Deliver

An explicit Deliver request for an implementation or fix runs through commit, push, PR, review and CI settlement, merge, required release or deployment, and consumer verification. Explicit plan-only, diagnosis-only, review-only, local-only, or PR-only requests retain their narrower endpoints. Routine edits and verification run natively; internal skill selection does not expand user authorization. For configuration and settings, consumer verification means reading the value back where it is consumed, such as the config file, a CLI readback, or a model list; live-process adoption, restarts, and inference canaries are out of scope unless requested.

What it adds

Deliver coordinates useful Compound Engineering (CE) stages and the requested endpoint. Automatically select a CE stage when it helps, resolving the dependency when that stage is needed. CE owns review settlement and CI/PR monitoring; Deliver continues its settled result to the authorized terminal boundary.

How it works

Choose the stages that fit the work: planning, diagnosis, structured implementation, review, or shipping. Use compound-engineering:ce-commit-push-pr when creating a PR or pushing user-requested commits to an existing PR. Before every push, review the exact candidate’s complete cumulative change-set and affected lifecycle through the whole-candidate gate. Existing Thermos correctness/security and maintainability passes alongside Codex review fulfill this coverage; their completed findings feed the same CE review owner. Standalone Thermos review remains optional when publication is outside the requested scope. Reuse unchanged proof and keep a compact receipt binding candidate identity, coverage, and findings dispositions. Required pre-push evidence blocks publication; publication-dependent release, install, and consumer checks remain pending with a named delivery-tail owner.

When TYPESAFE_API_KEY is present, Jev is the default for model and effort selection, workflow choice, evidence selection, work priority, and review triage throughout delivery. Explicit choices and privacy restrictions take precedence; uncertainty or service failure falls back to ordinary judgment.

For delegation, choose model and reasoning effort together: Opus 5.5 at medium in Claude Code, GPT-6.1 Sol at medium in Codex. Sol owns implementation, debugging, planning, and ambiguous or multi-file work even when small or bounded. Use Fable 5.1 for harder Claude Code work. Raise Sol 6.1 to high, xhigh, or max for justified demanding Codex work; reserve Astra for a rare, explicitly justified escalation after a concrete residual failure or quality gap on Sol 6.1. Omitting the model means the child inherits. Native children handle ordinary subtasks, and visible tasks require an explicit user request.

Illustrative delivery outline:

> Use Deliver to fix the retry path in the webhook worker through consumer verification.
model=gpt-6.1-sol  effort=medium
scope=bounded-change  review_and_ci=CE
tail=CE-disposition -> merge -> required-release/deployment -> consumer-check
stop=report observed delivery state and verification

Scope

One software change or PR belongs here. Routine parallel work uses native children; explicitly requested fleet/account allocation or a delegated remote agent belongs to Orchestrate. Contracts, retrospectives, and cleanup are available on demand.

Source

Ships in the railyard plugin.

Proof point

Record the observed PR and merge commit, verify reachability from the intended base, and complete required release or deployment steps. For a plugin, that includes required marketplace pins, a supported manager update, and installed-runtime verification. A local pass, open PR, merge, and deployed result prove different boundaries; an intermediate result does not finish an explicit Deliver request.