Prompt Section Extractor

Parse named prompt sections from explicit boundary markers.

Requirements

• Define extract_sections(text).
• While outside a section, a complete line <<<name>>> starts a section. Do not trim the line before recognizing a marker.
• A name is any nonempty text between the delimiters, preserved verbatim (including spaces). Names are case-sensitive; exact END is reserved as the closing marker. <<<>>> is not a valid opener.
• A section ends with a line exactly <<<END>>>.
• Return {"sections": dict, "errors": list}.
• Store section content without the boundary lines.
• Join interior content lines with "\n". The line that holds <<<END>>> is not part of the body (so a one-line body has no trailing newline). Blank lines before END are preserved as empty lines in the joined body.
• Duplicate section names should produce duplicate:name. On duplicates, keep the first completed section body and append duplicate:<name> while keeping the first stored value unchanged.
• Text outside sections that's not blank should produce outside_text (emit at most once for the whole parse).
• Unterminated sections should produce unterminated:name and must not store a partial body in sections.
• Inside a section, every line except the exact closing marker is content, including opener-looking lines. There is no nesting or escape syntax.
• An unmatched closing marker outside a section is nonblank outside text. A closing marker with extra spaces is ordinary content while a section is active.
• Emit errors in discovery order: a duplicate is discovered at its closing marker; an unterminated section is reported only at end of input. An unfinished duplicate adds unterminated:name, not another duplicate error.

Example

A newline-separated prompt yields one completed section without its boundary markers:

text = "<<<system>>>\nBe brief.\n<<<END>>>\n"
assert extract_sections(text)["sections"] == {"system": "Be brief."}

Completed bodies survive later errors, and only completed duplicates count as duplicates:

text = (
    "outside\n<<<a>>>\none\n<<<END>>>\n"
    "<<<a>>>\ntwo\n<<<END>>>\n"
    "<<<a>>>\npartial"
)
assert extract_sections(text) == {
    "sections": {"a": "one"},
    "errors": ["outside_text", "duplicate:a", "unterminated:a"],
}

The line <<<nested>>> within a would be part of its body, not a second section. Conversely, a literal body line equal to <<<END>>> cannot be represented in this format without adding an escaping rule in a follow-up.

Constraints

• Use explicit state tracking.
• Input is a string with LF or CRLF line endings; returned bodies normalize separators to "\n". Do not trim content-line spaces.
• Preserve internal newlines inside sections.
• Keep errors deterministic.