Retro runs at the end of a session, started by name: /agent-harness:retro in Claude Code or $agent-harness:retro in Codex. It lists what slowed the session down, with the moment each item happened, finds the cause, and builds the fix into the setup.
The output is fixes made and checked, not a list of lessons. It prefers code (a script, hook or test), then an improved instruction or route, and adds a new rule only when the session gave it a receipt.
What You Get
- A list of this session’s friction items, each tied to a moment
- A cause for each item from five classes
- Reversible repository fixes made and verified where they apply
- Exact proposed text for changes outside the repository, such as user-level settings
- A one-line report per item: incident, cause, fix, verification
Common Use Cases
- After an agent ran tests against the wrong server twice
- After the owner corrected the same mistake in one session
- Recording a correction to an installed skill in the project’s own instructions
- Deciding that a problem was a one-off and needs no rule
What Makes It Unique
Retro refuses to invent rules. An item with no moment behind it is dropped, a new rule needs a dated receipt, and code comes before prose. Corrections to a third-party skill go in the project’s own instructions, since an installed copy is overwritten by updates. It edits only reversible repository files and proposes global changes with exact text for the owner to apply.
Under The Hood
The skill is one instruction file with no scripts, and it runs only when you name it. The agent reviews the session, then edits repository files such as instruction files, routes, hooks and scripts, and checks them by running them in the shell, for example piping a sample payload into a hook. It leaves the user’s own settings and global instructions alone unless the owner says so, and it drafts an upstream issue only when asked.