update-madagents

Update MadAgents

Update an existing install in place. The core idea is a 3-way merge per file:

• base = the pristine original of the version the user installed — reconstructed by rendering that version's templates from git history (not a local copy the user could edit).
• current = what's in the repo now (base + the user's edits).
• new = the current version, freshly rendered.

git merge-file current base new then does the right thing per file: unchanged → take new; user-edited only → keep theirs; both → merge (clean keeps both; overlapping → conflict).

Uses the same per-provider adapters/<provider>/render.sh as install, so base and new render identically — the only differences come from the version change itself.

Procedure

1. Detect the install and read the manifest

TARGET="<repo path>"; MANIFEST="$TARGET/.madagents/install.json"
test -f "$MANIFEST" || { echo "No manifest — see Legacy fallback"; }
get(){ python3 -c "import json,sys;print(json.load(open('$MANIFEST')).get(sys.argv[1],''))" "$1"; }
COMMIT="$(get source_commit)"; INSTALLED="$(get version)"      # the commit id is all we need
PROVIDER="$(get provider)"; [ -n "$PROVIDER" ] || PROVIDER=claude_code   # legacy installs = claude_code
DOCS_DST="$TARGET/.madagents/madgraph_docs"; DOCS="$DOCS_DST"   # derived from the target, not stored
# Provider-specific instruction file + its block-payload name:
if [ "$PROVIDER" = codex ]; then IFILE="$TARGET/AGENTS.md"; IBLOCK="AGENTS.block.md"
else IFILE="$TARGET/.claude/CLAUDE.md"; IBLOCK="CLAUDE.block.md"; fi

2. Locate source, compare versions

SOURCE="$(git -C "$PWD" rev-parse --show-toplevel)"
RENDER="$SOURCE/install/data/madagents/adapters/$PROVIDER/render.sh"
NEW_VERSION="$(git -C "$SOURCE" describe --tags --always --dirty)"
NEW_COMMIT="$(git -C "$SOURCE" rev-parse HEAD)"
[ "$NEW_COMMIT" = "$COMMIT" ] && { echo "Already at $INSTALLED — up to date."; exit 0; }
echo "Updating: $INSTALLED  ->  $NEW_VERSION"