openharness

/**
 * Reading a transcript for session search: line by line from a byte offset, through the same
 * incremental normalizer each engine uses for its live view.
 *
 * Transcripts are large (a long Codex rollout runs to gigabytes) and nearly all of it is tool output.
 * The reader looks at the first bytes of each line and drops the output records before they are
 * decoded or parsed, and never holds more than one line in memory — a 200 MB command dump costs a
 * seek, not a parse.
 */

import { open } from 'node:fs/promises'

import type { LiveEvent } from '../normalize.js'
import { lineToEvents, newTurnState } from '../normalize.js'
import { AgyNormalizer } from '../../engines/agy/normalizer.js'
import { AmpNormalizer } from '../../engines/amp/normalizer.js'
import { CodexNormalizer } from '../../engines/codex/normalizer.js'
import { CommandCodeNormalizer } from '../../engines/commandcode/normalizer.js'
import { CopilotNormalizer } from '../../engines/copilot/normalizer.js'
import { CursorNormalizer } from '../../engines/cursor/normalizer.js'
import { GrokNormalizer } from '../../engines/grok/normalizer.js'
import { MuseNormalizer } from '../../engines/muse/normalizer.js'
import { PiNormalizer } from '../../engines/pi/normalizer.js'

export type LineNormalizer = (line: string) => LiveEvent[]

/**
 * A fresh normalizer for one pass over one transcript, or null for an engine whose history is not a
 * JSONL file (OpenCode, Kilo, Hermes and Devin keep theirs in a database; a terminal has none).
 * Fresh on every pass: a pass starts at a turn boundary, where no earlier state is needed.
 */
export function lineNormalizer(engine: string, sessionId: string): LineNormalizer | null {
  switch (engine) {
    case 'claude': {
      const state = newTurnState()
      return (line) => {
        const events = lineToEvents(line, state)
        const asked = claudeSideAsk(line)
        return asked ? [{ type: 'user_message', payload: { content: asked } }, ...events] : events
      }
    }
    case 'codex': {
      // Sub-agent threads are resolved for the live view's cards; search only needs the text.
      const normalizer = new CodexNormalizer('live', () => null)
      return (line) => normalizer.ingest(line)
    }
    case 'cursor': {
      const normalizer = new CursorNormalizer('live', sessionId)
      return (line) => normalizer.ingest(line)
    }
    case 'muse': return ingestWith(new MuseNormalizer())
    case 'amp': return ingestWith(new AmpNormalizer())
    case 'grok': return ingestWith(new GrokNormalizer())
    case 'agy': return ingestWith(new AgyNormalizer())
    case 'copilot': return ingestWith(new CopilotNormalizer())
    case 'pi': return ingestWith(new PiNormalizer('live'))
    case 'commandcode': return ingestWith(new CommandCodeNormalizer('live'))
    default: return null
  }
}

/**
 * The person's words that Claude Code writes outside its prompt records. The live view leaves them out
 * on purpose, since neither opens a turn there, but both are things a person remembers asking:
 *  - a message typed while the agent was working, delivered mid-turn as a `queued_command` attachment
 *    (sub-agent hand-backs arrive the same way and are filed under the agent by `askKind`; task
 *    notifications are not prompts and are left out);
 *  - a `/goal`, whose text is recorded only in the bookkeeping line that turns its Stop hook on.
 */
export function claudeSideAsk(line: string): string | null {
  const queued = line.includes('"queued_command"')
  if (!queued && !(line.includes('"isMeta":true') && line.includes(GOAL_HEAD))) return null
  let raw: {
    type?: unknown; isMeta?: unknown; message?: { content?: unknown }
    attachment?: { type?: unknown; commandMode?: unknown; prompt?: unknown; origin?: { kind?: unknown } }
  }
  try { raw = JSON.parse(line) } catch { return null }
  if (queued) {
    const attachment = raw.attachment
    if (attachment?.type !== 'queued_command' || attachment.commandMode !== 'prompt') return null
    // "Goal set: …", which Claude queues for itself: the goal's own line below already holds it.
    if (attachment.origin?.kind === 'auto-continuation') return null
    return promptText(attachment.prompt)
  }
  if (raw.type !== 'user' || raw.isMeta !== true) return null
  const goal = GOAL_SET.exec(promptText(raw.message?.content) ?? '')
  return goal ? goal[1].trim() || null : null
}

const GOAL_HEAD = 'A session-scoped Stop hook is now active with condition: '
const GOAL_SET = /^A session-scoped Stop hook is now active with condition: "([\s\S]*)"\. Briefly acknowledge/

