supercode-interchange 0.4.17

//! Helpers shared by more than one harness codec.

use super::*;

/// D7: repair tool-result/tool-call ORDERING for the Claude Code loader.
///
/// Real Claude Code transcripts are well-formed by `uuid`/`parentUuid` tree
/// structure but are NOT guaranteed to be well-formed in raw file order: async
/// tool-result delivery (or an edited turn) can write a `tool_result` `user`
/// line BEFORE the assistant `tool_use` line that owns it, even though the
/// parent/child tree itself is fine. The active-branch projection restores
/// parent-before-child order, but a result can still trail a later assistant
/// turn instead of sitting next to the assistant that owns it. Both OpenAI-
/// and Anthropic-shaped wire APIs reject either shape on `resume`. Measured
/// on a real corpus: 21/450 sessions (4.7%), 89 inverted pairs.
///
/// This reorders `messages` so every OWNED `Role::Tool` result (its
/// `tool_call_id` matches a `tool_calls` entry on some `Role::Assistant`
/// message anywhere in the list) sits immediately after the `Role::Assistant`
/// message that owns it, while leaving every other message's relative order
/// untouched. Orphan tool results — no matching call anywhere in the list —
/// are left in their ORIGINAL position, untouched; they are never moved. It
/// is a pure reorder: same message count, same multiset of messages, in/out.
///
/// Tool-call ids (`toolu_*`) are globally unique within a Claude Code session
/// — each appears exactly once as a call and once as its result — so a
/// simple id -> owning-assistant map is sufficient; no special-casing is
/// needed for sidechain (`isSidechain`) messages, since `push_claude_*`
/// already pushes those inline with their own distinct ids.
///
/// Results whose matching call is missing entirely (no owner found) are left
/// in place untouched — `ensure_tool_results_paired` (which runs right after
/// this) is responsible for synthesizing a placeholder result for any call
/// that ends up unanswered; this pass never drops or fabricates anything.
pub(super) fn reorder_tool_results_after_calls(messages: &mut Vec<ChatMessage>) {
    // 0. First pass: which tool_call_ids are actually "owned" — emitted by
    //    some assistant message anywhere in the list — and the position of
    //    that owning assistant. Owned as `String` (not borrowed) so this map
    //    can outlive the later `messages.drain(..)`.
    let mut owner_positions: HashMap<String, usize> = HashMap::new();
    for (index, m) in messages.iter().enumerate() {
        if m.role == Role::Assistant {
            for c in m.tool_calls() {
                if !c.id.is_empty() {
                    owner_positions.entry(c.id.clone()).or_insert(index);
                }
            }
        }
    }

    // Fast, cheap detection of "nothing to do": every owned result must be
    // in the contiguous tool-result block immediately following its owning
    // assistant. Checking only result-before-owner inversions is insufficient
    // after Claude's active-branch projection: that projection can put the
    // owner first while leaving its result behind a later assistant turn.
    // Mere orphans never set this flag. A canonical session returns with
    // `messages` byte-for-byte unchanged, mirroring
    // `ensure_tool_results_paired`'s own no-op guard.
    let mut contiguous_owner = None;
    let needs_reorder =
        messages
            .iter()
            .enumerate()
            .any(|(message_index, message)| match message.role {
                Role::Assistant => {
                    contiguous_owner = Some(message_index);
                    false
                }
                Role::Tool => match message
                    .tool_call_id
                    .as_deref()
                    .and_then(|id| owner_positions.get(id))
                    .copied()
                {
                    Some(owner) => Some(owner) != contiguous_owner,
                    None => {
                        // An orphan or unlinked tool message interrupts the
                        // owner's contiguous result block but never moves by
                        // itself.
                        contiguous_owner = None;
                        false
                    }
                },
                _ => {
                    contiguous_owner = None;
                    false
                }
            });
    if !needs_reorder {
        return;
    }

