Conversation Logging

All conversations in EleAgent are persisted for post-analysis. This is a core Phase 1 deliverable — the recorded interactions will inform what MCP components and data pipelines to build in Phase 2.

What Is Logged

Data	Storage Location	Format
Conversation state	/artefacts/state/home/.claude/	Internal session files
Transcript logs	/artefacts/logs/	JSONL (one JSON object per line)
Session artefacts	/artefacts/sessions/<key>/	Mixed (CSV, text, etc.)
Exported conversations	/artefacts/conversations/export/	Structured exports
Audio recordings	/artefacts/sessions/<key>/	Audio files with transcriptions

JSONL Transcript Logs

Each AI interaction is logged as a JSONL (JSON Lines) entry in /artefacts/logs/. These structured logs capture:

• User messages (original query text)
• Agent responses (full output including tool calls)
• Tool executions (what commands/scripts the agent ran)
• Token usage (input/output token counts)
• Timing data (query duration)
• Session metadata (session ID, project path)

Audio Recording

The web UI includes a built-in audio recording feature for capturing test sessions. When a testing supervisor observes an engineer using the system, they can record the discussion to capture:

• Verbal observations about the agent's behavior
• Points of confusion or frustration
• Suggestions for improvement
• Context that isn't reflected in the chat (e.g., "I expected it to check the Excel file first")

Audio recordings are stored in the session's artefacts directory and can be transcribed automatically for analysis.

Persistence

All logged data is stored on the host filesystem via the artefacts volume mount:

volumes:
  - "${ARTEFACTS_PATH:-./artefacts}:/artefacts"

Data persists across:

• Container restarts
• Image rebuilds
• Stack redeployments

WARNING

Back up the artefacts directory regularly. It contains all conversation history and is the primary input for Phase 2 analysis.

Purpose: Informing Phase 2

The central insight of the phased approach is: let practice dictate the tooling, not the other way around.

Instead of guessing which MCP servers or RAG pipelines to build, Phase 1 records how engineers actually use the system. Post-analysis of these logs will reveal:

• Most common query types — what do engineers ask most?
• Frequently accessed files — which documents are queried repeatedly?
• Pain points — where does the agent take too long or produce incorrect results?
• Cross-referencing patterns — which data sources are commonly combined?
• Tool usage patterns — which Python libraries and shell tools are used most?

This data directly informs the selection and design of MCP components and data pipelines for Phase 2.