abstractcore 2.32.0__tar.gz → 2.34.0__tar.gz

{abstractcore-2.32.0 → abstractcore-2.34.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: abstractcore
-Version: 2.32.0
+Version: 2.34.0
 Summary: Unified interface to all LLM providers with essential infrastructure for tool calling, streaming, and model management
 Author-email: Laurent-Philippe Albou <contact@abstractcore.ai>
 Maintainer-email: Laurent-Philippe Albou <contact@abstractcore.ai>
@@ -49,7 +49,7 @@ Requires-Dist: uvicorn[standard]<1.0.0,>=0.23.0
 Requires-Dist: python-multipart<1.0.0,>=0.0.6
 Requires-Dist: sse-starlette<4,>=1.8
 Requires-Dist: cryptography>=41.0.0
-Requires-Dist: abstractvoice>=0.19.5
+Requires-Dist: abstractvoice>=0.20.0
 Requires-Dist: abstractvision>=0.3.34
 Requires-Dist: abstractmusic>=0.1.17; python_version >= "3.10"
 Requires-Dist: abstract3d>=0.3.2; python_version >= "3.10"
@@ -69,7 +69,7 @@ Requires-Dist: accelerate>=1.0.0; extra == "apple"
 Requires-Dist: sentence-transformers<6.0.0,>=5.1.0; extra == "apple"
 Requires-Dist: numpy<3.0.0,>=2.1.0; extra == "apple"
 Requires-Dist: Pillow<13.0.0,>=12.1.1; extra == "apple"
-Requires-Dist: abstractvoice[all-apple]>=0.19.5; extra == "apple"
+Requires-Dist: abstractvoice[all-apple]>=0.20.0; extra == "apple"
 Requires-Dist: omnivoice>=0.1.5; extra == "apple"
 Requires-Dist: abstractvision[all-apple]>=0.3.34; extra == "apple"
 Requires-Dist: abstractmusic[all-apple]>=0.1.17; extra == "apple"
@@ -87,7 +87,7 @@ Requires-Dist: gguf<1.0.0,>=0.10.0; extra == "gpu"
 Requires-Dist: sentence-transformers<6.0.0,>=5.1.0; extra == "gpu"
 Requires-Dist: numpy<3.0.0,>=1.20.0; python_version < "3.13" and extra == "gpu"
 Requires-Dist: numpy<3.0.0,>=2.1.0; python_version >= "3.13" and extra == "gpu"
-Requires-Dist: abstractvoice[all-gpu]>=0.19.5; extra == "gpu"
+Requires-Dist: abstractvoice[all-gpu]>=0.20.0; extra == "gpu"
 Requires-Dist: omnivoice>=0.1.5; extra == "gpu"
 Requires-Dist: abstractvision[all-gpu]>=0.3.34; extra == "gpu"
 Requires-Dist: abstractmusic[all-gpu]>=0.1.17; extra == "gpu"
@@ -196,17 +196,19 @@ Dynamic: license-file
 Unified LLM Interface
 > Write once, run everywhere
 
-What's new in 2.32.0: decision models, text in and calibrated typed decisions out
-(`generate(document, output="decision", questions=...)`, `POST /v1/decisions`) over hosted Jev, any
-System One-compatible endpoint and local Kev / Lev / Laya checkpoints; one llama.cpp weights instance
-per GGUF with a per-model slot pool for concurrent requests; the reserved context computed from free
-memory and settable per model; and every cached asset (adapters included) listed in the catalogue.
-See the [changelog](CHANGELOG.md#2320---2026-11-02).
-
-What's new in 2.31.1: `/v1/chat/completions` reuses a model already resident in the process (a host's
-runtime pool) instead of loading a second llama.cpp copy per request; the download catalog lists the
-LFM2 and LFM2.5 families (GGUF and MLX, the 24B-A2B models included). See the
-[changelog](CHANGELOG.md#2311---2026-10-28).
+What's new in 2.34.0: a local model that no call has used for 60 minutes is unloaded from memory and loads
+again on its next call; a model kept loaded (Keep loaded) never is, and no running call is interrupted
+(`process_residency.set_idle_unload(seconds)`, `0` turns it off). A provider key you send to the server's discovery
+routes now reaches only the provider you name, never every provider those routes check. The web console's themes
+match ui-kit 0.9.14. See the [changelog](CHANGELOG.md#2340---2026-11-03).
+
+What's new in 2.33.0: GGUF prompt-cache snapshots drop an unused 0.5 GB copy each, stay under a byte budget
+(`gguf_prompt_cache_max_gb`), and are never evicted under a prefill that is reading them; a model that is not
+downloaded raises `ModelNotDownloadedError` naming it; an OpenAI-compatible key check refuses a wrong key even where
+the model list is public (OVH); hosted decision models (`typesafe`, `systemone`) are connection families, listed and
+routed as decision models, and a local decision model answers its own decisions; a provider profile this version
+cannot read is kept and reported instead of resetting the whole config. See the
+[changelog](CHANGELOG.md#2330---2026-11-03).
 
