fix(eval): parse code fences like lib.sh in the outputs eval

test-convert-outputs.sh's Python fence model read fences only at column 0. With an indented opening fence and a column-0 closer - valid CommonMark that GitHub renders, lint accepts and convert_openclaw splits correctly - the model missed the opener and treated the closer as its own opener, so it reported a torn fenced block across SOUL.md/AGENTS.md and failed the Check Tools workflow on an agent file that was fine.

Mirror lib.sh's fence_open_p / fence_closes_p (up to three spaces, same character, a closer at least as long as the opener with nothing but whitespace after the run) so the eval sees the same blocks the converter does.

test-convert-fence-model.sh builds a throwaway repo with one such fixture and runs the real eval; on the previous model it fails with the false torn-block report, and with this change all tools validate. The full roster still passes with zero manifest drift.

Author: Tong-bit-art

.github/workflows/check-tools.yml

@@ -36,6 +36,9 @@ jobs:
         # maintainers regenerate at landing) and enforced on pushes to main.
         run: bash scripts/test-convert-outputs.sh ${{ github.event_name == 'pull_request' && '--drift=advisory' || '' }}
 
+      - name: Parse indented code fences like lib.sh in the outputs eval
+        run: bash scripts/test-convert-fence-model.sh
+
       - name: Validate agent selection (install.sh --agent / --agents-file)
         run: bash scripts/test-agent-selection.sh
 

scripts/test-convert-fence-model.sh

@@ -0,0 +1,63 @@
+#!/usr/bin/env bash
+#
+# test-convert-fence-model.sh — the converted-outputs eval must parse source
+# code fences the way lib.sh and GitHub do.
+#
+# The eval's Python model read fences only at column 0, while CommonMark (and
+# lib.sh's fence_open_p / fence_closes_p, aligned to it in #855 and #1028)
+# allows up to three spaces of indentation and does not constrain a closer by
+# the opener's indent. With an indented opening fence and a column-0 closer,
+# the model missed the opener and treated the closer as its own opener, so it
+# reported the following lines as a torn fenced block across SOUL.md/AGENTS.md
+# — a hard failure in the Check Tools workflow for an agent file that lint,
+# GitHub and convert_openclaw all handle correctly.
+#
+# Builds a throwaway repo around the real eval and one such fixture, then
+# requires the eval to validate it cleanly. `--update` keeps the check on a
+# fresh repo without a committed manifest.
+#
+# Usage: ./scripts/test-convert-fence-model.sh
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+scratch="$(mktemp -d "${TMPDIR:-/tmp}/agency-fence-model.XXXXXX")"
+trap 'rm -rf "$scratch"' EXIT
+
+mkdir -p "$scratch/repo/scripts" "$scratch/repo/engineering"
+cp "$SCRIPT_DIR/convert.sh" "$SCRIPT_DIR/lib.sh" "$SCRIPT_DIR/build-hermes-plugin.py" \
+   "$SCRIPT_DIR/test-convert-outputs.sh" "$scratch/repo/scripts/"
+
+cat > "$scratch/repo/divisions.json" <<'EOF'
+{
+  "divisions": {
+    "engineering": {"label": "Engineering", "icon": "Code", "color": "#3B82F6"}
+  }
+}
+EOF
+
+cat > "$scratch/repo/engineering/fence-fixture.md" <<'EOF'
+---
+name: Fence Model Fixture
+description: Fixture agent with an indented opening code fence
+color: blue
+---
+## Identity
+  ```text
+  code sample
+```
+## Core Mission
+mission text
+EOF
+
+cd "$scratch/repo"
+if ! bash scripts/test-convert-outputs.sh --update > "$scratch/eval.log" 2>&1; then
+  cat "$scratch/eval.log" >&2
+  echo "FAIL: the outputs eval rejected a GitHub-valid indented opening fence" >&2
+  exit 1
+fi
+if grep -q 'torn across' "$scratch/eval.log"; then
+  echo "FAIL: the outputs eval reported a torn fence for correctly split output" >&2
+  exit 1
+fi
+echo "PASS: the outputs eval parses indented code fences like lib.sh and GitHub"

scripts/test-convert-outputs.sh

@@ -330,7 +330,14 @@ elif not colour_bad:
 # block in half and leaves each file holding a dangling fence, which renders as
 # broken markdown for every user of that integration (#849). So every fenced
 # block in a source must land intact in exactly one of the two files.
-SPLIT_FENCE = re.compile(r"^(`{3,}|~{3,})(.*)$")
+#
+# This model must read fences exactly as lib.sh's fence_open_p / fence_closes_p
+# do (and as GitHub renders): up to three spaces of indentation, same character,
+# a closer at least as long as the opener with nothing but whitespace after it.
+# Reading fences at column 0 only missed an indented opener and paired its
+# column-0 closer with the following lines, reporting a tear the converter had
+# not made.
+OPEN_FENCE = re.compile(r"^( {0,3})(`{3,}|~{3,})")
 
 def body_lines(text):
     """Mirror lib.sh's get_body, including `$(...)`'s trailing-newline strip."""
@@ -346,13 +353,17 @@ def body_lines(text):
     return out
 
 def fence_blocks(lines):
-    """Inclusive (opener, closer) index pairs; closer = last line if unterminated."""
+    """Inclusive (opener, closer) index pairs; closer = last line if unterminated.
+
+    Mirrors lib.sh fence_open_p / fence_closes_p so the eval sees the same
+    blocks the converter does.
+    """
     res, marker, mlen, start = [], "", 0, None
     for i, line in enumerate(lines):
-        m = SPLIT_FENCE.match(line)
+        m = OPEN_FENCE.match(line)
         if not m:
             continue
-        tok, rest = m.group(1), m.group(2)
+        tok, rest = m.group(2), line[m.end():]
         if not marker:
             marker, mlen, start = tok[0], len(tok), i
         elif tok[0] == marker and len(tok) >= mlen and not rest.strip():