recusal-openai 0.1.0__tar.gz

recusal_openai-0.1.0/.gitignore

@@ -0,0 +1,44 @@
+# Python
+__pycache__/
+*.py[cod]
+*.egg-info/
+.eggs/
+build/
+dist/
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.venv/
+venv/
+env/
+
+# Secrets, never commit
+.env
+.env.*
+*.key
+*.pem
+*.p12
+*service-account*.json
+secrets/
+
+# OS / IDE
+.DS_Store
+Thumbs.db
+.vscode/
+.idea/
+
+# Local hook activation, the repo ships .claude/settings.json.example only;
+# copying it to settings.json is a deliberate per-clone step, never committed
+.claude/settings.json
+.claude/settings.local.json
+
+# Local working docs, not part of the published project
+linkedin.md
+NOTES.md
+RUNBOOK.local.md
+ESSAY.local.md
+SOURCES.local.md
+*.local.md
+
+# Engineering Harness local state, internal planning notes
+.claude/harness/

recusal_openai-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 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 2026 Philip Paz
+
+   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.

recusal_openai-0.1.0/PKG-INFO

@@ -0,0 +1,124 @@
+Metadata-Version: 2.4
+Name: recusal-openai
+Version: 0.1.0
+Summary: Recusal gate for the OpenAI API: judge every function call of a Responses or Chat Completions loop you own, refuse before it runs, and optionally put each decision on a hash-chained audit log.
+Project-URL: Homepage, https://github.com/philpaz/recusal
+Project-URL: Documentation, https://github.com/philpaz/recusal/tree/main/adapters/openai
+Project-URL: Repository, https://github.com/philpaz/recusal
+Project-URL: Issues, https://github.com/philpaz/recusal/issues
+Project-URL: Changelog, https://github.com/philpaz/recusal/blob/main/CHANGELOG.md
+Author-email: Philip Paz <philip.paz@gmail.com>
+License-Expression: Apache-2.0
+License-File: LICENSE
+Keywords: agent-governance,ai-agents,guardrails,openai,recusal,tool-use
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Topic :: Security
+Classifier: Topic :: Software Development :: Quality Assurance
+Classifier: Typing :: Typed
+Requires-Python: >=3.10
+Requires-Dist: openai<4,>=3.34.0
+Requires-Dist: recusal<1,>=0.14.0
+Description-Content-Type: text/markdown
+
+# recusal-openai
+
+The [Recusal](https://github.com/philpaz/recusal) gate for a loop you own on the
+[OpenAI API](https://platform.openai.com/docs/api-reference): every function call the
+model proposes is judged by your policy before your code runs it, and a refused call
+never runs. It covers the Responses API and Chat Completions with one decision function.
+
+This is a separate distribution from `recusal`. The kernel has zero runtime dependencies
+and a frozen compatibility perimeter; this package carries the OpenAI SDK dependency and
+its own version, so an SDK release never forces a release of the kernel.
+
+**Status: not yet published to PyPI.** Until it is, install it from the repository:
+
+```bash
+pip install "recusal-openai @ git+https://github.com/philpaz/recusal#subdirectory=adapters/openai"
+```
+
+Requires Python 3.10+ (the SDK's minimum), `recusal>=0.14.0` and `openai>=3.34.0`. It is
+tested on every pull request against exactly `openai==3.34.0`, and weekly against the
+newest release.
+
+## Use
+
+```python
+from openai import OpenAI
+from recusal import Finding
+from recusal_openai import gate_function_calls
+
+
+def run_sql(sql: str) -> str: ...
+
+
+def policy(tool_name, tool_input):
+    sql = tool_input.get("sql", "")
+    if tool_name == "run_sql" and sql.lstrip().upper().startswith("DELETE"):
+        return [Finding.fail("no_deletes", severity="CRITICAL", message="DELETE is not allowed")]
+    return []
+
+
+client = OpenAI()
+response = client.responses.create(model=..., input=..., tools=...)
+answers = gate_function_calls(response, {"run_sql": run_sql}, policy)
+# pass ``answers`` as the next request's input
+```
+
+`gate_tool_calls(completion, tools, policy)` does the same for Chat Completions and
+returns `role: "tool"` messages. A completion requested with `n > 1` carries alternative
+messages, so pass `choice=<list position>`; without it the gate refuses the completion
+before anything runs.
+
+Each proposed call gets exactly one answer under its own id: the tool's result, or
+Recusal's reason. A policy that raises refuses the call (fail closed).
+
+The gate binds the decoded arguments to the tool's signature with its defaults applied,
+judges that invocation, and calls the tool with exactly those arguments. A default the
+model omitted is therefore judged at the value the function would use.
+
+## Put every decision on the record
+
+```python
+from recusal.audit import AuditLog
+from recusal_openai import audited_function_calls
+
+log = AuditLog(path="decisions.jsonl")
+answers = audited_function_calls(response, tools, policy, log)
+```
+
+`audited_tool_calls` is the Chat Completions form. Each call the model proposes, run or
+refused, is one hash-chained entry labeled `openai.responses.function_call` or
+`openai.chat_completions.tool_call` with the API's own call id and a SHA-256 fingerprint
+of the arguments (never the arguments). If the log cannot be written the call is refused.
+Check the log with `recusal audit verify decisions.jsonl`.
+
+The audited functions call the plain ones, so there is one decision path.
+
+## What the gate refuses, and what it does not see
+
+- **Refused with a reason, never run:** arguments that are not valid JSON or not a JSON
+  object, a tool the application does not have, arguments that do not bind to the tool,
+  and a tool whose signature cannot be judged by name (variadic or positional-only
+  parameters, an unsupported annotation, a default that is not JSON data).
+- **Refused by raising, before anything in the turn runs:** an output item or tool call
+  of a type the gate does not handle (`custom_tool_call`, a hosted tool call, a type a
+  later SDK adds), a call without an id, and a duplicate id. A turn refused this way
+  writes no audit entry.
+- **Not seen at all:** hosted tools (web search, file search, code interpreter, image
+  generation, remote MCP) run on the provider's side and never reach your loop.
+
+## Compatibility
+
+This package is outside the kernel's frozen perimeter (`STABILITY.md`, Integrations). Its
+public names are the ones in `recusal_openai.__all__`. When an SDK release changes the
+response types the gate reads, this package gets a release and the kernel does not.
+
+A runnable offline demonstration, with SDK-validated responses and no API key, is
+[`examples/openai_responses_gate.py`](https://github.com/philpaz/recusal/blob/main/examples/openai_responses_gate.py)
+and
+[`examples/openai_responses_audit.py`](https://github.com/philpaz/recusal/blob/main/examples/openai_responses_audit.py).