 AbstractCore is an offline-capable, open-source-first LLM infrastructure layer
 for Python applications. It gives you one `create_llm(...)` API across local

{abstractcore-2.32.0 → abstractcore-2.34.0}/README.md

@@ -9,17 +9,19 @@
 Unified LLM Interface
 > Write once, run everywhere
 
-What's new in 2.32.0: decision models, text in and calibrated typed decisions out
-(`generate(document, output="decision", questions=...)`, `POST /v1/decisions`) over hosted Jev, any
-System One-compatible endpoint and local Kev / Lev / Laya checkpoints; one llama.cpp weights instance
-per GGUF with a per-model slot pool for concurrent requests; the reserved context computed from free
-memory and settable per model; and every cached asset (adapters included) listed in the catalogue.
-See the [changelog](CHANGELOG.md#2320---2026-11-02).
-
-What's new in 2.31.1: `/v1/chat/completions` reuses a model already resident in the process (a host's
-runtime pool) instead of loading a second llama.cpp copy per request; the download catalog lists the
-LFM2 and LFM2.5 families (GGUF and MLX, the 24B-A2B models included). See the
-[changelog](CHANGELOG.md#2311---2026-10-28).
+What's new in 2.34.0: a local model that no call has used for 60 minutes is unloaded from memory and loads
+again on its next call; a model kept loaded (Keep loaded) never is, and no running call is interrupted
+(`process_residency.set_idle_unload(seconds)`, `0` turns it off). A provider key you send to the server's discovery
+routes now reaches only the provider you name, never every provider those routes check. The web console's themes
+match ui-kit 0.9.14. See the [changelog](CHANGELOG.md#2340---2026-11-03).
+
+What's new in 2.33.0: GGUF prompt-cache snapshots drop an unused 0.5 GB copy each, stay under a byte budget
+(`gguf_prompt_cache_max_gb`), and are never evicted under a prefill that is reading them; a model that is not
+downloaded raises `ModelNotDownloadedError` naming it; an OpenAI-compatible key check refuses a wrong key even where
+the model list is public (OVH); hosted decision models (`typesafe`, `systemone`) are connection families, listed and
+routed as decision models, and a local decision model answers its own decisions; a provider profile this version
+cannot read is kept and reported instead of resetting the whole config. See the
+[changelog](CHANGELOG.md#2330---2026-11-03).
 
 AbstractCore is an offline-capable, open-source-first LLM infrastructure layer
 for Python applications. It gives you one `create_llm(...)` API across local

{abstractcore-2.32.0 → abstractcore-2.34.0}/abstractcore/config/manager.py

@@ -36,6 +36,7 @@ from .provider_profiles import (
     normalize_provider_family,
     normalize_string_list,
     profile_id_from_virtual_provider,
+    provider_profile_from_dict,
     provider_profiles_from_dict,
     split_api_key_value,
 )
@@ -1289,8 +1290,10 @@ class ConfigurationManager:
           - `capability_defaults` keys other than version/routes/seeded, and
             route keys this version cannot parse (a future modality); a route
             row's unknown fields already survive as `options`;
-          - `provider_profiles` keys other than `profiles`, and unknown columns
-            of a profile row that is still present (a deleted row stays deleted).
+          - `provider_profiles` keys other than `profiles`, profile rows this
+            version cannot load (a provider family it does not know), and
+            unknown columns of a profile row that is still present (a deleted
+            row stays deleted).
 
         Known keys are never carried: a known row or field absent from `mine`
         is a deliberate delete (`clear-default`, `delete-provider`) or a value
@@ -1345,6 +1348,12 @@ class ConfigurationManager:
             if isinstance(mine_rows, dict):
                 known_columns = {f.name for f in fields(ProviderProfile)}
                 for profile_id, raw_row in raw_profiles["profiles"].items():
+                    if profile_id not in mine_rows:
+                        try:
+                            provider_profile_from_dict(str(profile_id), raw_row)
+                        except (ValueError, TypeError):
+                            mine_rows[profile_id] = copy.deepcopy(raw_row)  # not loaded: kept as stored
+                        continue
                     mine_row = mine_rows.get(profile_id)
                     if not isinstance(raw_row, dict) or not isinstance(mine_row, dict):
                         continue

{abstractcore-2.32.0 → abstractcore-2.34.0}/abstractcore/config/provider_profiles.py

@@ -276,6 +276,20 @@ def provider_profiles_from_dict(data: Any) -> ProviderProfilesConfig:
 
     profiles: Dict[str, ProviderProfile] = {}
     for raw_id, raw_profile in rows.items():
-        profile = provider_profile_from_dict(str(raw_id), raw_profile)
+        try:
+            profile = provider_profile_from_dict(str(raw_id), raw_profile)
+        except (ValueError, TypeError) as exc:
+            # A row this Core cannot load (a provider family it does not know, written by a
+            # newer Core, 2026-11-03: `typesafe` read by 2.31) is skipped and said, never fatal:
+            # raising here made the whole config read as corrupt (a `.corrupt-*.bak` copy and
+            # DEFAULTS for the session). The save keeps the row as stored
+            # (`ConfigurationManager._carry_unknown_store_keys`).
+            import logging
+
+            logging.getLogger(__name__).warning(
+                "provider profile %r is not loaded by this AbstractCore (%s); it stays in the config file unchanged",
+                str(raw_id), exc,
+            )
+            continue
         profiles[profile.id.lower()] = profile
     return ProviderProfilesConfig(profiles=profiles)

