v12.4.0 — 100% functionality through every install path

[12.4.0] — 2026-05-26 — 100% functional through every install path

npx connect, bash install.sh, docker run, docker compose up — same
4 services, same MCP HTTP API, same dashboard. No path is missing pieces
anymore.

Added — MCP Streamable HTTP transport

• src/server.py now supports MCP_TRANSPORT=http (also accepts
  streamable-http). Boots a uvicorn ASGI app exposing:
  • POST/GET/DELETE /mcp — MCP Streamable HTTP (spec 2025-03-26)
  • GET /healthz — DB-independent liveness probe
• Same Server instance powers both stdio and HTTP, so all 60+ MCP tools
  are reachable over the network — IDEs, sidecars, k8s pods can all
  connect without a stdio bridge.
• stateless=False keeps in-process session multiplexing; clean
  SIGTERM via uvicorn's timeout_graceful_shutdown=5 matches Docker
  stop-grace.

Added — scheduler service (cron-style)

• docker/scheduler_daemon.py — stdlib-only scheduler that runs
  orphan-backfill 4×/day and check-updates weekly inside the same
  image. Replaces the mcuadros/ofelia external container and its
  Docker-socket mount.
• npm wrapper 1.3.0 ships matching scheduler services on macOS
  (StartCalendarInterval) and Linux (OnCalendar systemd timers).

Added — full single-image stack

• tam-entrypoint shim routes docker run … mcp|dashboard|reflection|scheduler|all
  through the supervisor.
• Default TAM_SUPERVISOR_SERVICES=mcp,dashboard,reflection,scheduler
  → docker run -p 3737:3737 -p 37737:37737 -v vol:/data ghcr.io/.../total-agent-memory
  brings the whole product up with one command.
• EXPOSE 3737 37737; HEALTHCHECK probes both /healthz endpoints.
• ML caches pinned to the data volume (HF_HOME, TRANSFORMERS_CACHE,
  TORCHINDUCTOR_CACHE_DIR) — prevents torch._dynamo from crashing on
  containers with tmpfs /tmp < 100 MB.
• Default OLLAMA_URL=http://host.docker.internal:11434 — single-image
  finds host Ollama out of the box on Docker Desktop / Windows / WSL2.
  Graceful degradation to FTS+embeddings if Ollama unreachable.

Added (v12.3 carry-over — never tagged separately)

• docker/tam_supervisor.py — stdlib supervisor (prefix logging,
  exponential-backoff restarts, signal propagation).
• /healthz endpoint on the dashboard.
• Auto-initialise memory.db on first boot inside the container.
• launchagents/com.total-agent-memory.dashboard.plist (installed by
  install.sh on macOS).
• .github/workflows/docker.yml — multi-arch (amd64+arm64) build+push
  to ghcr on every git tag, with /healthz + /api/stats + docker
  healthcheck smoke before publish.
• .github/workflows/smoke.yml — install-matrix smoke on every PR.
• tests/smoke/install-docker-pull.sh, tests/smoke/install-sh-macos.sh.
• Plist placeholders unified to __INSTALL_DIR__ / __MEMORY_DIR__ /
  __HOME__. Old plists hardcoded __HOME__/claude-memory-server/...,
  breaking clones with any other directory name.

Changed

• docker-compose.yml mcp service is back (was removed in v12.3 when
  HTTP transport didn't exist). Same image as dashboard/reflection,
  command: ["mcp"].
• scheduler compose service now runs in the main image
  (command: ["scheduler"]) instead of mcuadros/ofelia — no more
  Docker socket mount in the stack.
• install.sh Step 5 (old dashboard-service.sh install call) removed
  — it created a com.claude-total-memory.dashboard plist that fought
  the canonical com.total-agent-memory.dashboard plist for port 37737.

Fixed

• bash install.sh under any clone name now writes valid plist/unit
  paths instead of references to a non-existent ~/claude-memory-server/.
• bash install.sh --uninstall cleans up the new dashboard plist too.