    // 1. Second pass: route messages into the "spine" (everything that stays
    //    at its own position — non-tool messages AND orphan tool results)
    //    versus owned tool results (pulled out, to be reattached right after
    //    their owner). Record, for each spine index that's an assistant, the
    //    set of tool_call_ids it owns.
    let mut spine: Vec<ChatMessage> = Vec::with_capacity(messages.len());
    let mut call_owner: HashMap<String, usize> = HashMap::new();
    // Buffer of (original_position, message) for every OWNED tool result,
    // built alongside the spine; a result can reference a call emitted later
    // in file order, so owner spine-index is resolved in a later step once
    // `call_owner` is complete.
    let mut owned_results: Vec<(usize, ChatMessage)> = Vec::new();

    let drained: Vec<ChatMessage> = std::mem::take(messages);
    for (orig_pos, msg) in drained.into_iter().enumerate() {
        if msg.role == Role::Tool {
            let is_owned = msg
                .tool_call_id
                .as_deref()
                .map(|id| !id.is_empty() && owner_positions.contains_key(id))
                .unwrap_or(false);
            if is_owned {
                owned_results.push((orig_pos, msg));
                continue;
            }
            // Orphan: no matching call anywhere. Treat exactly like a
            // non-tool message for placement — it joins the spine at its
            // current position and is never moved.
            spine.push(msg);
            continue;
        }
        if msg.role == Role::Assistant {
            let spine_idx = spine.len();
            for c in msg.tool_calls() {
                if !c.id.is_empty() {
                    call_owner.entry(c.id.clone()).or_insert(spine_idx);
                }
            }
        }
        spine.push(msg);
    }

    // 2. Resolve each owned result's owner spine-index now that `call_owner`
    //    is complete, then bucket results by owner spine-index. Every result
    //    here was routed as "owned" because its id was found in `owned_ids`,
    //    which was built from the exact same `tool_calls()` scan that
    //    populates `call_owner` below, so the lookup is guaranteed to hit.
    let mut buckets: HashMap<usize, Vec<(usize, ChatMessage)>> = HashMap::new();
    for (orig_pos, msg) in owned_results.into_iter() {
        let id = msg
            .tool_call_id
            .as_deref()
            .filter(|id| !id.is_empty())
            .expect("routed as owned, so tool_call_id must be a non-empty owned id");
        let idx = *call_owner
            .get(id)
            .expect("owned id must have an owning assistant in call_owner");
        buckets.entry(idx).or_default().push((orig_pos, msg));
    }
    // Keep each bucket's results in their original relative file order.
    for v in buckets.values_mut() {
        v.sort_by_key(|(pos, _)| *pos);
    }

    // 3. Rebuild: emit each spine message (which now includes orphans at
    //    their original position, untouched) in order; immediately after
    //    emitting an assistant message that owns one or more tool results,
    //    emit its owned results, in original relative order.
    let mut out: Vec<ChatMessage> = Vec::with_capacity(spine.len() + buckets.len());
    for (idx, msg) in spine.into_iter().enumerate() {
        out.push(msg);
        if let Some(results) = buckets.remove(&idx) {
            for (_, r) in results {
                out.push(r);
            }
        }
    }
    *messages = out;
}

/// Guarantee every assistant `tool_calls` entry is answered by a following tool
/// result. Interrupted/aborted turns leave a tool call with no result, which
/// many chat-completions endpoints reject when the conversation is replayed.
/// We insert a synthetic placeholder result immediately after the assistant
/// turn so the transcript stays valid for continuation. (Orphan results — a
/// tool message with no preceding call — do not occur in practice and are left
/// untouched.)
pub(super) fn ensure_tool_results_paired(messages: &mut Vec<ChatMessage>) {
    let answered: HashSet<String> = messages
        .iter()
        .filter(|m| m.role == Role::Tool)
        .filter_map(|m| m.tool_call_id.clone())
        .collect();

