v1.6 — The instrumented pipeline

v1.5 made the AI part of your app an evaluated artifact. v1.6 asks the ruder follow-up: what did all that evaluating cost, and would you notice if it doubled?

Until today the honest answer was no. opchain could plan, build, audit, ship, and monitor your app — and could not tell you what any of that cost. Like a contractor who does beautiful work and throws away every receipt. v1.6 fixes the receipts. The theme, in one sentence:

You can’t steer what you can’t see.

The two new skills

• oc-cost-ops (/oc-cost) — LLM cost attribution at the resolution that actually matters: the skill phase. Not “you spent $40 this month” but “the build loop’s evaluator rounds are 31% of sprint cost.” On top of attribution: budget gates in the checkpoint (a phase that blows its budget fails the gate the same way a failing test does) and model-tier routing — Haiku for mechanical, repetitive phases; Opus where a wrong judgment is the expensive part. It also adds a cost-regression gate that runs beside oc-prompt-ops’s score gate, because a prompt change that improves the score 2% and triples the cost is not an improvement. It’s a subscription.
• oc-telemetry-ops (/oc-telemetry) — opt-in usage metering that records which skills and phases actually run. It’s local-first: everything lands in .checkpoints/usage.sqlite, in your repo, where you can open it with sqlite3 and check our math. Anonymized aggregates — and only aggregates — feed the public /dashboard. Default off, and content-free by schema: the tables have no column that could hold a prompt. It answers “which skills get used” without ever learning what you used them for.

Wire 1.1 — the checkpoint protocol grows three fields

Instrumentation needs somewhere to live, and in opchain everything durable lives in the checkpoint. The checkpoint protocol moves to wire 1.1 with three additive, optional fields:

• cost — what each phase spent, written by oc-cost-ops
• eval_scores — scored quality over time, not just a pass/fail verdict
• telemetry_handle — the pointer that lets metering stay local

Both "1.0" and "1.1" validate. Existing checkpoints don’t break, don’t need migrating by hand, and get swept up opportunistically by oc-migration-ops. If your resume flow worked yesterday, it works today — the fields simply start appearing.

The ripples

As usual, the new skills don’t sit off to the side; the rest of the catalog learns to use them:

• oc-bug-check and oc-code-auditor now emit eval_scores against a stable rubric. The binary verdict and letter grade are unchanged — but your code quality now has a trend line, and trend lines are where slow rot goes to get caught.
• oc-monitoring-ops gains an AI-app monitoring template: token rate, cost rate, eval drift, and hallucination/refusal flags. Your error budget, meet your actual budget.
• oc-orchestrator’s /oc-ops next is now cost-aware: within a priority rank, over-budget checkpoints sort first. “What should I work on?” now includes “the thing quietly on fire, financially.”
• oc-prompt-ops’s cost_per_eval placeholder is finally wired to measured numbers instead of estimates.

The catalog goes from 22 to 24 skills, in lockstep at 1.6.0.

The through-line

v1.5’s bet was that the AI part of an app should be measured like code. v1.6 extends the same bet to the pipeline itself: every phase now reports what it spent and how it scored, in a file you own, on your disk.

One warning from experience: instrumentation has a reliable side effect — it embarrasses you. We’ve already pointed both new skills at opchain’s own development history, and we’ll publish what they found, receipts included. Some of the receipts are unflattering. That’s rather the point.

Browse the skill library, watch the numbers land on the dashboard, or install opchain and run /oc-cost on your own pipeline.