{abstractcore-2.32.0 → abstractcore-2.34.0}/abstractcore/config/route_engines.py

@@ -63,7 +63,7 @@ __all__ = [
 # AbstractCore's AbstractVoice floor (it has `abstractvoice.engine_runtime`, the
 # public runtime probe, since 0.15.0). Mirrors the abstractvoice pins in
 # pyproject.toml; the release stager raises both together.
-ABSTRACTVOICE_ENGINE_RUNTIME_FLOOR = "0.19.5"
+ABSTRACTVOICE_ENGINE_RUNTIME_FLOOR = "0.20.0"
 
 # Voice providers that run remotely: AbstractVoice needs nothing beyond its
 # core install for them, and whether they are configured is not a runtime

{abstractcore-2.32.0 → abstractcore-2.34.0}/abstractcore/console/theme_sync.py

@@ -26,6 +26,24 @@ from typing import Any, Dict, List, Optional, Sequence, Tuple
 
 GENERATED_MODULE = "themes.py"
 
+#: The ui-kit release the console's theme copy must come from at least (backlog 0869).
+#: `themes.py` records the kit version it was generated from; a test compares the two
+#: without the abstractuic checkout, so a stale copy fails CI everywhere. Raise this
+#: with every re-sync.
+KIT_THEME_FLOOR = "0.9.14"
+
+
+def kit_version(kit_src: Path) -> str:
+    """The ui-kit version beside ``ui-kit/src`` (``ui-kit/package.json``); refuses to guess."""
+    package = kit_src.parent / "package.json"
+    try:
+        version = str(json.loads(package.read_text(encoding="utf-8"))["version"]).strip()
+    except (OSError, ValueError, KeyError) as e:
+        raise ValueError(f"cannot read the ui-kit version from {package}: {e}") from e
+    if not version:
+        raise ValueError(f"{package} has an empty version")
+    return version
+
 _SPEC_RE = re.compile(
     r"\{\s*id:\s*\"(?P<id>[a-z0-9-]+)\",\s*label:\s*\"(?P<label>[^\"]+)\",\s*"
     r"group:\s*\"(?P<group>dark|light)\",\s*swatches:\s*\[(?P<swatches>[^\]]*)\]",
@@ -102,14 +120,19 @@ def parse_theme_css_blocks(theme_css_text: str) -> List[Tuple[str, str]]:
     return blocks
 
 
+# `:root` alone or with attribute qualifiers (`:root[data-density="dense"]`): a token rule.
+# `:root .af-session-card` (a descendant) or `:root.theme-x` is a component or theme rule.
+_TOKEN_ROOT_RE = re.compile(r":root(\[[^\]]*\])*")
+
+
 def _is_token_block(selector: str, body: str) -> bool:
-    """A responsive-token block: ``html {...}``, a non-theme ``:root...{...}``,
+    """A responsive-token block: ``html {...}``, a bare ``:root`` (or ``:root[attr]``) ``{...}``,
     or an ``@media``/``@supports`` block holding only such ``:root`` rules."""
-    if selector == "html" or (selector.startswith(":root") and ".theme-" not in selector):
+    if selector == "html" or _TOKEN_ROOT_RE.fullmatch(selector):
         return "{" not in body
     if selector.startswith(("@media", "@supports")):
         inner = [sel.strip() for sel in re.findall(r"([^{}]+)\{", body)]
-        return bool(inner) and all(sel.startswith(":root") and ".theme-" not in sel for sel in inner)
+        return bool(inner) and all(_TOKEN_ROOT_RE.fullmatch(sel) for sel in inner)
     return False
 
 
@@ -240,6 +263,7 @@ def generate_themes_module(kit_src: Path) -> str:
     light_ids = [s["id"] for s in specs if s["group"] == "light"]
     css_text = "\n\n".join(block for _, block in blocks)
     provenance = {
+        "kit_version": kit_version(kit_src),
         "theme_ts_sha256": hashlib.sha256(theme_ts.encode("utf-8")).hexdigest(),
         "theme_css_sha256": hashlib.sha256(theme_css.encode("utf-8")).hexdigest(),
     }

{abstractcore-2.32.0 → abstractcore-2.34.0}/abstractcore/console/themes.py

@@ -289,8 +289,9 @@ KIT_THEME_SPECS = [
 KIT_LIGHT_THEME_IDS = ["catppuccin-latte", "rose-pine-dawn", "one-light", "everforest-light", "solarized-light", "light"]
 
 KIT_SOURCE_PROVENANCE = {
+    "kit_version": "0.9.14",
     "theme_ts_sha256": "5ae99c5b4c018983702cdbef688428ebe36972136d20d9269fdd5cf4d5ec48cc",
-    "theme_css_sha256": "2f1d216d5c21126972058dd60686a9a098623bd9b59188cc95cbcee5cdcd8444"
+    "theme_css_sha256": "83fe3a16b95177d39f7d5394c1d62fe127960b843d53c5deae143fa86a5745f7"
 }
 
