intentseal-core 0.1.0__tar.gz

intentseal_core-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+

intentseal_core-0.1.0/NOTICE

@@ -0,0 +1,4 @@
+IntentSeal
+Copyright 2026 Sankalp Wanjari
+
+This product is licensed under the Apache License, Version 2.0 (see LICENSE).

intentseal_core-0.1.0/PKG-INFO

@@ -0,0 +1,73 @@
+Metadata-Version: 2.4
+Name: intentseal-core
+Version: 0.1.0
+Summary: IntentSeal Core Engine: prompt-injection firewall and action guard for LLM agents
+Author-email: Sankalp Wanjari <sankalpwanjari85@gmail.com>
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/sankalp2515/intentseal
+Project-URL: Source, https://github.com/sankalp2515/intentseal
+Project-URL: Documentation, https://github.com/sankalp2515/intentseal/blob/main/docs/INTEGRATION.md
+Project-URL: Issues, https://github.com/sankalp2515/intentseal/issues
+Keywords: prompt injection,llm security,ai agents,guardrails,firewall,rag
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Security
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.11
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: NOTICE
+Requires-Dist: httpx>=0.27
+Requires-Dist: pydantic>=2.6
+Requires-Dist: pyyaml>=6.0
+Requires-Dist: python-dotenv>=1.0
+Requires-Dist: anyascii>=0.3
+Provides-Extra: engine
+Requires-Dist: beautifulsoup4>=4.12; extra == "engine"
+Requires-Dist: pypdf>=4; extra == "engine"
+Requires-Dist: pdfplumber>=0.11; extra == "engine"
+Requires-Dist: python-docx>=1.1; extra == "engine"
+Requires-Dist: ftfy>=6.1; extra == "engine"
+Requires-Dist: confusable-homoglyphs>=3.2; extra == "engine"
+Requires-Dist: detect-secrets>=1.4; extra == "engine"
+Provides-Extra: classifiers
+Requires-Dist: onnxruntime>=1.17; extra == "classifiers"
+Requires-Dist: tokenizers>=0.15; extra == "classifiers"
+Requires-Dist: huggingface_hub>=0.23; extra == "classifiers"
+Requires-Dist: numpy>=1.26; extra == "classifiers"
+Provides-Extra: ocr
+Requires-Dist: rapidocr-onnxruntime>=1.3; extra == "ocr"
+Requires-Dist: pillow>=10; extra == "ocr"
+Provides-Extra: postgres
+Requires-Dist: psycopg[binary]>=3.2; extra == "postgres"
+Provides-Extra: anthropic
+Requires-Dist: anthropic>=1.12; extra == "anthropic"
+Provides-Extra: all
+Requires-Dist: intentseal-core[classifiers,engine,ocr]; extra == "all"
+Dynamic: license-file
+
+# intentseal-core
+
+The engine behind [IntentSeal](https://pypi.org/project/intentseal/), a prompt-injection firewall for AI agents:
+content extraction for every format (PDF, Word, HTML, email, Markdown, JSON, code, images via OCR), normalisation and
+decoding, a three-tier detector (structural rules, a local classifier, an AI judge), session risk and data-flow
+tracking, the action guard and the decision log.
+
+Most users install the SDK instead:
+
+```bash
+pip install intentseal                 # SDK, remote mode
+pip install "intentseal[embedded]"     # SDK with this engine in-process
+```
+
+Extras of this package: `engine` (extraction, normalisation, secret scan), `classifiers` (local ONNX classifier),
+`ocr`, `postgres`, `anthropic`, `all`.
+
+It also installs the `intentseal` command: `intentseal init` writes a starter policy; `intentseal providers check`
+lists the configured model providers and tests each key (key values are never printed).
+
+Source and documentation: https://github.com/sankalp2515/intentseal · Licensed under the Apache License 2.0.

