oc-cost-ops

Topics in the source

• Cost Ops
• /oc-cost — Command Reference
• Executable eval-cost commands
• How This Skill Fits the Build Pipeline
• The cost checkpoint field (wire 1.1)
• Principle 1: Attribute, don't estimate
• Principle 2: Route to the cheapest tier that holds quality
• Principle 3: Cost is a regression dimension

On first invocation, read references/orchestrator.md and follow its welcome protocol. Make LLM spend a first-class, attributable number in the pipeline. v1.5 added four AI-native skills; the predictable next question was "what did that cost me?" — and before v1.6 the honest answer across opchain was "we don't precisely know." Cost Ops closes that loop: every skill phase that calls a model can attribute its spend, the oc-cost-ops checkpoint carries a budget that gates when tripped, and every model choice can be checked against a tier-routing recommendation (don't run Opus on a phase Haiku handles). This is not a tri-agent harness and not a billing system. It's the instrumentation layer that sits underneath the model-facing skills — oc-claude-api owns the request surface and the canonical pricing, oc-prompt-ops owns the eval suite, oc-telemetry-ops owns local metering — and Cost Ops turns the token counts they already produce into attributed dollars, budgets, and routing advice. Pricing facts come from oc-claude-api, not memory. Per-token prices, Batch-API economics (50% of standard), and cache read/write multipliers are sourced there. Current model IDs: Fable 5 (claude-fable-5), Opu

This is an excerpt. Refer to the complete source and its current requirements.