/** A prompt's text: a string, or the text blocks of a content array. */
function promptText(prompt: unknown): string | null {
  if (typeof prompt === 'string') return prompt.trim() ? prompt : null
  if (!Array.isArray(prompt)) return null
  const text = prompt
    .map((block) => block && typeof block === 'object' && (block as { type?: unknown }).type === 'text' ? (block as { text?: unknown }).text : null)
    .filter((part): part is string => typeof part === 'string' && part.trim() !== '')
    .join('\n')
  return text || null
}

function ingestWith(normalizer: { ingest(line: string): LiveEvent[] }): LineNormalizer {
  return (line) => normalizer.ingest(line)
}

/**
 * Records that carry only tool output, reasoning or accounting, told apart by their opening bytes.
 * Skipping them is safe because the turn collector reads prompts, answer text and tool calls only;
 * `transcript.spec.ts` checks the turns come out the same with and without the skip.
 */
const SKIP_HEAD: Record<string, RegExp> = {
  codex: /"type":"(?:function_call_output|custom_tool_call_output|reasoning|token_count|local_shell_call_output|web_search_call_output)"|^\{"timestamp":"[^"]*","type":"token_usage_record"/,
  // A user record whose content is a tool result. A prompt can never open with one.
  claude: /^\{[^]*?"type":"user"[^]*?"content":\[\{"tool_use_id"|^\{[^]*?"type":"user"[^]*?"content":\[\{"type":"tool_result"/,
}

/** How much of a line is read before deciding whether to skip it. */
const HEAD_BYTES = 1024
/** A longer line is not a prompt or an answer anyone wrote; it is dropped unread. */
const MAX_LINE_BYTES = 16 * 1024 * 1024
const CHUNK_BYTES = 1024 * 1024

export function skipPredicate(engine: string): ((head: string) => boolean) | null {
  const pattern = SKIP_HEAD[engine]
  return pattern ? (head) => pattern.test(head) : null
}

const TIMESTAMP = /"timestamp"\s*:\s*"([^"]{10,40})"/

/** When a transcript line was written, from its own `timestamp` field; null when it has none. */
export function lineTime(line: string): number | null {
  const match = TIMESTAMP.exec(line)
  if (!match) return null
  const at = Date.parse(match[1])
  return Number.isFinite(at) ? at : null
}

export interface LineVisit {
  text: string
  /** Byte offset of the line's first byte. */
  offset: number
}

/**
 * Calls `visit` for every complete line from `start`, and resolves with the offset just past the last
 * complete line. A final line without its newline is left for the next pass: the engine may still be
 * writing it. `visit` may return a promise to pace the reader.
 */
export async function forEachLine(
  path: string,
  start: number,
  visit: (line: LineVisit) => void | Promise<void>,
  options: { skip?: ((head: string) => boolean) | null; shouldStop?: () => boolean } = {},
): Promise<{ end: number }> {
  const handle = await open(path, 'r')
  try {
    const buffer = Buffer.allocUnsafe(CHUNK_BYTES)
    let position = start
    let lineStart = start
    let parts: Buffer[] = []
    let length = 0
    let decided = false
    let dropping = false
    for (;;) {
      if (options.shouldStop?.()) return { end: lineStart }
      const { bytesRead } = await handle.read(buffer, 0, CHUNK_BYTES, position)
      if (bytesRead === 0) return { end: lineStart }
      let cursor = 0
      while (cursor < bytesRead) {
        const newline = buffer.indexOf(10, cursor)
        const stop = newline === -1 || newline >= bytesRead ? bytesRead : newline
        if (!dropping) {
          const piece = buffer.subarray(cursor, stop)
          if (length + piece.length > MAX_LINE_BYTES) {
            dropping = true
            parts = []
            length = 0
          } else {
            parts.push(Buffer.from(piece))
            length += piece.length
            if (!decided && options.skip && (length >= HEAD_BYTES || stop < bytesRead)) {
              decided = true
              const head = Buffer.concat(parts, length).subarray(0, HEAD_BYTES).toString('utf8')
              if (options.skip(head)) {
                dropping = true
                parts = []
                length = 0
              }
            }
          }
        }
        if (stop === bytesRead) break
        // A complete line.
        if (!dropping && length > 0) {
          const text = Buffer.concat(parts, length).toString('utf8')
          await visit({ text, offset: lineStart })
        }
        lineStart = position + stop + 1
        parts = []
        length = 0
        decided = false
        dropping = false
        cursor = stop + 1
      }
      position += bytesRead
    }
  } finally {
    await handle.close()
  }
}