Delegate bounded coding tasks to Aider and review diffs

Hands a bounded coding task to a separate implementer - Aider - then lets you review what it produced and land it yourself. You write the brief and own the judgment; Aider does the typing in its own run; you verify and commit. The loop needs only a shell command and file access, so any comparable orchestrator can drive it.

Aider commits by default, and two of its defaults would destroy the reviewable diff this skill exists to produce: --auto-commits (default True) commits Aider's own edits after each exchange, and --dirty-commits (default True) commits pre-existing uncommitted work before Aider starts editing. The relay always passes --no-auto-commits and --no-dirty-commits, neither configurable through it - driving aider by hand instead requires passing both flags manually, or the work lands as commits nobody reviewed. The relay also passes --no-gitignore, since Aider otherwise writes .aider* into .gitignore on startup and dirties the tree that is about to be read.

The loop is five steps. Steps 1, 4, and 5 require judgment; 2 and 3 are mechanical:

1. Write the brief. Aider sees only the text sent to it plus the files in its editing scope - no chat history or shared context. Include the goal, current state, what to change, what to leave untouched, the project's actual gates, and a report contract. Keep one task per brief.
2. Dispatch via the bundled helper, which wraps Aider's headless --message-file mode, captures the run, and writes result.json. The brief rides --message-file rather than argv, so it stays out of the host process list and clear of the OS argument size cap. The relay writes artifacts under the system temp dir by default and never commits.
3. Wait for completion. The helper blocks until Aider finishes. A pre-run usage error exits 2 and writes no result; a missing aider exits 127 and writes status: "aider_unavailable". Aider itself exits 0 even when it never reached a model, so the relay scans the run for Aider's own endpoint and authentication errors and reports status: "failed" when it finds one - a failed status with an endpoint-related error is a configuration problem, not a coding failure.
4. Review - do not trust the self-report. Re-run the project's gates yourself, read the diff against the brief starting with touchedFiles, run relevant guard skills if installed, and round-trip migrations plus grep for dangling references after removals or renames. Aider's --auto-lint is on by default, so it may have already run a linter and fixed its own complaints - that is Aider's lint, not the project's gates, so run those anyway.
5. Land it. The implementer edits the working tree; the orchestrator commits, only after the gates pass and the diff holds. If rework is needed, send a delta brief with --resume-last, then review again.

Use it to delegate a bounded coding task to Aider and get a clean, unreviewed-diff-free result back.

Do NOT use it when the task is small enough to do inline, when the aider CLI is not installed or no model is configured for it, or when the implementer should manage its own commits - Aider can, but this skill deliberately turns that off, since the diff is the deliverable.

Prerequisites: install Aider via python -m pip install aider-chat or the standalone installer, configure a model (Aider reads provider keys like OPENAI_API_KEY or ANTHROPIC_API_KEY from the environment, or its own config), confirm aider --version succeeds, and work in or point --cd at the target git repository. Omitting --model uses Aider's own configured default.

Aider also talks to any OpenAI-compatible endpoint - llama.cpp's server, Ollama, vLLM, LM Studio, or anything serving the same API - by pairing --model with --api-base, e.g. --model openai/<served-model-name> --api-base http://127.0.0.1:<port>/v1. The openai/ prefix is required to tell Aider to speak the OpenAI protocol; a placeholder non-empty OPENAI_API_KEY is still needed even though the server ignores its value; and local models often fail Aider's default diff edit format, so --edit-format whole trades tokens for reliability, paired with a tight --file scope to keep whole-file rewrites cheap. A local endpoint that isn't running looks like a hang, not an error, until the relay's --timeout watchdog fires and reports status: "timeout".

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

Other flags: --model <name>, --api-base <url>, --file <path> (repeatable) to scope the edit surface, --read <path> for context only, --read-only for a dry run with no files modified, --resume-last to continue the previous chat, and --timeout 2h for a hard watchdog limit.

No account or provider registration is needed for local use - Aider is a pip install, the endpoint is the user's own, and OPENAI_API_KEY only has to be non-empty. The relay pins flags that would otherwise reach the network on their own: --no-check-update, --no-analytics (Aider's own default is random, which opts some sessions in by itself), and --no-detect-urls (without which Aider offers to scrape any URL in the brief, and --yes-always accepts that offer silently). --no-suggest-shell-commands closes the remaining network path. What stays outside the relay's control is the brief itself - instructions telling Aider to install a package or call an API are still carried out, and --auto-lint runs the repository's own tooling. The relay passes --yes-always because a headless run cannot answer a prompt, and that consent covers every prompt Aider would otherwise raise, not just file edits. File selection (--file, --read, --subtree-only) sets chat context and token cost, not a security boundary - it does not confine what Aider can reach. --read-only maps to Aider's --dry-run; the relay does not independently verify the claim, instead reporting what git status --porcelain shows and warning if a read-only run left the tree changed. Aider has no session ids - its resume unit is the .aider.chat.history.md file kept in the repository, so --resume-last maps to --restore-chat-history and resume is per-worktree, not per-user.

Orchestrating agents or developers who want to delegate a bounded coding task to Aider, including against local or self-hosted models, while keeping review and commit authority themselves.