Delegate bounded coding tasks to Aider and review diffs

Lets an orchestrating coding agent hand a bounded task to Aider as a separate implementer, then review and commit the result itself. The critical fact about Aider is that it commits by default: --auto-commits (default True) commits its own edits after each exchange, and --dirty-commits (default True) commits the orchestrator's pre-existing uncommitted work before editing even starts. The bundled relay always passes --no-auto-commits and --no-dirty-commits - neither is configurable through it - plus --no-gitignore, because Aider otherwise writes .aider* into .gitignore on startup and dirties the tree being reviewed. Dispatch wraps Aider's headless --message-file mode, so the brief never rides the command line (no process-list exposure, no OS argument-size cap) and writes a structured result.json. The relay also pins --yes-always (auto-confirms every prompt, since a headless run cannot answer one) together with --no-suggest-shell-commands, --no-analytics, --no-check-update, and --no-detect-urls, so nothing in the dispatch path reaches the network or runs a suggested shell command unread. Resume maps to Aider's own per-repository chat history file (.aider.chat.history.md) via --resume-last, so resume is per-worktree, not per-user. It also talks to any OpenAI-compatible local endpoint (llama.cpp, Ollama, vLLM, LM Studio) by pairing --model openai/<served-model-name> with --api-base, using --edit-format whole for local models that struggle with Aider's default search/replace diff format. result.json reports status (completed/failed/timeout/aborted/aider_unavailable), exitCode, aiderVersion, finalMessage, touchedFiles, readOnly, resumed, error, and stderrTail.

Use it to delegate a bounded coding task to a separate Aider implementer, reviewing its diff yourself afterward, when the user has explicitly asked for delegation to Aider. Do NOT use it when: the task is small enough to do inline and delegation overhead isn't worth it; the aider CLI is not installed or no model is configured for it; or you want the implementer to manage its own commits - Aider can do that, but this skill deliberately turns it off because the diff, not the commit, is the deliverable.

The brief is written to brief.txt and delivered through --message-file, so it never rides argv and has no size cap; it should state the goal, current state, the specific change, a DO NOT TOUCH list, the project's real gate commands, and an explicit REPORT section, since Aider will not volunteer a structured summary otherwise. --file <path> (repeatable) sets Aider's editing scope and --read <path> supplies read-only context - both are chat-context and token-cost controls, not a security boundary, since Aider has no sandbox. Dispatch:

node "<skill-dir>/scripts/relay.mjs" --brief brief.txt --cd /path/to/repo

Outputs land in a run directory (temp by default): brief.txt, final.txt (Aider's report), stderr.txt, and result.json, published atomically. touchedFiles (git status --porcelain lines) is the review starting point; the implementer never runs git add, git commit, or git push.

Wraps the Aider CLI (installed via python -m pip install aider-chat), which reads provider keys from the environment (OPENAI_API_KEY, ANTHROPIC_API_KEY) or its own config, and can target any OpenAI-compatible local server through --api-base (llama.cpp, Ollama, vLLM, LM Studio) using an openai/<served-model-name> model string. Requires Node 18+ and git for the relay, and reads a project's own CONVENTIONS.md, style guide, or CLAUDE.md/AGENTS.md (passed with --read) to discover binding conventions and real gate commands.

Orchestrating coding agents that need to delegate a scoped implementation task to Aider - including to a self-hosted or local model - while keeping review-and-commit authority. Adapted from the community amElnagdy/delegate-skills repository under the MIT license; this catalog entry is a docs-only import, since the executable scripts/relay.mjs is not bundled here.