion.
- !Desktop screenshot replies are cleaner and no longer dump raw artifact markup into chat.
v0.1.1
STABLEMar 23, 2026Summary
Computer supervision, approval-backed desktop control, runtime health checks, and sharper chat, trace, and workflow surfaces.
HIGHLIGHTS
- +Added Computer Control with approval-required, trusted, observe-only, pause, and emergency-stop session handling.
- +Desktop and computer tools now respect session policy, pending approvals, action caps, timeouts, and allowlist targets.
- +Added saved tool test cases so operators can keep sample inputs and expected outputs inside the Tools Workshop.
- +Improved Chat with auto-created conversations, starter prompts, routing chips, and clearer failure handling.
- +Improved Execution Trace with category filters, better labels, and expandable raw event details.
- +Improved Workflows with clearer run summaries, status badges, and rerun-focused operator feedback.
- +Added runtime health checks for Ollama, model availability, Python, and `protoc`.
OPERATOR NOTES
- !Approval-required desktop actions now stage a pending action in Computer Control instead of running immediately.
- !A one-shot approval only arms the next desktop action, which keeps broad computer sessions bounded.
- !Follow-up messages like `continue` now stay inside the active computer session instead of being re-routed as unrelated tasks.
- !Older Python tools that import `requests` can now run from the staged runtime without a separate package install.
v0.1.0
STABLEMar 22, 2026Summary
First packaged desktop release with workflow runs, execution traces, and local installer support.
HIGHLIGHTS
- +Workflow runs now track real execution status instead of assuming success.
- +Conversations and workflows expose execution trace timelines.
- +Agents and tools support version snapshots and rollback.
- +Runtime dependency handling is better on Windows, including `protoc`, Python, and tool staging.
- +Release packaging now stages a portable build with checksums and install assets.
OPERATOR NOTES
- !This release is still local-first and expects local model/runtime setup such as Ollama and Python.
- !Existing development databases may require a clean reset if older migration checksums no longer match the current repo history.
- !Packaged Windows builds include `tool_scripts` so first-run tool seeding works outside the repo layout.