intentseal_core-0.1.0/README.md

@@ -0,0 +1,21 @@
+# intentseal-core
+
+The engine behind [IntentSeal](https://pypi.org/project/intentseal/), a prompt-injection firewall for AI agents:
+content extraction for every format (PDF, Word, HTML, email, Markdown, JSON, code, images via OCR), normalisation and
+decoding, a three-tier detector (structural rules, a local classifier, an AI judge), session risk and data-flow
+tracking, the action guard and the decision log.
+
+Most users install the SDK instead:
+
+```bash
+pip install intentseal                 # SDK, remote mode
+pip install "intentseal[embedded]"     # SDK with this engine in-process
+```
+
+Extras of this package: `engine` (extraction, normalisation, secret scan), `classifiers` (local ONNX classifier),
+`ocr`, `postgres`, `anthropic`, `all`.
+
+It also installs the `intentseal` command: `intentseal init` writes a starter policy; `intentseal providers check`
+lists the configured model providers and tests each key (key values are never printed).
+
+Source and documentation: https://github.com/sankalp2515/intentseal · Licensed under the Apache License 2.0.

intentseal_core-0.1.0/intentseal_core/__init__.py

@@ -0,0 +1,15 @@
+"""IntentSeal Core Engine."""
+
+import os
+
+__version__ = "0.1.0"
+
+
+def adopt_legacy_env() -> None:
+    """Settings were named ASG_* before 8 Oct 2026 (D-038). An old name still works when the new one is unset."""
+    for name, value in list(os.environ.items()):
+        if name.startswith("ASG_"):
+            os.environ.setdefault("INTENTSEAL_" + name[4:], value)
+
+
+adopt_legacy_env()

intentseal_core-0.1.0/intentseal_core/__main__.py

@@ -0,0 +1,5 @@
+import sys
+
+from .cli import main
+
+sys.exit(main())

