Sources

AgentLens reads the session files that the coding agents write themselves. It never writes to them (SQLite files are opened read-only), and a record it does not understand produces a warning and is skipped, never a failure.

Source ID	Agent	Where it reads
claude-code	Claude Code	the session files under ~/.claude/projects (or $CLAUDE_CONFIG_DIR)
opencode	OpenCode	the SQLite database under ~/.local/share/opencode, and the JSON files of versions before 1.2.0 under storage/; a session in both is counted once
copilot-cli	GitHub Copilot CLI	~/.copilot/session-state and the session-store.db next to it (or $COPILOT_HOME)
vscode-copilot	GitHub Copilot Chat in VS Code and VS Code Insiders	the chatSessions files in the user directory, including the .jsonl operation log

agentlens scan shows what was found and what it holds, with a reason where a source was not found.

What each source reports

The clients differ in what they record, and AgentLens shows the difference instead of filling gaps:

• Claude Code reports tokens per request with the cache split, the static context the client adds to every request, tool calls with their outcome, subagents, compactions and the git branch. It is the only source that reports the static context, so agentlens context and the idle cost of unused depend on it.
• OpenCode reports tokens per request and the cost it computed itself (shown as is).
• Copilot CLI reports tokens and premium requests per request from session-store.db; a session without rows in the store is read from the session.usage_record events of newer versions (one per request, measured), and failing that from the session totals, which cannot be split per request. When both the store and the records exist, the store is used and the records are not counted again. The records carry no subagent marker as far as seen, so their requests count as main-agent requests. It also records hooks and permission prompts.
• VS Code Copilot reports prompt and completion tokens per request without a cache split, plus the share of system instructions and tool definitions in the prompt, which agentlens context uses as reported token counts.

A value a client does not record is n/a, never 0.

Adding an OpenTelemetry export

An export adds what the local files do not hold, such as the cost Claude Code computed per request and the skill each request is attributed to. A session that is in the local data and in an export is counted once.