Skip to content

armgit / avibe

Public Imported from GitHub
0

Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.

fix(codex): disable native synchronous user input at launch (#1954)

Call White03e42c205f3d
diff --git a/docs/plans/codex-native-user-input.md b/docs/plans/codex-native-user-input.md
new file mode 100644
index 000000000..5e2e844f3
--- /dev/null
+++ b/docs/plans/codex-native-user-input.md
@@ -0,0 +1,69 @@
+# Codex Native User-Input Policy
+
+## Ownership
+
+Codex owns native tool registration. Avibe owns which native interaction
+surfaces it exposes in the app-server processes it launches. Model capability
+metadata and host interaction policy are separate concerns.
+
+Avibe passes `-c tools.experimental_request_user_input.enabled=false` through
+the existing final app-server configuration overrides, in both Direct and
+Gateway mode. Earlier backend extra arguments cannot override this policy.
+No user configuration, model catalog, prompt, route, or model selection is
+rewritten. Existing unsupported synchronous request handling remains a
+compatibility fallback, not the mechanism for hiding the tool.
+
+This change does not restart an existing app-server or install a different
+Codex binary. It applies when Avibe next launches an app-server normally.
+
+## Asynchronous Gap
+
+In Codex 0.153.2 the setting above removes `request_user_input` from the
+model-visible tool list, but does **not** remove `request_user_input_async`.
+The latter is registered independently when the model catalog advertises
+`request_user_input_async` or its legacy name `send_user_message_async`.
+The removed `features.send_async_message` flag does not disable it.
+`features.default_mode_request_user_input` controls the synchronous tool's
+available modes, not asynchronous registration.
+
+The preferred completion is an upstream tool-registration opt-out, consumed
+as a process argument by Avibe. The implementation must not replace the
+complete model catalog to disguise a host policy as model capability.
+The upstream request is
+[openai/codex#43821](https://github.com/openai/codex/issues/43821).
+Until an applicable upstream version supports this, asynchronous suppression
+remains unresolved; this change is not a complete fix for missing async question
+UI.
+
+Upstream currently accepts issue reports and analysis, not external code PRs.
+Related reports
+[#43753](https://github.com/openai/codex/issues/43753),
+[#43803](https://github.com/openai/codex/issues/43803), and
+[#43057](https://github.com/openai/codex/issues/43057)
+describe missing or disappearing asynchronous question interfaces.
+Those reports do not establish the same root cause or a maintainer commitment
+to an opt-out.
+
+## Evidence and Boundaries
+
+- Eight credential-free loopback Responses captures using the actual 0.153.2
+ binary covered Default and Plan modes, with no input override, synchronous
+ disable, the removed async flag, and both flags plus the Default-mode flag.
+ Every capture kept async input exposed. Only synchronous disable removed
+ synchronous input. The catalog bytes were identical across all cases.
+- The captures preserved the remaining tool names and
+ `parallel_tool_calls=true`. No real model calls or model-directed tool
+ executions occurred. This does not prove background job completion.
+- Transport unit coverage requires the explicit policy value and verifies
+ host overrides follow conflicting backend arguments.
+- The opt-in native prompt contract also checks the model-visible synchronous
+ tool is absent and ordinary execution tools remain. The shared native harness
+ applies these checks across all six cases: legacy migration, prompt refresh,
+ process restart, automatic/manual compaction, retention pressure, and forks.
+ The existing baseline/overlay assertions and cleanup are preserved.
+ Its fixture is eligible for async input without requiring that upstream
+ limitation to persist forever.
+- Native contracts require `CODEX_PROMPT_CONTRACT_BINARY` to select a real
+ binary. Ordinary CI runs unit coverage and skips this opt-in contract.
+- No global async shutdown, prompt prohibition, model-capability filtering,
+ gateway-only workaround, custom Codex fork, deployment, or user-state reset.
diff --git a/modules/agents/codex/transport.py b/modules/agents/codex/transport.py
index 6a6cbf743..cc21df2b1 100644
--- a/modules/agents/codex/transport.py
+++ b/modules/agents/codex/transport.py
@@ -54,6 +54,9 @@ AVIBE_APP_SERVER_CONFIG_OVERRIDES = (
"features.tool_suggest=false",
"features.workspace_dependencies=false",
"skills.include_instructions=false",
+ # Host-owned questions. Codex 0.153.2 gates the synchronous tool only;
+ # asynchronous exposure still needs upstream support (openai/codex#43821).
+ "tools.experimental_request_user_input.enabled=false",
)
diff --git a/tests/test_codex_prompt_delivery_contract.py b/tests/test_codex_prompt_delivery_contract.py
index 0e0e37a63..a6e6f7d63 100644
--- a/tests/test_codex_prompt_delivery_contract.py
+++ b/tests/test_codex_prompt_delivery_contract.py
@@ -71,6 +71,16 @@ def _developer_texts(body):
]
+def _tool_names(tools):
+ names = set()
+ for tool in tools:
+ if tool.get("type") == "namespace":
+ names.update(_tool_names(tool.get("tools", [])))
+ elif name := tool.get("name"):
+ names.add(name)
+ return names
+
+
@asynccontextmanager
async def _native_server(tmp_path, *, configured_instructions=None):
requests = []
@@ -132,7 +142,9 @@ async def _native_server(tmp_path, *, configured_instructions=None):
"supported_in_api": True,
"priority": 1,
"support_verbosity": False,
- "experimental_supported_tools": [],
+ # Eligible for the independent async path: the host's
+ # synchronous opt-out must not depend on catalog edits.
+ "experimental_supported_tools": ["request_user_input_async"],
"truncation_policy": {"mode": "tokens", "limit": 10000},
"base_instructions": "You are a test assistant.",
"model_messages": {"collaboration_modes": {"default": CATALOG_PROMPT}},
@@ -189,6 +201,11 @@ async def _native_server(tmp_path, *, configured_instructions=None):
try:
await harness.native.start()
yield harness
+ for model_request in requests:
+ names = _tool_names(model_request["tools"])
+ assert "request_user_input" not in names
+ if not any(item.get("type") == "compaction_trigger" for item in model_request["input"]):
+ assert {"exec_command", "write_stdin"} <= names
finally:
await harness.native.stop()
server.close()
diff --git a/tests/test_codex_transport.py b/tests/test_codex_transport.py
index 4485036f5..fac61bb17 100644
--- a/tests/test_codex_transport.py
+++ b/tests/test_codex_transport.py
@@ -62,6 +62,8 @@ class CodexTransportHealthTests(unittest.IsolatedAsyncioTestCase):
"features.plugins=true",
"-c",
"features.multi_agent=true",
+ "-c",
+ "tools.experimental_request_user_input.enabled=true",
],
)
@@ -107,6 +109,8 @@ class CodexTransportHealthTests(unittest.IsolatedAsyncioTestCase):
"features.plugins=true",
"-c",
"features.multi_agent=true",
+ "-c",
+ "tools.experimental_request_user_input.enabled=true",
*_forced_config_args(),
),
)
@@ -208,6 +212,7 @@ class CodexTransportHealthTests(unittest.IsolatedAsyncioTestCase):
"features.plugins=false",
"features.terminal_visualization_instructions=false",
"skills.include_instructions=false",
+ "tools.experimental_request_user_input.enabled=false",
}.issubset(disabled)
)
# ``agents`` is a role-definition table in older supported Codex