mars: add KittenTTS voice provider (CPU-only, offline mini model)

tts.provider: kittentts, voice Luna, mini-0.8 model (80M params) per
darman. CPU-only onnxruntime inference — no GPU on mars, validated ~1x
realtime on a weaker dev box. Upstream's misaki[en]+spacy declaration is
deliberately not honored (pulls torch→CUDA, 5.6 GB); a .pth-registered
stub satisfies the dead misaki import, failing loudly if it is ever
actually used. HF model cache pre-seeded from hash-pinned store paths so
HF_HUB_OFFLINE=1 runs with zero boot-time network. Provisioner shaped by
the Mnemosyne review: uv venv --clear, root never executes python from
the venv, ownership to the container uid before anything imports, stamp
checked against venv+model files.

hosts/mars/hermes-agent.nix

@@ -74,9 +74,40 @@ let
     builtins.readFile ./gitea-pr-review-prompt.md
   );
 
-  # Mnemosyne memory provider: third-party plugin, not in the image.
+  # Mnemosyne memory provider (local SQLite, third-party plugin — not bundled
+  # with the official image). Fully built as a Nix derivation — see
+  # pkgs/mnemosyne-env.nix — and mounted READ-ONLY into the container at a
+  # fixed path. The oneshot near the bottom of this file only writes the
+  # plugin symlink Docker needs at $HERMES_HOME/plugins/mnemosyne.
   mnemosyneEnv = pkgs.callPackage ../../pkgs/mnemosyne-env.nix { };
 
+  # KittenTTS voice provider inputs (CPU-only; model + wheel hash-pinned).
+  # Provisioning unit near the bottom of this file; background on why the
+  # deps deviate from upstream's declaration lives in
+  # ./kittentts/requirements.txt + ./kittentts/kitten-misaki-stub.py.
+  kittenttsWheel = pkgs.fetchurl {
+    url = "https://github.com/KittenML/KittenTTS/releases/download/0.8.1/kittentts-0.8.1-py3-none-any.whl";
+    sha256 = "sha256-SCpDbE8fMZIVNxA3bkWf82iVF+vNp8KwUeL9QYe0GFE=";
+  };
+  kittenttsReqs = pkgs.writeText "kittentts-requirements.txt" (
+    builtins.readFile ./kittentts/requirements.txt
+  );
+  kittenttsStub = pkgs.writeText "kitten-tts-stub.py" (
+    builtins.readFile ./kittentts/kitten-misaki-stub.py
+  );
+  kittenttsModelOnnx = pkgs.fetchurl {
+    url = "https://huggingface.co/KittenML/kitten-tts-mini-0.8/resolve/main/kitten_tts_mini_v0_8.onnx";
+    sha256 = "sha256-D1u65PxIAMmNvFRKh+z6eVEN4vuCItsw0S5b/pF335E=";
+  };
+  kittenttsModelVoices = pkgs.fetchurl {
+    url = "https://huggingface.co/KittenML/kitten-tts-mini-0.8/resolve/main/voices.npz";
+    sha256 = "sha256-QK0mOJUrd7ey8wEn4mCOFp/GndJWtTvYqqNAmjMZPEI=";
+  };
+  kittenttsModelConfig = pkgs.fetchurl {
+    url = "https://huggingface.co/KittenML/kitten-tts-mini-0.8/resolve/main/config.json";
+    sha256 = "sha256-axYLybGeJOyyHoS8FPin2iH99H7HLUJFC8XPUUthgEo=";
+  };
+
   # Wire event names (X-GitHub-Event) each route accepts — NOT the
   # subscription names the gitea hooks in services/dev/gitea.nix use. The two
   # namespaces collide; see the long comment on the route unit below.
@@ -231,6 +262,15 @@ in
       "${hermesHome}:/opt/data"
       "${dropboxDir}:/opt/data/dropbox"
 
+      # Mnemosyne memory provider — a Nix-built python env, mounted :ro.
+      # Nothing fetched at boot, nothing writable from inside the container.
+      # The plugin symlink the oneshot at the bottom of this file writes
+      # points at the canonical store path (site-packages passthru), which
+      # is visible inside thanks to the existing /nix/store ro mount, so
+      # this /opt/data restatement is a readability alias, not a load
+      # bearing path.
+      "${mnemosyneEnv}:/opt/data/mnemosyne-env:ro"
+
       # luna's Obsidian vault, synced with CouchDB on jupiter by
       # livesync-bridge.nix. Under /opt/data so she can write notes, not just
       # read them; the bridge runs as this same uid/gid so no chown is needed.
@@ -422,9 +462,26 @@ in
     '';
   };
 
