Part three of the Handmade Claude Code campaign: the agent learns where it is and remembers what happened. A loop with tools is a capable stranger; what makes it a colleague is context — the project's own instructions, the files the user points at, the environment it runs in — and memory: a conversation that can be picked up again, and one that survives running out of room.

This session continues the agent you built in parts one and two, in the same repository. Nothing about how it is started changes; what changes is what it knows before the first model call, and what it keeps after the last.

The command and the protocol, unchanged

agent: <command>     started as:  <command> -C <dir> [--yes] -p "<prompt>"

Project instructions

Before the first model call, the agent walks from the working directory up to the filesystem root and collects, in every directory, the file AGENTS.md — or CLAUDE.md when there is no AGENTS.md beside it. Their contents go into the system prompt, outermost directory first, innermost last.

Inside such a file a line that is exactly @<path> imports another file: its contents are inlined in place, the path relative to the importing file. Imports nest (at least two levels deep); a file already imported is not imported twice.

Mentions

A token @<path> in the prompt attaches that file: the user message carries the prompt as typed and the file's contents (a second text block, or the same block — your call). A path that does not exist is left as typed.

Where and when

The system prompt says where and when the agent is: the absolute path of the working directory, today's date as YYYY-MM-DD, and — when the working directory is inside a git repository — the current branch name.

Sessions

Every headless run is a session, named by the session_id the JSON output carries. Where and how you store it is yours (a file per session is the usual answer); what it must allow is:

• --resume <session_id> -p "<prompt>" — the new prompt is appended to that session's conversation and the whole of it goes to the model.
• --continue -p "<prompt>" — the same, with the most recent session of this working directory.

Compaction

Contexts fill up. .agent/settings.json may set a budget:

{"compaction": {"input_tokens": 100000}}

When the input_tokens the last reply reported exceeds the budget, the next model call is not the next turn — it is a compaction call: the conversation so far plus one user message asking the model to summarise it. The reply's text becomes the new conversation — one user message carrying the summary (worded however you like) — and only then does the turn the user asked for go to the model, on top of the summary. The old messages are gone from every request after that. The default budget is yours; the checks set a small one.

The judges

The quality panel — architecture, performance, code quality, test quality, technical governance and DX review — sits on the rungs where its subject is decided, one verdict per judge per rung, on top of the rung's own points. There is no closing review: what you build is judged as you build it, and a rung you never reach is a verdict you never get.

The ladder

• Set up and carry parts one and two forward (10)
• AGENTS.md is the system prompt (20)
• Up the tree (20)
• Imports (20)
• Mentions (20)
• Where and when (10)
• Resume a session (30)
• Continue the last session (20)
• Compaction (40)