    // Nothing missing? Leave the vector byte-for-byte unchanged.
    let any_missing = messages.iter().any(|m| {
        m.tool_calls()
            .iter()
            .any(|c| !c.id.is_empty() && !answered.contains(&c.id))
    });
    if !any_missing {
        return;
    }

    let mut out: Vec<ChatMessage> = Vec::with_capacity(messages.len() + 4);
    for msg in messages.drain(..) {
        let synth: Vec<ChatMessage> = msg
            .tool_calls()
            .iter()
            .filter(|c| !c.id.is_empty() && !answered.contains(&c.id))
            .map(|c| {
                let mut m = ChatMessage::tool_result(
                    c.id.clone(),
                    c.function.name.clone(),
                    "[no tool result recorded — turn interrupted]".to_string(),
                );
                // TR-10: an interrupted call never executed to completion —
                // never a candidate for `ReductionKind::ToolInputElided`
                // (the "still-pending calls are never input-elided"
                // boundary).
                crate::mark_tool_error(&mut m);
                m
            })
            .collect();
        out.push(msg);
        out.extend(synth);
    }
    *messages = out;
}

/// Whether `msg` is excluded from every replay/export path — the frozen
/// cross-format discipline (`docs/interop/opencode-pi-spec.md` §2.1/§2.2/§6):
/// a message marked `compacted_out` (pre-compaction history a source harness
/// itself no longer replays) or `pi_exclude_from_context` (a pi `!!` bash
/// escape) must never reach `to_jsonl`/`to_jsonl_spliced` for ANY target
/// format, not just the one that produced the marker — so a translated
/// compacted session replays the same sliced context the source harness
/// would, instead of double-including history plus its own summary.
pub(super) fn is_replay_excluded(msg: &ChatMessage) -> bool {
    msg.metadata.get("compacted_out").map(String::as_str) == Some("true")
        || msg
            .metadata
            .get("pi_exclude_from_context")
            .map(String::as_str)
            == Some("true")
}

/// STRICT-VERBATIM line split for `raw` capture (IX-1) — the counterpart to
/// [`non_empty_lines`]'s blank-skipping/trimming PARSE view.
///
/// Splits only on `\n` (never treats `\r\n` specially), so a CRLF source's
/// `\r` survives as part of the returned line's own content; blank lines and
/// trailing-whitespace-only lines are kept verbatim rather than dropped or
/// trimmed. This is what makes `Session.raw` — populated from this at every
/// raw-capture site — byte-for-byte faithful to the source for ANY input, not
/// just well-formed LF JSONL with no blank lines.
///
/// Returns `(lines, ends_with_newline)`. `lines.join("\n")` alone cannot
/// distinguish a source that ended with a trailing newline from one that
/// didn't (both split into the same line list), so `ends_with_newline`
/// records that fact out-of-band — [`join_lines_verbatim`] is the exact
/// inverse, given both. `text == ""` yields `(vec![], false)`: an empty
/// source has zero lines, not one blank line.
pub(super) fn split_lines_verbatim(text: &str) -> (Vec<&str>, bool) {
    if text.is_empty() {
        return (Vec::new(), false);
    }
    let ends_with_newline = text.ends_with('\n');
    let body = if ends_with_newline {
        &text[..text.len() - 1]
    } else {
        text
    };
    (body.split('\n').collect(), ends_with_newline)
}

/// The exact inverse of [`split_lines_verbatim`]: reconstruct the original
/// source bytes from its verbatim lines plus the trailing-newline flag.
pub(super) fn join_lines_verbatim(lines: &[String], ends_with_newline: bool) -> String {
    let mut out = lines.join("\n");
    if ends_with_newline {
        out.push('\n');
    }
    out
}

/// PARSE-iteration view (blank/whitespace-only lines skipped, each line
/// trimmed, `\r` line endings normalized away by [`str::lines`]) — used
/// everywhere a loader walks lines looking for JSON *records*, where a blank
/// line is simply not a record and must not become a spurious parse
/// failure/empty entry. Deliberately NOT used for `raw` capture any more
/// (IX-1) — see [`split_lines_verbatim`] for that.
pub(super) fn non_empty_lines(text: &str) -> impl Iterator<Item = &str> {
    text.lines().map(str::trim).filter(|l| !l.is_empty())
}

pub(super) fn inject_first_jsonl_top_level(out: &mut String, key: &str, extension: Value) {
    let Some(line_end) = out.find('\n') else {
        return;
    };
    let Ok(mut record) = serde_json::from_str::<Value>(&out[..line_end]) else {
        return;
    };
    let Some(object) = record.as_object_mut() else {
        return;
    };
    object.insert(key.to_string(), extension);
    out.replace_range(..line_end, &record.to_string());
}

