codex-sdk-python 0.104.0__tar.gz → 0.105.0__tar.gz

{codex_sdk_python-0.104.0 → codex_sdk_python-0.105.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: codex-sdk-python
-Version: 0.104.0
+Version: 0.105.0
 Summary: Python SDK for the Codex CLI agent with async threads, streaming events, and structured outputs
 Keywords: codex,sdk,python,api,cli,agent,async,streaming
 Author: Vectorfy Co
@@ -44,7 +44,7 @@ Embed the Codex agent in Python workflows. This SDK wraps the bundled `codex` CL
       <td><strong>Lifecycle</strong></td>
       <td>
         <a href="#ci-cd"><img src="https://img.shields.io/badge/CI%2FCD-Active-16a34a?style=flat&logo=githubactions&logoColor=white" alt="CI/CD badge" /></a>
-        <img src="https://img.shields.io/badge/Release-0.104.0-6b7280?style=flat&logo=pypi&logoColor=white" alt="Release 0.104.0 badge" />
+        <img src="https://img.shields.io/badge/Release-0.105.0-6b7280?style=flat&logo=pypi&logoColor=white" alt="Release 0.105.0 badge" />
         <a href="#license"><img src="https://img.shields.io/badge/License-Apache--2.0-0f766e?style=flat&logo=apache&logoColor=white" alt="License badge" /></a>
       </td>
     </tr>
@@ -73,7 +73,7 @@ Embed the Codex agent in Python workflows. This SDK wraps the bundled `codex` CL
 </div>
 
 - Runtime dependency-free: uses only the Python standard library.
-- Codex CLI binaries are downloaded separately; use `scripts/setup_binary.py` from the repo or install the Codex CLI and set `codex_path_override`.
+- Codex CLI binaries can be sourced either from the vendored `src/codex_sdk/vendor/` tree (when present) or from a system-installed `codex` binary (`codex_path_override` supported).
 - Async-first API with sync helpers, streaming events, and structured output.
 - Python 3.8/3.9 support is deprecated and will be removed in a future release; use Python 3.10+.
 
@@ -168,6 +168,8 @@ python examples/app_server_turn_session.py
 python examples/config_overrides.py
 python examples/hooks_streaming.py
 python examples/notify_hook.py
+python examples/pydantic_ai_run_compat.py
+python examples/pydantic_ai_run_stream_compat.py
 ```
 
 <a id="features"></a>
@@ -185,6 +187,15 @@ python examples/notify_hook.py
 | ![Abort](https://img.shields.io/badge/Abort-Signals-ef4444?style=flat&logo=gnubash&logoColor=white)                | Cancel running turns via `AbortController` and `AbortSignal`.           |
 | ![Telemetry](https://img.shields.io/badge/Telemetry-Logfire%20Spans-f97316?style=flat&logo=simpleicons&logoColor=white) | Optional spans if Logfire is installed and initialized.                 |
 
+### Execution surfaces and transports
+
+| Surface | Primary transport | Typical use |
+| ------- | ----------------- | ----------- |
+| `Codex` + `Thread` | `codex exec --experimental-json` (JSONL over stdio) | Conversational turns, streaming item events, structured output (`run_json` / `run_pydantic`). |
+| `AppServerClient` | `codex app-server` (JSON-RPC over stdio) | Full app-server protocol: thread management, turn sessions, approvals, account/config/skills APIs. |
+| `CodexModel` (PydanticAI provider) | `codex app-server` (JSON-RPC over stdio) | Tool-call planning + text generation through app-server turn sessions with incremental stream events. |
+| `codex_handoff_tool` (PydanticAI tool) | `Thread.run(...)` on top of `codex exec` | Delegate repository-aware tasks from another model to Codex. |
+
 <a id="configuration"></a>
 ## ![Configuration](https://img.shields.io/badge/Configuration-Options%20%26%20Env-0ea5e9?style=for-the-badge&logo=json&logoColor=white)
 
@@ -312,6 +323,10 @@ For richer integrations (thread fork, requirements, explicit skill input), use t
 protocol. The client handles the initialize/initialized handshake and gives you access to
 JSON-RPC notifications.
 
+`AppServerClient` starts `codex app-server` and multiplexes notifications + server-initiated
+requests (for example approval requests in `turn_session(...)`) on top of the same stdio
+connection.
+
 ```python
 import asyncio
 from codex_sdk import AppServerClient, AppServerOptions
@@ -427,6 +442,11 @@ If you are working from source and the vendor directory is missing, run
 `python scripts/setup_binary.py` to fetch and assemble the platform `@openai/codex`
 artifacts into `src/codex_sdk/vendor/`.
 
