macOSDMG
Apple silicon
M-series Macs · signed and notarized
- File
- wisp-science_0.20.0_aarch64.dmg
- SHA256
5123746ab24edb0735622f3abb0c216498060935a560a4bebe2113fa37b1d6a0
CURRENT RELEASE · WISP SCIENCE
Wisp Science v0.20.0
Choose the installer for your device, then verify the filename and SHA256 before installing.
01 · DOWNLOAD
Download installers
Installers are served by GitHub Releases. Use the SHA256 value to verify the downloaded file.
macOSDMG
Intel
Intel Macs · signed and notarized
- File
- wisp-science_0.20.0_x64.dmg
- SHA256
3db9e61a7e3c1e2a062509ae6e2ce953126985746fde71b102474b5ac06fa0eb
WindowsEXE
Setup EXE
Standard Windows installer · unsigned
- File
- wisp-science_0.20.0_x64-setup.exe
- SHA256
83fd9189bdf6bc6b7cb8f9fb52fa82d4725bc845d6bf4f28121446946d49dcc0
WindowsMSI
MSI package
Enterprise-style deployment · unsigned
- File
- wisp-science_0.20.0_x64_en-US.msi
- SHA256
94a8e57cb50d5632c6561359c362cab1fed6ef6eac9b6a5b645a6dbd7569eb6b
02 · RELEASE NOTES
What changed
wisp-science v0.20.0 puts context management under explicit user control. Automatic history rewriting is gone: Wisp warns near the context limit, while /compact archives the full conversation before reducing model context. This release also adds a read-only explore subagent, migrates bundled scientific skills to Wisp's native Run/tool contract, separates progress, reasoning, and tool activity in chat, and improves per-reply token accounting and model settings.
What's changed
- Stop automatic token/age compaction; warn at 80% and let users run archive-first /compact from the CLI or desktop chat.
- Cap shell, Python, and R results when they enter model context while retaining full UI output; archive isolated read-only exploration traces under .wisp/subagents/.
- Migrate bundled GPU, figure, paper, literature, PDF, customization, and skill-authoring workflows away from the unavailable host.* SDK to explicit Wisp Runs and tools.
- Add the journal-club-ppt workflow and catalog-wide checks for legacy host contracts and GPU Run lifecycles.
- Render progress commentary, collapsed reasoning, and consecutive tool groups in wire order; show the current in-flight step even while Steps is collapsed.
- Display and persist per-reply input, output, cache, and reasoning tokens across supported providers.
- Add model-aware reasoning-effort choices, automatic known context/output limits, and drag-to-reorder model profiles.
- Stop repeated tool-call loops, drop historical OpenAI-compatible reasoning replay, and reject binary files in read with actionable guidance.
- Stabilize SQLite snapshots and write transactions, resolve Run cancellation races, and fix PPTX/XLSX preview sizing, PNG cropping, and cross-session side chat.