Test-fix loop
Aider's canonical recipe — failing test in, fix attempt out, retry once with the parser's error fed back, then bail.
This is the most common coding-agent pattern: feed a failing test to the model, apply the fix, and if anything failed to apply, hand the structured failure message back to the model for one retry. After that, give up.
import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";
import { applyEdits } from "editkit";
import { readFile, writeFile } from "node:fs/promises";
async function attempt(prompt: string) {
const { text } = await generateText({
model: openai("gpt-4o"),
system: SEARCH_REPLACE_PROMPT,
prompt,
});
const results = await applyEdits(text, (p) => readFile(p, "utf8"));
for (const r of results) if (r.ok) await writeFile(r.path, r.after);
return results.filter((r) => !r.ok);
}
const failures = await attempt(
`Fix the code so this test passes:\n\n${testOutput.slice(0, 2000)}`,
);
if (failures.length) {
await attempt(
`Your previous edit failed:\n${failures
.map((f) => `${f.path}: ${f.message}`)
.join("\n")}\n\nTry again.`,
);
}Why "one retry, then bail"
If the model can't recover from its own structured failure message, two more rounds won't help — they'll just burn tokens. The aider design treats the failure message as a single high-signal correction. Match that.
See it run
The agent-loop example
runs this exact pattern offline with a deterministic mock LLM. The first attempt fails on
purpose (model invents a function name); the second succeeds after seeing the structured
search-not-found error.
bun run --filter editkit-example-agent-loop demo