DIGX Antigravity Agents
Deploys the Claude Manager + Implementer protocol into a repository: Claude plans, adjudicates and verifies; the Antigravity CLI implements; a fresh Claude subagent reviews.
The value is not the delegation — it is that the delegation is checkable. An implementer that reports success, an empty fingerprint, and a gate suite with no gates in it all look identical to a clean run from the outside. Everything installed here exists to make one of those distinguishable.
Route on what you were given
| Situation | Do this |
|---|---|
agy --version fails, or first use on this machine |
Read reference/setup.md |
setup — install or authenticate the CLI |
Read reference/setup.md |
No argument, repo has no .agy/ |
Read reference/install.md, then install |
No argument, repo has .agy/ |
Report status: config, surfaces, ledger, gates |
install / adopt / setup request |
Read reference/install.md, then install |
status |
.agy/tripwire check + ( . .agy/config && bash $AGY_GATES ) + tail the ledger |
doctor — something is off |
Read reference/dispatch-traps.md first |
| A run is hanging, was WATCHDOG-KILLED, or died before its result event | reference/dispatch-traps.md, trap 4 |
| A run reports PROBLEMS but the work is committed and the gates are green | reference/dispatch-traps.md, trap 6 |
protocol — how do I run a task |
Read reference/protocol.md |
| Writing the plan, not just running it | reference/protocol.md, planning dispatch |
| Reviewers keep dying mid-review, or reviews are stacking up | reference/protocol.md, sizing the review |
| About to run a task in a repo already set up | Read reference/protocol.md |
Read the reference file before acting. These scripts encode failures that are invisible when you get them wrong — that is the whole point of them.
Installing
Prerequisite: agy, node, git and bash on PATH, with agy
authenticated. If agy --version does not answer, stop and read
reference/setup.md — the installer's final check will fail anyway, and it is
cheaper to find out now.
bash ~/.claude/skills/agy-agents/scripts/install --dry-run # always first
bash ~/.claude/skills/agy-agents/scripts/install --project NAME --milestone m1
Run from the repository root. It is idempotent and never overwrites an edited
.agy/config or .agy/gates without --force.
Adoption is the same command: it finds an existing SDD workspace and an
existing gate runner and binds to them rather than duplicating them. Details,
flags and the full config reference: reference/install.md.
Then do the two things the installer cannot do.
- Write the real gates. It emits a starter from whatever toolchain it detected. A generated gate suite is a guess, and gates are the only thing standing between a confident report and a broken tree.
- Fill every
{{marker}}indispatch-context.md. Only you know what goes in them, and that file is read by every implementer and every reviewer on the plan — an unfilled constraint is inherited by all of them and reads as no constraint..agy/dispatchrefuses to run until they are gone.
Running a task
Never call agy directly — three of its behaviours make a broken run report
success. Use .agy/dispatch <N>, which handles all three.
The loop, the reviewer contract, the fix loop and task sizing are in
reference/protocol.md. The installed repo gets its own thin
agy-task-cycle skill holding just that project's bindings.
Verifying the harness itself
bash ~/.claude/skills/agy-agents/scripts/selftest
Builds a throwaway repo and exercises the whole harness against a stubbed
agy — no quota spent, nothing outside its temp dir touched. Run
it after editing anything under assets/. It aborts rather than run if agy
resolves to the real CLI.