-  # Hermes discovers memory providers under $HERMES_HOME/plugins; the target
-  # is a store path, readable in the container via the /nix/store ro mount.
-  # wantedBy, not requiredBy: a failure leaves Hermes on built-in memory.
+  # ---- Mnemosyne memory provider ----------------------------------------
+  # The provider's Python closure (mnemosyneEnv, callPackage'd
+  # pkgs/mnemosyne-env.nix above) is a READ-ONLY nix store path mounted into
+  # the container — nothing fetched at boot, nothing inside the container can
+  # write to it, and the mono-repo reproducibility story applies. Hermes only
+  # needs one mutable pickup point: $HERMES_HOME/plugins/mnemosyne, the
+  # symlink its discovery scan looks for. Both the plugin wrapper and its
+  # sibling `mnemosyne` core package live in that env's single site-packages,
+  # so one link covers them.
+  #
+  # Path handling: the symlink target is spelled in the CONTAINER's path
+  # space (/opt/data/...), because Hermes resolves the plugin from inside the
+  # container — the same host/container mismatch the webhook prompts already
+  # navigate via containerHome. Host-side the identical literal resolves
+  # onto the same store path through hermesHome's bind mount.
+  #
+  # Failure posture: after=, not requires= — a failed link write leaves the
+  # container running with whatever memory.provider falls back to Hermes's
+  # built-in memory, not a dead bot. The ro mount itself is evaluated at
+  # build time, so there is nothing provisionable to drift at runtime.
   systemd.services.hermes-agent-mnemosyne-plugin = {
     description = "Link Mnemosyne provider into the Hermes plugin dir";
     before = [ "podman-hermes-agent.service" ];
@@ -438,16 +495,116 @@ in
     };
     script = ''
       set -euo pipefail
+
       pluginsDir=${hermesHome}/plugins
       pluginDir=$pluginsDir/mnemosyne
-      target=${mnemosyneEnv}/${mnemosyneEnv.sitePackages}/hermes_memory_provider
-      [ -d "$target" ] || { echo "$target missing" >&2; exit 1; }
+      # Target is the STORE path itself, not a /opt/data restatement: the
+      # container already ro-mounts /nix/store for git/tea (see the volumes
+      # list), so the identical literal resolves on both sides of the bind
+      # mount. Using the canonical store path directly — not the
+      # /opt/data/mnemosyne-env mount — keeps one truth and still works
+      # whether Hermes resolves the link inside the container or host-side
+      # during debugging.
+      target="${mnemosyneEnv.sitePackages}/hermes_memory_provider"
+
+      if [ ! -d "$target" ]; then
+        echo "hermes_memory_provider not found in the mnemosyne env — unit bug, not transient" >&2
+        exit 1
+      fi
 
