Documentation › Support
Explain a failure with AI
One click: AI reads a red build's log and tells you what broke, why, and how to fix it. Production failures can be explained automatically.
A red build leaves a long log. On a failed build, Explain this failure ✨ sends the relevant part of it to an AI model (Claude, by Anthropic) and shows, in plain English:
What failed : the step and the error.
Why : the most likely cause, with the module, file, line, field or Python package when the log names it.
How to fix : concrete steps for your repository (a line for requirements.txt , the XML to correct, the test to change…).
Confidence : high, medium or low. Low means the log doesn't say enough: check it yourself or ask support.
Where to find it
The Explain button on a red build's line in the project's Builds tab and in the branch's History : it opens the branch's Logs tab with the explanation.
The branch's Logs tab, above the log of a failed build (a failed deploy, failed tests on development…).
The progress page of a failed Odoo install on your server .
A failed restore in the branch's History : Explain on its line. The build it restored into keeps running, so the AI reads only the part of that build's log the restore wrote ( Log opens the log there). A failed database import is a build of its own: its Explain opens that build's explanation.
A failed backup in the Backups tab: a backup keeps no log, only its error output (for example pg_dump 's message), and Explain reads that. A failed backup that kept no error output says so: there is nothing to explain.
The lines it relies on are links: line 42 opens the log at that line and highlights it. Copy copies the explanation, and Still stuck? Ask support opens a support request with the explanation and a link to the build already filled in: add what you tried and send it.
What is sent, and what isn't
Never the whole log: the end of it (the step that failed) and the lines around its errors, about 12,000 tokens at most, with the Odoo version, edition, stage, branch and the build's error.
Before anything leaves the platform, secrets are removed: passwords, API keys and tokens (GitHub, Stripe, AWS, Anthropic…), Bearer headers, credentials in addresses, private keys, your project's master and database passwords, and email addresses.
Customer data that database errors print (the values of DETAIL: Key (…)=(…) , Failing row contains , a failed query's values, long text and JSON) is replaced too.
The log is treated as data, never as instructions: text in your log can't change what the AI does, and its answer only links to your log and to odoo.com, github.com, python.org or pypi.org.
Cost and limits
An explanation is kept with the build: opening it again, by you or a teammate, is instant and free. Explain again asks for a new one. Each project has a daily number of explanations set by the platform administrator; the button says when today's are used up. Who can use it is who can read the build's logs (the stage's role, see Collaborators and roles ). With an API token, reading an explanation already given needs the read scope ( GET /api/builds/<id>/explanation ), and asking for one needs builds:write (or admin ), like starting a build ( POST /api/builds/<id>/explanation ).
Production failures, explained automatically
When the platform administrator turns it on, a failed production build (and staging, if chosen) is explained right away, in the background: the answer is already there when you open the build. The failure email and the bell notification also get a two-sentence summary (what failed and the first step to fix it) with a link to the full explanation. The email waits up to 25 seconds for it and is sent without it when the AI is slower, unavailable or today's limit is used, so it is never held back longer. Phone and desktop push notifications keep their short generic text: they never show build errors on a lock screen.
👍 or 👎 Tell us whether the explanation helped: platform administrators see the answers that didn't, to improve them.
AI can be wrong. Read the cited log lines before you change production, and test the fix on a development or staging branch first.
Platform administrators : the build doctor uses the Anthropic API key of the help chat ( Platform → AI & help → Help chat ). Without it customers don't see the button and you see a reminder instead. Platform → AI & help → Build doctor turns it on or off, picks the model (Claude Haiku 4.5 by default: fast and cheap for reading logs; Sonnet 5 or Opus 5.5 for harder failures), sets the daily limits per project and for the platform, turns on Explain production failures automatically (off by default) with the stages it covers, and shows the cost and the feedback (automatic explanations are marked). Automatic explanations count toward the same daily limits.