recusal_openai-0.1.0/README.md

@@ -0,0 +1,98 @@
+# recusal-openai
+
+The [Recusal](https://github.com/philpaz/recusal) gate for a loop you own on the
+[OpenAI API](https://platform.openai.com/docs/api-reference): every function call the
+model proposes is judged by your policy before your code runs it, and a refused call
+never runs. It covers the Responses API and Chat Completions with one decision function.
+
+This is a separate distribution from `recusal`. The kernel has zero runtime dependencies
+and a frozen compatibility perimeter; this package carries the OpenAI SDK dependency and
+its own version, so an SDK release never forces a release of the kernel.
+
+**Status: not yet published to PyPI.** Until it is, install it from the repository:
+
+```bash
+pip install "recusal-openai @ git+https://github.com/philpaz/recusal#subdirectory=adapters/openai"
+```
+
+Requires Python 3.10+ (the SDK's minimum), `recusal>=0.14.0` and `openai>=3.34.0`. It is
+tested on every pull request against exactly `openai==3.34.0`, and weekly against the
+newest release.
+
+## Use
+
+```python
+from openai import OpenAI
+from recusal import Finding
+from recusal_openai import gate_function_calls
+
+
+def run_sql(sql: str) -> str: ...
+
+
+def policy(tool_name, tool_input):
+    sql = tool_input.get("sql", "")
+    if tool_name == "run_sql" and sql.lstrip().upper().startswith("DELETE"):
+        return [Finding.fail("no_deletes", severity="CRITICAL", message="DELETE is not allowed")]
+    return []
+
+
+client = OpenAI()
+response = client.responses.create(model=..., input=..., tools=...)
+answers = gate_function_calls(response, {"run_sql": run_sql}, policy)
+# pass ``answers`` as the next request's input
+```
+
+`gate_tool_calls(completion, tools, policy)` does the same for Chat Completions and
+returns `role: "tool"` messages. A completion requested with `n > 1` carries alternative
+messages, so pass `choice=<list position>`; without it the gate refuses the completion
+before anything runs.
+
+Each proposed call gets exactly one answer under its own id: the tool's result, or
+Recusal's reason. A policy that raises refuses the call (fail closed).
+
+The gate binds the decoded arguments to the tool's signature with its defaults applied,
+judges that invocation, and calls the tool with exactly those arguments. A default the
+model omitted is therefore judged at the value the function would use.
+
+## Put every decision on the record
+
+```python
+from recusal.audit import AuditLog
+from recusal_openai import audited_function_calls
+
+log = AuditLog(path="decisions.jsonl")
+answers = audited_function_calls(response, tools, policy, log)
+```
+
+`audited_tool_calls` is the Chat Completions form. Each call the model proposes, run or
+refused, is one hash-chained entry labeled `openai.responses.function_call` or
+`openai.chat_completions.tool_call` with the API's own call id and a SHA-256 fingerprint
+of the arguments (never the arguments). If the log cannot be written the call is refused.
+Check the log with `recusal audit verify decisions.jsonl`.
+
+The audited functions call the plain ones, so there is one decision path.
+
+## What the gate refuses, and what it does not see
+
+- **Refused with a reason, never run:** arguments that are not valid JSON or not a JSON
+  object, a tool the application does not have, arguments that do not bind to the tool,
+  and a tool whose signature cannot be judged by name (variadic or positional-only
+  parameters, an unsupported annotation, a default that is not JSON data).
+- **Refused by raising, before anything in the turn runs:** an output item or tool call
+  of a type the gate does not handle (`custom_tool_call`, a hosted tool call, a type a
+  later SDK adds), a call without an id, and a duplicate id. A turn refused this way
+  writes no audit entry.
+- **Not seen at all:** hosted tools (web search, file search, code interpreter, image
+  generation, remote MCP) run on the provider's side and never reach your loop.
+
+## Compatibility
+
+This package is outside the kernel's frozen perimeter (`STABILITY.md`, Integrations). Its
+public names are the ones in `recusal_openai.__all__`. When an SDK release changes the
+response types the gate reads, this package gets a release and the kernel does not.
+
+A runnable offline demonstration, with SDK-validated responses and no API key, is
+[`examples/openai_responses_gate.py`](https://github.com/philpaz/recusal/blob/main/examples/openai_responses_gate.py)
+and
+[`examples/openai_responses_audit.py`](https://github.com/philpaz/recusal/blob/main/examples/openai_responses_audit.py).