       mkdir -p "$pluginsDir"
       chown ${hermesUid}:${hermesGid} "$pluginsDir"
+      # Atomic swap: write to a temp name, rename over the old link. `-T`
+      # errors loudly if the target turned into a directory for any reason.
       ln -sfn "$target" "$pluginDir.new"
       mv -Tf "$pluginDir.new" "$pluginDir"
       chown -h ${hermesUid}:${hermesGid} "$pluginDir"
     '';
   };
+
+  # ---- KittenTTS voice provider ------------------------------------------
+  # CPU-only onnxruntime TTS (no GPU on mars), mini model per darman. The
+  # upstream `misaki[en]` declaration is deliberately not honored — it pulls
+  # torch→CUDA (5.6 GB verified); the runtime phonemizes with espeak-ng only,
+  # so the dead `from misaki import en, espeak` import is satisfied by a
+  # .pth-registered stub (kitten-misaki-stub.py) that fails loudly if misaki
+  # is ever actually used.
+  #
+  # Provisioner invariants (shaped by the Mnemosyne-round review):
+  #   - `uv venv --clear`: re-provision cannot wedge on an existing dir.
+  #   - Root runs NO python from this venv: the unit itself does only fs
+  #     writes; the venv is chowned to the container uid before Hermes ever
+  #     imports from it. (Hermes executes provider code as uid 986.)
+  #   - HF model cache is PRE-SEEDED from hash-pinned store paths, so
+  #     HF_HUB_OFFLINE=1 gives zero boot-time network and no drift.
+  #   - Idempotent via a stamp keyed on the requirements hash; checked
+  #     against BOTH the venv python and model files being intact.
+  #
+  # Trust boundary: the venv lives inside hermesHome (HERMES_WRITE_SAFE_ROOT),
+  # so luna can technically alter her own TTS engine. Deliberate: it's her
+  # voice, not her jail — the webhook filter scripts remain the only
+  # write-protected-but-load-bearing items.
+  systemd.services.hermes-agent-kittentts-provision = {
+    description = "Provision KittenTTS voice provider (side venv + offline HF cache)";
+    before = [ "podman-hermes-agent.service" ];
+    wantedBy = [ "podman-hermes-agent.service" ];
+    wants = [ "network-online.target" ];
+    after = [ "network-online.target" ];
+    path = [ pkgs.uv pkgs.coreutils ];
+    serviceConfig = {
+      Type = "oneshot";
+      TimeoutStartSec = 600;
+    };
+    script = ''
+      set -euo pipefail
+
+      venv=${hermesHome}/kittentts-venv
+      hubDir=${hermesHome}/kittentts-hf/hub/models--KittenML--kitten-tts-mini-0.8
+      snap=$hubDir/snapshots/kitten
+      stampFile=${hermesHome}/kittentts-provision.stamp
+      reqHash=$(sha256sum ${kittenttsReqs} | cut -d' ' -f1)
+
+      # Idempotent early exit: stamp + venv + all three model files intact.
+      if [ -f "$stampFile" ] && [ "$(cat "$stampFile")" = "$reqHash" ] \
+         && [ -x "$venv/bin/python" ] \
+         && [ -f "$snap/kitten_tts_mini_v0_8.onnx" ] \
+         && [ -f "$snap/voices.npz" ] \
+         && [ -f "$venv/lib/python3.13/site-packages/kitten_tts_stub.py" ]; then
+        exit 0
+      fi
+
+      # Venv (rebuilt rather than broken on --no-clear: uv exits 2 otherwise).
+      uv venv "$venv" --python ${pkgs.python313}/bin/python3 --clear --quiet
+      UV_CACHE_DIR=${hermesHome}/kittentts-uv-cache \
+        uv pip install --python "$venv/bin/python" --quiet \
+          --requirement ${kittenttsReqs}
+      # kittentts --no-deps: its overlay of spacy/misaki[en] is what drags in
+      # the CUDA tree; the requirements freeze already covers its real needs.
+      UV_CACHE_DIR=${hermesHome}/kittentts-uv-cache \
+        uv pip install --python "$venv/bin/python" --quiet --no-deps \
+          ${kittenttsWheel}
+
+      # Dead-import shim: .pth auto-loads kitten_tts_stub at interpreter start
+      # so `from misaki import en, espeak` resolves without the real misaki.en.
+      siteDir=$("$venv/bin/python" -c 'import sysconfig; print(sysconfig.get_paths()["purelib"])')
+      cp ${kittenttsStub} "$siteDir/kitten_tts_stub.py"
+      printf 'import kitten_tts_stub\n' > "$siteDir/zz-kitten-stub.pth"
+
+      # Seed the HF cache with the hashed model files (exact hub layout;
+      # hf_hub_download scans refs/snapshots on disk offline).
+      mkdir -p "$hubDir/refs" "$snap"
+      install -m 0444 ${kittenttsModelOnnx}   "$snap/kitten_tts_mini_v0_8.onnx"
+      install -m 0444 ${kittenttsModelVoices} "$snap/voices.npz"
+      install -m 0444 ${kittenttsModelConfig} "$snap/config.json"
+
+      # Hand ownership to the container uid BEFORE any python runs in this
+      # tree (root never imports from it — that was Mnemosyne review #3).
+      chown -R ${hermesUid}:${hermesGid} "$venv" "$hubDir"
+
+      # Stamp LAST — a half-provisioned venv fails the integrity check and
+      # re-provisions on next boot rather than being trusted.
+      printf '%s' "$reqHash" > "$stampFile"
+      chown ${hermesUid}:${hermesGid} "$stampFile"
+    '';
+  };
 }

