avirtual

Wirescope — Development skill for Claude Code

Development community

A transparent analytical proxy for the Claude Code & Codex CLIs — see the real per-turn cost, prompt-cache warmth, and refusals on the wire, and keep caches warm.

How to install Wirescope

This entry records only its repository, not the path inside it, so there is no exact command to give. Open avirtual/wirescope and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Wirescope does

A transparent analytical proxy for the Claude Code & Codex CLIs — see the real per-turn cost, prompt-cache warmth, and refusals on the wire, and keep caches warm.

Alternatives in Development

System Cache Cleanup — Clean temporary files and caches to free disk space: $ARGUMENTS 23.4k ★
Mono Color Skill — One-ink editorial print image skill — warm paper, halftone photography, active negative space, and restrained 3k ★
10 Context And Prompts — Prompt 10: Wire Up System Prompt, Context Gathering & Memory System 2.3k ★

README

wirescope

**A transparent, analytical forward-proxy for the Claude Code and Codex CLIs — that can also reshape the wire.** It sits between your agent CLI and the model backend, observing what only the wire can see (the real per-turn bill, prompt-cache warmth, refusals, the assistant's text as it streams) and, when you ask it to, **trimming the context the CLI sends** — stripping `CLAUDE.md` from a subagent, cutting a bloated tool roster, keeping a session's cache warm. Point a CLI at it with one environment variable. Observation is automatic and invisible; the wire-reshaping is deterministic and flag-gated — a few cache-coherent transforms run by default, the `[wirescope:…]` directive shaping is opt-in, and `WIRESCOPE_PASSTHROUGH=1` turns all of it off for a byte-faithful forward.

   claude / codex CLI  ──►  wirescope (localhost)  ──►  api.anthropic.com
                              │                          chatgpt backend
                              ├─ prices every turn (the real bill)
                              ├─ tracks prompt-cache warmth + keeps it warm
                              ├─ surfaces refusals the CLI hides
                              ├─ streams text + receipts to subscribers
                              └─ reshapes the request on opt-in directives

Why it exists — the thesis

**Cost is context carriage, not "the model thinking."** In one measured 18-turn trivial session, context *carriage* cost \$2.72 while the generated output cost \$0.07 — a **38×** ratio. A single "2+2" turn shipped 31 tools (~23k tokens) and used none of them. "It's cached, so it's cheap" is backwards: the right baseline for junk context is **zero**, and caching still bills ~10% of it every turn while it occupies bandwidth and the context window.

wirescope was built first to **find and price that waste** — and then to **act on it on the wire**, in the one place a CLI's own knobs can't reach: a *subagent's* context. Native `--tools` and `omitClaudeMd` onl