AI recommendations
Generate AI fix recommendations from failed-job logs, give feedback on what helped, and let spike detection watch for sudden regressions.
The Recommendations page in the sidebar turns your pipeline telemetry into concrete optimisation advice. TrimCI feeds recent failed-job log excerpts and run statistics to a self-hosted AI model running on TrimCI's own server — your logs are never sent to a third-party AI provider — and returns prioritized cards with evidence and suggested actions. Everyone in the organization can read the recommendations; only owners and admins can generate new ones.
Simplified illustration — the real screens hold more detail, but the labels and steps match the app.
Step by step
-
Open Recommendations
Click "Recommendations" in the sidebar. Every member of the organization sees the full list; the "Generate recommendations" panel at the top is shown only to owners and admins. Generation needs pipeline data, so connect a provider and let the first sync finish before your first run.
-
Generate a batch
Choose the analysis window under "Analyze last" — 7, 30, or 60 days (30 is the default) — and press "Generate". TrimCI collects failure statistics and the stored log tails of recent failed jobs, redacts secrets and access tokens from them, and sends the sanitized excerpts to the local AI model.
-
Wait for the results
A status banner shows "Generating recommendations…" and refreshes itself every few seconds; a run typically takes 20–120 seconds. Only one batch can run at a time, and the page reloads automatically when the batch is ready or has failed. A failed batch does not count against your monthly quota.
-
Read the cards
Recommendations arrive grouped by severity, from critical to low. Each card carries a category, a confidence badge, the affected repositories (linked to the dashboard), and a "Show details" section with the evidence behind the advice and a numbered list of suggested actions. Cards older than two weeks are flagged as possibly outdated.
-
Give feedback
Answer "Was this helpful?" with a thumbs up or down at the bottom of any card — you can press "Change" to revise your vote. Feedback helps you track which advice your team has already evaluated.
-
Dismiss and restore
The close button in a card's corner dismisses it — the card is hidden, never deleted. Tick "Show dismissed" in the filter bar to see hidden cards and press "Restore" to bring one back. The same bar filters the list by severity and category.
Automatic spike detection
On the Business plan (and during the trial, which includes Business limits) TrimCI also watches for trouble on its own. Every hour it compares your failure rate over the last 24 hours with the 24 hours before that. When both windows have at least 5 runs and the rate jumps materially — at least 20 percentage points higher and at least doubled, with the current rate above 5% — it generates a recommendation batch automatically, at most once per 24 hours. Cards from these batches carry an "Auto" badge. On Free and Team plans this monitoring is off; generate recommendations manually instead.
Generation is limited per calendar month by plan: 5 batches on Free, 50 on Team, and 150 on Business. The panel shows how many generations you have left, and the counter resets on the 1st. There is also a one-minute cooldown between generations, purely to absorb double-clicks.
Privacy by design: the AI model is self-hosted on TrimCI's infrastructure, so your job logs never leave the server or reach an external AI provider. Before anything is shown to the model, known secret and token patterns are redacted from the log excerpts, and only a short tail of each failed job's log is stored in the first place — purged after 30 days.