 KIT_ROOT_CSS = ':root {\n  color-scheme: dark;\n\n  /* Colors */\n  --bg-primary: #1a1a2e;\n  --bg-secondary: #16213e;\n  --bg-tertiary: #0f3460;\n  --text-primary: #eee;\n  --text-secondary: #aaa;\n  --text-muted: #666;\n  --accent: #e94560;\n  --success: #27ae60;\n  --warning: #f39c12;\n  --error: #e74c3c;\n  --info: #60a5fa;\n\n  /* Fonts (no remote imports; prefer system fonts) */\n  --font-mono: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, "Liberation Mono", "Courier New", monospace;\n  --font-sans: system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, sans-serif;\n\n  /* Typography / density (controlled by thin-client Settings; avoid per-app drift) */\n  --font-scale: 1;\n  --header-density: 1;\n\n  --font-size-xxs: calc(10px * var(--font-scale));\n  --font-size-xs: calc(11px * var(--font-scale));\n  --font-size-sm: calc(12px * var(--font-scale));\n  --font-size-md: calc(13px * var(--font-scale));\n  --font-size-base: calc(14px * var(--font-scale));\n  /* Headings are FLUID (responsive workstream 2026-09-30): they reach their\n   * desktop size at ~1000px wide and shrink slightly below, so a phone is\n   * never bigger than a laptop. Sizes up to lg stay fixed px (legibility\n   * floor; apps use lg as their 16px touch-input size, so it must not dip).\n   * Everything still multiplies by the user\'s --font-scale. */\n  --font-size-lg: calc(16px * var(--font-scale));\n  --font-size-xl: calc(clamp(16px, 14px + 0.4vw, 18px) * var(--font-scale));\n  --font-size-2xl: calc(clamp(18px, 15px + 0.7vw, 22px) * var(--font-scale));\n  /* Role sizes: reading text, form fields, monospace. Coarse pointers raise\n   * their floors (see the Responsive tokens block). */\n  --font-size-body: var(--font-size-base);\n  --font-size-input: var(--font-size-md);\n  --font-size-code: var(--font-size-sm);\n  /* Select arrow drawn on touch, where native select appearance is off\n   * (see the RESPONSIVE LAYER). In the base token block so consoles that\n   * sync only :root tokens receive it by name; mid-grey reads on light and\n   * dark themes alike, so no per-theme variant. */\n  --af-select-chevron: url("data:image/svg+xml,%3Csvg xmlns=\'http://www.w3.org/2000/svg\' width=\'12\' height=\'8\' viewBox=\'0 0 12 8\'%3E%3Cpath d=\'M1 1.5l5 5 5-5\' fill=\'none\' stroke=\'%238a93a3\' stroke-width=\'1.6\' stroke-linecap=\'round\' stroke-linejoin=\'round\'/%3E%3C/svg%3E");\n\n  --line-height-base: 1.5;\n\n  /* Sizes */\n  --header-height: calc(48px * var(--header-density));\n  --footer-height: calc(32px * var(--header-density));\n  --header-pad-x: 16px;\n  --header-pad-y: calc(8px * var(--header-density));\n\n  /* Radius */\n  --radius-sm: 4px;\n  --radius-md: 8px;\n  --radius-lg: 10px;\n\n  /* General UI surfaces */\n  --ui-surface-1: rgba(0, 0, 0, 0.16);\n  --ui-surface-2: rgba(255, 255, 255, 0.06);\n  --ui-surface-3: rgba(0, 0, 0, 0.25);\n  --ui-border-1: rgba(255, 255, 255, 0.10);\n  --ui-border-2: rgba(255, 255, 255, 0.14);\n  --ui-shadow-1: 0 10px 30px rgba(0, 0, 0, 0.35);\n\n  /* Entity-semantic tokens (one visual language for summoned-entity surfaces\n   * across the entity app, gateway console, and flow — c594 contract).\n   * Dark values are the shipped truth operators learned in the Observer\n   * entity view (abstractobserver/src/entity/entity.css); the light-theme\n   * variants live in the grouped light block below. Named for MEANING, not\n   * color — values may tune per theme, the vocabulary is the contract.