external-ai

Write the review prompt to a temp file with absolute paths and an explicit output file instruction, then launch the CLI (codex exec, agy -p, kiro-cli chat, or claude -p) in the background with stdout and stderr redirected. Wait for the CLI-specific completion signal and collect results from the output file first.

Use Codex for line-by-line verification and long deep investigations, since it excels at exact code cross-referencing but has heavier setup. Use agy for quick independent reviews, cross-cutting investigation, and long-form generation, as it finishes in tens of seconds to five minutes.

Codex can end a session without returning a final assistant message, leaving stdout empty even though the tokens-used sentinel appears in stderr. Instruct it in the prompt to write results to a file via apply_patch, and lower reasoning effort to medium.

No. kiro-cli returns exit code 0 even when tool approvals fail or executed shell commands fail; only missing auth or empty input returns 1. Detect completion by checking the result file exists and scanning ANSI-stripped stderr for rejection warnings.

Avoid delegation for tasks under two minutes, interactive design discussions, simple questions answerable via web search, and code containing confidential information, since prompts are sent to external APIs and incur per-token API costs.