zeuikli

Output Compress — Development skill for Claude Code

Development community

Tiered, opt-in output compression skill with a deterministic fidelity gate (Claude Code Agent Skill + Codex AGENTS.md variant).

How to install Output Compress

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

What Output Compress does

Tiered, opt-in output compression skill with a deterministic fidelity gate (Claude Code Agent Skill + Codex AGENTS.md variant). Caveman-derived.

Alternatives in Development

Cs2 Knife — Create a high-fidelity procedural Three.js 3D reconstruction of 14.1k ★
Fidelity — Measure how faithfully a clone reproduces a site — pixel-diff plus motion-fidelity into one 0-100 score, a let 3.6k ★
New Op — Guide for adding a new variant to the Op enum (e.g. Op::Tag, Op::Link) 201 ★

README

output-compress

Tiered, opt-in output compression with a deterministic fidelity gate. See `SKILL.md` for the full rules (Claude Code / Agent Skills format) or `AGENTS.md` for the same rules in Codex's format. `USAGE.md` covers day-to-day workflows, log analysis, and auto-activation wiring (§7).

What's in the box (v1.5.0)

Three compression levels — lite / full / ultra (caveman-derived syntax); deletion only, never rewriting, with negation-first ordering to avoid the most common gate failure (SKILL.md §1, §4).
Model-tier compression caps — the level is capped by who reads the text (small models lite, mid-tier full, frontier ultra; uncalibrated models start at lite) (SKILL.md §2).
Never-compress whitelist (7 categories) — code/paths/URLs/env vars, all numbers and dates, negations + quantifier bounds with their clause, project structured tags (CUSTOM_TAG_PATTERNS), safety-critical statements, contract fields as whole blocks, and hedges/qualifiers (the "decontextualization" axis) (SKILL.md §3).
Deterministic fidelity gate — scripts/fidelity-check.py regex/set-diffs the whitelist between original and compressed text; no LLM self-judgment. --coverage pre-check tells you when compressing isn't worth it; --log appends JSONL records with saving_pct and the grounded_pct rewriting-detector axis (SKILL.md §4).
Delegated compression (optional) — hand the deletion judgment to a cheap worker model while the orchestrator keeps the gate (scaling-paradox grounding: bigger compressor models are less faithful) (SKILL.md §4b).
Pace/usage coupling (optional) — scripts/usage-pacer.py emits AHEAD/ON_PACE/BEHIND burn-rate verdicts plus absolute-threshold escalation (80%/95% used) and handoff states (HANDOFF_PREP/HANDOFF_HALT); provider feeders claude-usage-fetch.py / codex-usage-fetch.py supply real usage numbers, and codex-handoff.py offers opt-in packet-backed han