Review Conversation

Transparent conversation transcript viewer for SPARTA stress test sessions. Every turn, grade, and evaluation visible — no black boxes.

Quick Start

# List available session files
./run.sh files

# Show summary of latest session file
./run.sh list

# Show full transcript of session #3
./run.sh show -n 3

# Show specific session by ID
./run.sh show -i stress_20260222_175146

# Show only disagreements (student vs teacher)
./run.sh compare

# Export to markdown for human annotation
./run.sh export -o review.md

# NDJSON stream for automation
./run.sh json-stream

Commands

Show Options

What You See Per Turn

For persona turns (Margaret/Jennifer asking):

• Speaker, action type (QUERY/FOLLOW_UP), timestamp
• Persona evaluation verdict: SATISFACTORY / INCOMPLETE / WRONG / FLAW_CAUGHT / FLAW_MISSED
• Evaluation reasoning (why the persona accepted or rejected)

For system turns (SPARTA answering):

• Full response text with entity gate decisions
• Self-grade: letter grade + composite score + iteration count
• QRA citation count
• Issues flagged during self-grading
• Rationale from semantic grader

What You See Per Session

• Session grade: 8-dimension rubric breakdown with visual bars
• Student vs Teacher comparison: side-by-side grades, confidence, model names
• Agreement signal: AGREE (green) or DISAGREE (red)
• Improvement delta: first → final score, outer/inner iteration counts
• QRA citation metrics: verified/total, grounding average

Data Sources

Environment Variables

Semantic Search via /memory (search, ingest)

Composes with /memory for BM25 + semantic embedding + multi-hop graph traversal. Far superior to bespoke text search for natural language queries.

Workflow

# 1. After a stress test, ingest sessions to /memory
./run.sh ingest

# 2. Search semantically (natural language)
./run.sh search "conversations where Brandon missed a fake control"
./run.sh search "GPS spoofing questions that got resolved"
./run.sh search "low-scoring sessions with iteration improvement"

# 3. Or use --semantic on find (combines structured filters + semantic)
./run.sh find --semantic "fake control missed" --grade F

How It Works

ingest extracts problem/solution pairs from each session and stores them with tags (grade_A, difficulty_complex, resolution_resolved, etc.) for filtered recall.

Graceful Degradation

If /memory is unavailable (no ArangoDB, no embedding service), semantic commands return empty results and fall back to structured filters. The skill never crashes — it just loses the semantic layer.

Agent-Human Collaboration (find)

The find command is the primary collaboration surface. When the human asks a question about conversations, the agent translates it into structured filters:

# "Show me a conversation where Brandon didn't flag a non-existent control"
./run.sh find --eval flaw_missed --resolution no_coverage

# "Show me where Margaret failed the first response but iteration saved it to 90%"
./run.sh find --first-eval "!satisfactory" --min-composite 0.9

# "Which complex questions got fully resolved?"
./run.sh find --difficulty complex --resolution resolved

# "Find conversations about SV-SP-1"
./run.sh find --text "SV-SP-1"

# "Show disagreements on flawed questions"
./run.sh find --difficulty flawed --disagree

# "How many F-grade sessions are there?"
./run.sh find --grade F --format count

# "Show low-scoring sessions in Rich detail"
./run.sh find --max-composite 0.5 --format show

Find Filters

All filters are AND-combined. Default output is brief (pasteable markdown).

Translation Guide for Agents

When the human asks in natural language, translate to filters:

Agent In-Chat Display (brief)

The brief command outputs plain markdown (no Rich/ANSI codes) that agents can paste directly into chat:

# Summary table (paste in chat)
./run.sh brief

# Single session compact view
./run.sh brief -n 3
./run.sh brief -i stress_20260222

Visualization with /create-figure (data)

Export JSON files that /create-figure consumes directly:

# Export all chart data
./run.sh data -o /tmp/charts/

# Then visualize with /create-figure
cd .pi/skills/create-figure
./run.sh radar --input /tmp/charts/radar.json --output radar.pdf
./run.sh heatmap --input /tmp/charts/heatmap.json --output heatmap.pdf
./run.sh metrics --input /tmp/charts/grade_distribution.json --type bar

# Export only radar chart data
./run.sh data -c radar -o /tmp/charts/

JSON files produced:

• radar.json — 8-dimension rubric scores per session (for radar charts)
• heatmap.json — difficulty x grade distribution (for heatmaps)
• grade_distribution.json — grade letter counts (for bar charts)
• resolution_distribution.json — resolution outcome counts (for bar/pie)
• agreement.json — agreed/disagreed/no_teacher counts (for bar/pie)

For deeper analytics, pipe json-stream output to /analytics:

./run.sh json-stream > sessions.jsonl
cd .pi/skills/analytics
./run.sh describe sessions.jsonl
./run.sh chart sessions.jsonl --name grade --for-figure

Integration with /sparta-stress-test

Run a stress test, then immediately review:

# Run 25 conversations
cd .pi/skills/sparta-stress-test
./run.sh run --count 25

# Review results
cd .pi/skills/review-conversation
./run.sh list
./run.sh compare
./run.sh show -n 1