Appearance
Signal
Signal is Liddr's digest experience. Instead of asking a question, you open Signal to see what mattered across your connected sources — already gathered, categorized, and linked to the evidence behind it.
A digest is organized into four cards — What changed, Risks, Decisions, and Recommendations — so you can skim it in about thirty seconds and know where to look next.
The four cards
Every Signal digest routes recent activity into four cards:
- What changed — the concrete movement: PRs merged, issues completed or created, docs published, deployments shipped.
- Risks — things that need attention: blocked issues, security-group or IAM changes, and other signals that tend to mean trouble.
- Decisions — choices that were made and recorded, surfaced so they don't get lost in the stream.
- Recommendations — synthesized suggestions for what to do next, derived from the rest of the digest rather than detected directly from a single event.
Each item in a card is one sentence of context. Empty cards are normal and expected — if nothing risky happened, the Risks card is simply empty.
Source chips
Every item carries its own source chips — small attributions showing where the item came from, such as GITHUB, PLAT-1990, INC-241, or ADR-018. When the underlying event has a real URL, the chip links straight to it; when it doesn't, the chip is shown as a plain reference. The chips are how a digest stays accountable: each line points back at the activity that produced it.
Trust the chips
Signal never fabricates URLs, names, numbers, or incidents. If an item cites a ticket or PR, that reference came from real activity in your sources — follow the chip to confirm.
How detection works across sources
Signal is built on a deterministic layer that runs before any model is involved:
- Detect. Source-specific detectors turn raw activity (a merged PR, a blocked Jira issue, an AWS deployment) into typed signal events.
- Score. Events are scored for relevance so the highest-signal items rise to the top.
- Cluster. Related events are grouped — for example, "3 PRs merged in the same repo today" becomes one item rather than three.
- Categorize. Each event is auto-routed to a card by its type. Merges and completions land in What changed; blocked work and risky infrastructure changes land in Risks; recorded choices land in Decisions.
- Narrate. The model writes each item's one-sentence summary and authors the Recommendations card, working only from the clustered, pre-categorized data and the source references attached to it.
Because the detection, scoring, and categorization are deterministic, the model's job is narration, not judgment about what happened — which is what keeps the digest grounded.
Grounded by construction
The cards are assembled from detected events and their source references. If the model call fails for any reason, Signal falls back to a deterministic rendering of the same clustered data, so you still get a usable, source-linked digest.
Where Signal appears
Signal is its own surface in the Liddr app. Each run is recorded as a digest run — capturing when it ran, the time window it covered, the four-card payload, the events that contributed, and the approximate cost — so a digest is reproducible and auditable after the fact.
Staying grounded across any source
Signal works against any connected source that produces signals — the more you connect, the fuller the picture. Because every item traces back to a detected event and its chips, the digest reflects your real activity, not a generic summary. To broaden what Signal can surface, connect more sources; see Ingest connectors.