intentseal_core-0.1.0/intentseal_core/action_guard.py

@@ -0,0 +1,174 @@
+"""CORE-9 Action guard: deterministic checks on every tool call before it runs.
+
+Checks, in order (first hard violation wins; all observations are recorded):
+  1. tool allowed for this agent
+  2. pinned fields: values come from trusted session context, never from the model
+  3. outbound scan: canaries / secrets / registered values in any argument
+  4. recipient allow-list (send-type tools)
+  5. URL domain allow-list (fetch-type tools) - stops data leaving through a URL
+  6. path stays inside the sandbox (file tools)
+  7. provenance: high-risk target that appears only in untrusted content
+  8. session risk: high-risk tool while the session carries risk from untrusted content
+"""
+
+from __future__ import annotations
+
+import fnmatch
+import time
+from dataclasses import dataclass, field
+from email.utils import getaddresses
+from typing import Any
+
+from .outbound import scan_args
+from .policy import AgentPolicy, ToolPolicy
+from .session import Session, domain_of
+from .types import ActionResult, Decision
+
+
+@dataclass
+class ActionCheck:
+    rule: str
+    decision: Decision
+    reason: str
+    labels: list[str] = field(default_factory=list)
+
+
+def _recipient_ok(addr: str, allowed: list[str]) -> bool:
+    a = addr.strip().lower()
+    if a.count("@") != 1 or not a.split("@")[0] or not a.split("@")[1]:
+        return False
+    for rule in allowed:
+        r = rule.lower()
+        if (r.startswith("@") and a.endswith(r)) or a == r:
+            return True
+    return False
+
+
+def recipients(value: Any) -> list[str]:
+    """Addresses a mailer would deliver to, parsed the way mail libraries parse them (display names, angle
+    brackets, comma/semicolon lists). A piece that does not parse to an address is kept as is, so it fails."""
+    raw = ", ".join(map(str, value)) if isinstance(value, (list, tuple)) else str(value)
+    out: list[str] = []
+    for piece in raw.replace(";", ",").split(","):
+        if not piece.strip():
+            continue
+        parsed = [a for _, a in getaddresses([piece])]
+        out += parsed if parsed and all(parsed) else [piece]
+    # getaddresses splits "Name <x@a> @b" into two entries; any leftover fragment must also pass.
+    for _, a in getaddresses([raw.replace(";", ",")]):
+        if a and a not in out:
+            out.append(a)
+    return out
+
+
+def _domain_ok(url: str, allowed: list[str]) -> bool:
+    d = domain_of(url)
+    if not d:
+        return False
+    return any(d == a.lower() or d.endswith("." + a.lower()) or fnmatch.fnmatch(d, a.lower()) for a in allowed)
+
+
+def _path_ok(path: str) -> bool:
+    p = path.replace("\\", "/")
+    return not (p.startswith("/") or ".." in p.split("/") or (len(p) > 1 and p[1] == ":"))
+
+
+def check_action(tool: str, args: dict[str, Any], policy: AgentPolicy, session: Session,
+                 registered_secrets: list[str] | None = None) -> ActionResult:
+    t0 = time.perf_counter()
+    args = dict(args or {})
+    checks: list[ActionCheck] = []
+    tp: ToolPolicy | None = policy.tools.get(tool)
+
+    if tp is None:
+        if policy.unknown_tools == "block":
+            checks.append(ActionCheck("tool-not-allowed", Decision.BLOCK, f"'{tool}' is not an allowed tool for this agent",
+                                      ["Tool Abuse"]))
+        tp = ToolPolicy()
+
+    # 2. pinned fields (rewrite, never trust the model's value)
+    overrides: dict[str, Any] = {}
+    for f in tp.pinned:
+        if f in session.trusted:
+            if str(args.get(f, "")).strip() != str(session.trusted[f]).strip():
+                overrides[f] = {"requested": args.get(f), "pinned": session.trusted[f]}
+            args[f] = session.trusted[f]
+    if overrides:
+        checks.append(ActionCheck("pinned-field-override", Decision.ALLOW,
+                                  f"pinned field(s) {', '.join(overrides)} set from trusted context", ["Tool Abuse"]))
+
+    # 3. outbound secrets / canaries
+    hits = scan_args(args, policy.canary_values(), registered_secrets or [])
+    if hits:
+        kinds = sorted({h.kind for h in hits})
+        checks.append(ActionCheck("outbound-secret", Decision.BLOCK,
+                                  f"arguments contain protected data ({', '.join(kinds)}) in {', '.join(sorted({h.field for h in hits}))}",
+                                  ["Secret Extraction", "Credential Theft"]))
+
+    # 4. recipients
+    if tp.allowed_recipients and tp.recipient_arg in args:
