vibeOS for OpenCode

Cost-aware control plane for OpenCode Desktop.

v0.23.0 — Disintegration Milestone

Compact footer format: 🦠 brain | Deepseek | V4 Pro | $12.57 | VibeMaX ⚡ Budget VibeMaX is now the default optimization mode. Model display names cleaned up (V4 Pro, Sonnet, 2.5 Flash). Install: npx vibeostheog setup --project or npx vibeostheog setup Keeps expensive models on strategy, routes implementation to cheaper tiers, surfaces savings in real time.

For teams, vibeOS adds practical guardrails: delegation enforcement, flow and TDD controls, pattern learning, stress-aware routing, VibeBoX decision tracking, reporting, and remote API protection for the core algorithms.

Delegation Enforcement

Every write/edit/notebookedit on the brain tier is intercepted, cost-estimated, and blocked with a visible enforcement note. Work must be delegated to medium or cheap. This is the primary savings mechanism.

Savings per Delegation

Every blocked brain-tier write/edit saves at least $0.026 (Opus -> Sonnet). The running total is tracked in ~/.claude/delegation-state.json and displayed in the live footer.

Model Tiers

Benchmarked on the DeepSeek v4 family. Prices based on 700 input + 300 output tokens per turn.

DeepSeek Chat costs $0/turn when routed through the Direct DeepSeek provider (no OpenRouter markup).

Source: src/lib/pricing.ts. Conservative estimates — savings are always understated.

Optimization Modes

Policy Comparison — Sorted by Quality Descending

VibeUltraX — Cascade pipeline: MagicCoder:7b (local Ollama) proposes, v4 Flash reviews, v4 Pro refines. Benchmarked at 107% of Brain quality at 58% cost (local inference is free, only Flash/Pro API calls cost).

VibeQMaX (Quality Max) — Routes strategic turns through v4 Pro with full thinking, strict enforcement, strict flow checks, and quality TDD. Write/edit turns delegated to cheaper tiers per enforcement rules. Effective blended cost ~$0.00029/turn (50% of Raw Brain).

Raw Brain — v4 Pro (no framework). Baseline for all comparisons. 100% quality at 1.00x cost.

VibeMaX (ML-Optimized, v0.23 Default) — Intelligent cost-quality sweet spot. Routes through v4 Flash (medium) and uses a random forest classifier (29 trees, gini-split, trained on telemetry) to decide each turn. Classifies on 11 derived features: message length, code block density, urgency, complexity, repetition, question ratio, and more. Benchmarked at ~75% of Brain quality at 18% of cost.

Budget — DeepSeek Chat. Direct routing. ~40% quality at 0.00x cost. 100% savings.

Cost vs Quality Visual

Quality
  baseline  . Raw Brain . VibeQMaX
    107%        |   . VibeUltraX
  ~75%      |   . VibeMaX (default)
  ~55%      |   . Speed
  ~40%      |   . Budget
              |
              +--------------------------------
              1.0x  0.50x 0.32x 0.18x 0.00x
                          Cost Multiplier

Configuration Per Mode

Benchmark Details

All tests run with DeepSeek v4 family. Quality scores measured against Raw Brain (v4 Pro, full thinking, no vibeOS overhead). VibeMaX quality benchmark derived from real session telemetry with bootstrap confidence intervals (36 bootstrap samples). Pareto frontier computed from 70 holdout scenarios across 170 training samples via hyperparameter sweep. VibeUltra is the first mode that beats Raw Brain on both accuracy and cost — Pareto-dominant.

Benchmarked on 1000 simulated questions across 20 runs, using model accuracies from MMLU-Pro / GPQA Diamond with real error correlation data.

Features

How It Works

8 hooks into OpenCode Desktop:

Local Models (Ollama)

To use MagicCoder:7b or other local models with vibeOS and OpenCode Desktop:

Setup

1. Install Ollama — curl -fsSL https://ollama.ai/install.sh | sh
2. Pull MagicCoder — ollama pull magicoder:7b
3. Add provider — In OpenCode settings, add an Ollama provider (default: http://localhost:11434)
4. Detect — Run trinity rebuild — MagicCoder:7b appears in the model dropdown
5. Assign slot — trinity set cheap magicoder:7b

Minimum Hardware

Note: Local models run entirely on your machine. vibeOS treats them as any other OpenCode provider — pricing shows $0.00/turn.

Install

npx vibeostheog setup --project        # per-project
npx vibeostheog setup                   # global ~/.config/opencode/
npx vibeostheog setup --help             # full usage

One-command setup: deploys plugin files and registers in opencode.json. Restart OpenCode Desktop.

Local dev checkout:

{
  "plugin": ["/absolute/path/to/theSaver-oc/src/index.js"]
}

Commands

trinity help for full reference. Commands register in the TUI sidebar.

Report commands: report-save, report-list, report-read, research-audit

Live Footer

— 🦠 brain | Deepseek | V4 Pro | $12.57 | VibeMaX ⚡ Budget

Tier icon + lowercase quality (🦠 brain / ⚙ medium / ⚡ cheap), provider label, cleaned model name, total savings, VibeBrand + flash icon + mode. Persisted in ~/.claude/delegation-state.json.

Architecture

Plugin Source

Single-file runtime src/index.js (5529+ lines). TypeScript source of truth at src/vibeOS-lib/*.ts and src/utils/*.ts. Build: npm run build (tsc compile + sync-ts-build + deploy script).

State Files (~/.claude/)

Local vs Remote

Fully functional locally: Model tier classification, static pricing, stress scoring, context budget, turn classification, TDD skeleton generation, flow enforcement, savings ledger, session metrics, reports, footer, dashboard, smart cache, VibeBoX fallback.

Requires remote API (api-token): Bootstrap token exchange, advanced VibeBoX with full session history, dynamic per-prompt delegation, cross-session calibration, live pricing fetch beyond static map, learned subagent routing.

When the remote API is unreachable, the plugin degrades gracefully to rule-based local algorithms. Core enforcement features continue working.

VibeBoX Decision Engine

7 sub-regimes (INIT, DIVERGENT, EXPLORING, REFINING, CONVERGING, CLOSED, LOOPING). Classification via entropy trends, action consistency, feature contradiction, embedding drift. 11 derived features per turn. 4 loop intervention levels. PIVOT/SWITCH detection. Outcome tracking from satisfaction signals.

Regime -> mode mapping via syncControlSettings():

Stress > 1.5 escalates any regime to quality.

Remote API Server

src/vibeOS-api-server/ — Fastify + SQLite at api.vibetheog.com. Endpoints: delegation check, tier routing, stress scoring, VibeBoX analysis/calibration, TDD skeleton gen, pattern observation, pricing fetch, context compression. Auth via VIBEOS_API_TOKEN. Client: src/vibeOS-api-server/client.js with automatic local fallback.

Dashboard

SolidJS SPA at src/dashboard/. Build: npm run build:dashboard (vite). Served by MCP server or standalone. SSE /events for real-time push.

Environment Variables

Troubleshooting

trinity help is the canonical command reference. This README stays high-level so command details follow the code without a rewrite.