orca/src/shared/harness-injected-user-turns.ts

// Why: agent harnesses (Claude Code and its forks) inject machinery into the
// conversation as user-role turns — background task notifications, system
// reminders, inter-agent messages, slash-command envelopes, local-command
// output, interruption and compaction notices. These fire user-prompt hooks
// and land in transcripts, but they are not something the user typed, so
// prompt-derived UI must not surface them.
//
// We match only tags we have observed from harnesses, never a broad kebab
// shape: a real prompt starting with a custom `<my-element>` or a Grok
// `<user_query>` envelope is a genuine user turn, and misclassifying it would
// hide the turn (drop it from transcripts, demote its session title, or leave
// the agent visibly done after an interrupt).
const LEADING_TAG_NAME = /^<([a-z][a-z0-9-]*)(?:[\s>]|$)/

// Consumers must only treat tags we have observed from harnesses as machinery;
// arbitrary kebab tags can be genuine user code.
const KNOWN_HARNESS_TAG_NAMES = new Set([
  'agent-message',
  'bash-input',
  'bash-stderr',
  'bash-stdout',
  'command-args',
  'command-message',
  'command-name',
  'cross-session-message',
  'fork-boilerplate',
  'local-command-caveat',
  'local-command-stderr',
  'local-command-stdout',
  'mcp-polling-update',
  'mcp-resource-update',
  'system-reminder',
  'task-notification',
  'teammate-message',
  'user-memory-input',
  'user-prompt-submit-hook'
])

// Injected turns identified by a leading string rather than a known tag name:
// the harness only emits <channel> in its attributed `<channel source=…>` form
// (a bare <channel> is a real RSS/XML paste), plus prose deliveries and notices.
const COMPACT_CONTINUATION_PREFIX = 'this session is being continued from a previous conversation'
const HARNESS_INJECTED_TURN_PREFIXES = [
  '<channel source=',
  '[request interrupted',
  'a message arrived from ',
  'another claude session sent a message',
  'no response requested.',
  'caveat: the messages below were generated by the user while running local commands',
  COMPACT_CONTINUATION_PREFIX
]

// Why: classification only inspects leading tags/prefixes. Cap the toLowerCase
// copy so vault-scan / prompt-seed paths stay O(1) on multi-KB pastes.
const HARNESS_CLASSIFY_HEAD_LIMIT = 256
const HARNESS_CLASSIFY_LEADING_WS_LIMIT = 64

/** True only for observed harness shapes. Match on trimmed, lowercased text.
 *  Unknown kebab tags stay user turns — only tags we have observed count. */
export function isKnownHarnessInjectedUserTurnText(text: string): boolean {
  const normalized = normalizedHarnessTurnHead(text)
  if (!normalized) {
    return false
  }
  const tagName = LEADING_TAG_NAME.exec(normalized)?.[1]
  if (tagName && KNOWN_HARNESS_TAG_NAMES.has(tagName)) {
    return true
  }
  return HARNESS_INJECTED_TURN_PREFIXES.some((prefix) => normalized.startsWith(prefix))
}

/** True only for the observed post-compaction continuation prompt. */
export function isCompactContinuationUserTurnText(text: string): boolean {
  return normalizedHarnessTurnHead(text).startsWith(COMPACT_CONTINUATION_PREFIX)
}

function normalizedHarnessTurnHead(text: string): string {
  let start = 0
  const wsScanEnd = Math.min(text.length, HARNESS_CLASSIFY_LEADING_WS_LIMIT)
  while (start < wsScanEnd && isAsciiWhitespace(text.charCodeAt(start))) {
    start += 1
  }
  if (start >= text.length) {
    return ''
  }
  const headEnd = Math.min(text.length, start + HARNESS_CLASSIFY_HEAD_LIMIT)
  return text.slice(start, headEnd).toLowerCase()
}

function isAsciiWhitespace(code: number): boolean {
  return code === 32 || code === 9 || code === 10 || code === 13 || code === 12
}