Gemini DevOps Agent

This is the dev-only execution layer for using the Gemini plan where it is most useful: bounded build, test, deploy, log, and replay work with external truth in exit codes, reports, screenshots, traces, metrics, and rollout status. Codex and Claude remain the architects and engineers. Antigravity CLI runs constrained DevOps recipes and reports evidence back.

The practical change from the original Gemini CLI idea is now mandatory: use Antigravity CLI (agy), not Gemini CLI, and pin the session to Gemini 3.5 Flash (High).

Implementation evidence

The scan covered the current BestBuds checkout and the adjacent Reaktor framework checkout. These are the source-of-truth surfaces the tool references:

Two older memory paths have drifted in the current checkout: n8n/BUILD_MATRIX.md and cloud/observability/README.md are not present. The replacement source of truth is the set above.

Execution model

Codex / Claude
  solutioning, architecture, source edits, final judgment
        |
        | bounded handoff packet
        v
tools/gemini-devops/handoff.mjs
  recipe catalog + prompt pack + safety class + model pin
        |
        | agy --model "Gemini 3.5 Flash (High)" --add-dir ../reaktor
        v
Antigravity DevOps Executor
  reads source-of-truth docs, runs recipe commands, classifies failures
        |
        v
Gradle / Dagger / npm / Wrangler / Fastlane / Maestro / Playwright /
Karate / k6 / Keploy / k3s / Pulumi / Supabase
        |
        v
evidence report: exit codes, logs, screenshots, traces, reports, URLs

This is intentionally not a general coding agent. Antigravity should not invent architecture, edit unrelated files, add dependencies, push git state, or decide production safety. It should run a known recipe and return evidence.

Repo tooling

The checked-in tool lives at:

tools/gemini-devops/
  README.md
  handoff.mjs
  recipes.json
  agent-card.json
  prompts/system.md

The wrapper resolves agy from /Users/ovd/.local/bin/agy if the shell PATH does not include it. It pins Gemini 3.5 Flash (High) by default and adds the sibling Reaktor repo with --add-dir /Users/ovd/dev/reaktor. If any other model is requested, the wrapper refuses the run.

List recipes:

node tools/gemini-devops/handoff.mjs list

Generate a prompt packet without starting Antigravity:

node tools/gemini-devops/handoff.mjs prompt web.reaktor.checks \
  --goal "Validate the current Reaktor Web docs and workbench changes."

Run Antigravity:

node tools/gemini-devops/handoff.mjs run web.reaktor.checks \
  --goal "Validate the current Reaktor Web docs and workbench changes." \
  --timeout 30m

Run interactively when human approval will be needed:

node tools/gemini-devops/handoff.mjs run mobile.ios.smoke \
  --interactive \
  --goal "Run the iOS dev-login smoke after the chat persistence changes."

Mutating recipes are blocked unless the caller passes --allow-mutating after operator approval. That includes Cloudflare deploys, k3s rollouts, mobile distribution, live fixture writes, load tests, Keploy capture, and docs deploy.

Test the handoff layer:

npm run test:gemini-devops

The test suite uses Node's built-in test runner and a fake local agy binary. It validates the recipe catalog, approval gates, prompt generation, pinned-model rejection, A2A discovery/message/JSON-RPC packets, and the exact Antigravity argument shape without spending Gemini tokens.

A2A orchestration

A2A helps at the boundary between the high-quality solutioning agents and the cheap routine executor. Codex or Claude can hand a structured A2A task to the Gemini-backed DevOps executor instead of spending its own context watching Gradle, Wrangler, Maestro, Karate, k6, Keploy, or k3s loops.

The local shape is packet generation, not a long-running server:

node tools/gemini-devops/handoff.mjs a2a-card

That command generates an A2A agent card with one skill per checked-in recipe. It declares local HTTP+JSON and JSON-RPC interfaces, the Gemini 3.5 Flash (High) model pin, and a Reaktor-specific recipe extension that carries safety class and approval metadata.

Codex or Claude can then generate a task request:

node tools/gemini-devops/handoff.mjs a2a-message k3s.status \
  --client claude \
  --goal "Report node, pod, and service health. Do not mutate anything."

For clients that prefer JSON-RPC:

node tools/gemini-devops/handoff.mjs a2a-jsonrpc web.reaktor.checks \
  --client codex \
  --goal "Run the Reaktor Web browser regression checks and return artifacts."

