The smallest serious coding agent has a 443-line edit tool, and that is the point

Half of pi's shipped tools are switched off when a session starts ^[7][1]. That is the audit worth running: `ls`, `grep`, `find` and `truncate` are in the repo and never reach the model unless someone opts in ^[7], so the surface an ordinary run puts in front of the model is the same four it began with ^[8]. Repo size tells you what a maintainer agreed to support. It does not tell you what the agent can reach in a session nobody configured.

The 443-line edit tool is the case against counting lines at all ^[3]. It is long because it applies multiple disjoint edits in one call, matches each against the original file rather than incrementally, detects and restores line endings, and serialises concurrent mutations ^[9]. That is one tool with four behaviours a shorter implementation does not have, not four tools. Compaction is in the codebase on the same logic, because long sessions need it, and the eval package because a harness you cannot measure is one you cannot improve ^[10].

The most portable thing in the writeup is smaller than any of that. pi checks whether the model's output was cut off by the token limit, and if it was, refuses every tool call in that message, on the grounds that a truncated argument list still parses sometimes ^[11]. The author reports his own harness records that same stop reason in three places and branches on it in none, a bug from a family that had already caught him twice ^[12].

He nearly filed the opposite story. The first note was that the minimalism thesis had eroded, and there was a paragraph half-written arguing the refusals had lapsed ^[18]; both grep matches dissolved when he read what they actually matched ^[6]. By his own count he has now been wrong that way six times on one project, including a metric that reported zero problems twice and a regex that silently matched nothing ^[13].

The evaluation side is the one place where the piece has corroboration rather than assertion. pi's eval package and the author's, built independently, landed on the same five primitives: isolate each run in a temp directory, persist the session as an artifact, index runs in a JSONL file, compare configurations against each other, and pin provider and model into the results ^[14]. The divergence is instructive too: pi's graders assert on tool calls by name, arguments, status and return value, while his grade the end state, which can report that a task succeeded without saying whether the agent got there acceptably ^[16].

So the label is carried by the refusal list, not the file count, and that list has not moved in a year ^[17]. Background bash was left out because it means managing processes an operator cannot see, subagents because they put decisions where nobody can inspect them ^[5]. An agent with four shallow tools is not minimal in that sense. It is just early.