kcode-audio-transcript-plan 0.1.0

#![forbid(unsafe_code)]

use kcode_audio_ingress::{ConfirmationState, RecordingState, RecordingStatus};
use serde_json::{Value, json};
use uuid::Uuid;

const ESTIMATED_CHARACTERS_PER_TOKEN: u64 = 4;
const INGRESS_CONTEXT_DIVISOR: u64 = 4;

/// A bounded category and library-owned diagnostic for planning failures.
#[derive(Clone, Debug, Eq, PartialEq)]
pub enum Error {
    InvalidInput(&'static str),
    Conflict(&'static str),
    Internal(&'static str),
}

impl std::fmt::Display for Error {
    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        let message = match self {
            Self::InvalidInput(message) | Self::Conflict(message) | Self::Internal(message) => {
                message
            }
        };
        formatter.write_str(message)
    }
}

impl std::error::Error for Error {}

/// The complete deterministic piece plan for one recording status.
#[derive(Debug, Eq, PartialEq)]
pub struct TranscriptPlan {
    pieces: Vec<TranscriptPiece>,
}

impl TranscriptPlan {
    /// Plans a completed transcript using one quarter of the effective context.
    ///
    /// A recording that is not complete produces an empty plan.
    pub fn new(recording: &RecordingStatus, effective_context_tokens: u64) -> Result<Self, Error> {
        let maximum_piece_characters = maximum_piece_characters(effective_context_tokens)?;
        let RecordingState::Complete { transcript } = &recording.state else {
            return Ok(Self { pieces: Vec::new() });
        };
        let texts = split_transcript(transcript, maximum_piece_characters)?;
        let count = u32::try_from(texts.len())
            .map_err(|_| Error::Internal("audio transcript contains too many pieces"))?;
        let pieces = texts
            .into_iter()
            .enumerate()
            .map(|(index, text)| {
                let index = u32::try_from(index)
                    .map_err(|_| Error::Internal("audio transcript piece index exceeds u32"))?;
                Ok(TranscriptPiece {
                    id: audio_piece_id(recording.id, index),
                    index,
                    count,
                    estimated_tokens: estimate_tokens(&text),
                    text,
                })
            })
            .collect::<Result<_, Error>>()?;
        Ok(Self { pieces })
    }

    /// Returns the ordered, complete piece plan.
    pub fn pieces(&self) -> &[TranscriptPiece] {
        &self.pieces
    }
}

/// One deterministic transcript-only piece.
///
/// Fields are readable by consumers, while construction remains owned by
/// [`TranscriptPlan`].
#[derive(Debug, Eq, PartialEq)]
#[non_exhaustive]
pub struct TranscriptPiece {
    pub id: String,
    pub index: u32,
    pub count: u32,
    pub text: String,
    pub estimated_tokens: u64,
}

impl TranscriptPiece {
    /// Reproduces the existing audio ingress metadata for this piece.
    pub fn metadata(&self, recording: &RecordingStatus) -> Value {
        json!({
            "kind":"audio-transcript",
            "recordingId":recording.id.to_string(),
            "sha256":recording.sha256,
            "originalFilename":recording.original_filename,
            "extension":file_name_extension(&recording.original_filename),
            "mimeType":"audio/wav",
            "sizeBytes":recording.size_bytes,
            "sourceCreatedAt":recording.recorded_at.to_rfc3339(),
            "pieceIndex":self.index,
            "pieceCount":self.count,
            "speakerConfirmationState":recording.correction_packet.as_ref().map(|packet| match packet.confirmation_state {
                ConfirmationState::Unconfirmed => "unconfirmed",
                ConfirmationState::AutomaticallyTrained => "automatically_trained",
                ConfirmationState::Confirmed => "confirmed",
            }),
        })
    }

    /// Reproduces the existing History input for this piece.
    pub fn formatted_text(&self, recording: &RecordingStatus) -> Result<String, Error> {
        let speaker_mapping = confirmed_speaker_mapping(recording)?;
        Ok(format!(
            "Vnote final transcript piece\n\nRecording began: {}\nRecording SHA-256: {}\nOriginal filename: {}\nExtension: {}\nMIME type: audio/wav\nSize: {} bytes\nTranscript piece: {} of {}{}\n\n{}",
            recording.recorded_at.to_rfc3339(),
            recording.sha256,
            recording.original_filename,
            file_name_extension(&recording.original_filename),
            recording.size_bytes,
            self.index + 1,
            self.count,
            speaker_mapping,
            self.text,
        ))
    }
}