\n   * Consumption rule: these are MARK colors (node rings, lanes, diamonds,\n   * chrome); as text they clear 4.5:1 only on primary/secondary surfaces\n   * (scar reads below AA as text on several dark themes) — entity-colored\n   * labels pair the mark with --text-primary text, and identity/bond/accent\n   * are one warm family: never color-alone coding between them. */\n  /* Code/JSON syntax tokens (dark set = the VSCode-dark palette the shared\n   * JSON viewers shipped as literals; promoted to tokens 2026-07-11 so light\n   * themes can override — the literals stay as fallbacks in component CSS\n   * for consumers loading components without theme.css). */\n  --syntax-key: #9cdcfe;\n  --syntax-string: #ce9178;\n  --syntax-number: #b5cea8;\n  --syntax-keyword: #569cd6;\n\n  --entity-identity: #e7b45a;\n  --entity-memory: #6ea8d8;\n  --entity-diary: #7bc98c;\n  --entity-standing: #c084dd;\n  --entity-scar: #e05555;\n  --entity-bond: #f0c060;\n  --entity-accent: #e8a54a;\n\n  /* Account-kind row tints (.af-row--admin|user|entity, .af-kind-chip--*):\n   * faint washes over whatever surface the row sits on. Users stay neutral,\n   * admin takes the theme accent, entities the entity-standing violet. Text\n   * on a tinted row keeps >= 4.5:1 in every theme (check_modal_rows.mjs\n   * computes it). The light-theme strengths live in the grouped light block. */\n  --af-row-tint-admin: color-mix(in srgb, var(--accent) 7%, transparent);\n  --af-row-tint-user: transparent;\n  --af-row-tint-entity: color-mix(in srgb, var(--entity-standing) 8%, transparent);\n  /* Text on a tinted row: the theme\'s primary, and a secondary nudged toward\n   * the foreground so it keeps 4.5:1 over the wash (.af-row__muted). */\n  --af-row-text: var(--text-primary);\n  --af-row-text-muted: color-mix(in srgb, var(--text-secondary) 65%, #ffffff);\n  --af-row-mark-admin: var(--accent);\n  --af-row-mark-user: var(--ui-border-2);\n  --af-row-mark-entity: var(--entity-standing);\n\n  /* Semantic tints (hue-matched to the base colors above — subtle/border\n   * variants must never drift to a different hue family than their base) */\n  --accent-subtle: rgba(233, 69, 96, 0.12);\n  --info-subtle: rgba(96, 165, 250, 0.12);\n  --success-subtle: rgba(39, 174, 96, 0.12);\n  --warning-subtle: rgba(243, 156, 18, 0.12);\n  --error-subtle: rgba(231, 76, 60, 0.12);\n\n  /* Semantic borders (used for selection + pills) */\n  --accent-border: rgba(233, 69, 96, 0.35);\n  --info-border: rgba(96, 165, 250, 0.35);\n  --success-border: rgba(39, 174, 96, 0.35);\n  --warning-border: rgba(243, 156, 18, 0.35);\n  --error-border: rgba(231, 76, 60, 0.35);\n  --muted-border: rgba(148, 163, 184, 0.22);\n\n  /* Chips / badges */\n  --ui-chip-bg: rgba(0, 0, 0, 0.22);\n  --ui-chip-border: rgba(255, 255, 255, 0.12);\n  --ui-chip-text: rgba(255, 255, 255, 0.92);\n\n  /* Small overlay surfaces (tabs, pills) */\n  --ui-overlay-bg: rgba(0, 0, 0, 0.18);\n  --ui-overlay-bg-hover: rgba(255, 255, 255, 0.08);\n  --ui-overlay-border: rgba(255, 255, 255, 0.14);\n  --ui-pill-bg: rgba(0, 0, 0, 0.22);\n  --ui-pill-border: rgba(255, 255, 255, 0.12);\n\n  /* Markdown/code surfaces */\n  --ui-code-bg: rgba(0, 0, 0, 0.18);\n  --ui-code-border: rgba(255, 255, 255, 0.08);\n  --ui-code-block-bg: rgba(0, 0, 0, 0.25);\n  --ui-code-block-border: rgba(255, 255, 255, 0.08);\n\n  /* Compatibility aliases used by multiple thin clients */\n  --bg: var(--bg-primary);\n  --bg-surface: var(--bg-secondary);\n  --bg-elevated: var(--bg-tertiary);\n  --bg-card: var(--bg-secondary);\n  --panel: var(--bg-secondary);\n  --panel2: var(--bg-tertiary);\n  --text: var(--text-primary);\n  --muted: var(--text-secondary);\n  --danger: var(--error);\n  --ok: var(--success);\n\n  /* Strokes */\n  --border: var(--bg-tertiary);\n  --border-default: rgba(255, 255, 255, 0.12);\n  --border-accent: rgba(233, 69, 96, 0.55);\n}'

