Repository navigation
v0.25.2-rc2
Pre-releaseImportant
Release candidate for v0.25.2. A pre-release for testing before the final ships. Installers and package managers keep offering the last final release unless you ask for this one.
GAIA v0.25.2 Release Notes
What's New
- GAIA chats show what fills the context window, and skills that work together now combine in one turn. (PRs #5746, #5422)
- A write request relayed from another window names where the file will land. (PR #5783)
Bug Fixes
A selection of the user-visible fixes; the full list is in the changelog below.
gaia initinstalls the GAIA agent even when Lemonade or model setup fails, and the Windows installer keeps%VARS%in your user PATH. (PRs #5802, #5809)- An update replaces a running agent and restarts it, and a crashed Lemonade check is reported as a failed check. (PRs #5747, #5779)
- GAIA no longer reports Lemonade ready when the context-size re-check fails, and cloud models no longer set the local context ceiling. (PRs #5793, #5813)
- A request to run something now gets an answer back when nothing ran, instead of silence. (PR #5753)
- Claude models keep their reasoning across tool calls, because thinking blocks are sent back with their tool turn. (PR #5072)
- One malformed tool call no longer ends the turn — non-object tool arguments are recovered. (PR #5732)
- Shell checks understand Windows paths with backslashes, and
ghrepo names and API routes are no longer treated as file paths. (PRs #5794, #5751) run_pythonruns in a scratch folder when the repo around it is not approved. (PR #5768)- Text files saved in Windows ANSI (CP1252) open in
read_fileand in document search, as do UTF-16 text files; unreadable Excel files give an actionable error and titled sheets read correctly. (PRs #5766, #5761, #5812) - Files with an uncommon extension can be found by name. (PR #5754)
- Document indexing no longer overflows on dense tables, and vision-extracted blocks stay whole when a document is split into sections. (PRs #5742, #5817)
- Agent UI follow-through — a tool row waiting for your approval says so with no timer; an agent that stopped answering is not blamed on the model server; the model chip follows a switch made in another window; Review changes keeps a last line without a newline on its own row; always-on skills are labelled instead of offering Use; the oldest cached chat agents are evicted. (PRs #5786, #5785, #5748, #5749, #5731, #5741)
- The desktop app shows "needs you" in the tray for prompts from another window, and the Linux window keeps its system title bar. (PRs #5752, #5807)
- The terminal UI says "send it again" when the agent crashes instead of naming a daemon API call. (PR #5750)
- A follow-up email request can name a thread shown last turn, the email agent streams its
/queryanswer, and the suspicious-mail scan reports a mailbox refresh error. (PRs #5811, #5767, #5782) - Skills prefer stable releases unless you ask for prereleases, and a manifest with no version is not flagged as drifted. (PRs #5755, #5778)
- Memory searches with a date-only end date include that final day. (PR #5818)
- Failures are no longer hidden —
gaia mcpcommands exit 1 on failure, TTS self-tests exit non-zero when synthesis fails, agent package loading errors surface, connector token-parse errors are kept, and undo-history save failures no longer mask the query error. (PRs #5777, #5788, #5745, #5795, #5734) - Settings list recommended Fireworks models first. (PR #5781)
- Eval fixes — an invalid
--baseline-refis an error, and contracted negations ("doesn't") are recognized in claims. (PRs #5784, #5808) - On macOS, the agent's shell sees your login PATH when GAIA is opened from Finder or the Dock, so tools installed with Homebrew or similar are found. (PR #5867)
- Chat sessions are kept when saving fails, instead of being lost. (PR #5862)
- Scratchpad measures the real database size and refuses to re-create a table that conflicts with an existing one. (PR #5856)
- File search stays out of other apps' data on macOS. (PR #5854)
gaia initreplaces a release candidate's agent build with the published one, so final-release agent fixes reach people who updated from an RC. (PR #5831)- Free disk space is reported correctly, without the agent detouring through a shell. (PR #5849)
- Goals in the Agent UI — deleting a goal no longer reports a failure, and on the desktop app, where goal runs are not supported yet, a failed background goal run now stops and says so instead of retrying in a loop. (PRs #5839, #5841)
- Undo says when macOS is asking for permission instead of hanging. (PR #5846)