recusal_openai-0.1.0/pyproject.toml

@@ -0,0 +1,54 @@
+# The Recusal adapter for the OpenAI API, a separate distribution from `recusal`.
+#
+# Why separate: `recusal` has zero runtime dependencies and a frozen compatibility
+# perimeter (STABILITY.md). An adapter follows a framework that releases on its own
+# schedule, so it carries that dependency and its own version here, and a framework's
+# breaking change never forces a release of the kernel.
+#
+# Build backend pinned exactly, the same pin as the kernel's pyproject.toml: a floating
+# backend is unreviewed code in the release path.
+[build-system]
+requires = ["hatchling==1.27.0"]
+build-backend = "hatchling.build"
+
+[project]
+name = "recusal-openai"
+dynamic = ["version"]
+description = "Recusal gate for the OpenAI API: judge every function call of a Responses or Chat Completions loop you own, refuse before it runs, and optionally put each decision on a hash-chained audit log."
+readme = "README.md"
+# The OpenAI SDK's own minimum.
+requires-python = ">=3.10"
+license = "Apache-2.0"
+license-files = ["LICENSE"]
+authors = [{ name = "Philip Paz", email = "philip.paz@gmail.com" }]
+keywords = ["recusal", "openai", "ai-agents", "agent-governance", "guardrails", "tool-use"]
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Developers",
+    "Operating System :: OS Independent",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3 :: Only",
+    "Topic :: Security",
+    "Topic :: Software Development :: Quality Assurance",
+    "Typing :: Typed",
+]
+dependencies = [
+    # 0.14.0 is the first release where `decide` and `recusal.invocation` are public.
+    "recusal>=0.14.0,<1",
+    # The lower bound is the version the pinned CI lane tests; the weekly latest lane
+    # (.github/workflows/compat.yml) tests the newest release.
+    "openai>=3.34.0,<4",
+]
+
+[project.urls]
+Homepage = "https://github.com/philpaz/recusal"
+Documentation = "https://github.com/philpaz/recusal/tree/main/adapters/openai"
+Repository = "https://github.com/philpaz/recusal"
+Issues = "https://github.com/philpaz/recusal/issues"
+Changelog = "https://github.com/philpaz/recusal/blob/main/CHANGELOG.md"
+
+[tool.hatch.version]
+path = "src/recusal_openai/__init__.py"
+
+[tool.hatch.build.targets.wheel]
+packages = ["src/recusal_openai"]

