git:20261012.5d5a3e1 to git:20261012.9adcef4

3 added, 7 removed. Audit A to A.

# AGENTS.md

<!-- Project instructions go here, outside the prompt-forge block. Keep them

short: only what an agent cannot infer from the code. -->

## Team knowledge (prompt-forge)

- This repository keeps a reviewed, shared record of what AI coding agents got

- wrong here, so that one teammate's agent does not repeat another's mistake.

- **Before a task**, read the `description:` line at the top of each file in

`knowledge/memory/` (a content search for `^description:` lists them all) and

open only the files that match the task. They hold facts the team has

verified. If an entry turns out to be wrong, say so and record it with

`record-lesson`.

- **At a checkpoint** (before committing, before opening a pull request, or when

reporting a task as done), if you hit a non-obvious failure that a teammate's

agent would likely repeat, run the `record-lesson` skill. Most tasks produce no

- lesson, and that is fine.

- - **Curation** runs only when a person asks for `curate-lessons`. It works on its

- own branch and lands through a reviewed pull request. Outside a curation run,

- agents do not edit `knowledge/memory/`.

+ lesson; then say nothing about it.

+ - **Curation** runs only when a person asks for `curate-lessons`. Outside a

+ curation run, agents do not edit `knowledge/memory/`.

- Skills live in `.agents/skills/` (and in `.claude/skills/` for Claude Code). If

your agent does not load Agent Skills natively, read a skill's `SKILL.md` when

the task matches its `description`.