{abstractcore-2.32.0 → abstractcore-2.34.0}/abstractcore/decision/pool.py

@@ -37,6 +37,22 @@ _LOAD_LOCKS: Dict[Key, threading.Lock] = {}
 _LOCK = threading.Lock()
 
 
+def serves_decisions_only(provider: Optional[str]) -> bool:
+    """Whether a provider serves decision models only (``typesafe``, ``systemone`` and endpoint
+    profiles of those families). Its model names (``jev-latest``) are the endpoint's own: never
+    one of the ``ALIASES`` of a resident model, and always decision models."""
+    name = str(provider or "").strip().lower()
+    if not name:
+        return False
+    try:
+        from ..providers.registry import get_provider_info
+
+        info = get_provider_info(name)
+    except Exception:
+        return False
+    return info is not None and list(info.supported_features or []) == ["decision"]
+
+
 def split_handle(handle: str) -> Tuple[Optional[str], str]:
     """``(provider, model)`` of a decision handle: the provider when the handle starts with
     one (``mlx/jaredpalmer/kev-4b``), else ``None`` and the handle whole."""

{abstractcore-2.32.0 → abstractcore-2.34.0}/abstractcore/embeddings/manager.py

@@ -936,6 +936,8 @@ class EmbeddingManager:
         self._load_lock = threading.RLock()
         self._model_cv = threading.Condition(self._load_lock)
         self._inflight_encodes = 0
+        # When an encode last began or ended on this manager: the idle unload's clock (ADR 0012).
+        self._last_used_at: Optional[float] = None
         if self.provider == "huggingface":
             self._load_model()
 
@@ -1287,11 +1289,13 @@ class EmbeddingManager:
         with self._model_cv:
             model = self._ensure_local_model()
             self._inflight_encodes += 1
+            self._last_used_at = time.time()
         try:
             yield model
         finally:
             with self._model_cv:
                 self._inflight_encodes -= 1
+                self._last_used_at = time.time()
                 self._model_cv.notify_all()
 
     def _sentence_transformers_source(self) -> "tuple[str, Dict[str, Any]]":
@@ -1304,7 +1308,7 @@ class EmbeddingManager:
         so nothing asks the Hub -- not even for a `refs/main` a pinned download
         may lack. A bare legacy name (`all-MiniLM-L6-v2`) is also looked up as
         `sentence-transformers/<name>`, as sentence-transformers itself does.
-        Not cached -> ModelNotFoundError with the plain "download it first".
+        Not cached -> ModelNotDownloadedError (a ModelNotFoundError) with the plain "download it first".
         Otherwise the id is passed through and may download on first use.
         """
         try:
@@ -1334,13 +1338,14 @@ class EmbeddingManager:
                 return str(snapshot), {"local_files_only": True}
 
         from ..config.manager import hf_download_first_hint
-        from ..exceptions import ModelNotFoundError
+        from ..exceptions import ModelNotDownloadedError
 
         looked_in = ", ".join(str(d) for d in cache_dirs) or "no cache directory found"
-        raise ModelNotFoundError(
+        raise ModelNotDownloadedError(
             f"Embedding model {names[0]!r} is not in the local Hugging Face cache (looked in: {looked_in}); "
             + hf_download_first_hint(names[0])
-            + "."
+            + ".",
+            model=names[0], provider=self.provider,
         )
 
     def _select_backend(self) -> EmbeddingBackend:

{abstractcore-2.32.0 → abstractcore-2.34.0}/abstractcore/exceptions/__init__.py

@@ -187,6 +187,19 @@ class ModelNotFoundError(ProviderError):
     pass
 
 
+class ModelNotDownloadedError(ModelNotFoundError):
+    """The named model is not on this computer, and loading never downloads (`offline.offline_first`).
+
+    A state, not a fault: the remedy is a download. `model` is the id to download and `provider`
+    the provider that serves it (`abstractcore models download <provider> <model>`). Callers read
+    these attributes; the message is for people, and its wording is not a contract."""
+
+    def __init__(self, message: str, *, model: "str | None" = None, provider: "str | None" = None):
+        super().__init__(message)
+        self.model = model
+        self.provider = provider
+
+
 class UnknownProviderError(ModelNotFoundError, ValueError):
     """The provider part of a provider/model id names no registered provider.
 
@@ -314,6 +327,7 @@ __all__ = [
     'SessionError',
     'ConfigurationError',
     'ModelNotFoundError',
+    'ModelNotDownloadedError',
     'UnknownProviderError',
     'ModelArtifactMismatchError',
     'format_model_error',

{abstractcore-2.32.0 → abstractcore-2.34.0}/abstractcore/providers/base.py

@@ -127,6 +127,7 @@ class PromptCacheStore:
         max_entries: int = 32,
         default_ttl_s: Optional[float] = None,
         on_evict: Optional[Any] = None,
+        on_set: Optional[Any] = None,
     ):
         self._max_entries = int(max_entries) if max_entries and int(max_entries) > 0 else 32
         self._default_ttl_s = default_ttl_s if default_ttl_s is None else float(default_ttl_s)
@@ -141,11 +142,19 @@ class PromptCacheStore:
         # deepcopy per key for the provider's lifetime once the key was
         # LRU-evicted, since nothing told the provider the key was gone).
         self._on_evict = on_evict
+        # Called as on_set(key, value) after every set(), outside the lock: a provider that bounds
+        # its entries by a measure the store cannot compute (GGUF snapshot BYTES) enforces it here,
+        # on every insert path at once.
+        self._on_set = on_set
 
     def set_on_evict(self, callback: Optional[Any]) -> None:
         """Register the eviction callback (see __init__). One callback, last wins."""
         self._on_evict = callback
 
+    def set_on_set(self, callback: Optional[Any]) -> None:
+        """Register the after-set callback (see __init__). One callback, last wins."""
+        self._on_set = callback
+
     def _fire_evictions(self, evicted: List[tuple]) -> None:
         """Invoke the eviction callback OUTSIDE self._lock (callbacks may take
         provider-side locks; never nest them under the store lock)."""
@@ -216,6 +225,12 @@ class PromptCacheStore:
                 evicted_key, evicted_entry = self._entries.popitem(last=False)
                 evicted.append((evicted_key, evicted_entry.value))
         self._fire_evictions(evicted)
+        cb = self._on_set
+        if cb is not None:
+            try:
+                cb(key, value)
+            except Exception as exc:
+                logger.warning(f"prompt cache on_set callback failed for key '{key}': {type(exc).__name__}: {exc}")
 
     def delete(self, key: str) -> bool:
         if not isinstance(key, str) or not key.strip():
@@ -224,6 +239,19 @@ class PromptCacheStore:
         with self._lock:
             return self._entries.pop(key, None) is not None
 