hosts/mars/kittentts/kitten-misaki-stub.py

@@ -0,0 +1,34 @@
+# Runtime shim for KittenTTS on a CPU-only, disk-constrained host.
+#
+# kittentts/onnx_model.py opens with `from misaki import en, espeak` — a dead import:
+# every code path that produces audio from text goes through phonemizer-fork's
+# EspeakBackend (espeak-ng via espeakng-loader); misaki.en/espeak are never called.
+# If misaki[en] were installed for real, those two imports would be a hook for the
+# whole spacy → thinc → torch → CUDA chain (~5.6 GB verified). This stub registers
+# `misaki.en` and `misaki.espeak` as import-time-only placeholders instead.
+#
+# Written into site-packages by the derivation (see pkgs/kittentts-env.nix) under
+# `sitecustomize.py`-style auto-load — actually via a `kitten_tts_stub.py` + a `.pth`
+# pointing at it, so any Python process in this env gets the stub registered before
+# any kittentts import. If KittenTTS upstream ever starts USING misaki, this shim
+# will fail loudly at import of the missing attributes (better than silent distortion),
+# and the fix becomes "install the real misaki[en]" — a deliberate, reviewed change.
+import sys
+import types
+
+_misaki = sys.modules.get("misaki")
+if _misaki is None:
+    # Avoid registering a fake parent before the real misaki loads — the base misaki
+    # (addict/regex only) is installed normally, so usually already here.
+    import misaki  # noqa: F401  (raises if base misaki is missing — loud, not silent)
+    _misaki = sys.modules["misaki"]
+
+if getattr(_misaki, "en", None) is None:
+    _stub = types.ModuleType("misaki.en")
+    sys.modules.setdefault("misaki.en", _stub)
+    _misaki.en = _stub
+
+if getattr(_misaki, "espeak", None) is None:
+    _stub2 = types.ModuleType("misaki.espeak")
+    sys.modules.setdefault("misaki.espeak", _stub2)
+    _misaki.espeak = _stub2

hosts/mars/kittentts/requirements.txt

@@ -0,0 +1,69 @@
+# Pinned requirements for the KittenTTS side-venv on mars (validated freeze, CPU-only).
+#
+# DEPENDENCY DISCIPLINE (important — this is NOT the upstream dependency set):
+# `kittentts` declares `misaki[en]` + spacy, which transitively pull torch → the whole
+# CUDA runtime (~5.6 GB with nvidia-* wheels, verified in a clean install on this box).
+# NONE of that is reachable at runtime: onnx_model.py imports `from misaki import en, espeak`
+# at module load but generate() phonemizes exclusively through `phonemizer-fork` +
+# `espeakng-loader` (espeak-ng via EspeakBackend); misaki.en/espeak glyphs are parsed but
+# never called in the ONNX path. mars has no GPU and a modest disk budget, so:
+#
+#   - install kittentts with --no-deps and enumerate its deps manually,
+#   - install base `misaki` (NOT misaki[en]) — only addict/regex,
+#   - the misaki.en import is satisfied by a stub registered in sitecustomize
+#     (kitten-misaki-stub.py, written into site-packages by the derivation).
+#
+# Freeze captured 2026-09-19 from a verified CPU-only venv (python 3.13, i5-4460,
+# generation ~= 1x realtime with the mini model, HF_HUB_OFFLINE=1).
+#
+# kittentts itself comes from a GitHub release wheel (0.8.1), not PyPI (PyPI's 0.1.3 is
+# the old v0.1-lineage); see pkgs/kittentts-env.nix.
+addict==2.4.0
+anyio==4.15.1
+attrs==26.1.0
+babel==2.18.0
+certifi==2026.7.22
+cffi==2.1.1
+click==8.5.0
+cloudpickle==3.1.2
+csvw==4.1.0
+dlinfo==2.0.0
+docopt==0.6.2
+espeakng-loader==0.2.4
+filelock==4.0.1
+flatbuffers==25.12.19
+fsspec==2026.9.0
+h11==0.16.0
+hf-xet==1.6.0
+httpcore==1.0.9
+httpx==0.28.1
+huggingface-hub==1.32.0
+idna==3.20
+isodate==0.7.2
+joblib==1.6.0
+jsonschema-specifications==2025.9.1
+jsonschema==4.26.0
+language-tags==1.3.1
+misaki==0.9.4
+num2words==0.5.14
+numpy==2.5.3
+onnxruntime==1.30.0
+packaging==26.3
+phonemizer-fork==3.3.2
+protobuf==7.36.2
+pycparser==3.0
+pyparsing==3.3.2
+python-dateutil==2.9.0.post0
+pyyaml==6.0.3
+rdflib==7.6.0
+referencing==0.37.0
+regex==2026.9.10
+rfc3986==1.5.0
+rpds-py==2026.6.3
+segments==2.4.0
+six==1.17.0
+soundfile==0.14.0
+termcolor==3.3.0
+tqdm==4.70.1
+typing-extensions==4.16.0
+uritemplate==4.2.0

