# Alerting rules for pi-prometheus.
#
# Load them from your Prometheus config:
#   rule_files:
#     - alerts.yml
#
# Every ratio divides by a series filtered with a positive-value matcher, so a
# session that has not reported yet, or that reports a zero denominator,
# produces no series at all instead of NaN or +Inf. Labels come from the
# file_sd target file, so session_id and cwd are always present.
groups:
  - name: pi-prometheus
    interval: 30s
    rules:
      - alert: PiContextFootprintHigh
        expr: |
          pi_context_static_tokens
            /
          (pi_context_window_tokens > 0)
            > 0.15
        for: 10m
        labels:
          severity: warning
        annotations:
          summary: 'Static context is {{ $value | humanizePercentage }} of the window in session {{ $labels.session_id }}'
          description: >-
            Every request in {{ $labels.cwd }} pays this before you type a word.
            Run /context-budget in that session to see which package, skill or
            context file is responsible, then uninstall it, narrow its tool set,
            or move the context file out of the project root.

      - alert: PiSessionNearCompaction
        expr: |
          pi_context_tokens
            /
          (pi_context_window_tokens > 0)
            > 0.85
        for: 5m
        labels:
          severity: warning
        annotations:
          summary: 'Session {{ $labels.session_id }} is at {{ $value | humanizePercentage }} of its context window'
          description: >-
            Compaction is close in {{ $labels.cwd }}, and a compaction loses
            detail from the conversation. Finish or checkpoint the current task,
            then start a fresh session with /new rather than letting the history
            be summarised away mid-task.

      - alert: PiToolErrorRateHigh
        expr: |
          sum by (job, instance, session_id, cwd, tool) (increase(pi_tool_errors_total[15m]))
            /
          (sum by (job, instance, session_id, cwd, tool) (increase(pi_tool_calls_total[15m])) >= 5)
            > 0.2
        for: 5m
        labels:
          severity: warning
        annotations:
          summary: 'Tool {{ $labels.tool }} is failing {{ $value | humanizePercentage }} of calls in session {{ $labels.session_id }}'
          description: >-
            Over the last 15 minutes more than one call in five to
            {{ $labels.tool }} failed, on at least five calls. The agent is
            burning tokens on retries. Check that tool's configuration in
            {{ $labels.cwd }}: wrong path, missing credential, or a permission
            prompt nobody answered.

      - alert: PiSpendRateHigh
        expr: |
          sum by (job, instance, session_id, cwd) (rate(pi_cost_usd_total[15m])) * 3600
            > 5
        for: 10m
        labels:
          severity: critical
        annotations:
          summary: 'Session {{ $labels.session_id }} is spending ${{ $value | printf "%.2f" }} per hour'
          description: >-
            Sustained for ten minutes in {{ $labels.cwd }}. Either the session is
            looping, or the context is large enough that every turn is expensive.
            Check the context usage panel first: a session near its window costs
            the full window on every turn.