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.

Step 6: Capture to Query Archaeology (Optional)

When to apply: Only for completed analyses (not partial) with confidence grade B or better.

After archiving, check if the analysis produced reusable patterns worth saving to .knowledge/query-archaeology/curated/ via helpers/knowledge/archaeology_helpers.py.

1. SQL patterns — If validated SQL queries exist in working/*.sql:

   • Ask: "Would you like to save any SQL patterns from this analysis?"
   • Offer to capture via capture_cookbook_entry(title, sql, dataset, tables, tags)
   • Only capture queries that passed tie-out or validation checks

2. Table knowledge — If the analysis revealed useful table metadata:

   • Offer to capture/update via capture_table_cheatsheet(table_name, dataset, grain, primary_key, common_filters, gotchas, common_joins)
   • Include grain, primary key, common filters, gotchas, and common joins

3. Join patterns — If the analysis used non-obvious joins:

   • Offer to capture via capture_join_pattern(tables, join_sql, cardinality, validated, dataset)
   • Record cardinality and whether the join was validated

Rules for this step:

• Ask the user: "Would you like to save any SQL patterns from this analysis?"
• If the user declines or there are no reusable patterns, skip silently
• Only offer for analyses with confidence grade B or better
• Never auto-capture without user confirmation

Verification Mode

When user wants to verify an existing archive (not create a new one):

1. Read .knowledge/analyses/index.yaml
2. Find the most recent entry (highest index, latest date)
3. Display what was captured:
   • Analysis ID and title
   • Date archived
   • Key findings (list them)
   • Metrics used
   • Agents invoked
   • Output files preserved
   • Tags
   • Confidence grade
4. Confirm dataset manifest was updated
5. Report archive stats (total analyses, last updated)

Do NOT create a new archive entry in verification mode.

Report format:

Archive verified: {title}
ID: {id}
Status: Successfully archived on {date}

Captured:
- {count} key findings
- {count} metrics tracked
- {count} agents used
- {count} output files preserved

Use `/history` to browse all past analyses.

Rules

1. Never overwrite an existing archive entry — always append
2. Key findings should be one sentence each, factual, with numbers where possible
3. Tags should be lowercase, no spaces (use hyphens)
4. If validation was not run, set confidence to null and note it
5. Archive even partial analyses — mark as partial: true
6. Verification requests don't create new entries — read and report existing archives
7. DO NOT create standalone analysis markdown files or archive directories — the archive system stores metadata in index.yaml only. Output files remain in their original locations (outputs/, working/) and are referenced by path in the output_files array. DO NOT copy or duplicate artifacts.

Edge Cases

• No outputs exist: Set output_files: [], do not invent file names. Archive with metadata only.
• Pipeline was interrupted: Archive what's available, mark as partial: true, document reason in a reason_incomplete field
• Duplicate question: Still archive — different runs may find different things
• Analysis index doesn't exist: Create it from template
• User says 'verify' but no archive exists: Report "No archive found. Would you like to create one?" and proceed to archive mode if confirmed
• Session state files don't exist: Use conversation history and user-provided metadata to build the archive entry
• User mentioned files but they don't exist: Only list files that actually exist in outputs/ or working/. If the user mentioned output files but they're not present, note in confirmation: "Files mentioned but not found: {list}"