the garage · notes on multi-model workflows & caching

Shift notes

Engineering write-ups on running multiple coding-agent models: model routing, parallel sessions, and prompt caching economics.

harnesses

Claudex: Running GPT-5.6 Sol Inside Claude Code

The recipe going around X, tested properly. What a harness really is, the CLIProxyAPI setup end to end, and the three things we measured that the thread never mentions.

 claude code internals

Every Way Claude Code Switches Models Without Asking You

Opus limit fallback, opusplan, fallback chains, Fable 5 safety rerouting, org clamps: all five automatic switches, what each costs, and how to take control.

 terminal craft

Claude Code + tmux: The Missing Manual for Persistent Agents

Keep agents alive through closed terminals and dropped SSH. The garage pattern, send-keys automation, and the PATH and window-ID gotchas nobody documents.

 codex

Codex CLI Model Switching: /model, config.toml & Reasoning Effort

The four levels of Codex model selection, the 0.134.0 profile-file change, effort from minimal to xhigh, and running Codex as a gear next to Claude.

 claude code

The Claude Code /model Command, Completely Explained

Every way to change models in Claude Code: /model, --model, ANTHROPIC_MODEL, settings.json. What persists, what resets, and the cache pitfall to avoid.

 prompt caching

Why Switching Models Mid-Session Breaks Your Prompt Cache

Prompt caches are model-scoped. One /model switch re-processes your whole context at full price. The math, the invalidation hierarchy, and the cache-safe fix.

 models

Haiku vs Sonnet vs Opus: Which Model for Which Coding Task

A practical routing guide: what to give each Claude tier, the upshift/downshift signals, and how to run several tiers together without penalties.

 workflow

How to Run Multiple Claude Code Sessions in Parallel

Terminal tabs, tmux, git worktrees, and headless runs. The full parallel-session setup, plus the pitfalls: shared services, rate limits, session sprawl.

 agents

Claude Code vs Codex CLI: Why the Answer Is Both

Different model families make different mistakes. Cross-review, design disagreements on demand, and quota arbitrage: the case for running both.

 reference

Anthropic Prompt Caching, Explained: TTLs, Pricing & the Prefix Rule

1.25×/2× writes, 0.1× reads, 5-minute vs 1-hour TTLs, break-even math, and the silent mistakes that zero your cache hit rate.

 workflow

A Cache-Safe Multi-Model Workflow for Claude Code & Codex

Haiku for grunt work, Sonnet for building, Opus for hard bugs. Three patterns to use them all without paying a cache rebuild on every switch.