f0rge M7 token audit (release binary, temp repo)
Loop: init -> issue -> task -> ci (fail then pass) -> wiki -> label -> milestone -> branch -> pr -> merge -> close -> --skill -> root help. All green, zero help needed.
| command | exit | stdout B | lines | stderr B |
|---|---|---|---|---|
init | 0 | 0 | 0 | 0 |
issue create --title --body --json | 0 | 163 | 1 | 0 |
issue list --json | 0 | 163 | 1 | 0 |
task create --title --json | 0 | 118 | 1 | 0 |
ci run (fail step) | 1 | 26 | 1 | 41 |
ci run (pass) | 0 | 28 | 1 | 0 |
wiki put docs/glossary.md - | 0 | 29 | 1 | 0 |
wiki list | 0 | 27 | 1 | 0 |
label create --name --color --description --json | 0 | 53 | 1 | 0 |
milestone create --name --json | 0 | 59 | 1 | 0 |
pr create --base main --head feat --title --json | 0 | 212 | 1 | 0 |
pr list | 0 | 18 | 1 | 0 |
pr merge 1 | 0 | 44 | 1 | 0 |
issue close 1 | 0 | 165 | 1 | 0 |
--skill / skill | 0 | 487 | 9 | 0 |
| root help | 0 | 155 | 5 | 0 |
Summary: every success is 0-1 lines stdout, silent stderr; fail is 1 line each side. Budgets hold: root 155/300B 5/12 lines, --skill 487/900B 9/30 lines, skill/skill.md 504/900B 17/30 lines, list --json 1 line/entity. No fixes needed.
How an agent operates f0rge:
f0rge --skill is the contract; f0rge + cmd -h is the map.
issue list --assignee you --state open, then issue view <id>.
Work in a worktree; verify with the repo's own ci steps.
pr create --base main --head <br> --issue <id>; owner merges.
Reconcile conflicts yourself; wiki put research/... what lasts.