Skip to content

v3.0.2

Choose a tag to compare

@github-actions github-actions released this 19 Oct 18:48
· 232 commits to main since this release
0a9de70

V3.0.2

Research through your own model router runs for real, and a reworked report's counts describe the report you approved.

  • Custom, OpenRouter and Ollama setups on a self-hosted server research your question. ./start.sh recognised only Google, Anthropic and OpenAI keys, so any other setup started in demo mode and every question came back as the same scripted report (#171, #177).
  • Approving a plan no longer fails the run on OpenAI-compatible and Anthropic models (#171, #177).
  • A rework no longer doubles a run's evidence, and "cited" counts against the revision you approve (#181).
  • The executor searches instead of scraping search-engine results pages (#178). Measured before shipping: citation support 96.9% over 190 judged claims on ten fixed questions (Gemini 2.5 Flash, judged by Claude Sonnet 4.5); one run, so not evidence of a change from 3.0.0's 96.4%.
  • Every build says which commit built it. The published server images report this tag's commit, and an image you build reports your checkout's (#170, #188).
  • Python 3.14 for the server and the desktop engine, Node 26 for the web image and CI, and Next.js 16.4.2 with updated sharp and source-map-js, past high-severity advisories.

macOS: the app is ad-hoc signed, not notarized. On first launch macOS says it could not verify the app; open it once with:

xattr -dr com.apple.quarantine "/Applications/Research Assistant.app"

Known issues — the full list is under Known in the changelog. The ones most likely to matter:

  • Desktop with Ollama and no API key: research fails at once with "No provider key found", and Settings offers no local model. Until the next release ships the fix (#202, merged), save any text as the Custom Endpoint key in Settings → Connections — Ollama ignores it.
  • More than five successful sign-ins to one account within 15 minutes lock it for the rest of the window (fixed on main, #200).
  • This release's bundle verifier needs Python 3.14; on an older Python, 3.0.1's verifier checks the same bundles.

Multi-arch container images (linux/amd64, linux/arm64) pushed to GHCR:

  • ghcr.io/adityamhaske/multi-agent-research-assistant-api:3.0.2
  • ghcr.io/adityamhaske/multi-agent-research-assistant-worker:3.0.2
  • ghcr.io/adityamhaske/multi-agent-research-assistant-frontend:3.0.2

latest points at this release.

Deploy the full stack on a free Oracle Cloud Ampere VM — see deploy/README.md.

What's Changed

  • release(v3.0.1): offer the published installers by @adityamhaske in #169
  • A mandatory release process, and an audit that runs it by @adityamhaske in #170
  • Fix gated runs failing on OpenAI-compatible routes, and every question getting the demo report by @adityamhaske in #171
  • Reset an agent's instructions to the shipped prompt from inside its editor by @adityamhaske in #179
  • An approval is not a 500 when its report cannot be filed into project memory by @adityamhaske in #180
  • fix(runs): a rework stores no second copy of the evidence, and the run page counts the latest draft by @adityamhaske in #181
  • Prove real-mode worker runs end to end, make loop ownership explicit, and label demo runs plainly by @adityamhaske in #177
  • Refuse search-engine results pages in read_webpage and steer the executor to web_search by @adityamhaske in #178
  • chore(deps-dev): bump brace-expansion from 1.1.18 to 1.1.21 in /frontend by @dependabot[bot] in #176
  • chore(deps): bump undici from 8.10.0 to 8.11.2 in /frontend by @dependabot[bot] in #175
  • chore(deps): bump the frontend-minor group in /frontend with 2 updates by @dependabot[bot] in #174
  • chore(deps): bump python from 3.11-slim-bookworm to 3.14-slim-bookworm in /backend in the backend-images group by @dependabot[bot] in #173
  • chore(deps): bump next from 16.3.5 to 16.3.6 in /frontend by @dependabot[bot] in #182
  • chore(deps): bump next from 16.3.6 to 16.3.8 in /frontend by @dependabot[bot] in #185
  • chore(deps): bump sharp from 0.35.4 to 0.35.5 in /frontend by @dependabot[bot] in #184
  • chore(deps): bump source-map-js from 1.2.1 to 1.2.2 in /frontend by @dependabot[bot] in #183
  • release(v3.0.2): prepare the release — notes, record, and the release evaluation by @adityamhaske in #186
  • fix(ci): test the server on the image's Python, and fail the commit that splits them (A10.06) by @adityamhaske in #187
  • fix(desktop): freeze the engine on the image's Python, so the suite tests both hosts by @adityamhaske in #189
  • fix(build): every build reports its own identity or unknown, never a stale stamp's by @adityamhaske in #188
  • release(v3.0.2): include #187-#189 in the record, and plan the checks they need by @adityamhaske in #192
  • fix(python): the advertised floor is the shipped Python 3.14, and A10.06 holds it there by @adityamhaske in #193
  • chore(deps): update numpy requirement from >=1.26 to >=2.5.4 in /backend by @dependabot[bot] in #190
  • chore(deps): bump the frontend-minor group across 1 directory with 3 updates by @dependabot[bot] in #191
  • chore(deps): bump node from 25-bookworm-slim to 26-bookworm-slim in /frontend in the frontend-images group across 1 directory by @dependabot[bot] in #172
  • fix(desktop): a packaged-app check whose bundle verification did not run fails CI by @adityamhaske in #194
  • fix(ci): build and test the frontend on the image's Node, and fail the commit that splits them (A10.07) by @adityamhaske in #196
  • release(v3.0.2): scope #192-#194, E1-E3 results, and the verifier's Python 3.14 floor by @adityamhaske in #195
  • release-audit(v3.0.2): scope the record PRs themselves (#195 and this one) by @adityamhaske in #197
  • fix(release-audit): A3.02 judges the site a release commit serves, not a run it cannot have by @adityamhaske in #198
  • fix(deps): patch postcss-selector-parser under the typography plugin (GHSA-rj75-hqrm-r3gf) by @adityamhaske in #199

Full Changelog: v3.0.1...v3.0.2

Desktop bundles for v3.0.2

These builds are not code-signed. Your OS will warn you on first launch:

OS What you will see What to do
macOS A block on first launch — the app is not signed with an Apple Developer ID or notarized Drag the app to Applications, then run xattr -dr com.apple.quarantine "/Applications/Research Assistant.app" in Terminal once
Windows SmartScreen warning More info → Run anyway
Linux AppImage is not executable chmod +x *.AppImage (.deb needs nothing)

Verify your download against SHA256SUMS before running it:

sha256sum -c SHA256SUMS --ignore-missing

Research bundles from v3.0.0 onward are format v2, and a verifier from before
v3.0.0 refuses them.
Check a bundle with the verify_bundle shipped in this
release; it still verifies every earlier (v1) bundle, unchanged.