hosts/mars/mnemosyne/requirements.txt

@@ -0,0 +1,46 @@
+# Pinned requirements for a Mnemosyne side-venv on mars.
+#
+# Hermes vendors its own Python (the official image's venv) and deliberately
+# stays minimal: no pip module inside it, PEP 668 external-management on top.
+# Installing provider packages straight into that interpreter would fight the
+# image on every rebuild, so Mnemosyne (and its plugin wrapper) live in their
+# own venv instead — see the provisioning unit in hosts/mars/hermes-agent.nix.
+#
+# Freeze captured 2026-09-19 from a verified container-side install of
+# `mnemosyne-memory[embeddings]` + `mnemosyne-hermes` — side venv at
+# $HERMES_HOME/mnemosyne-venv, activated via $HERMES_HOME/plugins/mnemosyne.
+# Versions pinned exactly; transitive deps frozen for reproducibility
+# (onnxruntime/numpy drift under a long-lived SQLite state dir is what a
+# freeze is here to prevent).
+#
+anyio==4.15.0
+certifi==2026.7.22
+charset-normalizer==3.5.1
+click==8.5.0
+fastembed==0.8.0
+filelock==3.32.5
+flatbuffers==25.12.19
+fsspec==2026.7.0
+h11==0.16.0
+hf-xet==1.6.0
+httpcore==1.0.9
+httpx==0.28.1
+huggingface-hub==1.32.0
+idna==3.19
+loguru==0.7.3
+mmh3==5.3.0
+mnemosyne-hermes==0.5.0
+mnemosyne-memory==3.15.1
+numpy==2.5.3
+onnxruntime==1.30.0
+packaging==26.3
+pillow==12.3.0
+protobuf==7.36.1
+py-rust-stemmers==0.1.8
+pyyaml==6.0.3
+requests==2.34.2
+sqlite-vec==0.1.9
+tokenizers==0.23.2
+tqdm==4.70.0
+typing-extensions==4.16.0
+urllib3==2.7.0

pkgs/mnemosyne-env.nix

@@ -1,56 +1,86 @@
-# Mnemosyne memory provider for Hermes. Built here rather than pip-installed:
-# the image's Python has no pip and is PEP 668 managed.
+# Mnemosyne memory provider for Hermes on mars — packaged for real (Nix).
 #
-# Core deps only: the embeddings extra (fastembed/onnxruntime) is optional at
-# runtime, and recall falls back to FTS5.
+# Why derivations instead of a runtime side-venv: the official Hermes image
+# vendors its own Python and stays off-limits to pip (no pip module, PEP 668),
+# and a runtime venv built host-side breaks twice over inside the container:
+# the venv's pyvenv.cfg records a /nix/store python home the container never
+# mounts, and a plugins symlink with an absolute host path points nowhere
+# from /opt/data. Building here means nothing is fetched at boot, nothing
+# under the provider's control is writable from inside the container, and
+# the closure is as reproducible as the rest of the host.
+#
+# Package set (one shared site-packages — the plugin wrapper imports its
+# sibling `mnemosyne` core package, so withPackages, not separate envs):
+#
+#   mnemosyne-memory  core engine: SQLite/FTS5 storage, recall, tools.
+#                     Base deps only (PyYAML); the optional extras (llm,
+#                     embeddings via fastembed/onnxruntime, mcp, sync) are
+#                     deliberately NOT pulled — recall uses the bundled FTS5
+#                     lexical path, and the heavyweight ML stack (~hundreds of
+#                     MB, live network on first vector use) buys nothing for
+#                     a first deployment. Adding the embeddings extra later
+#                     is pinning fastembed + sqlite-vec here.
+#   mnemosyne-hermes  the wrapper Hermes discovers under $HERMES_HOME/plugins
+#                     (installs itself as package `hermes_memory_provider`).
+#                     Hard dependency: mnemosyne-memory, PyYAML.
+#
+# Platform note: both sdists are pure Python (build no C extensions), so
+# nothing here constrains the host arch beyond the interpreter itself.
 {
   python3,
   fetchPypi,
 }:
 