/// Remove the last conversational turn from `messages`: everything from the
/// last `user` message to the end (the user prompt plus the assistant's
/// response and any tool calls/results it triggered).
pub(super) fn remove_last_turn(messages: &mut Vec<ChatMessage>) {
    if let Some(idx) = messages.iter().rposition(|m| m.role == Role::User) {
        messages.truncate(idx);
    } else {
        messages.clear();
    }
    // IX-6 fix: the new tail exposed by `truncate` may still carry
    // `__codex_open_turn` from when it was marked (it was NOT the last
    // message at that time — items after it, now removed by the rollback,
    // intervened). A bare `function_call` arriving after the rollback is a
    // genuinely NEW turn and must get its own message, not merge into this
    // stale marked tail — close it out here so `push_codex_item`'s
    // adjacency check (`out.last()` + marker) can't be fooled by the
    // truncation re-exposing it.
    if let Some(last) = messages.last_mut() {
        last.metadata.remove("__codex_open_turn");
    }
}

fn truncate_messages(messages: &mut Vec<ChatMessage>, message_limit: usize) {
    truncate_messages_with_anchor(messages, message_limit, Vec::new());
}

pub(super) fn truncate_messages_with_anchor(
    messages: &mut Vec<ChatMessage>,
    message_limit: usize,
    preceding_users: Vec<ChatMessage>,
) {
    let limit = message_limit.max(1);
    // Anchor the numeric tail at its own human boundary. Choosing the newest
    // users anywhere in the window is not enough: after several short prompts
    // followed by a tool-heavy turn, those users can all sit near the end while
    // the tail begins in older orphaned tool activity. The UI then correctly
    // hides that activity, so increasing `tail_messages` appears to load
    // nothing and the next prompt can collapse the visible history.
    //
    // Keep up to two users from before the raw tail boundary. The nearest one
    // makes the retained activity a real turn; the preceding one preserves
    // overlap when one turn alone exceeds the whole numeric window.
    let raw_tail_start = messages.len().saturating_sub(limit);
    let anchor_limit = if messages.len() < limit {
        limit.saturating_sub(messages.len()).min(2)
    } else {
        limit.min(2)
    };
    let mut anchor_candidates = preceding_users;
    anchor_candidates.extend(
        messages[..raw_tail_start]
            .iter()
            .filter(|message| message.role == Role::User)
            .cloned(),
    );
    let mut anchors = anchor_candidates
        .into_iter()
        .rev()
        .take(anchor_limit)
        .collect::<Vec<_>>();
    anchors.reverse();

    if messages.len() <= limit && anchors.is_empty() {
        return;
    }
    let tail_count = limit.saturating_sub(anchors.len()).min(messages.len());
    let tail_start = messages.len() - tail_count;
    let mut selected = Vec::with_capacity(anchors.len() + tail_count);
    selected.append(&mut anchors);
    selected.extend(messages[tail_start..].iter().cloned());
    debug_assert_eq!(selected.len(), limit.min(messages.len() + anchor_limit));
    *messages = selected;
}

pub(super) fn truncate_session_messages(session: &mut Session, message_limit: usize) {
    truncate_messages(&mut session.messages, message_limit);
}

/// Depth of a node in the parent forest (root = 0), bounded against cycles.
pub(super) fn depth_of(mut i: usize, parent_of: &[Option<usize>]) -> usize {
    let mut d = 0;
    let mut guard = 0;
    while let Some(p) = parent_of[i] {
        if p == i || guard > parent_of.len() {
            break;
        }
        i = p;
        d += 1;
        guard += 1;
    }
    d
}

/// Namespaced line-level extension carrying the one tool-result outcome state
/// Claude cannot represent natively. Keeping this narrower than the full Grok
/// portability envelope avoids changing unrelated target-message projection.
pub(super) const SUPERCODE_TOOL_OUTCOME_KEY: &str = "_supercode_tool_outcome";

pub(super) fn restore_tool_outcome_extension(value: &Value, message: &mut ChatMessage) {
    if !crate::is_tool_error(message)
        && value
            .get(SUPERCODE_TOOL_OUTCOME_KEY)
            .and_then(Value::as_str)
            == Some("unknown")
    {
        crate::mark_tool_outcome_unknown(message);
    }
}

