Metrics: Real vs Estimated

Note

Generated from ONBOARDING.md — the design authority for this project. Edit the source file, not this page.

Metric	Status	Source
Tokens in (non-cached)	✅ Real	post_api_request.usage.input_tokens
Tokens out	✅ Real	post_api_request.usage.output_tokens
Cache read tokens	✅ Real	post_api_request.usage.cache_read_tokens
Cache write tokens	✅ Real	post_api_request.usage.cache_write_tokens
Reasoning tokens	✅ Real	post_api_request.usage.reasoning_tokens
API call latency	✅ Real	post_api_request.api_duration (seconds → ms)
Tool call latency	✅ Real	post_tool_call.duration_ms
Model name	✅ Real	post_api_request.response_model or model
Provider name	✅ Real (verbatim)	post_api_request.provider
Platform	✅ Real	on_session_start.platform
Cron job ID	✅ Real (parsed)	session_id regex extraction
Session duration	✅ Real (wall time)	started_at → ended_at (last turn)
Tool success/failure	✅ Real	post_tool_call.status; result-JSON fallback (truthy error, terminal exit_code, success is False) on older Hermes
Subagent count	✅ Real (proxy)	subagent_stop hook count
Cost (USD)	⚠️ Estimated	Local pricing table × token counts
Tokens when usage=None	⚠️ Estimated, flagged	approx_input_tokens + chars/4, row marked estimated=1
Subagent cost (global)	✅ Real	Child runs fire own hooks → independent runs rows
Subagent cost (per-cron-job)	✅ Real (attributed)	subagent_edges (v11) + recursive CTE in db.spend_by_scope("cron_job", ...)
MoA aggregator cost	✅ Real (re-attributed)	Aggregator usage from post_api_request, priced under the aggregator’s real provider/model resolved from the preset (moa.py)
MoA reference-model cost	❌ Not available	References run via auxiliary call_llm, which fires no hooks — tokens unrecoverable (recorded MoA cost is a lower bound)

Cost is always an estimate — computed from a local pricing table, not from a provider billing API. Users can override prices via ~/.hermes/telemetry/pricing.yaml.