recusal_openai-0.1.0/smoke_test.py

@@ -0,0 +1,167 @@
+"""Prove the INSTALLED recusal-openai wheel gates real OpenAI SDK response objects.
+
+Run it from a directory outside the repository, with an interpreter that has the built
+wheels installed:
+
+    python /path/to/adapters/openai/smoke_test.py
+
+It fails if the package it imports is the source tree and not an installed distribution:
+a check run from the checkout would prove nothing about the wheel. The responses are built
+with the SDK's own ``model_validate``; no API key or network is used.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import sys
+import tempfile
+from importlib import metadata
+from pathlib import Path
+
+# Python puts this script's directory first on the import path. Drop it, so nothing in
+# the source tree can stand in for the installed package.
+_HERE = Path(__file__).resolve().parent
+sys.path[:] = [entry for entry in sys.path if Path(entry or ".").resolve() != _HERE]
+
+import recusal_openai  # noqa: E402
+from openai.types.chat import ChatCompletion  # noqa: E402
+from openai.types.responses import Response  # noqa: E402
+from recusal_openai import (  # noqa: E402
+    SURFACE_CHAT,
+    SURFACE_RESPONSES,
+    audited_function_calls,
+    audited_tool_calls,
+    gate_function_calls,
+    gate_tool_calls,
+)
+
+import recusal  # noqa: E402
+from recusal import Finding  # noqa: E402
+from recusal.audit import AuditLog, verify  # noqa: E402
+
+EXECUTED: list[str] = []
+CALLS = [("c1", "SELECT 1"), ("c2", "DELETE FROM t")]
+
+
+def run_sql(sql: str) -> str:
+    """Record SQL text without connecting to a database."""
+    EXECUTED.append(sql)
+    return f"recorded: {sql}"
+
+
+TOOLS = {"run_sql": run_sql}
+
+
+def policy(tool_name, tool_input):
+    if str(tool_input.get("sql", "")).upper().startswith("DELETE"):
+        return [Finding.fail("no_deletes", severity="CRITICAL", message="DELETE is not allowed")]
+    return []
+
+
+def response() -> Response:
+    return Response.model_validate(
+        {
+            "id": "resp_offline",
+            "object": "response",
+            "created_at": 0,
+            "model": "offline",
+            "output": [
+                {
+                    "type": "function_call",
+                    "id": f"fc_{ident}",
+                    "call_id": ident,
+                    "name": "run_sql",
+                    "arguments": json.dumps({"sql": sql}),
+                    "status": "completed",
+                }
+                for ident, sql in CALLS
+            ],
+            "parallel_tool_calls": True,
+            "tool_choice": "auto",
+            "tools": [],
+        }
+    )
+
+
+def completion() -> ChatCompletion:
+    return ChatCompletion.model_validate(
+        {
+            "id": "chatcmpl_offline",
+            "object": "chat.completion",
+            "created": 0,
+            "model": "offline",
+            "choices": [
+                {
+                    "index": 0,
+                    "finish_reason": "tool_calls",
+                    "message": {
+                        "role": "assistant",
+                        "content": None,
+                        "tool_calls": [
+                            {
+                                "id": ident,
+                                "type": "function",
+                                "function": {
+                                    "name": "run_sql",
+                                    "arguments": json.dumps({"sql": sql}),
+                                },
+                            }
+                            for ident, sql in CALLS
+                        ],
+                    },
+                }
+            ],
+        }
+    )
+
+
+def main() -> None:
+    for module in (recusal, recusal_openai):
+        origin = Path(module.__file__).resolve()
+        assert "site-packages" in origin.parts, f"{module.__name__} is not installed: {origin}"
+    assert metadata.version("recusal-openai") == recusal_openai.__version__
+    requires = metadata.requires("recusal-openai") or []
+    assert any(r.startswith("recusal") and "openai" not in r for r in requires), requires
+
+    # Negative control: a policy that refuses nothing runs the DELETE, so the refusals
+    # below are the policy's, reached through the gate.
+    EXECUTED.clear()
+    gate_function_calls(response(), TOOLS, lambda *_: [])
+    assert EXECUTED == ["SELECT 1", "DELETE FROM t"], EXECUTED
+
+    EXECUTED.clear()
+    answers = {a["call_id"]: a["output"] for a in gate_function_calls(response(), TOOLS, policy)}
+    assert EXECUTED == ["SELECT 1"], EXECUTED
+    assert answers["c1"] == "recorded: SELECT 1" and "DELETE is not allowed" in answers["c2"]
+
+    EXECUTED.clear()
+    messages = {
+        m["tool_call_id"]: m["content"] for m in gate_tool_calls(completion(), TOOLS, policy)
+    }
+    assert EXECUTED == ["SELECT 1"], EXECUTED
+    assert "DELETE is not allowed" in messages["c2"], messages
+
+    for audited, proposal, surface in (
+        (audited_function_calls, response(), SURFACE_RESPONSES),
+        (audited_tool_calls, completion(), SURFACE_CHAT),
+    ):
+        EXECUTED.clear()
+        with tempfile.TemporaryDirectory() as folder:
+            log = AuditLog(path=os.path.join(folder, "audit.jsonl"))
+            audited(proposal, TOOLS, policy, log)
+            assert EXECUTED == ["SELECT 1"], EXECUTED
+            by_id = {entry["action"]["tool_use_id"]: entry["action"] for entry in log.entries}
+            assert by_id["c2"]["decision"] == "deny" and by_id["c2"]["surface"] == surface, by_id
+            assert by_id["c1"]["decision"] != "deny", by_id
+            intact, problems = verify(log.entries)
+            assert intact, problems
+
+    print(
+        f"recusal-openai {recusal_openai.__version__} with recusal "
+        f"{recusal.__version__}: installed wheel gates and records"
+    )
+
+
+if __name__ == "__main__":
+    main()

