Skip to content

Harness coverage

TokenShift supports six coding agent harnesses today — Claude Code, Codex, GitHub Copilot CLI, Cursor, opencode, and Devin CLI.

Output compression, telemetry and session-history backfill work on all six — on macOS, Linux and Windows alike. Three things differ.

HarnessFull-output recoveryToken usage reportingHook config
Claude CodeYesPer turnShared with the harness
CodexNo *Per turnShared with the harness
Copilot CLIYesAt session endOwn file
CursorNo *Per turnShared with the harness
opencodeYesPer turnOwn file
Devin CLIYesPer turn †Shared with the harness

* Codex and Cursor sandbox the proxy, so there is no local cache to replay from. Compression and telemetry are unaffected — see Recovery cache scope.

† Devin CLI’s hooks carry no model or usage figures. TokenShift reads both from the session store the CLI keeps locally, so usage arrives per turn like the other harnesses. Its swe-2 models are not in the pricing feed yet (Cognition lists them as free in the CLI through 2026-10-10), so those turns carry tokens and a zero dollar cost. Turns on other models, such as Gemini, price normally.

  • Full-output recovery — whether tokenshift recover can pull back the original, uncompressed output of a call TokenShift compressed.
  • Token usage reporting — when the model’s own token counts reach your telemetry.
  • Hook config — whether TokenShift keeps its hook wiring in its own file, or shares the harness’s. Shared is fine in normal use; worth knowing if another tool edits it too.

Every row above is a property of the harness, not of the operating system. Where a difference between operating systems appears, it will be called out here.


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