- Docs — Full access still asks before a gated shell command, script shell commands time out after 10 minutes,
max_stepsdefaults to 50, the eval startup reservation is 240 s, email install examples drop--trust, and the stateful email agent examples show their auth headers. (PRs #5769, #5758, #5759, #5763, #5780, #5760)
Contributors
Thanks to the community members who contributed to this release:
- @xubenqiang — finding files by name outside the extension allowlist, email streaming and refresh errors, installer and skills version checks, and eval baseline and doc fixes (PRs #5754, #5767, #5782, #5779, #5778, #5784, #5769, #5763)
- @maheshrijal — stable-release preference for skills,
gaia mcpexit codes and agent package loading errors (PRs #5755, #5777, #5745) - @jipeng6036-del — recovering from non-object tool arguments, dense-table indexing and undo-history error handling (PRs #5732, #5742, #5734)
- @TejasS1233 — VLM blocks kept whole in section chunking and date-only
time_toin memory search (PRs #5817, #5818) - @amandeavor — documenting the
max_stepsdefault, the script shell timeout and auth headers in the email agent examples (PRs #5759, #5758, #5760) - @xueyilan137-maker — excluding cloud models from the local context ceiling and keeping connector token parse errors (PRs #5813, #5795)
- @huong8373tt-beep — contracted negation in eval claims and a steadier hub pip test (PRs #5808, #5814)
- @erSpell — CP1252 and UTF-16 text files in document search (PR #5761)
- @AyobamiH — labelling always-on skills in the Agent UI (PR #5731)
- @chris00234 — evicting least-recently-used cached chat agents and keeping chat sessions when saving fails (PRs #5741, #5862)
- @akshat-lakhera — not reporting Lemonade ready when the context re-check fails and measuring the real scratchpad database size (PRs #5793, #5856)
- @d4rshnn — recommended Fireworks models first in settings (PR #5781)
- @dannyward630 — tests for Lemonade cleanup boundaries (PR #5757)
- @KrBeTect2025 — TTS self-tests exiting non-zero on failure (PR #5788)
- @swastiktawate — dropping
--trustfrom email install examples (PR #5780)
Full Changelog
67 commits since v0.25.1:
f404f5b2d— fix(agent): report free disk space correctly, without a shell detour (#5849)2e724c47d— fix(agent): the agent's shell sees the user's login PATH on macOS (#5867)515e5c016— fix(tools): keep file search out of other apps' data on macOS (#5854)9475554f8— fix(init): a final release replaces a release candidate's agent build (#5831)c8fd6b377— fix(ui): undo says when macOS is asking, instead of hanging (#5846)8412f5c71— fix(ui): a failed background goal run stops and says why (#5841)2160bc5d9— fix(ui): deleting a goal no longer reports a failure (#5839)8a78027fd— fix(scratchpad): measure the real database size and refuse a conflicting re-create (#5856)c02cb28d1— fix: preserve chat sessions when saving fails (#5862)5f9b9af0a— Release v0.25.2 (#5850)e2a195c7b— feat(skills): contributor-triage skill for PR and issue triage (#5816)ca6e51b6a— docs(email): add auth headers to stateful agent examples in SKILL.md (#5760)e44fd39c0— fix(memory): include final day for date-only time_to (#5818)e7ddce1de— fix(ui): a relayed write prompt names where the file lands (#5783)13945c69e— fix(skills): tell the model to combine skills, add pairs_with companions (#5422)c46be4737— chore(deps): bump the root-npm-dependencies group with 2 updates (#5796)172384f1e— chore(deps): bump the agent-ui-dependencies group in /src/gaia/apps/webui with 6 updates (#5798)0d0882b81— chore(deps): bump anthropics/claude-code-action from 1.0.252 to 1.0.261 in the github-actions group (#5800)05a07660f— fix(agent): send back an answer to a run request when nothing ran (#5753)cc0e9d022— docs: Full access in the Agent UI still asks before a gated shell command (#5769)0dc738008— fix(email): surface the mailbox refresh error in the suspicious scan (#5782)e21d704be— fix(rag): handle CP1252 and UTF-16 text files (#5761)59dcdb14b— docs(eval): the startup reservation is 240s, not 120s (#5763)5b842d93d— docs(sdk): document max_steps framework default of 50 (#5759)c6a4ef025— docs(guides): document 10-minute timeout for script shell commands (#5758)cbb2b3886— fix(rag): keep VLM blocks atomic in section-based chunking (#5817)0a81860d1— test(hub): stabilize pip stuck detection test (#5814)05f74ccd8— fix(eval): recognize contracted negation in claims (#5808)e18cbea23— test(installer): cover Lemonade cleanup boundaries (#5757)8fc1ed77e— fix(eval): treat an invalid --baseline-ref as an error, not first adoption (#5784)e3090254e— fix(lemonade): don't report ready when the context re-check fails (#5793)5eee6693b— fix(settings): prioritize