Archive Analysis

Skill: Archive Analysis

Purpose

Save a completed analysis to the knowledge system's analysis archive for future recall. Captures key findings, metrics used, agents invoked, and output file paths so that past work can be referenced in future sessions.

When to Use

• After completing an L3+ analysis (post-validation)
• After /run-pipeline completes successfully
• User says "save this analysis" or "archive this"
• Automatically triggered at the end of Step 18 (Close the Loop)
• VERIFICATION MODE: User says "verify the archive" or "show me what was captured" — read existing entry, don't create new

Instructions

Step 0: Determine Mode

Check if this is an archive request (create new entry) or a verification request (read existing entry).

Verification signals:

• User says "verify", "check", "show me what was captured", "did it save", "was it archived"
• User mentions pipeline "auto-saved" or "already completed"
• No new findings mentioned — just wants to see what's already there

If verification: Skip to Verification Mode (below). Otherwise, proceed to Step 1.

Step 1: Gather Analysis Metadata from Session Context

Record only what exists: findings, metrics, agents, and files come from the files and session history listed below, or from what the user stated. Anything not found is null or [], never invented.

Where to look:

1. Pipeline state — Read working/session_state.yaml if it exists (contains pipeline progress, agents run, resume instructions)
2. Active dataset — Read .knowledge/active.yaml for dataset ID
3. Output files — Scan outputs/ and working/ directories for artifacts created in this session (use ls -lt to find recent files)
4. Validation report — Look for working/validation_report.md or working/validation_summary.json for confidence grade
5. User's stated findings — If user provides key findings in their request, use those verbatim

What to extract:

1. Title: Derive from the original question or business context (check session history for the initial analytical question)
2. Question: The original user question (look back in conversation for the first analytical request)
3. Question level: From the Question Router classification (L1-L5) — check if router was invoked
4. Dataset ID: From .knowledge/active.yaml
5. Key findings:
   • If validation report exists, extract from there
   • Otherwise, check working/analysis_summary.md or outputs/narrative.md
   • If nothing exists, use findings the user mentioned in their archive request
   • Format as single-sentence bullets with numbers
6. Metrics used:
   • Check validation report or analysis summary for metric references
   • Scan SQL queries in working/*.sql for column names
   • Match against .knowledge/metrics/index.yaml if available
7. Agents used:
   • Read working/session_state.yaml if pipeline was used
   • Check conversation history for agent invocations (look for "Task" tool uses)
   • List in execution order
8. Output files:
   • Run ls -lt outputs/ working/ | head -20 to find recent files
   • Filter for files created during this session (check timestamps)
   • Include both deliverables (outputs/) and working files (working/)
   • ONLY list files that actually exist — check the directory before listing
   • If no files exist or user didn't mention any outputs, set output_files: []
9. Tags: Auto-generate from:
   • Keywords in the original question (mobile, checkout, seasonal, etc.)
   • Metric names used
   • Dataset name
   • Analysis type (funnel, root-cause, segmentation, etc.)
10. Confidence:
    • Read from working/validation_report.md or working/validation_summary.json
    • If validation was not run, set to null and note it

If actual files don't exist: Create the archive with whatever metadata is available. Mark it as partial: true if deliverables are missing.

Step 2: Create Archive Entry

1. Read the schema: .knowledge/analyses/_schema.yaml to understand required/optional fields
2. Generate a unique ID: analysis_{YYYYMMDD}_{HHMMSS} (use current timestamp)
3. Build the entry dict following the schema structure

Example entry format:

- id: analysis_20260403_232623
  title: "Mobile checkout conversion drop investigation"
  date: "2026-04-03"
  dataset: {active_dataset}
  question: "What caused the conversion rate drop and how does it relate to the mobile checkout flow?"
  question_level: L4
  findings:
    - "Mobile checkout conversion fell 2.1pp (5.4% → 3.3%) in March; desktop was flat"
    - "82% of the drop is concentrated at the payment step on iOS"
    - "Sessions with a payment error retry at 11% vs. 64% baseline"
  metrics:
    - conversion_rate
    - checkout_completion
    - funnel_drop_off
    - device_segmentation
  agents:
    - question-framing
    - data-explorer
    - descriptive-analytics
    - root-cause-investigator
    - validation
    - chart-maker
  output_files:
    - outputs/checkout_conversion_analysis.png
    - outputs/conversion_funnel_chart.png
    - working/checkout_funnel_analysis.png
  tags:
    - conversion
    - mobile
    - checkout
    - funnel-analysis
    - root-cause
  confidence: B
  partial: false

Step 3: Append to Index

1. Read .knowledge/analyses/index.yaml
2. If file doesn't exist, create it from template:

   analyses: []
   total_analyses: 0
   last_updated: null

3. Append the new entry to the analyses list
4. Increment total_analyses
5. Update last_updated to current date (YYYY-MM-DD)
6. Write back to index.yaml

Step 4: Update Dataset Stats

1. Read .knowledge/datasets/{active}/manifest.yaml
2. Increment analysis_count
3. Update last_used to current date
4. Write back

Step 5: Confirm

Report to user:

Analysis archived: {title}
ID: {id}
Findings: {count} key findings captured
Use `/history` to browse past analyses.