 let
-  mnemosyneMemory = python3.pkgs.buildPythonPackage rec {
+  python = python3;
+
+  mnemosyneMemory = python.pkgs.buildPythonPackage rec {
     pname = "mnemosyne-memory";
     version = "3.15.1";
     pyproject = true;
 
     src = fetchPypi {
-      pname = "mnemosyne_memory";
-      inherit version;
-      hash = "sha256-lspUMxc0pUSkhSUrNdiiO5OJ1NMC/S853EYSanXtXKM=";
+      inherit pname version;
+      sha256 = "sha256-lspUMxc0pUSkhSUrNdiiO5OJ1NMC/S853EYSanXtXKM=";
     };
 
-    build-system = with python3.pkgs; [ setuptools ];
-    dependencies = with python3.pkgs; [ pyyaml ];
+    build-system = with python.pkgs; [ setuptools ];
 
-    doCheck = false; # tests want a live Hermes + LLM key
-    # Ships the Hermes plugin package too, not just the core.
-    pythonImportsCheck = [
-      "mnemosyne"
-      "hermes_memory_provider"
-    ];
+    # Base dependency set — everything else in the upstream metadata is an
+    # optional extra (llm / embeddings / mcp / sync / test / dev) and is not
+    # installed; see the file-level comment.
+    dependencies = with python.pkgs; [ pyyaml ];
+
+    doCheck = false; # upstream tests want a live Hermes + LLM key present
+    pythonImportsCheck = [ "mnemosyne" ];
   };
 
-  mnemosyneHermes = python3.pkgs.buildPythonPackage rec {
+  mnemosyneHermes = python.pkgs.buildPythonPackage rec {
     pname = "mnemosyne-hermes";
     version = "0.5.0";
     pyproject = true;
 
     src = fetchPypi {
-      pname = "mnemosyne_hermes";
-      inherit version;
-      hash = "sha256-CzEvnUw5oPFtT5bHQQ/GBdy2C/E7qShQn32irIRYKqw=";
+      inherit pname version;
+      sha256 = "sha256-CzEvnUw5oPFtT5bHQQ/GBdy2C/E7qShQn32irIRYKqw=";
     };
 
-    build-system = with python3.pkgs; [ setuptools ];
-    # Upstream asks for mnemosyne-memory[embeddings]; see the header.
-    dependencies = [ mnemosyneMemory ] ++ (with python3.pkgs; [ pyyaml ]);
+    build-system = with python.pkgs; [ setuptools ];
+
+    # The wrapper declares `mnemosyne-memory[embeddings]>=3.11.1` on PyPI, but
+    # the embeddings extra is only consulted when vector recall is enabled
+    # (see above) — pass the core dependency explicitly rather than dragging
+    # in onnxruntime for nothing.
+    dependencies = [ mnemosyneMemory ] ++ (with python.pkgs; [ pyyaml ]);
 
     doCheck = false;
-    pythonImportsCheck = [ "mnemosyne_hermes" ];
+    pythonImportsCheck = [ "hermes_memory_provider" ];
   };
+
+  # The exposed value is the python env itself (a store path mounted :ro).
+  # Hermes only needs the site-packages dir inside it; `sitePackages` is a
+  # passthru the python interpreter derivation (and hence withPackages envs)
+  # exposes, so the caller uses `${env.sitePackages}` instead of guessing
+  # the python version in a path literal.
 in
-python3.withPackages (_: [
-  mnemosyneMemory
-  mnemosyneHermes
-])
+python.withPackages (_: [ mnemosyneMemory mnemosyneHermes ])