#[doc(hidden)]
pub fn percent_decode_path(encoded: &str) -> Option<String> {
    fn hex(byte: u8) -> Option<u8> {
        match byte {
            b'0'..=b'9' => Some(byte - b'0'),
            b'a'..=b'f' => Some(byte - b'a' + 10),
            b'A'..=b'F' => Some(byte - b'A' + 10),
            _ => None,
        }
    }

    let bytes = encoded.as_bytes();
    let mut decoded = Vec::with_capacity(bytes.len());
    let mut index = 0usize;
    while index < bytes.len() {
        if bytes[index] == b'%' {
            let high = *bytes.get(index + 1)?;
            let low = *bytes.get(index + 2)?;
            decoded.push(hex(high)? * 16 + hex(low)?);
            index += 3;
        } else {
            decoded.push(bytes[index]);
            index += 1;
        }
    }
    String::from_utf8(decoded).ok()
}

// ---- shared helpers -------------------------------------------------------

pub(super) fn push_text(buf: &mut String, v: Option<&Value>) {
    if let Some(Value::String(s)) = v {
        if !buf.is_empty() {
            buf.push('\n');
        }
        buf.push_str(s);
    }
}

/// Extract a Claude `tool_result` block's content, preserving non-text items
/// instead of silently dropping them:
///
/// - text blocks are concatenated;
/// - PARITY-11 (nested images, real-session-confirmed — the everyday "Read a
///   PNG / screenshot tool output" shape): `image` blocks are captured into
///   the returned `content_parts`-shaped `Vec<Value>` via
///   [`claude_image_block_to_part`] — the SAME base64/url conversion the
///   top-level `image` content-block path (`push_claude_user`) already uses
///   — instead of being flattened to the bare `[image]` marker text that used
///   to make the data unrecoverable from every writer. An unconvertible
///   source (D5 discipline — a Files-API `{"type":"file",...}` reference,
///   etc.) folds [`UNCONVERTIBLE_IMAGE_MARKER`] into the text instead of
///   vanishing, exactly like the top-level path;
/// - `tool_reference` blocks become `[tool_reference: <name>]`;
///
/// and if the block yields no text/images at all, fall back to the record's
/// `toolUseResult` field (string used directly, structured value serialized),
/// which is where Claude Code stores the actual result in many cases.
///
/// Returns `(text, images)`; callers that only need the old text-only
/// behavior can ignore the second element — every caller MUST fold non-empty
/// `images` into the resulting `ChatMessage.content_parts` themselves (this
/// function has no `ChatMessage` to attach to).
pub(super) fn extract_tool_result_content(
    content: Option<&Value>,
    tool_use_result: Option<&Value>,
) -> (String, Vec<Value>) {
    let mut parts: Vec<String> = Vec::new();
    let mut images: Vec<Value> = Vec::new();
    match content {
        Some(Value::String(s)) => {
            if !s.is_empty() {
                parts.push(s.clone());
            }
        }
        Some(Value::Array(items)) => {
            for item in items {
                match item.get("type").and_then(Value::as_str) {
                    Some("text") => {
                        if let Some(t) = item.get("text").and_then(Value::as_str) {
                            parts.push(t.to_string());
                        }
                    }
                    Some("image") => match claude_image_block_to_part(item) {
                        Some(part) => images.push(part),
                        None => parts.push(UNCONVERTIBLE_IMAGE_MARKER.to_string()),
                    },
                    Some("tool_reference") => {
                        let name = item.get("tool_name").and_then(Value::as_str).unwrap_or("?");
                        parts.push(format!("[tool_reference: {name}]"));
                    }
                    _ => {
                        if let Some(s) = item.as_str() {
                            parts.push(s.to_string());
                        }
                    }
                }
            }
        }
        Some(other) => parts.push(other.to_string()),
        None => {}
    }

    let joined = parts.join("\n");
    if !joined.trim().is_empty() || !images.is_empty() {
        return (joined, images);
    }
    // Empty tool_result content — recover from toolUseResult.
    match tool_use_result {
        Some(Value::String(s)) => (s.clone(), images),
        Some(v) => (v.to_string(), images),
        None => (joined, images),
    }
}