+        parsed = recipients(args[tp.recipient_arg])
+        bad = [r for r in parsed if not _recipient_ok(r, tp.allowed_recipients)]
+        if bad or not parsed:
+            checks.append(ActionCheck("recipient-not-allowed", Decision.BLOCK,
+                                      f"recipient outside allowed list ({len(bad)} address(es))", ["Tool Abuse"]))
+
+    # 5. URL domains
+    if tp.allowed_domains and tp.url_arg in args:
+        if not _domain_ok(str(args[tp.url_arg]), tp.allowed_domains):
+            checks.append(ActionCheck("domain-not-allowed", Decision.BLOCK,
+                                      f"URL domain '{domain_of(str(args[tp.url_arg]))}' is not approved", ["Tool Abuse"]))
+
+    # 6. sandbox paths
+    if tp.path_arg and tp.path_arg in args and not _path_ok(str(args[tp.path_arg])):
+        checks.append(ActionCheck("path-outside-sandbox", Decision.BLOCK, "path leaves the workspace", ["Tool Abuse"]))
+
+    # 7. provenance of high-risk targets
+    if tp.risk == "high":
+        for key in (tp.recipient_arg, tp.url_arg):
+            if key in args and session.from_untrusted_only(str(args[key])):
+                checks.append(ActionCheck("target-from-untrusted-content",
+                                          Decision.ESCALATE if tp.escalate_if_uncertain else Decision.BLOCK,
+                                          f"'{key}' value appears only in untrusted content, not in the user's request",
+                                          ["Indirect Prompt Injection", "Tool Abuse"]))
+
+    # 8. session risk on high-risk tools
+    if tp.risk == "high" and session.risk >= policy.session.escalate_at:
+        hard = session.risk >= policy.session.block_at
+        checks.append(ActionCheck("high-risk-tool-in-risky-session", Decision.BLOCK if hard else Decision.ESCALATE,
+                                  f"session risk {session.risk:.2f} after untrusted content",
+                                  sorted(session.labels) or ["Multi-Step Jailbreaks"]))
+
+    # 8b. explicit intent (reported after the more specific rules above) (docs/13 L1): a share/destructive tool the user never asked for is held for approval
+    if tp.intent_terms and not session.user_asked_for(tp.intent_terms):
+        checks.append(ActionCheck("action-not-requested", Decision.ESCALATE,
+                                  f"the user's messages do not ask for this action ({', '.join(tp.intent_terms[:4])}…)",
+                                  ["Tool Abuse"]))
+
+    # 9. data flow (CORE-17): where each argument's text came from; flagged content into a high-risk tool
+    taint = session.taint.trace(args, policy.taint.min_coverage) if policy.taint.enabled else []
+    flagged = sorted({h.arg for h in taint if h.flagged})
+    if tp.risk == "high" and flagged and policy.taint.flagged_to_high_risk in ("escalate", "block"):
+        checks.append(ActionCheck("tainted-from-flagged-content",
+                                  Decision.BLOCK if policy.taint.flagged_to_high_risk == "block" else Decision.ESCALATE,
+                                  f"argument(s) {', '.join(flagged)} copy text from content the gateway flagged",
+                                  ["Indirect Prompt Injection", "Tool Abuse"]))
+
+    order = {Decision.BLOCK: 3, Decision.ESCALATE: 2, Decision.NEUTRALISE: 1, Decision.ALLOW: 0}
+    worst = max(checks, key=lambda c: order[c.decision], default=None)
+    decision = worst.decision if worst else Decision.ALLOW
+    enforced = decision if policy.mode == "enforce" else Decision.ALLOW
+    labels = sorted({l for c in checks for l in c.labels})
+    reason = "; ".join(c.reason for c in checks) or "within policy"
+    notice = None
+    if enforced == Decision.BLOCK:
+        notice = f"Action '{tool}' was blocked by the security policy: {worst.reason}. Continue the user's task without it."
+    elif enforced == Decision.ESCALATE:
+        notice = f"Action '{tool}' needs human approval ({worst.reason}). It was not performed; tell the user it is awaiting approval."
+    if decision != Decision.ALLOW:
+        session.add_risk(0.5 if decision == Decision.BLOCK else 0.3, labels)
+    return ActionResult(decision=decision, enforced=enforced,
+                        final_args=args if enforced == Decision.ALLOW else None,
+                        pinned_overrides=overrides, labels=labels, taint=[h.as_dict() for h in taint],
+                        policy_rule=worst.rule if worst else "allow", reason=reason, notice=notice,
+                        latency_ms=round((time.perf_counter() - t0) * 1000, 2))

