get_correction_history

Instructions

Implementation Reference

• packages/touch/src/contextpulse_touch/mcp_server.py:182-227 (handler)

  Actual implementation of get_correction_history tool. Queries the activity database for 'correction_detected' events from the 'keys' modality, formats them into a human-readable string with timestamp, original text, corrected text, confidence, correction type, and seconds after paste.

  @mcp_app.tool()
  def get_correction_history(limit: int = 20) -> str:
      """Get recent Voice correction detections.
  
      Shows what words/phrases the user corrected after voice dictation,
      along with confidence scores and timestamps. These corrections are
      automatically fed back to Voice's vocabulary for self-improving dictation.
  
      Args:
          limit: Maximum number of corrections to return (default 20).
      """
      limit = max(1, min(limit, 200))
  
      conn = _get_db()
      if not conn:
          return "No activity database found."
  
      try:
          rows = conn.execute(
              "SELECT timestamp, payload FROM events "
              "WHERE modality = 'keys' AND event_type = 'correction_detected' "
              "ORDER BY timestamp DESC LIMIT ?",
              (limit,),
          ).fetchall()
          conn.close()
  
          if not rows:
              return "No corrections detected yet. Corrections are captured when you edit Voice-dictated text."
  
          lines = [f"=== Correction History (last {len(rows)}) ===\n"]
          for row in rows:
              p = json.loads(row["payload"])
              ts = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime(row["timestamp"]))
              orig = p.get("original_text", "?")
              corr = p.get("corrected_text", "?")
              conf = p.get("confidence", 0)
              ctype = p.get("correction_type", "?")
              secs = p.get("seconds_after_paste", 0)
  
              lines.append(f"[{ts}] {orig!r} -> {corr!r}")
              lines.append(f"  Type: {ctype}, Confidence: {conf:.0%}, {secs:.1f}s after paste")
              lines.append("")
  
          return "\n".join(lines)
      except Exception as e:
          return f"Error reading corrections: {e}"

• packages/touch/src/contextpulse_touch/mcp_server.py:182-183 (registration)

  Tool registered via @mcp_app.tool() decorator on the FastMCP server ('ContextPulse Touch') in the touch package.

  @mcp_app.tool()
  def get_correction_history(limit: int = 20) -> str:

• glama/server.py:480-496 (registration)

  Stub registration in the Glama.ai registry. Returns a message telling users to install ContextPulse locally; this is a discovery-only stub, not the real implementation.

  @mcp_app.tool()
  def get_correction_history(limit: int = 20) -> str:
      """Return recent voice transcription corrections detected by the user or auto-detected.
  
      Returns correction events: original transcript, corrected text, confidence
      delta, timestamp, and whether the correction was manual or auto-suggested.
  
      USE WHEN: training the vocabulary, analyzing systemic Whisper errors, or
      debugging why a specific term keeps mis-transcribing.
      NOT FOR: vocabulary management — use add_to_vocabulary / remove_from_vocabulary.
  
      BEHAVIOR: pure read. No side effects.
  
      PARAMETERS:
        limit: max results, ordered newest-first. Range 1-100. Default 20.
      """
      return _LOCAL_ONLY_MSG

• glama/server.py:480-496 (handler)

  Glama stub handler for get_correction_history. Returns _LOCAL_ONLY_MSG, as this server is a registry discovery stub, not the actual local daemon.

  @mcp_app.tool()
  def get_correction_history(limit: int = 20) -> str:
      """Return recent voice transcription corrections detected by the user or auto-detected.
  
      Returns correction events: original transcript, corrected text, confidence
      delta, timestamp, and whether the correction was manual or auto-suggested.
  
      USE WHEN: training the vocabulary, analyzing systemic Whisper errors, or
      debugging why a specific term keeps mis-transcribing.
      NOT FOR: vocabulary management — use add_to_vocabulary / remove_from_vocabulary.
  
      BEHAVIOR: pure read. No side effects.
  
      PARAMETERS:
        limit: max results, ordered newest-first. Range 1-100. Default 20.
      """
      return _LOCAL_ONLY_MSG

• packages/touch/src/contextpulse_touch/mcp_server.py:23-28 (helper)

  Helper function _get_db() that opens a sqlite3 connection to the activity database (ACTIVITY_DB_PATH), used by get_correction_history to fetch correction events.

  def _get_db() -> sqlite3.Connection | None:
      if not _DB_PATH.exists():
          return None
      conn = sqlite3.connect(str(_DB_PATH), timeout=5)
      conn.row_factory = sqlite3.Row
      return conn