+`scripts/setup_binary.py` behavior:
+- Resolves `CODEX_NPM_VERSION` when set; otherwise resolves latest `@openai/codex`.
+- Tries legacy `@openai/codex-sdk` package first, then falls back to per-target `@openai/codex@<version>-<suffix>` artifacts.
+- Rebuilds `src/codex_sdk/vendor/` for all supported targets and validates current-platform binary presence.
+
 <a id="auth"></a>
 ## ![Auth](https://img.shields.io/badge/Auth%20%26%20Credentials-Access-2563eb?style=for-the-badge&logo=gnubash&logoColor=white)
 
@@ -633,6 +653,8 @@ Example scripts under `examples/`:
 - `notify_hook.py`: notify script for CLI callbacks.
 - `pydantic_ai_model_provider.py`: Codex as a PydanticAI model provider.
 - `pydantic_ai_handoff.py`: Codex as a PydanticAI tool.
+- `pydantic_ai_run_compat.py`: minimal compatibility harness for `Agent.run(...)`.
+- `pydantic_ai_run_stream_compat.py`: minimal compatibility harness for `Agent.run_stream(...)`.
 
 <a id="sandbox"></a>
 ## ![Sandbox](https://img.shields.io/badge/Sandbox-Permissions%20%26%20Safety-1f2937?style=for-the-badge&logo=gnubash&logoColor=white)
@@ -655,6 +677,9 @@ Additional controls:
 
 This SDK offers two ways to integrate with PydanticAI:
 
+Note: starting in `0.105.0`, `CodexModel` is app-server-only and no longer supports
+the legacy `CodexModel(codex=...)` fallback path.
+
 ### 1) Codex as a PydanticAI model provider
 
 Use `CodexModel` to delegate tool-call planning and text generation to Codex, while PydanticAI executes tools and validates outputs.
@@ -682,11 +707,20 @@ print(result.output)
 ```
 
 How it works:
+- `CodexModel` keeps a persistent app-server session and starts/reuses threads based on
+  `thread_reuse_mode` (`"run"` by default, `"always"` optional).
 - `CodexModel` builds a JSON schema envelope with `tool_calls` and `final`.
 - Codex emits tool calls as JSON strings; PydanticAI runs them.
 - If `allow_text_output` is true, Codex can place final text in `final`.
-- Streaming APIs (`Agent.run_stream_events()`, `Agent.run_stream_sync()`) are supported; Codex
-  emits streamed responses as a single chunk once the turn completes.
+- Streaming APIs emit incremental events while app-server turn notifications arrive.
+
+Provider options:
+- `performance_profile`: `"balanced"` (default) or `"max"`.
+- `thread_reuse_mode`: `"run"` (default) or `"always"`.
+
+Lifecycle:
+- `CodexModel` owns an app-server client by default; call `await model.close()` in
+  long-lived processes to release resources.
 
 Safety defaults (you can override with your own `ThreadOptions`):
 - `sandbox_mode="read-only"`
@@ -749,38 +783,49 @@ If `logfire` is installed and initialized, the SDK emits spans:
 If Logfire is missing or not initialized, the span context manager is a no-op.
 
 <a id="architecture"></a>
-<a id="acheature"></a>
 ## ![Architecture](https://img.shields.io/badge/Architecture-Stack%20map-1f2937?style=for-the-badge&logo=serverless&logoColor=white)
 
-### System components
+### System components and transport paths
 
 ```mermaid
 flowchart LR
   subgraph App[Your Python App]
     U[User Code]
-    T[Thread API]
+    T[Codex + Thread API]
+    A[AppServerClient API]
+    PM[PydanticAI CodexModel]
+    PH[PydanticAI codex_handoff_tool]
   end
 
   subgraph SDK[Codex SDK]
     C[Codex]
     E[CodexExec]
-    P[Event Parser]
+    P[Thread Event Parser]
+    R[JSON-RPC Client]
   end
 
-  subgraph CLI[Bundled Codex CLI]
+  subgraph CLI[Codex CLI]
     X["codex exec --experimental-json"]
+    S["codex app-server"]
   end
 
   FS[(Filesystem)]
   NET[(Network)]
 
   U --> T --> C --> E --> X
+  U --> A --> R --> S
+  U --> PM --> R
+  U --> PH --> C
   X -->|JSONL events| P --> T
+  S -->|JSON-RPC messages| R --> A
+  S -->|JSON-RPC messages| R --> PM
   X --> FS
   X --> NET
+  S --> FS
+  S --> NET
 ```
 
-### Streaming event lifecycle
+### Thread API streaming lifecycle (`codex exec`)
 