recusal_openai-0.1.0/src/recusal_openai/__init__.py

@@ -0,0 +1,43 @@
+"""Recusal for the OpenAI API: judge every function call in a loop you own.
+
+    from recusal_openai import gate_function_calls
+
+    response = client.responses.create(model=..., input=..., tools=...)
+    answers = gate_function_calls(response, {"run_sql": run_sql}, policy)
+    # pass ``answers`` as the next request's input
+
+``gate_tool_calls(completion, tools, policy)`` is the same gate for Chat Completions.
+``policy(tool_name, tool_input)`` is an ordinary Recusal policy returning findings. A
+refused call is answered with Recusal's reason under the call's own id, and its tool does
+not run. ``audited_function_calls`` and ``audited_tool_calls`` are the same loops with
+each decision on a hash-chained audit log.
+
+The decision is made by ``recusal.decide``. This package adds only what is the OpenAI
+API's: where the calls are in a response, how each is answered, and its call id.
+"""
+
+from ._audit import SURFACE_CHAT, SURFACE_RESPONSES, audited_function_calls, audited_tool_calls
+from ._gate import (
+    PASSIVE_OUTPUT_TYPES,
+    Policy,
+    Tools,
+    gate_function_calls,
+    gate_tool_calls,
+    judge_and_run,
+)
+
+__version__ = "0.1.0"
+
+__all__ = [
+    "PASSIVE_OUTPUT_TYPES",
+    "SURFACE_CHAT",
+    "SURFACE_RESPONSES",
+    "Policy",
+    "Tools",
+    "__version__",
+    "audited_function_calls",
+    "audited_tool_calls",
+    "gate_function_calls",
+    "gate_tool_calls",
+    "judge_and_run",
+]

recusal_openai-0.1.0/src/recusal_openai/_audit.py

@@ -0,0 +1,79 @@
+"""The audited loop: the same decision, with every call on a hash-chained ``AuditLog``.
+
+``audited_function_calls`` and ``audited_tool_calls`` call ``gate_function_calls`` and
+``gate_tool_calls`` with the recording arguments passed through to ``recusal.decide``,
+which writes the entry before returning. There is no second decision path. If the log
+cannot be written the decision becomes a deny, so nothing runs unrecorded.
+
+Each entry carries the surface of the API shape that adjudicated the call and the API's own
+call id, and a SHA-256 fingerprint of the arguments, never the arguments themselves.
+
+What each entry is bound to:
+
+- a call that is judged: the bound invocation the tool is called with, defaults included;
+- a call refused before judgment (arguments that are not a JSON object, an unknown tool, a
+  signature the gate cannot judge, arguments that do not bind): the arguments as proposed.
+
+What is not on the log: a turn the gate refuses as a whole by raising (an item type it does
+not handle, a call without an id, a duplicate id). It raises before anything is judged, so
+it runs nothing and writes no entry. A hosted tool never reaches the loop at all.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Optional
+
+from openai.types.chat import ChatCompletion
+from openai.types.responses import Response
+
+from recusal.audit import AuditLog
+
+from ._gate import Policy, Tools, gate_function_calls, gate_tool_calls
+
+#: The surface labels: which API shape's loop adjudicated the call.
+SURFACE_RESPONSES = "openai.responses.function_call"
+SURFACE_CHAT = "openai.chat_completions.tool_call"
+
+
+def audited_function_calls(
+    response: Response,
+    tools: Tools,
+    policy: Policy,
+    log: AuditLog,
+    *,
+    actor: Optional[str] = None,
+    control: Optional[dict[str, Any]] = None,
+) -> list[dict[str, Any]]:
+    """Responses API: ``gate_function_calls`` with every call recorded under its ``call_id``."""
+    return gate_function_calls(
+        response,
+        tools,
+        policy,
+        audit=log,
+        surface=SURFACE_RESPONSES,
+        actor=actor,
+        control=control,
+    )
+
+
+def audited_tool_calls(
+    completion: ChatCompletion,
+    tools: Tools,
+    policy: Policy,
+    log: AuditLog,
+    *,
+    choice: Optional[int] = None,
+    actor: Optional[str] = None,
+    control: Optional[dict[str, Any]] = None,
+) -> list[dict[str, Any]]:
+    """Chat Completions: ``gate_tool_calls`` with every call recorded under its ``id``."""
+    return gate_tool_calls(
+        completion,
+        tools,
+        policy,
+        choice=choice,
+        audit=log,
+        surface=SURFACE_CHAT,
+        actor=actor,
+        control=control,
+    )

