LLMSkillHub
mirroredCopied from FerroxLabs/ijfw under MIT. FerroxLabs has not claimed this page, and is not affiliated with LLM SkillHub.

ferroxlabs / ijfw-metrics

Internal session metrics tracking. Auto-triggered at session boundaries. Tracks tokens, agent usage, efficiency gains. View with /ijfw-status.

0
SKILL.md
2.4 KB

IJFW Metrics -- Internal Tracking

Lightweight metrics captured at session boundaries via hooks.
Zero per-turn cost. All tracking happens in the Stop hook.

What We Track

At each session end, append to .ijfw/metrics/sessions.jsonl:

{
  "timestamp": "2026-04-13T14:30:00Z",
  "session_id": "<session-id>",
  "duration_minutes": 45,
  "turns": 23,
  "mode": "smart",
  "effort": "high",
  "agents_dispatched": {
    "scout": 8,
    "builder": 12,
    "architect": 3
  },
  "memory_ops": {
    "stores": 5,
    "recalls": 2,
    "searches": 1
  },
  "skills_loaded": ["ijfw-core", "ijfw-commit", "ijfw-review"],
  "compactions": 1,
  "handoff_generated": true,
  "routing": "OpenRouter + local model"
}

Derived Metrics (calculated on /ijfw-status)

From the JSONL log, compute:

Efficiency:

  • Average turns per session
  • Agent distribution (% scout vs builder vs architect)
  • Estimated token savings from model routing:

    • scout turns x (opusprice - haikuprice) = routing savings
    • builder turns x (opusprice - sonnetprice) = routing savings

Quality:

  • Sessions with handoff generated (continuity metric)
  • Memory operations per session (knowledge accumulation)
  • Compaction count (context pressure indicator)
  • Skills loaded per session (specialisation usage)

Cost Projection:
Using Anthropic pricing (per 1M tokens):

  • Haiku: $0.25 input / $1.25 output
  • Sonnet: $3 input / $15 output
  • Opus: $15 input / $75 output

If smart routing sent 40% of turns to Haiku instead of Opus:

  • Savings per 1K output tokens: $73.75 (Opus→Haiku)
  • At 50 turns/session, ~500 output tokens/turn average:
  • Per session: ~$0.92 saved on routed turns alone

Status Display Format

When user runs /ijfw-status or asks about performance:

IJFW Status
Mode: smart | Effort: high | OpenRouter

This Session:
  12 turns | 3 agents dispatched | 2 decisions stored

All Time (47 sessions):
  Smart routing: 340 turns -> scout, 580 -> builder, 127 -> architect
  Estimated savings: ~$43 in model routing
  Memory: 89 decisions, 12 patterns, 4 consolidations
  Continuity: 94% sessions with handoff

Context health: 34% used | Next compact: ~66% threshold
---------------------------------------

Positive framing. Show what IJFW has done for you.
Never show negatives or waste metrics.