/// Parses only canonical `audio:<uuid>:<u32>` piece identities.
pub fn parse_audio_piece_id(value: &str) -> Option<(Uuid, u32)> {
    let mut components = value.split(':');
    if components.next()? != "audio" {
        return None;
    }
    let recording_id = Uuid::parse_str(components.next()?).ok()?;
    let piece_index = components.next()?.parse::<u32>().ok()?;
    if components.next().is_some() || audio_piece_id(recording_id, piece_index) != value {
        return None;
    }
    Some((recording_id, piece_index))
}

fn audio_piece_id(recording_id: Uuid, piece_index: u32) -> String {
    format!("audio:{recording_id}:{piece_index}")
}

fn maximum_piece_characters(effective_context_tokens: u64) -> Result<usize, Error> {
    let piece_tokens = effective_context_tokens / INGRESS_CONTEXT_DIVISOR;
    if piece_tokens == 0 {
        return Err(Error::InvalidInput(
            "effective ingress context must contain at least four tokens",
        ));
    }
    let characters = piece_tokens
        .checked_mul(ESTIMATED_CHARACTERS_PER_TOKEN)
        .ok_or(Error::InvalidInput(
            "effective ingress context is too large",
        ))?;
    usize::try_from(characters)
        .map_err(|_| Error::InvalidInput("effective ingress context exceeds platform limits"))
}

fn split_transcript(
    transcript: &str,
    maximum_piece_characters: usize,
) -> Result<Vec<String>, Error> {
    let mut remaining = transcript.trim();
    if remaining.is_empty() {
        return Err(Error::Internal("completed audio transcript is empty"));
    }

    let mut pieces = Vec::new();
    while remaining.chars().count() > maximum_piece_characters {
        let cutoff = remaining
            .char_indices()
            .nth(maximum_piece_characters)
            .map(|(index, _)| index)
            .unwrap_or(remaining.len());
        let prefix = &remaining[..cutoff];
        let minimum = prefix
            .char_indices()
            .nth(maximum_piece_characters / 2)
            .map(|(index, _)| index)
            .unwrap_or(0);
        let boundary = prefix
            .rfind("\n\n")
            .filter(|index| *index >= minimum)
            .or_else(|| prefix.rfind('\n').filter(|index| *index >= minimum))
            .unwrap_or(cutoff);
        let piece = remaining[..boundary].trim();
        if piece.is_empty() {
            return Err(Error::Internal("could not split audio transcript"));
        }
        pieces.push(piece.to_owned());
        remaining = remaining[boundary..].trim();
    }
    if !remaining.is_empty() {
        pieces.push(remaining.to_owned());
    }
    Ok(pieces)
}

fn estimate_tokens(value: &str) -> u64 {
    (value.chars().count() as u64).div_ceil(ESTIMATED_CHARACTERS_PER_TOKEN)
}

fn confirmed_speaker_mapping(recording: &RecordingStatus) -> Result<String, Error> {
    let Some(packet) = recording.correction_packet.as_ref() else {
        return Ok(String::new());
    };
    if packet.confirmation_state != ConfirmationState::Confirmed {
        return Err(Error::Conflict(
            "audio speaker labels require exact human confirmation before ingress",
        ));
    }

    let mut lines = Vec::new();
    for chunk in &packet.chunks {
        for observation in &chunk.observations {
            let confirmed = observation
                .confirmed_full_name
                .as_deref()
                .ok_or(Error::Internal(
                    "confirmed audio packet omitted an observation label",
                ))?;
            let local_label = serde_json::to_string(&observation.local_label)
                .map_err(|_| Error::Internal("could not render confirmed audio speaker mapping"))?;
            let confirmed = serde_json::to_string(confirmed)
                .map_err(|_| Error::Internal("could not render confirmed audio speaker mapping"))?;
            let candidate = observation
                .identified_full_name
                .as_deref()
                .map(serde_json::to_string)
                .transpose()
                .map_err(|_| Error::Internal("could not render confirmed audio speaker mapping"))?
                .unwrap_or_else(|| "null".into());
            lines.push(format!(
                "- chunk {}/{}, source {:.3}-{:.3}s: localLabel={}, confirmedFullName={}, classifierCandidate={}",
                chunk.chunk_index + 1,
                chunk.chunk_count,
                chunk.audio_start_ms as f64 / 1_000.0,
                chunk.audio_end_ms as f64 / 1_000.0,
                local_label,
                confirmed,
                candidate,
            ));
        }
    }
    if lines.is_empty() {
        return Err(Error::Internal(
            "confirmed audio packet contains no speaker observations",
        ));
    }
    Ok(format!(
        "\n\nHuman-confirmed speaker-label data (authoritative; do not infer alternatives):\n{}",
        lines.join("\n")
    ))
}

