Skip to content
repo

release-engineering

Deploy. Verify. Converge. Then ship.

2 skills1 human gate15–30 min/session
Install this pack Read the reference

From a clean build to a verified production release

The contract

Use it when
A build-loop PR is adversarial-review-clean and ready to go to production.
You type
Run the release loop for this build.
You provide
A merged, adversarial-review-clean inner build-loop output.
You receive
A release readiness record — e2e results, telemetry snapshot, security review — and a convergence-verified prod ship.

Where you decide

The agent pauses at these points. You choose whether to continue, redirect, or stop.

Approve the production releaseDecision 1 of 1

What changes when you install this

After installing release-engineering, completed build-loop output goes through release-loop before reaching production. The release-lead agent deploys to an ephemeral environment, runs e2e tests, observes telemetry, and feeds deployed findings back to the inner loop — no human relay. You review at one gate: the prod ship.

Requires:

The journey

1. Trigger and confirm the deploy

  • Agent does: activates release-loop on completion of the inner build loop (work-loop + adversarial review clean); deploys the integrated whole to an ephemeral environment.
  • You do: watch the initial deploy log to confirm the right branch and config are targeted; if the deploy target looks wrong, stop it early — a mid-cycle redirect costs more than catching it at the first log line.
  • Output: a running ephemeral deployment.
  • State: read-only

2. Validate and converge

  • Loop does: runs end-to-end tests against the deployed environment, observes telemetry, feeds deployed findings back to the inner loop, redeploys after each inner-loop fix, and iterates until convergence — e2e clean, telemetry stable.
  • You do: check in at the end of each outer loop iteration — skim the e2e results and telemetry snapshot; flag anomalies the agent might not catch (an assertion too weak to detect a real failure, a telemetry spike marked as noise); provide judgment on what “stable” means for your service.
  • Output: a converged deployed state — e2e clean, telemetry stable.
  • State: read-only

3. Ratify the release readiness record

  • Agent does: generates the release readiness record — e2e results, telemetry snapshot, security review on the deployed diff, deferred items, and any borderline gates.
  • You do: read the full release readiness record, not just the summary; the borderline gates section matters most — these are the agent’s “close enough” calls you may decide differently.
  • You decide: approve the prod ship — ratify if satisfied, or reject with a one-line reason to re-enter the loop.
  • Output: a prod-ship decision; after this gate the change reaches real users or real data.
  • State: confirmed-write

What good output looks like

  1. YouUse release-loop for the merged export-filter change. Deploy it to an isolated ephemeral environment, converge on end-to-end results and telemetry, then prepare the release-readiness record. Do not deploy to production without my approval.
  2. AgentThe target is the merged, independently reviewed change. The environment is isolated from production data and traffic, and teardown is available. Starting the ephemeral deployment.
  3. AgentFirst iteration: the empty-result export flow failed end to end. Other checks and telemetry remained stable. I returned the defect to the inner loop, verified the reviewed fix, and redeployed.
  4. AgentSecond iteration: all release end-to-end checks pass; latency and errors remain within the recorded service objectives; the deployed-diff security review is clean; and the rollback rehearsal passes. The release-readiness record links the evidence. Borderline gates: none. Production remains unchanged.
  5. YouApprove the production release.

Human gates

For each gate, everything you need to make a confident decision.

  • Approve the production release

    Trigger
    After the deployed whole converges — e2e clean, telemetry stable, security review done
    Time
    15–30 minutes
    What to check, good, bad, consequence

    What to check

    • Is the release readiness record complete? (e2e results, telemetry snapshot, security review)
    • Are there any borderline gates in the record? (The agent surfaces these — don't skip them.)
    • Is there anything in the diff that wasn't in the scope of this release?
    • Is the rollback path documented and tested?

    What good looks like

    A release readiness record showing convergence — e2e clean, telemetry stable, security reviewer clean, no borderline gates.

    What bad looks like

    A record that notes borderline gates you wave through. Or a record that omits the security diff review.

    Consequence of skipping

    This gates the production release. The agent cannot ship to production without your ratification. After this gate, the change reaches real users or real data — irreversible.

Typical session

Agent turns
varies
Human gates
1
Wall-clock time
15–30 min

Install

agentbundle install --pack release-engineering

Skills in this pack

  • release-loop1 gate

    The release supervisor. Deploys to ephemeral env, runs e2e, feeds findings back to the inner loop, iterates to convergence, then surfaces for prod ship approval.

  • define-slo

    Authors an OpenSLO v1 YAML document for a service. Produces the SLO artifact the release-loop PRR gate reads to populate the error-budget status field.

Technical reference & next steps