recommended Fireworks models (#5781)c30ad2792— fix(skills): prefer stable releases unless prereleases are requested (#5755)dda6d5bff— fix(rag): actionable errors for unreadable Excel files; read titled sheets right (#5812)cf32d2fc5— fix(email): a follow-up can name a thread shown last turn (#5811)e5661858a— fix(installer): keep %VARS% in the user PATH when install.ps1 adds GAIA (#5809)a1a846e23— fix(electron): Linux desktop window keeps its system title bar and controls (#5807)cfc7e8c27— fix(init): install the GAIA agent even when Lemonade or model setup fails (#5802)65f5700c2— fix(shell): check a Windows path written with backslashes (#5794)401e61f69— fix(agent): run_python runs in scratch when the repo around the cwd is unapproved (#5768)38f8047e6— fix(tools): read_file opens CP1252 (Windows ANSI) text files (#5766)34ec0fcab— fix(ui): a tool row held for your approval says so, with no timer (#5786)ca6460852— fix(ui): an agent that stopped answering isn't blamed on the model server (#5785)bfe6a85b0— fix(electron): show "needs you" in the tray for prompts from another window (#5752)c17ea45e0— fix(shell): gh repo names and API routes are not file paths (#5751)558b307a3— fix(tui): say "send it again" when the agent crashes, not a daemon API call (#5750)22cafbba9— fix(ui): keep a line without a trailing newline on its own row in Review changes (#5749)7bd560f5c— fix(ui): the model chip and footer follow a switch made in another window (#5748)9ccd888b6— fix(installer): an update replaces a running agent and restarts it (#5747)84e41ce76— fix(ui): show what fills the context window for GAIA chats (#5746)afd6941db— fix(llm): send thinking blocks back with their tool turn on the Claude provider (#5072)ebff3c527— fix(tools): find a file named with an extension outside the allowlist (#5754)e03e8d209— fix(ui): label always-on skills instead of offering Use (#5731)acac3d1cd— docs(hub): drop --trust from email install examples (#5780)d7b2c7202— fix(llm): exclude cloud models from local context ceiling (#5813)26a2ae1ba— fix(connectors): preserve unexpected token response parsing errors (#5795)00e9c0d99— fix(audio): tts self-tests exit non-zero when synthesis fails (#5788)df39b208b— fix(installer): report a crashed Lemonade check as a check failure (#5779)362ebb2dc— fix(skills): an omitted manifest version is not version drift (#5778)2827acdd1— fix(gaia): preserve query errors when undo history cannot be saved (#5734)fa95dfa17— fix(mcp): exit 1 when a gaia mcp command fails (#5777)e1fe4b4b2— fix(email): stream the email agent's /query answer (#5767)70834de88— fix(cli): surface agent package loading errors (#5745)951fc4052— fix(rag): index dense no-break-space tables without embedding overflow (#5742)6d3f07f9f— fix(ui): evict least-recently-used cached chat agents (#5741)63538f4da— fix(agents): recover from non-object tool arguments (#5732)0a7fc1e3c— docs(plans): one-tag release train (#5172) (#5492)
Full Changelog: v0.25.1...v0.25.2
Installing this release candidate
Desktop app and terminal
Download the installer for your platform from the assets below:
- Windows —
gaia-0.25.2-rc.2-win-x64-setup.exe: one installer for the desktop app, the terminal, or both.gaia-agent-ui-0.25.2-rc.2-x64-setup.exeis the desktop app on its own - macOS —
gaia-agent-ui-0.25.2-rc.2-arm64.dmg(Apple Silicon only) - Linux —
gaia-agent-ui-0.25.2-rc.2-amd64.deb(Debian/Ubuntu) orgaia-agent-ui-0.25.2-rc.2-x86_64.AppImage(any distro)
The terminal on its own:
- macOS —
gaia-0.25.2-rc.2-darwin-arm64.pkg(Apple Silicon) orgaia-0.25.2-rc.2-darwin-x64.pkg(Intel) - Linux —
gaia_0.25.2-rc.2_amd64.deb(Debian/Ubuntu) orgaia-0.25.2-rc.2.x86_64.rpm(Fedora/RHEL/SUSE). They install as version0.25.2~rc.2, so the final release upgrades over them
Installer checksums are in gaia-installers-SHA256SUMS.txt. The terminal hub is also attached as standalone binaries (gaia-<os>-<arch>, checksums in gaia-tui-SHA256SUMS.txt).
The GAIA agent itself is attached as gaia-agent-<os>-<arch>, built from this candidate's commit. This candidate's gaia init and gaia hub install gaia install it from here rather than from the Agent Hub, and say so if it is missing.
One-line installer
Installs this candidate's Python CLI, terminal and GAIA agent together:
curl -fsSL https://amd-gaia.ai/install.sh | GAIA_RELEASE_TAG=v0.25.2-rc2 sh$env:GAIA_RELEASE_TAG = "v0.25.2-rc2"; irm https://amd-gaia.ai/install.ps1 | iexPython CLI
pip install --pre amd-gaia # newest pre-release
pip install amd-gaia==0.25.2rc2 # exactly this onenpm
npm install -g @amd-gaia/agent-ui@nextFound a problem? Open an issue and mention v0.25.2-rc2.