fn file_name_extension(file_name: &str) -> String {
    file_name
        .rsplit_once('.')
        .and_then(|(stem, extension)| {
            (!stem.is_empty() && !extension.is_empty()).then_some(extension)
        })
        .map(|extension| format!(".{extension}"))
        .unwrap_or_else(|| "(none)".into())
}

#[cfg(test)]
mod tests {
    use super::*;
    use kcode_audio_ingress::{
        CorrectionChunk, CorrectionObservation, CorrectionPacket, ObservationKey, ParsedChunk,
    };

    fn completed_recording(transcript: impl Into<String>) -> RecordingStatus {
        let recorded_at = "2026-08-02T03:04:05Z".parse().unwrap();
        RecordingStatus {
            id: Uuid::parse_str("abcdef01-2345-6789-abcd-ef0123456789").unwrap(),
            user_id: "user".into(),
            sha256: "a".repeat(64),
            original_filename: "meeting.final.WAV".into(),
            size_bytes: 42,
            recorded_at,
            received_at: recorded_at,
            transcription_model: "transcription-model".into(),
            reconciliation_model: "reconciliation-model".into(),
            reconciliation_reasoning: "xhigh".into(),
            state: RecordingState::Complete {
                transcript: transcript.into(),
            },
            correction_packet: None,
        }
    }

    fn with_speaker_packet(
        mut recording: RecordingStatus,
        state: ConfirmationState,
        confirmed_name: Option<&str>,
    ) -> RecordingStatus {
        recording.correction_packet = Some(CorrectionPacket {
            recording_id: recording.id,
            user_id: recording.user_id.clone(),
            sha256: recording.sha256.clone(),
            original_filename: recording.original_filename.clone(),
            size_bytes: recording.size_bytes,
            recorded_at: recording.recorded_at,
            clean: true,
            chunk_count: 1,
            chunks: vec![CorrectionChunk {
                chunk_index: 0,
                chunk_count: 1,
                audio_start_ms: 0,
                audio_end_ms: 1_500,
                raw_gemini_response: "raw".into(),
                parsed: ParsedChunk {
                    utterances: Vec::new(),
                    notes: Vec::new(),
                    clip_valid: true,
                    clip_validity_reason: None,
                    speakers: Vec::new(),
                },
                observations: vec![CorrectionObservation {
                    local_label: "Speaker A".into(),
                    speaker_ordinal: 0,
                    observation_key: ObservationKey {
                        object_id: format!(
                            "kcode-audio-ingress/recording/{}/chunk/0",
                            recording.id
                        ),
                        piece_index: 0,
                    },
                    candidate: None,
                    identified_full_name: Some("Classifier Candidate".into()),
                    confirmed_full_name: confirmed_name.map(str::to_owned),
                }],
                clean: true,
            }],
            confirmation_state: state,
        });
        recording
    }

    #[test]
    fn quarter_context_plans_transcript_only_pieces() {
        let recording = completed_recording("a".repeat(801));
        let plan = TranscriptPlan::new(&recording, 400).unwrap();

        assert_eq!(plan.pieces().len(), 3);
        assert_eq!(
            plan.pieces()
                .iter()
                .map(|piece| piece.text.chars().count())
                .collect::<Vec<_>>(),
            vec![400, 400, 1]
        );
        assert_eq!(
            plan.pieces()
                .iter()
                .map(|piece| piece.estimated_tokens)
                .collect::<Vec<_>>(),
            vec![100, 100, 1]
        );
        assert_eq!(plan.pieces()[2].id, format!("audio:{}:2", recording.id));
    }

    #[test]
    fn splitting_prefers_late_paragraphs_and_counts_unicode_scalars() {
        let paragraph_recording =
            completed_recording(format!("{}\n\n{}", "a".repeat(250), "b".repeat(200)));
        let paragraph_plan = TranscriptPlan::new(&paragraph_recording, 400).unwrap();
        assert_eq!(paragraph_plan.pieces()[0].text, "a".repeat(250));
        assert_eq!(paragraph_plan.pieces()[1].text, "b".repeat(200));

        let unicode_recording = completed_recording("😀".repeat(5));
        let unicode_plan = TranscriptPlan::new(&unicode_recording, 4).unwrap();
        assert_eq!(
            unicode_plan
                .pieces()
                .iter()
                .map(|piece| piece.text.chars().count())
                .collect::<Vec<_>>(),
            vec![4, 1]
        );
    }

