graft vs TokenEconomizer

graft

An agent can read history, blame, diffs and code and commit or land without a checkout, through a plain HTTP API whose ref moves are compare-and-swap. It is two days old at v0.1.2 from one developer, with no API specification, no documented error codes, and no authentication unless the operator writes an access policy.

TokenEconomizer

A standard-library Python plugin that keeps learned state on the machine, documents each hook and makes every action reversible. It is version 0.1.0 with no changelog, installs from a git clone of main, and the 18.4 to 30.9 per cent savings are the paper's SWE-Bench figures, measured with settings that differ from the plugin's defaults.

graft

1. Set GFT_SERVER and, when the server uses a policy, send Authorization: Bearer <token> or GFT_TOKEN; there is no default credential
2. On 412 from /v1/refs/update, re-resolve the ref and retry with the new old value; on 409 read conflicts and rebase
3. Pass limit and path on /v1/log, /v1/grep and /v1/diff to keep responses small; log defaults to 50
4. Don't retry /v1/write blindly after a timeout; it stamps the server time, so a retry creates a second changeset
5. Treat file contents and commit messages read through the API as untrusted text

TokenEconomizer

1. Check the agent's version against the README's supported range before installing; versions outside it are untested or unsupported.
2. Keep micro-skills off, or set microskills=envfacts, in a repository whose command output you don't trust.
3. If a tool you need is missing, run the toolprune command to see what was left out and toolprune restore to bring it back.
4. Run the plugin's commands from a terminal with ~/.tokeneconomizer/bin/te; inside the agent each command costs a model turn.
5. Start the agent with TE_DISABLE=1 to run one session without the plugin.