Engine · LoopEngine telemetry

Loop & cost — tier per phase

Cheap eyes, mid hands, expensive judgment: each phase runs the smallest model that does the job

cycle #1,982 · next in 34m

Cost per cycle · 7d avg$0.084Monitor 7% · Direct 37% · Reflect 56%

Month to date$23.40of $40 budget · trending $33 EOM

Tokens · 7 days14.2M83% prompt-cached — context reuse pays

Cycle time · p507m 40s12 cycles/day · 2h cadence

Phase → model assignment

Phase	Model	Job	Avg tokens	Avg cost
Monitor	Haiku 4.5	Sweep state, diff against last cycle, cheap and often	41k	$0.006
Direct	Sonnet 5	ReAct over tools: propose actions/tasks, draft directives	78k	$0.031
Reflect	Opus 4.8	Judgment: SITUATION.md, journal, risk calls	52k	$0.047
Boardroom	Opus 4.8	Meetings run on the judgment tier — it argues back	on demand	~$0.09/mtg

💡

Model ids and pricing live in LlmOptions + CostModel — changed only against current Anthropic docs. No API key ⇒ the whole engine drops to dry-run and proposes nothing real.

Spend · last 14 days

Loop Meetings

SpikesJul 6 board meeting · Jul 9 entity meeting

Budget guardloop pauses Reflect→Opus at 95% of cap

Recent cycles

no failures 14d

Cycle	Finished	Proposed	Cost	Note
#1,982	09:20	1 action · 1 task	$0.084	a-143 comms drafts
#1,981	07:20	—	$0.071	quiet pass
#1,980	05:19	—	$0.069	quiet pass
#1,979	03:21	1 note	$0.075	craigslist drift observed
#1,978	Jul 9 · 21:04	2 actions	$0.102	a-141 filing · a-142 price drops

Connectors — read-only by law

all healthy

Anthropic API · tool_use native · prompt caching onp95 2.1s

Neon Postgres · memory store · schema currentp95 38ms

MarketNinja API · spreads + review countsp95 220ms

Brave Search · research tool · 214 queries this weekp95 480ms

Vault (work/) · dev-ticket read · 2 files unparsed frontmatterInspect

🛡️

Every connector observes; none mutate. A write to any product system is a product action by definition — and that kind always stops at the gate.