Clickfuse
GitHub Watch demo Live demo
ClickHouse + Trigger.dev Build Week

Ask one SRE question.
Get a proof board, not a paragraph.

Clickfuse is a chat agent that turns “Why did checkout latency spike after the 14:32 deploy?” into an evidence-backed incident board — anomaly timeline, service heatmap, before/after diff, suspect ladder, error-budget impact, and a verdict card.

Open live demo View source Watch video
  • Fixture mode works offline — no credentials required
  • Every panel opens a “Show evidence” drawer with SQL
  • MIT licensed prototype for one strong incident story

The problem

Incident response still starts with a wall of text.

Logs, traces, and dashboards exist — but when checkout breaks after a deploy, SREs still stitch the story together by hand. Generic “chat with your data” returns prose. Clickfuse returns a board you can defend in a war room.

Typical AI answer

“It looks like latency increased around the deploy. You may want to check the payment service and recent config changes…”

  • No visual proof
  • No query provenance
  • Hard to act on under pressure

Clickfuse board

Timeline · Heatmap · Diff · Suspects · Error budget · Verdict — each with Show evidence → SQL + window + confidence

  • Deterministic demo story
  • ClickHouse-backed panels
  • Trigger.dev task orchestration

Incident proof board

Six panels. One story. Full evidence.

The chat input starts the investigation. The response is the visual, interactive artifact — assembled progressively so the narrative lands before the last panel paints.

⏱

Anomaly timeline

p95 latency with rolling baseline bands and a hard deploy marker at 14:32. The break is visually obvious — at least 2× baseline.

▣

Service heatmap

Click a service to highlight it across the board. payment-service lights up red; peers stay cool gray/green.

Δ

Before / after diff

Span-level amplification from a single CTE query. payment spans jump ~8.3× while inventory and shipping stay flat.

⇅

Suspect ladder

Ranked root-cause candidates. payment-service v2.4.1 sits at #1 with the retry-timeout config diff called out.

⌀

Error-budget impact

Demo-visible burn so the business cost of the spike is part of the board, not a follow-up tab.

✓

Verdict + action

A clear close: roll back v2.4.1 or restore the 3s retry timeout — grounded in the panels above it.

Show evidence on every panel

Each drawer exposes the ClickHouse query, time window, row count, confidence, and Trigger.dev task context. Timeline includes both the raw http_logs anomaly-band query and a quantileMerge proof against latency_rollup_1m (AggregatingMergeTree State/Merge). 

SELECT minute, service, p95_ms,
  rolling_avg + 3 * rolling_stddev AS upper_band,
  p95_ms > rolling_avg + 3 * rolling_stddev AS is_anomaly
FROM baselines
WHERE rolling_avg IS NOT NULL
ORDER BY minute, service;

How it works

One question. Five tasks. A board you can trust.

  1. 01

    Ask the incident question

    Start from chat with the frozen demo sentence — or render the fixture board offline for judges and demos without credentials.

  2. 02

    Trigger.dev orchestrates the investigation

    The agent runs schema tasks: query-latency, query-heatmap, query-diff, rank-suspects, calculate-error-budget — each a first-class unit in the run trace.

  3. 03

    ClickHouse answers with depth

    MergeTree source tables, AggregatingMergeTree rollups with quantileState / quantileMerge, window-function anomaly bands, and before/after span diffs — all readonly.

  4. 04

    Board assembles progressively

    Skeleton panels fill as tasks complete. Open evidence drawers, click the payment hotspot, walk the suspect ladder, close on the verdict.

Seeded incident

The demo story is the product.

At 14:32, payment-service v2.4.1 changes retry timeout from 3s → 15s. Checkout p95 spikes ~2.6×. Payment spans amplify ~8.3×. Everything on the board exists to make that story impossible to miss.

“The agent found that the 14:32 payment deploy changed retry behavior, causing checkout latency to spike, and every visual is backed by ClickHouse evidence.”

Stack

Built for the hackathon judges — and for real SQL depth.

ClickHouse

Primary data layer. MergeTree logs & spans, deploy events, AggregatingMergeTree 1m rollups, parameterized readonly queries.

Trigger.dev

Chat agent with versioned prompt telemetry, task-backed tools, typed board parts, optional live streaming, and run-trace proof.

Next.js

Incident board UI with progressive skeleton assembly, fixture mode for offline judging, and live transport when credentials are set.

Fixture + smoke

pnpm smoke asserts the demo story is visually obvious before you hit record — deploy marker, 2× spike, payment #1, rollup path present.

Quick start

Public fixture-mode demo needs nothing. Local path is three commands.

pnpm install
pnpm smoke
pnpm dev

# then ask:
# Why did checkout latency spike after the 14:32 deploy?

See the board assemble in under a minute.

Open the public fixture demo, or clone the repo and run smoke checks before your own recording.

Launch demo Watch demo video Star on GitHub

Build Week prototype · MIT License · Not a general observability platform