Dashboard and stats

Two pages for the same question: “is the swarm healthy and affordable?”

Dashboard

Your landing page. Shows:

• Running agents — live count, per-mode
• Task status breakdown — how many in each column
• Today’s cost — rolling 24h total across all agents
• Draft PRs — PRs waiting for you to mark ready
• Renovate PRs — count awaiting action
• Recent agent runs — last 10 completed runs with cost

Everything is clickable. Every number is a filtered view.

Signals to watch

• human-required climbing → the orchestrator is escalating more than usual; either a bad run of tasks or a config drift
• Today’s cost spiking → check for a runaway agent on Agents page
• in-review static for days → reviewers (you) are the bottleneck

Stats

Deeper metrics. Aggregated from the audit log and stats.json.

Default view

Running totals:

• Lifetime cost
• Total agent runs
• Pass / fail ratio
• Average cost per task
• Most expensive tasks (top 10)

This week

Seven-day window. Use for:

• Weekly cost review (set a budget, watch the trend)
• Throughput tracking (tasks completed per day)
• Regression detection (fail rate going up over time)

Per-project

Toggle the project dropdown. Stats filter to one project:

• Hours of agent time
• Cost
• Pass rate
• Mean time in each status

Useful when one project is dominating your budget.

Cost attribution

Every result event from a headless agent includes cost and tokens. Sybra sums these into:

• Per agent run (agent_runs[].cost_usd on each task)
• Per task (rollup of all runs)
• Per day (~/.sybra/stats.json)
• Per month (rolled over at UTC midnight)

Interactive (tmux) sessions don’t emit result events. Their cost is unknown. If you use interactive mode heavily, stats understate your real spend.

Set a monthly budget

In ~/.sybra/config.yaml:

agent:
  max_cost_usd: 5.0   # per-run ceiling

No global monthly cap exists — Sybra only enforces per-run. Set per-run low (e.g. $0.50) and let escalations surface when tasks need more.

Export stats

CLI:

sybra-cli audit --days 7 --json > week.json

The audit log is append-only NDJSON at ~/.sybra/logs/audit/YYYY-MM-DD.ndjson. One entry per agent run and task change.

Read the raw audit log

tail -f ~/.sybra/logs/audit/$(date -u +%Y-%m-%d).ndjson | jq .

Events include type, task_id, agent_id, cost_delta, tokens, message.

Respond to anomalies

When a task’s cost exceeds p95 of similar tasks, the self-monitor flags it. You’ll see it on Dashboard’s “Recent agent runs” with an outlier badge. Click through to see:

• Why it was flagged (cost, turns, repeated failures)
• The run transcript
• Suggested corrective action (tighten prompt, lower max_turns, etc.)

Acting on these is how the swarm gets cheaper over time.