Skip to content

Latest commit

 

History

History
60 lines (47 loc) · 2.49 KB

File metadata and controls

60 lines (47 loc) · 2.49 KB

Changelog (curated highlights)

Full changelog lives in the private repo. This is a curated view showing the build trajectory -- grouped by feature area, not by date.

MCP Server

  • 10 tools exposed via FastMCP (stdio + streamable-HTTP transport)
  • Bearer auth middleware wrapping FastMCP's ASGI app for remote access
  • DNS-rebinding protection via ROTATOR_MCP_ALLOWED_HOSTS
  • Deployed as rotator-mcp.service (systemd) on VPS, Tailscale-fronted
  • Multi-host wiring: Windows, VPS shell, iPhone-via-SSH -- all share the same MCP endpoint

Voice Co-pilot

  • Phase A -- rotator call --device N: mic capture via sounddevice, Deepgram nova-3 streaming STT (ru+en), TranscriptBuffer with rolling state + markdown persistence to ~/.rotator/calls/<session>.md
  • Phase B -- call_consult MCP tool: dispatches every final utterance to Claude Code subprocess with a 5-layer composable system prompt. Textual TUI gains bubble panel (yellow border) for structured responses. ConversationState with rolling window of 30 turns.
  • Shared brain dispatch (call/consult.py) eliminates code duplication between local orchestrator and MCP tool paths

Telemetry

  • M1 -- @with_telemetry decorator wraps all 10 MCP tools. Records tool name, latency_ms, args_hash, timestamp to SQLite. Fail-soft: never lets observability break the system being observed.
  • M2 -- rotator_self_report MCP tool: Claude self-reports per-tool verdicts (helpful/neutral/waste) at session end. waste signal is the primary feedback for prompt improvement.
  • rotator usefulness today CLI: per-tool call counts, avg/p95 latency, self-report verdicts

Memory + Chronicles

  • MNEMO client with TTL cache, async httpx, 9 evidence URI schemes
  • Hierarchical chronicle builder (daily/weekly/monthly) -- reads events from MNEMO, synthesizes via LLM, posts structured chronicle back
  • Session notes: append-only markdown per conversation

Foundation (MVP week)

  • Architecture docs, SPEC, Q1-Q10 decisions, IMPLEMENTATION_PLAN
  • 6-layer prompt engine (legacy orchestrator path, now deprecated)
  • Bounded loop with citation guards and PII redaction
  • Skill extraction + evaluation pipeline with MNEMO sync
  • YAML fixture-based eval runner
  • Python package skeleton (hatchling, ruff, pytest, mypy)
  • GitHub Actions CI (ruff + pytest + no-openai grep)

Numbers at ship

  • 262 tests passing, ruff clean
  • ~3,300 LOC across 76 Python modules
  • 14,000+ MNEMO events
  • Built in 5 days (MVP) + 3 days (voice + telemetry)