/// Pull readable text out of a content value that may be a plain string or an
/// array of `{ "text": "..." }`-bearing blocks (any block type).
pub(super) fn extract_text_content(v: Option<&Value>) -> String {
    match v {
        Some(Value::String(s)) => s.clone(),
        Some(Value::Array(items)) => {
            let mut parts = Vec::new();
            for item in items {
                if let Some(t) = item.get("text").and_then(Value::as_str) {
                    parts.push(t.to_string());
                } else if let Some(s) = item.as_str() {
                    parts.push(s.to_string());
                }
            }
            parts.join("\n")
        }
        Some(other) => other.to_string(),
        None => String::new(),
    }
}

/// Extract Codex `input_image` content blocks from a `message` response_item's
/// `content` value into `content_parts` `image_url` entries — the inverse of
/// [`codex_message_content_blocks`]'s `input_image` emission (IX-5). Only a
/// block whose `image_url` is a non-empty string is recognized; anything else
/// (a missing/malformed `image_url`, a `file_id`-only reference) is left as
/// raw-only residue rather than synthesizing a corrupt/empty part — mirroring
/// the pi/opencode/Claude loaders' image-shape discipline.
pub(super) fn codex_extract_images(content: Option<&Value>) -> Vec<Value> {
    let Some(Value::Array(items)) = content else {
        return Vec::new();
    };
    items
        .iter()
        .filter(|item| item.get("type").and_then(Value::as_str) == Some("input_image"))
        .filter_map(|item| {
            let url = item.get("image_url").and_then(Value::as_str)?;
            if url.is_empty() {
                return None;
            }
            Some(serde_json::json!({"type": "image_url", "image_url": {"url": url}}))
        })
        .collect()
}

pub(super) fn value_to_arg_string(v: &Value) -> String {
    match v {
        Value::String(s) => s.clone(),
        other => other.to_string(),
    }
}

pub(super) fn push_gemini_user_parts(
    messages: &mut Vec<ChatMessage>,
    content_parts: Vec<Value>,
    timestamp: Option<&str>,
    source: &Value,
) {
    if content_parts.is_empty() {
        return;
    }
    let mut message = ChatMessage {
        role: Role::User,
        content: None,
        content_parts: Some(content_parts),
        tool_calls: None,
        tool_call_id: None,
        name: None,
        metadata: Default::default(),
    };
    if let Some(timestamp) = timestamp {
        message
            .metadata
            .insert("timestamp".into(), timestamp.into());
    }
    restore_gemini_message_extension(source, &mut message);
    messages.push(message);
}

pub(super) fn function_call(id: &str, name: &str, arguments: String) -> ToolCall {
    ToolCall {
        id: id.to_string(),
        kind: "function".to_string(),
        function: FunctionCall {
            name: name.to_string(),
            arguments,
        },
    }
}

pub(super) fn tool_message(tool_call_id: &str, content: String) -> ChatMessage {
    ChatMessage {
        role: Role::Tool,
        content: Some(content),
        content_parts: None,
        tool_calls: None,
        tool_call_id: Some(tool_call_id.to_string()),
        name: None,
        metadata: Default::default(),
    }
}

/// Emit a single assistant message combining accumulated text and tool calls.
/// A turn with neither (e.g. thinking-only) produces nothing.
pub(super) fn push_assistant(out: &mut Vec<ChatMessage>, text: String, calls: Vec<ToolCall>) {
    let has_text = !text.trim().is_empty();
    if !has_text && calls.is_empty() {
        return;
    }
    out.push(ChatMessage {
        role: Role::Assistant,
        content: has_text.then_some(text),
        content_parts: None,
        tool_calls: (!calls.is_empty()).then_some(calls),
        tool_call_id: None,
        name: None,
        metadata: Default::default(),
    });
}

