Delegate bounded coding tasks to Kimi Code and review diffs

Kimi Delegate hands a bounded coding task to the Kimi Code CLI (kimi) as a separate implementer working in its own session, while you stay the orchestrator writing the brief and reviewing the diff. In headless -p mode, Kimi always runs in auto permission mode and never asks for approval - it rejects --prompt combined with --yolo, --auto, or --plan, so the relay passes none of them and offers no --read-only or --full-access option. There is no CLI-enforced read-only mode for Kimi: inspecting touchedFiles and the diff after every run, not a flag, is the guarantee of what changed. The bundled helper wraps Kimi's headless prompt mode, captures the structured event stream, and writes a result.json; a pre-run usage error exits 2 with no result file, and a missing kimi binary exits 127 with status: "kimi_unavailable". Completion means the process exited and result.json exists - trust process state and the working tree over a progress display; Kimi's full report is the finalMessage field, also printed on stdout between the report markers.

Use it to delegate a bounded coding task you want implemented and then review yourself, or when the user explicitly asks for delegation to Kimi. Do not use it when the task is small enough to do inline, when the kimi CLI is not installed or authenticated, or when you need a CLI-enforced read-only implementer - headless Kimi has none.

Input is a self-contained brief - Kimi sees only the text sent plus what it can inspect in the workspace, no chat history - with the goal, current state, what to change, what to leave untouched, the project's actual gates, and a report contract telling Kimi not to commit; one task per brief. Prerequisites checked once: install Kimi Code (brew install kimi-code on macOS/Linux, or the native installer from the official docs), authenticate with kimi login (device-code flow) or /login in the TUI, and confirm kimi --version succeeds. Kimi uses default_model from its config.toml when --model is omitted; passing --model <alias> requires an alias the human already configured, not an invented one. Output is a result.json with the implementer's finalMessage and touchedFiles.

Wraps Kimi's headless prompt mode through a scripts/relay.mjs helper; that executable script is not bundled here - see the upstream project for the runtime. Extra workspace directories use repeatable --add-dir flags; a prior session resumes with --resume-last or --session <id> (delta brief only); --timeout caps a run (30-minute default; implementation briefs routinely need 1-2h). The relay writes its own artifacts under the system temp dir and never commits. Adapted from amElnagdy/delegate-skills (MIT).

Anyone orchestrating with an agent that can run a shell command and access files, who wants to hand a bounded coding task to Kimi Code as implementer, verify its diff and re-run the project's own gates rather than trusting Kimi's self-report, and land the reviewed result themselves - surfacing Kimi's own design decisions rather than absorbing them, and stopping to ask before a run expands past the brief.