recusal_openai-0.1.0/src/recusal_openai/_gate.py

@@ -0,0 +1,274 @@
+"""The gate for a loop the application owns: one decision function, two API shapes.
+
+An application that calls the OpenAI API directly reads the function calls out of the
+response, runs them, and sends the results back. That loop belongs to the application,
+which makes it the place to put the gate:
+
+- the Responses API: ``function_call`` output items, answered with ``function_call_output``
+  input items (``gate_function_calls``);
+- Chat Completions: ``tool_calls`` on the assistant message, answered with ``role: "tool"``
+  messages (``gate_tool_calls``).
+
+For every call the model proposes, the gate decodes the arguments the way the tool would
+receive them, judges the call with ``recusal.decide``, runs the tool only if it is not
+denied, and returns exactly one answer carrying that call's own id. Arguments that are not
+valid JSON or not a JSON object, a tool the application does not have, and a policy that
+raises are all refused with a reason, never run.
+
+Default-deny on what the loop does not understand: an output item or tool call of a type
+this gate does not handle (``custom_tool_call``, a hosted tool call, a type added in a later
+SDK release) raises, naming the type, instead of being skipped. The SDK's answer types do
+not require a call id, so the gate refuses a call without one and checks that every call
+got exactly one answer.
+
+Hosted tools (web search, file search, code interpreter, image generation, remote MCP) run
+on the provider's side and never reach this loop, so nothing here can gate them.
+"""
+
+from __future__ import annotations
+
+import copy
+import inspect
+import json
+from typing import Any, Callable, List, Optional
+
+from openai.types.chat import ChatCompletion
+from openai.types.responses import Response
+
+from recusal import Finding, decide
+from recusal.invocation import unsupported_signature as _unsupported_signature
+
+#: A Recusal policy: ``policy(tool_name, tool_input)`` returns a list of findings.
+Policy = Callable[[str, dict], List[Any]]
+#: The application's tools by the name the model calls them.
+Tools = dict[str, Callable[..., Any]]
+
+#: Response output items that cannot run anything, so the gate lets them pass. Everything
+#: else (custom_tool_call, hosted tool calls, a type a later SDK adds) raises.
+PASSIVE_OUTPUT_TYPES = frozenset({"message", "reasoning"})
+
+
+def _unjudgeable(name: str, plain: str, detail: str, proposal: Any, record: dict) -> str:
+    """The answer for a call the gate refuses before any policy runs.
+
+    Without recording arguments it is ``plain``, the text this gate has always returned.
+    With them (the audited loop) the refusal goes through ``decide`` like
+    every other, so it is one entry on the log, bound to the call as the model proposed it,
+    and the text the model reads is the reason on the record.
+    """
+    if not record:
+        return plain
+    bound = proposal if isinstance(proposal, dict) else {"input": proposal}
+
+    def unjudgeable(_name, _input):
+        return [Finding.fail("recusal_unjudgeable_input", severity="CRITICAL", message=detail)]
+
+    try:
+        _, reason = decide(name, bound, unjudgeable, safe_errors=True, **record)
+    except Exception as exc:  # noqa: BLE001, the refusal stands even if the record cannot be written
+        return f"Recusal failed closed (policy error): RECUSAL_POLICY_ERROR ({type(exc).__name__})"
+    return reason
+
+
+def judge_and_run(
+    name: str, arguments: str | None, tools: Tools, policy: Policy, **record: Any
+) -> str:
+    """Decide one proposed call and run it if allowed; return the text the model gets back.
+
+    The ONE decision path for both API shapes. The decoded arguments are bound to the
+    tool's signature with its defaults applied (``inspect.signature(...).bind`` then
+    ``apply_defaults``: argument binding, not type validation; no coercion happens on this
+    path), the policy judges a detached copy of that bound invocation, and the tool is
+    called with exactly those bound arguments, so a default the model omitted is judged at
+    the value the function would use. A tool whose signature cannot be judged by name
+    (variadic or positional-only parameters, an unsupported annotation, a default outside
+    the JSON domain) is refused, as is a call that does not bind (an unknown or missing
+    parameter; the refusal names the parameter, never a value).
+