The A2A message has two parts:

• a text prompt that Antigravity can run directly with agy;
• structured JSON containing the recipe id, safety class, allowed commands, expected artifacts, model, cwd, sibling Reaktor path, and approval state.

production-write and staging-write recipes are intentionally blocked from A2A packet generation unless the caller passes --allow-mutating after operator approval. This keeps A2A from becoming an unreviewed deploy tunnel.

When this moves into Reaktor Desktop, the same contract should be hosted as:

• GET /.well-known/agent-card.json for discovery;
• POST /message:send for blocking task execution;
• POST /message:stream once the run history panel can consume streaming status and artifact updates;
• a task store where each DevOps recipe run has a context id, task id, status, terminal result, and immutable artifact list.

The production executor should still call the same recipe catalog. A2A gives Claude/Codex a stable delegation interface; it does not expand what Gemini is trusted to decide.

Recipe matrix

Safety classes

Prompt constraints

The system prompt in tools/gemini-devops/prompts/system.md forces these rules:

• run only the selected recipe or harmless read-only inspection;
• do not edit files unless a recipe explicitly permits it;
• never run git commit/push/stash/reset;
• do not install dependencies without approval;
• do not run production deploys, rollouts, distributions, migrations, or secret changes without explicit approval;
• redact secrets and user identifiers in summaries;
• classify failures as environment/tooling, missing dependency, missing secret, device/simulator state, compile/typecheck, test regression, performance budget, live infra, flaky timeout, or needing Codex/Claude solutioning.

The output contract is fixed:

Recipe:
Goal:
Model:
CWD:
Sources read:
Commands:
Artifacts:
Result:
Failure class:
Next action for Codex/Claude:

That report shape is intentionally easy to lift into Reaktor Desktop run history later.

Handoff examples

Validate docs locally:

node tools/gemini-devops/handoff.mjs run docs.reaktor-web-build \
  --goal "Validate the new Gemini DevOps Agent docs page."

Ask Antigravity to gather live cluster status:

node tools/gemini-devops/handoff.mjs run k3s.status \
  --goal "Report node, pod, and service health. Do not mutate anything."

Prepare a prompt for a risky deploy without executing it:

node tools/gemini-devops/handoff.mjs prompt server.deploy-k3s \
  --goal "After Codex confirms the server fix, deploy the JVM server to k3s."

Run an approved docs deploy:

node tools/gemini-devops/handoff.mjs run docs.reaktor-web-deploy \
  --allow-mutating \
  --goal "Deploy this documentation page and verify the live route."

Productionization path

This dev-only layer is deliberately boring so it can become a Reaktor Desktop feature without redesign:

The first production version should not expose arbitrary shell. It should import this recipe catalog as a set of typed operation nodes with safety classes, inputs, expected artifacts, and permission gates. The "free-form agent" remains outside the trusted boundary.

What Antigravity should not own

Antigravity should not own:

• source architecture decisions;
• broad refactors;
• dependency upgrades;
• schema design;
• production deploy judgment;
• secret management;
• git state;
• final incident calls.

It should own:

• choosing the right checked-in recipe after a diff scan;
• running approved commands;
• watching logs;
• classifying failures;
• collecting artifacts;
• retrying only safe flaky steps;
• handing a tight diagnosis back to Codex or Claude.

That is the token-saving boundary: Antigravity burns the repetitive operational cycles, while Codex and Claude spend their context on design, source changes, and judgment.

Related docs

• DevOps - Use before changing the build/deploy/ops pipeline, the Kotlin platform model, deployment partitioning, gated deploy operations, or the reaktorDesktop Deploy pane; pairs with reaktor-cloud-pane (operator view) and secrets-management.
• Control Plane - Use when exposing Reaktor safely to humans and agents.
• Koog Workbench - Use before wiring real agent behavior into reaktorDesktop.
• Security Plan - Use before enabling deploy actions, remote agents, or privileged commands.
• Shadow Dev - Use before building production-like test and replay environments.
• Auth Architecture - Use before changing reaktor-auth or request authorization boundaries.

References

• Google Developers Blog, Transitioning Gemini CLI to Antigravity CLI (May 19, 2026).
• Google Antigravity, CLI migration documentation.
• A2A Protocol, specification: agent card discovery, supportedInterfaces, SendMessage, task status, artifacts, and .well-known/agent-card.json.
• Local verification: /Users/ovd/.local/bin/agy --version returned 1.0.5; /Users/ovd/.local/bin/agy models listed Gemini 3.5 Flash (High).