Cost Ops

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 is “what did that cost me?” — and today the honest answer across opchain is “we don’t precisely know.” Cost Ops closes that loop: every skill phase that calls a model can attribute its spend, every checkpoint can carry 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), Opus 4.8 (claude-opus-4-8), Sonnet 4.6 (claude-sonnet-4-6), Haiku 4.5 (claude-haiku-4-5). references/pricing-reference.md carries a dated snapshot and always defers to oc-claude-api as source of truth.

/oc-cost — Command Reference

COST OPS COMMANDS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  ATTRIBUTION
  /oc-cost              Summarize spend for the current project (by phase + model)
  /oc-cost attribute    Attribute a run's token counts → dollars, write checkpoint.cost
  /oc-cost report       Cost-per-shipped-feature report (feeds /showcase + /dashboard)

  BUDGETS & GATES
  /oc-cost budget       Set a per-phase / per-suite budget ceiling in the checkpoint
  /oc-cost gate         Run the budget + cost-regression gate (CI-friendly verdict)

  ROUTING
  /oc-cost route        Recommend a model tier per phase (Haiku/Sonnet/Opus/Fable)

  UTILITIES
  /checkpoint           Show checkpoint status

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  Type any command to begin. /oc-cost to see this again.

How This Skill Fits the Build Pipeline

oc-claude-api (pricing + token counts) ──┐
oc-prompt-ops (per-eval token counts)  ──┼──► oc-cost-ops ──► checkpoint.cost
oc-bug-check / oc-code-auditor (phase) ──┘        │            { total_usd, budget_usd,
                                                  │              by_phase, by_model }
                          ┌───────────────────────┼───────────────────────┐
                          ▼                       ▼                        ▼
                   /oc-cost gate            /oc-cost route          oc-telemetry-ops
                   budget + cost-           tier recommendation     aggregate → /dashboard
                   regression (CI)          per phase               cost-per-feature stats

Cost Ops reads token counts the model-facing skills already emit, multiplies by the oc-claude-api pricing, and writes the attributed result into the wire-1.1 cost checkpoint field. Downstream: oc-prompt-ops runs the cost-regression gate beside its score gate (a prompt change that holds quality but triples cost is a regression), and oc-telemetry-ops aggregates cost across runs for the public /dashboard.

The cost checkpoint field (wire 1.1)

Cost Ops is the owner of the cost field added to the checkpoint protocol in v1.6 (see oc-checkpoint-protocol § “Wire 1.1 extensions”):

"cost": {
  "currency": "USD",
  "total_usd": 12.34,
  "budget_usd": 50,
  "by_phase": { "spec": 3.10, "build": 8.40, "audit": 0.84 },
  "by_model": { "claude-opus-4-8": 9.2, "claude-haiku-4-5": 3.1 },
  "tokens": { "input": 1200000, "output": 240000 },
  "updated_at": "2026-06-25T12:00:00Z"
}

The validator warns (does not error) when total_usd > budget_usd — overspend is a signal. The gate decision (block the merge / deploy vs. warn) is policy Cost Ops applies in /oc-cost gate, documented in references/budget-gates.md.

Principle 1: Attribute, don’t estimate

Cost Ops never calls a model itself. It reads the token counts the model-facing skills already emit (oc-claude-api usage, oc-prompt-ops per-eval counts) and multiplies by the oc-claude-api price table. The result is attributed spend, broken down by phase and by model, written into the cost checkpoint field. Full method, the per-call → per-phase math, the worked example, and the read→merge→write rules: references/cost-attribution.md.

The two rollups must reconcile — Σ by_phase == Σ by_model == total_usd — which is how /oc-cost attribute catches a miscounted call. If a phase’s tokens are missing, it’s reported unknown, never guessed.

Principle 2: Route to the cheapest tier that holds quality

The cheapest model that holds quality on a phase is the right model for it — recommended, never silently applied (a quiet Opus→Haiku demotion is a quality regression dressed as a cost win). /oc-cost route emits a per-phase recommendation with the projected delta; the user applies it and the next eval_scores run confirms quality held. Decision tree + the opchain phase→tier table: references/model-tier-routing.md. Routing facts come from oc-claude-api.

Three levers, in order of usual impact: fix caching (often larger than a tier gap; caches are model-scoped), cap output (output is 5× input price at every tier), then downgrade tier. /oc-cost route checks cache-hit rate first and may recommend “keep the tier, fix caching” instead.

Principle 3: Cost is a regression dimension

Two gates run in CI beside the quality gates (references/budget-gates.md):

• Budget ceiling — cost.budget_usd per phase / per suite. Validator warns on overspend; strict mode blocks. Raising a budget is a logged decision.
• Cost regression — the gate oc-prompt-ops advertises: freeze a cost baseline, fail a PR when cost_per_eval spikes past regression_pct, even if quality held. It runs next to oc-prompt-ops’s score gate so a prompt/model change is judged on both axes — quality and cost — and neither moves silently.

Pricing

A dated snapshot of the oc-claude-api price table lives in references/pricing-reference.md (Fable 5 $10/$50, Opus 4.8 $5/$25, Sonnet 4.6 $3/$15, Haiku 4.5 $1/$5 per MTok; Batch = 50%; caching is the big lever). It is a snapshot — always defer to oc-claude-api as source of truth before quoting a price.

Boundaries (what oc-cost-ops does NOT own)

Checkpoint Integration

Location

{project-dir}/.checkpoints/oc-cost-ops.checkpoint.json

When to Write

Cross-Skill Reads

Principles

1. Attribute, don’t estimate. Cost is multiplied from real token counts and the oc-claude-api price table — not guessed.
2. The number lives in the checkpoint. cost is wire-1.1 protocol state, so spend survives across sessions and is auditable in the PR diff.
3. Cost is a regression dimension. A change that holds quality but triples spend is a regression; the cost gate runs beside the score gate.
4. Route to the cheapest tier that holds quality. Haiku for cheap repetitive phases, Opus for spec/audit/migration — recommended, never silently downgraded.
5. Pricing facts come from oc-claude-api. Never hard-code prices from memory; the snapshot is dated and defers to the source.