    #[test]
    fn canonical_audio_piece_ids_are_strict() {
        let recording_id = Uuid::parse_str("abcdef01-2345-6789-abcd-ef0123456789").unwrap();
        let canonical = format!("audio:{recording_id}:2");

        assert_eq!(parse_audio_piece_id(&canonical), Some((recording_id, 2)));
        assert!(parse_audio_piece_id(&format!("audio:{recording_id}:02")).is_none());
        assert!(parse_audio_piece_id("audio:ABCDEF01-2345-6789-ABCD-EF0123456789:2").is_none());
        assert!(parse_audio_piece_id(&format!("audio:{recording_id}:2:extra")).is_none());
        assert!(parse_audio_piece_id("audio:not-a-uuid:0").is_none());
    }

    #[test]
    fn metadata_reproduces_existing_keys_and_values() {
        let recording = completed_recording("Transcript");
        let plan = TranscriptPlan::new(&recording, 400).unwrap();

        assert_eq!(
            plan.pieces()[0].metadata(&recording),
            json!({
                "kind": "audio-transcript",
                "recordingId": "abcdef01-2345-6789-abcd-ef0123456789",
                "sha256": "a".repeat(64),
                "originalFilename": "meeting.final.WAV",
                "extension": ".WAV",
                "mimeType": "audio/wav",
                "sizeBytes": 42,
                "sourceCreatedAt": "2026-08-02T03:04:05+00:00",
                "pieceIndex": 0,
                "pieceCount": 1,
                "speakerConfirmationState": null,
            })
        );
    }

    #[test]
    fn legacy_rendering_reproduces_existing_text() {
        let recording = completed_recording("Transcript");
        let plan = TranscriptPlan::new(&recording, 400).unwrap();

        assert_eq!(
            plan.pieces()[0].formatted_text(&recording).unwrap(),
            format!(
                "Vnote final transcript piece\n\nRecording began: 2026-08-02T03:04:05+00:00\nRecording SHA-256: {}\nOriginal filename: meeting.final.WAV\nExtension: .WAV\nMIME type: audio/wav\nSize: 42 bytes\nTranscript piece: 1 of 1\n\nTranscript",
                "a".repeat(64)
            )
        );
    }

    #[test]
    fn confirmed_rendering_includes_the_complete_mapping() {
        let recording = with_speaker_packet(
            completed_recording("Transcript"),
            ConfirmationState::Confirmed,
            Some("Human Choice"),
        );
        let plan = TranscriptPlan::new(&recording, 400).unwrap();

        assert_eq!(
            plan.pieces()[0].formatted_text(&recording).unwrap(),
            format!(
                "Vnote final transcript piece\n\nRecording began: 2026-08-02T03:04:05+00:00\nRecording SHA-256: {}\nOriginal filename: meeting.final.WAV\nExtension: .WAV\nMIME type: audio/wav\nSize: 42 bytes\nTranscript piece: 1 of 1\n\nHuman-confirmed speaker-label data (authoritative; do not infer alternatives):\n- chunk 1/1, source 0.000-1.500s: localLabel=\"Speaker A\", confirmedFullName=\"Human Choice\", classifierCandidate=\"Classifier Candidate\"\n\nTranscript",
                "a".repeat(64)
            )
        );
        assert_eq!(
            plan.pieces()[0]
                .metadata(&recording)
                .get("speakerConfirmationState"),
            Some(&json!("confirmed"))
        );
    }

    #[test]
    fn unconfirmed_speaker_rendering_is_a_conflict() {
        let recording = with_speaker_packet(
            completed_recording("Transcript"),
            ConfirmationState::AutomaticallyTrained,
            None,
        );
        let plan = TranscriptPlan::new(&recording, 400).unwrap();

        assert!(matches!(
            plan.pieces()[0].formatted_text(&recording),
            Err(Error::Conflict(_))
        ));
        assert_eq!(
            plan.pieces()[0]
                .metadata(&recording)
                .get("speakerConfirmationState"),
            Some(&json!("automatically_trained"))
        );
    }
}