// ---- writers --------------------------------------------------------------

/// A placeholder timestamp for SYNTHESIZED records only — WAVE-2 item 1's
/// fallback (`docs/interop` build brief): every writer now emits a message's
/// REAL source timestamp (`metadata["timestamp"]`, the canonical ISO-8601
/// field every loader populates) when one is present. `SYNTH_TS` fires only
/// for a message with no source timestamp at all — a turn synthesized/
/// appended after import (the live agent loop, a splice's appended tail,
/// ...), which was never loaded from a real per-message timestamp to begin
/// with. Both tools tolerate identical timestamps; callers that need real
/// ones for a synthesized turn can post-process.
pub(super) const SYNTH_TS: &str = "2026-01-01T00:00:00.000Z";

/// The unix-millisecond twin of [`SYNTH_TS`], for OpenCode's numeric
/// `time.created`/`time.updated` fields.
pub(super) const SYNTH_TS_MS: i64 = 1_767_225_600_000;

/// `msg`'s real source timestamp (`metadata["timestamp"]`, canonical
/// ISO-8601 — WAVE-2 item 1) when present AND well-formed, else the
/// [`SYNTH_TS`] fallback. Validated via [`crate::sidecar::rfc3339_to_ms`] (a
/// parse, not just a presence check) so an absent, empty, or malformed
/// source value all degrade to the same documented fallback rather than
/// propagating garbage verbatim. Used by every writer that emits an
/// ISO-8601 timestamp field
/// (Claude Code, Codex, pi's entry-level `timestamp`).
pub(super) fn msg_timestamp_or_synth(msg: &ChatMessage) -> &str {
    match msg.metadata.get("timestamp") {
        Some(ts) if crate::sidecar::rfc3339_to_ms(ts).is_some() => ts.as_str(),
        _ => SYNTH_TS,
    }
}

/// A format-valid (v4-shaped) but deterministic uuid, derived from a counter.
pub(super) fn synth_uuid(n: usize) -> String {
    format!("00000000-0000-4000-8000-{n:012x}")
}

pub(super) fn push_jsonl(out: &mut String, value: &Value) {
    out.push_str(&value.to_string());
    out.push('\n');
}

/// Replay a raw (verbatim) JSONL `line` into `out`, rewriting `key` to
/// `new_id` when the line parses as a JSON object carrying that key — used
/// by A12's Claude Code splice, where the session id lives at the top level
/// of (almost) every record under `key = "sessionId"`. A line that fails to
/// parse, or parses but lacks `key`, is copied through byte-for-byte
/// (nothing to patch, so nothing is reserialized).
pub(super) fn push_spliced_line(out: &mut String, line: &str, new_id: Option<&str>, key: &str) {
    if let Some(new_id) = new_id {
        if let Ok(mut v) = serde_json::from_str::<Value>(line) {
            if v.get(key).is_some() {
                v[key] = Value::String(new_id.to_string());
                out.push_str(&v.to_string());
                out.push('\n');
                return;
            }
        }
    }
    out.push_str(line);
    out.push('\n');
}

impl Session {
    pub(super) fn cwd_string(&self) -> String {
        self.meta
            .cwd
            .as_ref()
            .map(|p| p.to_string_lossy().into_owned())
            .unwrap_or_else(|| ".".to_string())
    }

    /// `(raw_prefix_len, message_prefix_len)` for A12 splicing: how many
    /// leading `raw` lines / `messages` came from the imported log, as
    /// opposed to being appended after import.
    ///
    /// `imported_message_count` (see its doc comment) pins the message-side
    /// boundary directly. The raw-side boundary isn't separately tracked —
    /// [`Self::from_native_str`]'s turn-appending loop pushes exactly one
    /// `raw` line per appended message, so the two lists grow by the same
    /// `appended_count` from the same starting point, and
    /// `raw.len() - appended_count` recovers it without a second counter.
    pub(super) fn spliced_prefix_lens(&self) -> (usize, usize) {
        let message_prefix_len = self
            .imported_message_count
            .unwrap_or(self.messages.len())
            .min(self.messages.len());
        let appended_count = self.messages.len() - message_prefix_len;
        let raw_prefix_len = self.raw.len().saturating_sub(appended_count);
        (raw_prefix_len, message_prefix_len)
    }
}