intentseal_core-0.1.0/intentseal_core/classifier.py

@@ -0,0 +1,127 @@
+"""Tier 1 classifiers (CORE-4): cheap signals that decide what to escalate, never final decisions (D-011).
+
+- LocalClassifier: protectai/deberta-v3-base-prompt-injection-v2 (Apache-2.0), ONNX on CPU, chunked.
+- RemoteGuard: Llama Prompt Guard 2 via the router role `tier1_remote` (hosted, free tier).
+Both cache by content hash (CC2).
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import logging
+import os
+import threading
+import time
+from collections import OrderedDict
+from pathlib import Path
+
+from .config import repo_root
+
+log = logging.getLogger("intentseal.tier1")
+MODEL_REPO = os.environ.get("INTENTSEAL_TIER1_MODEL", "protectai/deberta-v3-base-prompt-injection-v2")
+
+
+class LocalClassifier:
+    def __init__(self, repo: str = MODEL_REPO, threads: int | None = None) -> None:
+        import numpy as np  # noqa: F401  (import check)
+        import onnxruntime as ort
+        from huggingface_hub import snapshot_download
+        from tokenizers import Tokenizer
+
+        os.environ.setdefault("HF_HOME", str(repo_root() / "data" / "hf_cache"))
+        path = Path(snapshot_download(repo, allow_patterns=["onnx/*", "*.json"]))
+        tok = path / "onnx" / "tokenizer.json"
+        self.tok = Tokenizer.from_file(str(tok if tok.exists() else path / "tokenizer.json"))
+        so = ort.SessionOptions()
+        so.intra_op_num_threads = threads or int(os.environ.get("INTENTSEAL_TIER1_THREADS") or "2")
+        self.sess = ort.InferenceSession(str(path / "onnx" / "model.onnx"), so, providers=["CPUExecutionProvider"])
+        self.inputs = {i.name for i in self.sess.get_inputs()}
+        cfg = json.loads((path / "config.json").read_text())
+        self.pos = next(int(k) for k, v in cfg["id2label"].items() if "INJ" in v.upper())
+        self.cls, self.sep = self.tok.token_to_id("[CLS]"), self.tok.token_to_id("[SEP]")
+        self._lock = threading.Lock()
+        self._cache: OrderedDict[str, float] = OrderedDict()   # bounded: CACHE_MAX entries
+
+    def score(self, text: str) -> float:
+        """Max probability over windows of up to 256 tokens (shorter windows keep CPU latency low)."""
+        import numpy as np
+
+        key = hashlib.sha256(text.encode("utf-8", "ignore")).hexdigest()
+        if key in self._cache:
+            return self._cache[key]
+        ids = self.tok.encode(text, add_special_tokens=False).ids
+        best, window, stride = 0.0, 254, 200
+        with self._lock:
+            for start in range(0, max(1, len(ids)), stride):
+                chunk = [self.cls] + ids[start:start + window] + [self.sep]
+                feed = {"input_ids": np.array([chunk], dtype=np.int64),
+                        "attention_mask": np.ones((1, len(chunk)), dtype=np.int64)}
+                if "token_type_ids" in self.inputs:
+                    feed["token_type_ids"] = np.zeros((1, len(chunk)), dtype=np.int64)
+                logits = self.sess.run(None, feed)[0][0]
+                p = np.exp(logits - logits.max())
+                best = max(best, float((p / p.sum())[self.pos]))
+                if start + window >= len(ids):
+                    break
+        self._cache[key] = best
+        while len(self._cache) > CACHE_MAX:
+            self._cache.popitem(last=False)
+        return best
+
+
+CACHE_MAX = 20000
+RETRY_S = 60.0
+_loaded: LocalClassifier | None = None
+_failure: tuple[float, str] | None = None     # (time, error type) of the last failed load
+_load_lock = threading.Lock()
+
+
+def local_classifier() -> LocalClassifier | None:
+    """The process-wide local classifier. A failed load is retried after RETRY_S (never cached for the life of the
+    process); while it is unavailable the engine sends content to the judge or holds it (fail closed)."""
+    global _loaded, _failure
+    if _loaded is not None:
+        return _loaded
+    with _load_lock:
+        if _loaded is None and (_failure is None or time.monotonic() - _failure[0] >= RETRY_S):
+            try:
+                _loaded, _failure = LocalClassifier(), None
+            except Exception as e:  # missing model / runtime
+                _failure = (time.monotonic(), type(e).__name__)
+                log.warning("Tier 1 local classifier unavailable: %s", type(e).__name__)
+    return _loaded
+
+
+def tier1_status() -> dict[str, str | bool]:
+    """For readiness checks: loaded, or the last load error."""
+    return {"loaded": _loaded is not None, "error": _failure[1] if _failure and _loaded is None else ""}
+
+
+class RemoteGuard:
+    """Prompt Guard 2 via the router (role `tier1_remote`). Returns None when unavailable."""
+
+    def __init__(self, router, role: str = "tier1_remote") -> None:
+        self.router = router
+        self.role = role
+        self._cache: dict[str, float] = {}
+
+    async def score(self, text: str) -> float | None:
+        from .llm.adapters import ChatRequest
+        from .llm.router import AllProvidersFailed
+
+        if self.role not in self.router.settings.roles:
+            return None
+        key = hashlib.sha256(text.encode("utf-8", "ignore")).hexdigest()
+        if key in self._cache:
+            return self._cache[key]
+        try:
+            r = await self.router.chat(self.role, ChatRequest(messages=[{"role": "user", "content": text[:2000]}],
+                                                              max_tokens=8, temperature=None))
+            val = float((r.content or "").strip())
+        except (AllProvidersFailed, ValueError):
+            return None
+        self._cache[key] = val
+        if len(self._cache) > CACHE_MAX:
+            self._cache.pop(next(iter(self._cache)))
+        return val