observation-extractor

Extract User Observations

Purpose

Extract user-reported issues from the voice transcript of an ADA capture session. This is the first step in voice-first analysis - the transcript is the ground truth of user observations.

Context

• ADA Binary Directory: {{$ADA_BIN_DIR}}
• Analysis Session Path: {{$ANALYSIS_SESSION_PATH}}
• Capture Session: {{$CAPTURE_SESSION}}
• Output Directory: {{$OUTPUT_DIRECTORY}}

Execution Principles

MUST:

1. You MUST process pagination only when the instruction guarded by <EXTERMELY_IMPORTANT></EXTERMELY_IMPORTANT> appeared in the outputs of ada query {{$CAPTURE_SESSION}} [subcommand] call.

MUST NOT:

1. You MUST NOT invent redirect to files when it is not asked to.
2. You MUST NOT persist JSON variable to the filesystem until it is asked to.
3. You MUST NOT assume pagination (instructions guarded by <EXTERMELY_IMPORTANT></EXTERMELY_IMPORTANT>) always happen for each ada query {{$CAPTURE_SESSION}} [subcommand] call.

Step 1. Get Session Info and Early Exit Check

Command: ${ADA_BIN_DIR}/ada query {{$CAPTURE_SESSION}} transcribe info --reading-model sonnet --format json --cache-dir {{$ANALYSIS_SESSION_PATH}}/cache

You MUST stop and respond the following JSON immediately if the command fails or word_count is 0.

{"status": "error", "error": "no_voice_recording", "issue_count": 0, "has_ambiguities": false, "fallback_suggestion": "Analyze using screenshots and trace events only"}

Command: ${ADA_BIN_DIR}/ada query {{$CAPTURE_SESSION}} time-info --reading-model sonnet --cache-dir {{$ANALYSIS_SESSION_PATH}}/cache

Set $SESSION_INFO to:

{
  "first_event_ns": "[first_event_ns from output]",
  "duration_sec": "[duration_sec from output]"
}

You MUST write the contents of $SESSION_INFO (a JSON object) to {{$OUTPUT_DIRECTORY}}/session_info.json.

Read the capture session's platform from its manifest:

Path: ~/.ada/sessions/{{$CAPTURE_SESSION}}/manifest.json