+    def evict(self, key: str) -> bool:
+        """Drop `key` as a capacity eviction: unlike delete(), the on_evict callback fires (the
+        provider frees what it keyed to the entry). For budgets the store cannot measure itself."""
+        if not isinstance(key, str) or not key.strip():
+            return False
+        key = key.strip()
+        with self._lock:
+            entry = self._entries.pop(key, None)
+        if entry is None:
+            return False
+        self._fire_evictions([(key, entry.value)])
+        return True
+
     def clear(self) -> None:
         with self._lock:
             self._entries.clear()
@@ -872,6 +900,7 @@ class BaseProvider(AbstractCoreInterface, ABC):
             # provider so per-key side state (snapshots, device pools) can be
             # released — silent eviction is a leak (2026-08-03 audit).
             on_evict=self._prompt_cache_store_evicted,
+            on_set=self._prompt_cache_store_after_set,
         )
 
         # Provider created successfully - no event emission needed
@@ -2953,6 +2982,45 @@ class BaseProvider(AbstractCoreInterface, ABC):
             if state is not None and "state" not in spec:
                 spec["state"] = state
 
+    def _answers_decisions_itself(self) -> bool:
+        """True when this instance's own model is a decision model (a local Kev / Lev / Laya
+        checkpoint on huggingface or mlx, or a decisions-only provider such as `typesafe`)."""
+        if getattr(self, "_decision_model", None) is not None or callable(getattr(self, "_build_decision_model", None)):
+            return True
+        from ..decision import is_decision_model
+        from ..decision.pool import serves_decisions_only
+
+        provider = str(getattr(self, "provider", "") or "").strip().lower()
+        if serves_decisions_only(provider):
+            return True
+        model = str(getattr(self, "model", "") or "").strip()
+        return provider in {"huggingface", "mlx"} and bool(model) and is_decision_model(model, provider)
+
+    def _decisions_stay_on_this_model(self, specs: List[Dict[str, Any]], output_routes: List[Any]) -> None:
+        """A decision model answers its own decisions: the `output.decision` default is for
+        instances that cannot decide (a text model). When the request names no decision
+        provider or model and this instance's model is a decision model, the default's
+        provider, model, endpoint and options are taken back off the spec, so
+        `_decision_route` picks this model (a `typesafe` default no longer sends a local Kev
+        instance's document to the hosted service). A route the request names still wins."""
+        decision_specs = [s for s in specs if str(s.get("modality") or "").lower() == "decision"]
+        decision_routes = [r for r in output_routes if getattr(r, "route_key", None) == "output.decision"]
+        if not decision_specs or len(decision_specs) != len(decision_routes):
+            return
+        if not self._answers_decisions_itself():
+            return
+        for spec, route in zip(decision_specs, decision_routes):
+            sources = dict(getattr(route, "field_sources", {}) or {})
+            if sources.get("provider") == "explicit" or sources.get("model") == "explicit":
+                continue
+            for key in ("provider", "model", "base_url"):
+                if sources.get(key) not in (None, "explicit"):
+                    spec.pop(key, None)
+            if sources.get("options") not in (None, "explicit"):
+                for key, value in dict(getattr(route, "options", {}) or {}).items():
+                    if spec.get(key) == value:  # the default's option, applied to the spec
+                        spec.pop(key, None)
+
     def _decision_load_kwargs(self) -> Dict[str, Any]:
         """Provider-specific options for ``load_decision_model`` (device, precision...)."""
         return {}
@@ -2999,15 +3067,17 @@ class BaseProvider(AbstractCoreInterface, ABC):
         A `model` written as `provider/model` names its own provider (it wins over the
         provider the route resolver copied from the `output.decision` default); a bare
         model picks the local engine (`decision.pool.parse_handle`); `kev-latest` /
-        `jev-latest` reuse the most recently used resident model.
+        `jev-latest` reuse the most recently used resident model, unless the route names a
+        provider that serves decisions only (`typesafe`, `systemone`): there they are the
+        hosted model's own name (`typesafe` + `jev-latest` is TypeSafe's Jev).
         """
-        from ..decision.pool import ALIASES, latest, parse_handle, split_handle
+        from ..decision.pool import ALIASES, latest, parse_handle, serves_decisions_only, split_handle
 
         route_provider = str(spec.get("provider") or "").strip().lower()
         route_model = str(spec.get("model") or "").strip()
         own_provider = str(getattr(self, "provider", "") or "").strip().lower()
         own_model = str(getattr(self, "model", "") or "").strip()
-        if route_model.lower() in ALIASES:
+        if route_model.lower() in ALIASES and not serves_decisions_only(route_provider):
             resident = latest()
             if resident is None:
                 raise ValueError(
@@ -3819,6 +3889,7 @@ class BaseProvider(AbstractCoreInterface, ABC):
                 is_acore_output = False
             else:
                 self._bind_decision_questions(maybe_specs, kwargs)
+                self._decisions_stay_on_this_model(maybe_specs, resolved_generate_route.output_routes)
                 kwargs["output"] = maybe_specs[0] if len(maybe_specs) == 1 else maybe_specs
 
         if is_acore_output:
@@ -7425,6 +7496,11 @@ class BaseProvider(AbstractCoreInterface, ABC):
     def _prompt_cache_backend_create(self) -> Optional[Any]:
         return None
 
