Telemetry

Bubo emits OpenTelemetry metrics and traces, namespaced llm_review.* and defined in src/bubo/telemetry/metrics.py. Telemetry is off by default. Enable it and point it at an OTLP/gRPC collector for live dashboards. For one-off numbers without a collector, read the same rollups over MCP, bubo report, or the static UI.

Enable

Configure [telemetry] in config/env.toml. enabled = false (the default) makes every metric and span a no-op. With enabled = true and no otlp_endpoint, instruments record in-process but nothing exports.

[telemetry]
enabled                 = true
otlp_endpoint           = "otel-collector:4317"   # OTLP/gRPC; empty = record but do not export
otlp_protocol           = "grpc"                  # grpc is the only supported protocol
service_name            = "bubo"
environment             = "prod"                  # low-cardinality label: dev / staging / prod
export_interval_seconds = 30
 
# Per-finding and outcome-sync metrics — drop either to cut emission volume.
emit_finding_events = true
emit_outcome_sync   = true
 
# LLM pricing for llm_review.cost.usd, USD per 1M tokens. All default to 0.
input_per_1m        = 2.50
output_per_1m       = 10.00
cached_input_per_1m = 0.25

Per-review cost = (input × input_per_1m + output × output_per_1m + cached × cached_input_per_1m) / 1e6. Leave the rates at 0 to skip cost.

Metrics

Every instrument carries repo. Run-scoped instruments also carry model, prompt_version, review_mode (poller/manual), status, dry_run, and tone.

Trim volume with two switches: emit_finding_events = false drops the per-finding counter; emit_outcome_sync = false keeps only the initial-post events on llm_review.findings.

Dashboards

• Throughput — rate(llm_review.runs[5m]) by repo, status.
• Saturation — histogram_quantile(0.95, llm_review.latency.queue_seconds_bucket). Rising means raise max_merge_requests_per_poll.
• Cost — sum(rate(llm_review.cost.usd[1d])) by (repo, model).
• Precision — llm_review.findings{status="resolved"} / llm_review.findings{status="posted"} by severity.
• Reliability — rate(llm_review.failures[5m]) by operation.

Traces

Each review is one trace. llm_review.run opens a child span per stage. A trace shows where a review’s wall-clock went — a slow review is almost always the agent stage.

llm_review.run              repo, mr_iid, sha, run_id
├─ llm_review.checkout      repo, sha
├─ llm_review.provenance    repo                       (governance only)
├─ llm_review.agent         repo, model, tokens_input/output/total, cost_usd, exit_code
└─ llm_review.post          repo, dry_run, findings_posted/planned/skipped

Any OTel backend (Tempo, Jaeger, Honeycomb, Grafana, Datadog) breaks latency, tokens, and cost down by stage and model. Spans export only when otlp_endpoint is set.

Cardinality

Metric attributes stay low-cardinality. MR IID, SHA, file path, line, fingerprint, and discussion ID live in SQLite or span attributes, never as metric labels. Spans carry the per-MR context for trace-level drilldown.

Reading metrics without a collector

Three read-only paths return the same rollups without OTLP, without writes, and without the Slack attachment. None require the poller to be running.

Over MCP

The MCP server exposes the metrics tools to any chat client or agent. Ask in plain language; the client picks the tool.

get_metrics(since_hours=48, project="acme/app") returns:

{
  "window_hours": 48,
  "project": "acme/app",
  "reviews_total": 124,
  "by_status": { "success": 70, "no_findings": 48, "failed": 5, "running": 1 },
  "findings_total": 74,
  "tokens_total_sum": 13125521,
  "cost_usd_sum": 25.61
}

get_governance_report returns the full nested report (the same document as bubo report, below). Every metrics tool reads SQLite and is safe to call anytime — only review_change writes.

Recipe — reproduce the overview metrics

The overview figures come from bubo report, the rollup also exposed as get_governance_report. Run it against your install:

# Full JSON report for a fixed window (the overview's month view).
bubo report --format json --since 2026-06-01 --until 2026-07-01 > report.json
 
# Rolling window, one project.
bubo report --format json --since-hours 720 --project acme/app

The JSON maps to the overview panels:

Export one tabular section as CSV instead of JSON:

bubo report --format csv --section audit --since-hours 720 > audit.csv
# --section noise_trend      → daily false-positive trend
# --section dispute_classes  → per-category dispute rates

Two overview numbers are not report fields. cost_usd_sum is the recorded provider cost; the overview’s all-in cost and its reviewer-time dollar value are estimates it documents in its own footnote. The changed/added/removed LoC and files-reviewed counts are derived from diff stats, not from bubo report.

Static HTML dashboard

bubo ui-export writes a self-contained static site — recent reviews, health, and the same report windows — with no server and no writes to the database.

bubo ui-export --out ./bubo-ui
open ./bubo-ui/index.html          # macOS; or xdg-open, or serve the directory

The bundle is index.html plus data.json, a read-only snapshot holding the 24h/7d/30d reports, the 50 most recent reviews, health, and the effective config. It loads over file:// or from any static host, and never touches reviewed_mrs.