harn-vm 0.8.145

//! Cross-channel fidelity regressions for the TypeScript tool-argument
//! parser: a tool value must decode to the SAME bytes whether the model
//! delivered it via a double/single-quoted string, a template literal, or a
//! heredoc. Two silent-corruption bugs are pinned here (both failed on
//! origin/main before the fix in `ts_value_parser.rs`):
//!
//!   1. A non-BMP scalar written as a UTF-16 surrogate pair (`😀`,
//!      the form `JSON.stringify` and most provider APIs emit) errored with
//!      `invalid \u escape` and the ENTIRE tool call was dropped.
//!   2. An unknown escape (`\d`, `\w`, `\b` regex; `\begin` LaTeX) inside a
//!      double/single-quoted string silently dropped the backslash (`"\d+"`
//!      -> `d+`), while the template-literal and heredoc channels preserved it.

use super::{json, parse_bare_calls_in_body, sample_tool_registry};

/// A non-BMP scalar delivered as a `😀` surrogate pair (😀, U+1F600)
/// must decode to that scalar, not drop the call. Source bytes literally
/// contain `😀` (non-raw Rust literal so `\\u` => backslash-u).
#[test]
fn surrogate_pair_escape_decodes_and_keeps_call() {
    let tools = sample_tool_registry();
    let src = "run({ command: \"hi \\uD83D\\uDE00 there\" })";
    let result = parse_bare_calls_in_body(src, Some(&tools));
    assert!(
        result.errors.is_empty(),
        "surrogate pair should not error: {:?}",
        result.errors
    );
    assert_eq!(result.calls.len(), 1, "call must not be dropped");
    assert_eq!(
        result.calls[0]["arguments"]["command"],
        json!("hi \u{1F600} there")
    );
}

/// The `\u{1F600}` brace form already decodes a full astral scalar — keep it
/// working (guards against a regression in the shared `parse_unicode_escape`).
#[test]
fn brace_form_nonbmp_escape_decodes() {
    let tools = sample_tool_registry();
    let src = r#"run({ command: "hi \u{1F600} there" })"#;
    let result = parse_bare_calls_in_body(src, Some(&tools));
    assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
    assert_eq!(
        result.calls[0]["arguments"]["command"],
        json!("hi \u{1F600} there")
    );
}

/// A lone high surrogate with no following low surrogate is genuinely invalid
/// and must still be rejected (no silent half-character, no panic).
#[test]
fn lone_high_surrogate_still_rejected() {
    let tools = sample_tool_registry();
    let src = "run({ command: \"x \\uD83D y\" })";
    let result = parse_bare_calls_in_body(src, Some(&tools));
    assert!(
        !result.errors.is_empty() || result.calls.is_empty(),
        "a lone surrogate must not silently produce a value: {:?}",
        result.calls
    );
}

/// Regex `\d+\w*` written through a double-quoted string must keep its
/// backslashes — same bytes the heredoc and template channels produce.
#[test]
fn unknown_escape_in_quoted_string_keeps_backslash() {
    let tools = sample_tool_registry();
    let result = parse_bare_calls_in_body(
        r#"edit({ action: "create", path: "re.py", content: "p = '\d+\w*'" })"#,
        Some(&tools),
    );
    assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
    assert_eq!(
        result.calls[0]["arguments"]["content"],
        json!("p = '\\d+\\w*'")
    );
}

/// Cross-channel byte-identity: the SAME regex content delivered via a
/// double-quoted string, a template literal, and a heredoc must decode to one
/// identical value. (Before the fix the quoted channel diverged: `d+w*`.)
#[test]
fn regex_content_identical_across_channels() {
    let tools = sample_tool_registry();
    let expected = json!("p = '\\d+\\w*'");

    let quoted = parse_bare_calls_in_body(
        r#"edit({ action: "create", path: "re.py", content: "p = '\d+\w*'" })"#,
        Some(&tools),
    );
    let template = parse_bare_calls_in_body(
        "edit({ action: \"create\", path: \"re.py\", content: `p = '\\d+\\w*'` })",
        Some(&tools),
    );
    let heredoc = parse_bare_calls_in_body(
        "edit({ action: \"create\", path: \"re.py\", content: <<EOF\np = '\\d+\\w*'\nEOF\n })",
        Some(&tools),
    );

    assert_eq!(quoted.calls[0]["arguments"]["content"], expected, "quoted");
    assert_eq!(
        template.calls[0]["arguments"]["content"], expected,
        "template"
    );
    assert_eq!(
        heredoc.calls[0]["arguments"]["content"], expected,
        "heredoc"
    );
}

