Clarifying one requirement retired the best argument for keeping this agent in bash

Behaviour had already left the language before the rewrite started. Prompts live in individual Markdown files with {{placeholder}} holes that the code fills, a choice the author made because prompts get edited constantly and because a stray `$` or backtick has to stay inert instead of being eaten by the shell ^[17]. After the migration he checked modification times, and the port had not changed a word of any prompt ^[18]. The tuned criteria and the order in which judgments get made live in those files, so switching languages replaced only the glue that assembles them ^[19].

The second decision was a workaround. Each kind of task is its own executable: verb on argv, JSON on stdin, JSON on stdout, adopted because bash has no modules and a process boundary beat having the pieces scribble on each other's variables ^[20]. The author calls that "Pure coping." ^[21] It also meant there was no big-bang rewrite ^[22].

Forking was never the complaint. 162 forks cost nothing next to a minute of model latency ^[10]. Expressiveness was the cost: every structure in the program either fit in a one-line `jq` expression or got split into three pieces, so what got written was the structure `jq` could state on one line ^[11]. He had three lists, the files a change actually touched, the files the agent itself had written, and the files approved in advance, and needed two differences between them ^[12]. The Python comprehensions are barely shorter than the `jq`; the difference is that they are what he would have written on the first attempt, while the `jq` took several tries ^[12].

One run died on line 567 with `1: command not found`, and he never located it; it went away when that section was rewritten for unrelated reasons ^[13]. The post does not say the agent gained test coverage before or during the port ^[23]. Rewriting 2,000 untested lines was one of the original arguments for staying ^[5].

The 86 millisecond difference is about 0.29 percent of the shortest run this program does ^[1]; one 30-second model call covers roughly 337 Python startups ^[4]. For that number to transfer, your program has to start the interpreter once and then spend its life waiting on something slow. Invoke the same script once per file from a build loop and a thousand calls add 86 seconds of startup ^[2].

Comments and docstrings are 43 percent of the Python file ^[3], and they are the notes recording which specific incident each safety check exists to prevent, which is exactly what the author had been afraid of losing ^[24]. "Evaluate this rewrite by total line count and it accomplished nothing," he wrote ^[16].

"My reasoning was valid; its inputs weren't, and I had spent almost no effort checking them," the author wrote ^[9]. That is the part worth copying. The language question was the less interesting one here, and the bug ledger says so: the expensive bugs were design errors that any language would have permitted ^[6]. A requirement recorded as hard turned out to be a preference about starting simply ^[7], and 86 milliseconds was the whole of what it had been protecting ^[8].