+    ``record`` goes straight to ``decide`` (``audit=``, ``surface=``, ``tool_use_id=``, ...),
+    so a loop that puts the decision on a log (``audited_function_calls``) decides,
+    and refuses, exactly as this one does. The entry for a judged call carries the
+    fingerprint of the bound invocation the tool is called with, defaults included; the
+    entry for a call refused before judgment is bound to the arguments as proposed.
+    """
+    try:
+        args = json.loads(arguments or "{}")
+    except ValueError:
+        detail = "arguments are not valid JSON"
+        return _unjudgeable(name, f"Recusal refused `{name}`: {detail}", detail, arguments, record)
+    if not isinstance(args, dict):
+        detail = "arguments are not a JSON object"
+        return _unjudgeable(name, f"Recusal refused `{name}`: {detail}", detail, args, record)
+    if name not in tools:
+        return _unjudgeable(
+            name, f"Recusal refused: unknown tool {name!r}", "unknown tool", args, record
+        )
+    func = tools[name]
+    why = _unsupported_signature(func)
+    if why is not None:
+        detail = f"its signature cannot be judged ({why})"
+        return _unjudgeable(name, f"Recusal refused `{name}`: {detail}", detail, args, record)
+    try:
+        bound = inspect.signature(func).bind(**args)
+    except TypeError as exc:  # the message names parameters only, never values
+        detail = f"arguments do not bind to the tool ({exc})"
+        return _unjudgeable(name, f"Recusal refused `{name}`: {detail}", detail, args, record)
+    bound.apply_defaults()
+    invocation = dict(bound.arguments)
+    try:
+        decision, reason = decide(
+            name, copy.deepcopy(invocation), policy, safe_errors=True, **record
+        )
+    except Exception as exc:  # noqa: BLE001, a buggy policy must not silently disable the gate
+        return f"Recusal failed closed (policy error): RECUSAL_POLICY_ERROR ({type(exc).__name__})"
+    if decision == "deny":
+        return reason
+    return str(func(**invocation))
+
+
+def _require_id(call_id: Any, what: str) -> str:
+    """The SDK does not require an id on the answer; the adapter does, or the answer is lost."""
+    if not isinstance(call_id, str) or not call_id:
+        raise ValueError(f"Recusal refused {what}: it has no call id to answer")
+    return call_id
+
+
+def _for_call(record: dict, call_id: str) -> dict:
+    """The recording arguments for one call of a turn: the turn's, plus the call's own id."""
+    return {**record, "tool_use_id": call_id} if record else {}
+
+
+def _check_ids(call_ids: list[str]) -> None:
+    """Refuse a turn with a duplicate id BEFORE any call in it runs: the turn's answers are
+    keyed by id, so a duplicate could not be answered once, and a refusal raised after
+    dispatch would leave a side effect behind for the caller's retry to repeat."""
+    if len(set(call_ids)) != len(call_ids):
+        raise ValueError(f"Recusal refused the turn: duplicate call ids {call_ids}")
+
+
+def _check_answered_once(call_ids: list[str], answered: list[str]) -> None:
+    """Every proposed call gets exactly one answer, and no answer invents a call."""
+    if sorted(call_ids) != sorted(answered):
+        raise ValueError(
+            f"Recusal refused the turn: calls {sorted(call_ids)} but answers {sorted(answered)}"
+        )
+    if len(set(call_ids)) != len(call_ids):
+        raise ValueError(f"Recusal refused the turn: duplicate call ids {call_ids}")
+
+
+def gate_function_calls(
+    response: Response, tools: Tools, policy: Policy, **record: Any
+) -> list[dict[str, Any]]:
+    """Responses API: judge every ``function_call`` item; return ``function_call_output`` items.
+
+    Raises TypeError for an output item type this adapter does not handle, and ValueError for
+    a call without an id. Pass the returned list as the next request's ``input``.
+    ``record`` is the recording arguments for the turn (``audit=``, ``surface=``, ...);
+    each call is recorded under its own ``call_id``. A turn refused as a whole raises
+    before anything is judged, so it runs nothing and writes no entry.
+    """
+    # pass 1: every item is one this adapter handles and carries a usable, unique id;
+    # nothing runs until the whole turn has passed, so a refused turn has run nothing
+    calls: list[tuple[str, str, Any]] = []
+    for item in response.output:
+        item_type = getattr(item, "type", None)
+        if item_type in PASSIVE_OUTPUT_TYPES:
+            continue
+        if item_type != "function_call":