/// Known escapes inside a quoted string still decode (no over-correction):
/// `\n` `\t` stay control chars, `\\` stays one backslash, `\"` stays a quote.
#[test]
fn known_escapes_in_quoted_string_unchanged() {
    let tools = sample_tool_registry();
    let result = parse_bare_calls_in_body(r#"run({ command: "a\nb\tc\\d\"e" })"#, Some(&tools));
    assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
    assert_eq!(
        result.calls[0]["arguments"]["command"],
        json!("a\nb\tc\\d\"e")
    );
}

/// Parse the `command` value of `run({ command: <literal> })`, returning the
/// decoded string or an `Err` describing why the call was dropped.
fn run_command(src: &str) -> Result<String, String> {
    let tools = sample_tool_registry();
    let result = parse_bare_calls_in_body(src, Some(&tools));
    if !result.errors.is_empty() {
        return Err(format!("{:?}", result.errors));
    }
    let call = result.calls.first().ok_or("call was dropped")?;
    call["arguments"]["command"]
        .as_str()
        .map(str::to_string)
        .ok_or_else(|| "command not a string".to_string())
}

/// A MALFORMED `\u`/`\x` escape (Perl `\x{...}`, a Windows path `\users`, a
/// short `\uAB`, `\uABCG`, a trailing `\x`) must NOT drop the whole tool call.
/// Before this fix `parse_string_literal` returned `Err` on any of these and
/// the model's edit/run silently never landed — the exact #3589 surrogate
/// signature, generalized to every other malformed-known-escape shape. The
/// degraded behavior keeps the `\u`/`\x` literal, byte-identical to the heredoc
/// and template-literal channels.
#[test]
fn malformed_known_escapes_keep_literal_not_dropped() {
    // (source value literal, expected decoded command)
    let cases = [
        (r#""m/\x{1F600}/""#, r"m/\x{1F600}/"), // Perl/PCRE hex-brace regex
        (r#""C:\xtra\data""#, r"C:\xtra\data"), // \x + non-hex, then \d unknown
        (r#""ends with \x""#, r"ends with \x"), // trailing \x
        (r#""bad \uAB stop""#, r"bad \uAB stop"), // short \u (2 hex)
        (r#""bad \uABCG end""#, r"bad \uABCG end"), // \u + non-hex 4th digit
        (r#""path C:\users\me""#, r"path C:\users\me"), // \u + non-hex (Windows)
    ];
    for (src_value, expected) in cases {
        let src = format!("run({{ command: {src_value} }})");
        let got =
            run_command(&src).unwrap_or_else(|err| panic!("call dropped for {src_value}: {err}"));
        assert_eq!(got, expected, "for source {src_value}");
    }
}

/// Cross-channel byte-identity for a malformed escape: Perl `\x{1F600}` must
/// decode to the SAME bytes via quoted string, template literal, and heredoc.
/// Before the fix the quoted channel dropped the call while the other two
/// preserved the content.
#[test]
fn malformed_escape_identical_across_channels() {
    let tools = sample_tool_registry();
    let expected = json!(r"m/\x{1F600}/");

    let quoted = parse_bare_calls_in_body(r#"run({ command: "m/\x{1F600}/" })"#, Some(&tools));
    let template = parse_bare_calls_in_body("run({ command: `m/\\x{1F600}/` })", Some(&tools));
    let heredoc = parse_bare_calls_in_body(
        "run({ command: <<EOF\nm/\\x{1F600}/\nEOF\n })",
        Some(&tools),
    );

    assert_eq!(quoted.calls[0]["arguments"]["command"], expected, "quoted");
    assert_eq!(
        template.calls[0]["arguments"]["command"], expected,
        "template"
    );
    assert_eq!(
        heredoc.calls[0]["arguments"]["command"], expected,
        "heredoc"
    );
}

/// WELL-FORMED `\xHH` and `\uHHHH` escapes still decode (no over-correction):
/// the degraded literal path must only engage when the escape is incomplete.
#[test]
fn well_formed_hex_and_unicode_escapes_still_decode() {
    assert_eq!(
        run_command(r#"run({ command: "\x41\x42" })"#).unwrap(),
        "AB"
    );
    // A literal multibyte scalar in source survives (push_scalar, unaffected).
    assert_eq!(
        run_command(r#"run({ command: "snow ☃ done" })"#).unwrap(),
        "snow \u{2603} done"
    );
    // `\u{...}` brace form still decodes a non-BMP scalar.
    assert_eq!(
        run_command(r#"run({ command: "emoji \u{1F600} done" })"#).unwrap(),
        "emoji \u{1F600} done"
    );
}