+    def _prompt_cache_store_after_set(self, key: str, cache_value: Any) -> None:
+        """Hook: `key` was just stored. Providers bounding the store by bytes enforce it here.
+        Default: nothing."""
+        _ = (key, cache_value)
+
     def _prompt_cache_store_evicted(self, key: str, cache_value: Any) -> None:
         """Hook: `key` left the prompt-cache store WITHOUT an explicit clear
         (LRU capacity eviction or TTL expiry).
@@ -9657,6 +9733,7 @@ Please provide a structured response."""
                 is_acore_output = False
             else:
                 self._bind_decision_questions(maybe_specs, kwargs)
+                self._decisions_stay_on_this_model(maybe_specs, resolved_generate_route.output_routes)
                 kwargs["output"] = maybe_specs[0] if len(maybe_specs) == 1 else maybe_specs
 
         if is_acore_output:

{abstractcore-2.32.0 → abstractcore-2.34.0}/abstractcore/providers/generation_cancel.py

@@ -239,11 +239,16 @@ import weakref as _weakref
 #: expiry is logged at ERROR and refuses the unload.
 DEFAULT_EJECT_DRAIN_TIMEOUT_S = 30.0
 
+#: How long a call that begins while an idle unload holds this instance's gate
+#: waits for it (core ADR 0012). The unload itself takes seconds; reaching this
+#: bound means something is stuck, and the call fails loudly instead of hanging.
+IDLE_GATE_WAIT_S = 120.0
+
 
 class _InflightCall:
     __slots__ = ("event", "started", "done", "__weakref__")
 
-    def __init__(self, event: threading.Event) -> None:
+    def __init__(self, event: Optional[threading.Event]) -> None:
         self.event = event
         self.started = _time.monotonic()
         self.done = threading.Event()
@@ -252,27 +257,49 @@ class _InflightCall:
 class InflightGenerations:
     """Registry of the generations currently running on one provider instance.
 
-    Only calls that carry a cancel event are registered (the host's, or a
-    private one the provider creates, see BaseProvider). Entries are weakly
+    Every call is registered and stamps `last_used_at`, with or without a cancel
+    event (the host's, or a private one the provider creates, see BaseProvider);
+    only a call with an event can be cancelled by an eject. Entries are weakly
     held: an abandoned, garbage-collected stream disappears by itself.
+
+    The idle gate (core ADR 0012): an idle unload closes it (`close_for_idle`), then
+    checks again that nothing runs before it ejects, so it never cancels a call; a
+    call that begins while it is closed waits for `reopen_after_idle` and then runs
+    (the provider reloads its model on demand).
     """
 
     def __init__(self) -> None:
         self._lock = threading.Lock()
+        self._gate = threading.Condition(self._lock)
+        self._closed_for_idle = False
         self._calls: "_weakref.WeakSet[_InflightCall]" = _weakref.WeakSet()
         # Wall-clock time a generation last began or ended on this instance: the
-        # "least recently used" of the residency failsafe (core ADR 0010).
+        # "least recently used" of the residency failsafe (core ADR 0010) and of
+        # the idle unload (ADR 0012).
         self.last_used_at: Optional[float] = None
 
-    def begin(self, event: Optional[threading.Event]) -> Optional[_InflightCall]:
-        if event is None:
-            return None
+    def begin(self, event: Optional[threading.Event]) -> _InflightCall:
         call = _InflightCall(event)
-        with self._lock:
+        with self._gate:
+            if self._closed_for_idle and not self._gate.wait_for(lambda: not self._closed_for_idle,
+                                                                 timeout=IDLE_GATE_WAIT_S):
+                raise RuntimeError(
+                    f"this model has been unloading for idleness for more than {IDLE_GATE_WAIT_S:.0f}s; "
+                    "the call was not started (send it again)")
             self._calls.add(call)
             self.last_used_at = _time.time()
         return call
 
+    def close_for_idle(self) -> None:
+        """Close the gate for an idle unload: calls that begin from now wait for the reopen."""
+        with self._gate:
+            self._closed_for_idle = True
+
+    def reopen_after_idle(self) -> None:
+        with self._gate:
+            self._closed_for_idle = False
+            self._gate.notify_all()
+
     def end(self, call: Optional[_InflightCall]) -> None:
         if call is None:
             return
@@ -301,7 +328,8 @@ class InflightGenerations:
                     "drain_timeout_s": timeout}
         t0 = _time.monotonic()
         for call in calls:
-            call.event.set()
+            if call.event is not None:  # a call without an event cannot be stopped: it is waited for
+                call.event.set()
         _log.warning(
             "%s/%s: %s — cancelled %d in-flight generation(s); waiting up to %.1fs for them to stop",
             provider, model, reason, len(calls), timeout,
@@ -327,6 +355,7 @@ class InflightGenerations:
 __all__ = [
     "CANCEL_KWARG",
     "DEFAULT_EJECT_DRAIN_TIMEOUT_S",
+    "IDLE_GATE_WAIT_S",
     "HttpCancelGuard",
     "InflightGenerations",
     "WATCH_EXIT_SLICE_S",