Harness coverage

TokenShift supports four coding agent harnesses today — Claude Code, Codex, GitHub Copilot CLI, and Cursor. Coverage isn’t identical across them: each harness exposes hooks differently, some sandbox the compressing proxy, and backfill maturity varies. This page scores that honestly, harness by harness, OS by OS.

One score per cell, out of 5 — see How the score works for what goes into it. Claude Code is the baseline: full, unsandboxed coverage with no real-time or backfill gaps.

Coverage matrix

Harness	macOS	Linux
Claude Code	5/5 — full coverage: hooks, backfill, token usage, compression.	5/5 — same as macOS.
Codex	4/5 — partial backfill support.	4/5 — same as macOS.
Copilot CLI	4/5 — per-turn tokens land via backfill, not real-time hooks.	4/5 — same as macOS.
Cursor	4/5 — backfill coverage, lower fidelity than the others.	4/5 — same as macOS.

How the score works

Four things combine into the one number:

Sandboxing — does the proxy run unsandboxed, recording stats in-process the instant a command completes, or sandboxed, handing stats off through a short-lived local file between hooks?
Backfill — does it exist, and how complete is the reconstruction: a full transcript-verified replay, or a coarser one?
Real-time fidelity — do the hooks themselves carry full per-turn token data, or only session-level totals?
Hook-file reliability — a dedicated hook file (no collision risk), or a shared one another tool could silently overwrite?

Score reference

Score	Means
5	Unsandboxed; backfill is a complete, transcript-verified replay; dedicated hook file.
4	One real gap — either sandboxed, a named backfill limitation, or a hook-file collision risk — not more than one.
3	Two or more of those gaps stack together.
1–2	No backfill and no real-time recovery — telemetry only, no depth.

How it works — what compression, telemetry, and the intent classifier each do.
Data contract — hook-event coverage, field by field.