+            raise TypeError(
+                f"Recusal refused output item of type {item_type!r}: "
+                "this adapter handles function_call items only"
+            )
+        # the type check above is on the item's ``type`` string, which a type checker cannot
+        # use to narrow the SDK's union of output items; from here it is a function call
+        call: Any = item
+        call_id = _require_id(getattr(call, "call_id", None), f"function_call {call.name!r}")
+        calls.append((call_id, call.name, call.arguments))
+    call_ids = [call_id for call_id, _, _ in calls]
+    _check_ids(call_ids)
+    # pass 2: judge and run each call, answering it under its own id
+    outputs = [
+        {
+            "type": "function_call_output",
+            "call_id": call_id,
+            "output": judge_and_run(name, arguments, tools, policy, **_for_call(record, call_id)),
+        }
+        for call_id, name, arguments in calls
+    ]
+    _check_answered_once(call_ids, [o["call_id"] for o in outputs])
+    return outputs
+
+
+def _chosen(completion: ChatCompletion, choice: Any):
+    """The one assistant message this turn dispatches: ``completion.choices[choice]``.
+
+    A completion requested with ``n > 1`` carries alternative messages, not one batch, so
+    one must be chosen explicitly; dispatching every alternative would run tool calls the
+    application never selected. ``choice`` is the list position. With one choice it may be
+    omitted. Every refusal here is raised before any call is judged or run.
+    """
+    choices = list(completion.choices or [])
+    if not choices:
+        raise ValueError("Recusal refused the completion: it has no choices")
+    if choice is None:
+        if len(choices) != 1:
+            raise ValueError(
+                f"Recusal refused the completion: it has {len(choices)} choices; pass "
+                "choice=<list position> to dispatch one assistant message"
+            )
+        return choices[0]
+    if isinstance(choice, bool) or not isinstance(choice, int):
+        raise TypeError(
+            f"Recusal refused the completion: choice must be an int list position, not "
+            f"{type(choice).__name__}"
+        )
+    if choice < 0 or choice >= len(choices):
+        raise ValueError(
+            f"Recusal refused the completion: choice {choice} is out of range for "
+            f"{len(choices)} choice(s)"
+        )
+    return choices[choice]
+
+
+def gate_tool_calls(
+    completion: ChatCompletion,
+    tools: Tools,
+    policy: Policy,
+    *,
+    choice: Optional[int] = None,
+    **record: Any,
+) -> list[dict[str, Any]]:
+    """Chat Completions: judge every ``tool_calls`` entry of ONE chosen assistant message;
+    return ``role: "tool"`` messages.
+
+    ``choice`` is the list position in ``completion.choices``; it may be omitted only when
+    the completion has exactly one choice. A completion with several choices (``n > 1``)
+    and no ``choice`` is refused before anything is judged: the choices are alternative
+    messages, and only the one the application selects is dispatched. Raises TypeError for
+    a tool call type this adapter does not handle (``custom``) or a non-int ``choice``, and
+    ValueError for a call without an id, no choices, or an out-of-range ``choice``. Append
+    the returned messages after the chosen assistant message in the next request.
+    ``record`` is the recording arguments for the turn, as in ``gate_function_calls``.
+    """
+    chosen = _chosen(completion, choice)
+    calls: list[tuple[str, str, Any]] = []
+    for call in chosen.message.tool_calls or []:
+        if call.type != "function":
+            raise TypeError(
+                f"Recusal refused tool call of type {call.type!r}: "
+                "this adapter handles function tool calls only"
+            )
+        call_id = _require_id(call.id, f"tool call {call.function.name!r}")
+        calls.append((call_id, call.function.name, call.function.arguments))
+    call_ids = [call_id for call_id, _, _ in calls]
+    _check_ids(call_ids)
+    outputs = [
+        {
+            "role": "tool",
+            "tool_call_id": call_id,
+            "content": judge_and_run(name, arguments, tools, policy, **_for_call(record, call_id)),
+        }
+        for call_id, name, arguments in calls
+    ]
+    _check_answered_once(call_ids, [o["tool_call_id"] for o in outputs])
+    return outputs

recusal_openai-0.1.0/tested-with.txt

@@ -0,0 +1 @@
+openai==3.34.0