Delegate bounded coding tasks to OpenCode and review the results

OpenCode Delegate hands a bounded coding task to the OpenCode CLI as a separate implementer. The orchestrator writes a brief and dispatches it with the bundled relay, which wraps opencode run, runs the brief under a chosen agent, captures the run, and writes a structured result.json. OpenCode does the typing in its own session with no shared context; the orchestrator reviews the diff and lands the commit. This is a docs-only import: the executable scripts/relay.mjs referenced throughout is not bundled here and has to be sourced upstream to actually run the loop.

Use it when the user has explicitly asked for delegation to OpenCode and the task is bounded enough for a single brief. Do not use it when the task is small enough to do inline, when opencode is not installed or not authenticated (opencode auth login), or when you only want a review - the read-only plan agent via --read-only covers that without a full write-capable dispatch. OpenCode has no safe default model: a bare opencode run errors, so a fresh run always needs --model provider/model or a --lane that sets one; naming the model is a two-owner decision where the human states which models are their flat-rate subscriptions (opencode models lists hundreds, most billed per token) and the orchestrator picks one of those per task.

The input is a brief - the goal, current state, what to change, what to leave untouched, the project's actual gate commands discovered from its own AGENTS.md/CLAUDE.md/Makefile, and a report contract - passed via --brief or piped on stdin. Required and optional flags: --model <provider/model> (required on a fresh run), --agent <name> (default build, write-capable), --read-only as a shortcut for the read-only plan agent, --variant <name> for provider reasoning effort, --no-auto to drop the relay's default auto-approve behavior, and --resume-last/--session <id> to continue a prior session with a delta brief (inheriting its model). The output is result.json, carrying status (completed, failed, timeout, aborted, or opencode_unavailable), agent, sessionId, finalMessage, cost (summed run cost in USD, or null), and touchedFiles - a git status --porcelain snapshot of the working root.

Wraps the OpenCode CLI, installed via npm i -g opencode-ai or the native installer from opencode.ai, with at least one model provider authenticated (opencode auth list shows the credentials). Permissions auto-approve by default because the relay passes --auto, so a headless run never blocks on a prompt no one can answer; --no-auto instead honors the agent's own allow/ask/deny permission config, which can hang a headless run if any action is set to ask. A read-only plan dispatch never receives --auto, so an auto-approved permission can't be used to sneak an edit past a review pass.

Orchestrating agents and engineers who already use OpenCode as an implementer and want a scripted brief-dispatch-review-commit loop instead of manual copy-paste, with model selection made explicit rather than left to a hidden default. The authorization model treats delegation as something the human opts into: the orchestrator never accepts OpenCode's result.json claims as evidence on their own, re-runs the project's real test/lint/build commands itself, reads the diff against the brief for scope creep or shortfall, and reports design decisions and non-blocking nitpicks back to the human rather than absorbing them silently.