Conversation Analysis

Analyze user messages from OpenCode sessions to identify recurring themes, communication patterns, and steering behaviours.

Critical Rules

• NEVER cat or read chunk files directly - they're huge and will explode your context

• Pass file paths to subagents - let them read and analyze independently

• Use parallel subagents - one per chunk, they run concurrently

• Subagents return structured JSON - you synthesize at the end

Workflow

Step 1: Run Extraction

~/.agents/skills/opencode-conversation-analysis/scripts/extract.sh

This script:

• Reads from ~/.local/share/opencode/opencode.db (SQLite) when available

• Falls back to ~/.local/share/opencode/storage/ for older OpenCode installs

• Extracts user messages from main sessions with metadata (session_id, title, timestamp, text)

• Filters out messages < 10 characters

• Chunks into ~320k char files (~80k tokens each)

• Outputs to /tmp/opencode-analysis/chunk_*.jsonl

Review the output summary to see how many chunks were created.

Step 2: Launch Parallel Subagents

For each chunk file, spawn a general subagent with this prompt template:

Read the file /tmp/opencode-analysis/chunk_N.jsonl which contains user messages from coding sessions (JSONL format with fields: session_id, session_title, timestamp, text).

Analyze these messages to identify recurring themes in how the user steers/guides AI coding assistants. Look for patterns like:

• How they give feedback
• How they correct mistakes
• How they scope/refine requests
• Communication style preferences
• Technical approaches they emphasize

For each theme you identify, provide:

1. Theme name (short, descriptive)
2. Description (1-2 sentences)
3. 2-3 direct quote examples from the messages

Return ONLY valid JSON in this format: { "themes": [ { "name": "Theme Name", "description": "Description of the pattern", "examples": ["quote 1", "quote 2"] } ] }

Launch ALL chunk subagents in parallel (single message, multiple Task tool calls).

Step 3: Synthesize Results

Once all subagents return:

• Collect all theme objects from all chunks

• Group similar themes (same name or overlapping descriptions)

• Merge examples from duplicate themes

• Rank themes by how many chunks they appeared in

• Pick the best 2-3 examples per theme

Step 4: Output Format

Present the final analysis as markdown with this structure:

Themes in How You Steer AI Coding Assistants

Analysis of N messages across M sessions (date range)

1. Theme Name

Description of the pattern.

Examples:

• "direct quote 1"
• "direct quote 2"
• "direct quote 3"

2. Next Theme

...

Output directly to the user - don't write to a file unless asked.

Customisation Options

The user may request:

• Different chunk sizes: Edit CHUNK_SIZE in extract.sh (default 320000 chars)

• Different message filter: Edit MIN_TEXT_LEN in extract.sh (default 10 chars)

• Include subagent sessions: Remove the parent_id IS NULL (SQLite) and parentID (legacy) filters in extract.sh

• Time period filtering: Add timestamp filtering in extract.sh

Storage Format Reference

See references/storage-format.md for details on OpenCode's conversation storage structure.