Omega Stenographer

Stenographer preserves conversation turns, extracts decisions and blockers, and builds searchable briefs without deleting the original exchanges.

Version 2.0 requires explicit session scope. Use the same session_id across ingest, brief, compaction, milestone and history calls. A global history search requires cross_session=true. Marking a milestone changes priority; it does not establish that a statement is true.

Install with python -m pip install .. Run omega-stenographer-mcp for explicit MCP ingestion, or python codex_capture.py to run the server with automatic visible-message capture from local Codex transcript files. The compatibility standalone entry point uses the same maintained implementation. No model API key is required for extraction or search.

Data defaults to ~/.omega-stenographer. OMEGA_CODEX_SESSIONS overrides the transcript directory; otherwise capture uses $CODEX_HOME/sessions or ~/.codex/sessions. The initial cutoff defaults to September 7, 2026 and can be set with OMEGA_CAPTURE_SINCE=YYYY-MM-DD. Older files are indexed at their current end, and new appended messages are subsequently captured. One OS lock permits a single observer across multiple MCP processes.

The observer excludes tool results, reasoning and configuration messages; persists source receipts and SQLite file offsets; retries partial lines; and delivers committed exchanges to Omega Brain through a durable local event bus. The two projects carry the same versioned runtime module so Stenographer also installs independently.

stenographer_capture_status reports progress and delivery errors. Start a helper with StenographerClient(session_id="example-task") and use ingest, get_brief, compact_guard, search or mark_milestone within that scope.

See current operating contract for migration, task isolation and backups. Older examples need explicit session IDs. Run python -m unittest discover -s tests -p "test_contracts.py" for standalone contract tests. SSWP software evidence uses the same task IDs and event bus.