 ```mermaid
 sequenceDiagram
@@ -803,21 +848,41 @@ sequenceDiagram
   Thread-->>Dev: turn.completed / turn.failed
 ```
 
-### PydanticAI model-provider loop
+### App-server lifecycle (`codex app-server`)
+
+```mermaid
+sequenceDiagram
+  participant Dev as Developer
+  participant App as AppServerClient
+  participant CLI as codex app-server
+
+  Dev->>App: start()
+  App->>CLI: spawn process
+  App->>CLI: initialize
+  CLI-->>App: initialize result
+  App->>CLI: initialized (notification)
+  Dev->>App: turn_session(thread_id, input)
+  App->>CLI: turn/start
+  CLI-->>App: thread/turn/item notifications
+  CLI-->>App: request (approval needed)
+  App-->>CLI: response (approve/reject)
+  CLI-->>App: turn completed notification
+  App-->>Dev: final turn payload
+```
+
+### PydanticAI model-provider loop (current implementation)
 
 ```mermaid
 sequenceDiagram
   participant Agent as PydanticAI Agent
   participant Model as CodexModel
-  participant SDK as Codex SDK
-  participant CLI as codex exec
+  participant App as Codex app-server
   participant Tools as User Tools
 
   Agent->>Model: request(messages, tools)
-  Model->>SDK: start_thread + run_json(prompt, output_schema)
-  SDK->>CLI: codex exec --output-schema
-  CLI-->>SDK: JSON envelope {tool_calls, final}
-  SDK-->>Model: ParsedTurn
+  Model->>App: thread/start (if needed)
+  Model->>App: turn/session(input + output_schema envelope)
+  App-->>Model: item/updated + turn/completed notifications
   alt tool_calls present
     Model-->>Agent: ToolCallPart(s)
     Agent->>Tools: execute tool(s)
@@ -833,12 +898,18 @@ sequenceDiagram
 flowchart LR
   Agent[PydanticAI Agent] --> Tool[codex_handoff_tool]
   Tool --> SDK[Codex SDK Thread]
-  SDK --> CLI[Codex CLI]
+  SDK --> CLI[codex exec]
   CLI --> SDK
   SDK --> Tool
   Tool --> Agent
 ```
 
+### Streaming semantics summary
+
+- `Thread.run_streamed()` and `Thread.run_streamed_events()` stream incremental `ThreadEvent` values as JSONL lines arrive from `codex exec`.
+- `AppServerClient.notifications()` streams incremental JSON-RPC notifications from `codex app-server`.
+- `CodexModel.request_stream(...)` emits incremental events from app-server turn notifications, with compatibility handling for current and legacy PydanticAI stream interfaces.
+
 <a id="testing"></a>
 ## ![Testing](https://img.shields.io/badge/Testing-Pytest%20%26%20Coverage-2563eb?style=for-the-badge&logo=pytest&logoColor=white)
 
@@ -847,9 +918,12 @@ This repo uses unit tests with mocked CLI processes to keep the test suite fast
 Test focus areas:
 - `tests/test_exec.py`: CLI invocation, environment handling, config flags, abort behavior.
 - `tests/test_thread.py`: parsing, streaming, JSON schema, Pydantic validation, input normalization.
+- `tests/test_app_server.py` and `tests/test_app_server_session.py`: JSON-RPC lifecycle, method wrappers, turn-session approvals/notifications.
 - `tests/test_codex.py`: resume helpers and option wiring.
+- `tests/test_config_overrides.py`: `--config` serialization and merge behavior.
 - `tests/test_abort.py`: abort signal semantics.
 - `tests/test_telemetry.py`: Logfire span behavior.
+- `tests/test_tool_envelope.py`: tool envelope parsing and schema normalization helpers.
 - `tests/test_pydantic_ai_*`: PydanticAI model provider and handoff integration.
 
 ### Run tests
@@ -891,10 +965,21 @@ uv run mypy src
 ## ![CI/CD](https://img.shields.io/badge/CI%2FCD-Overview-1F4B99?style=for-the-badge&logo=gnubash&logoColor=white)
 
 This repository includes GitHub Actions workflows under `.github/workflows/`.
-The CI pipeline runs linting, type checks, and `pytest --cov=codex_sdk`.
-Release automation creates GitHub releases from `CHANGELOG_SDK.md` when you push a
-`vX.Y.Z` tag or manually dispatch the workflow, then the publish workflow uploads
-the package to PyPI on release publish.
+
+Current workflows:
+- `ci.yml`: lint (`black`, `isort`, `flake8`), type-check (`mypy`), and tests (`pytest --cov=codex_sdk`) on Python 3.10/3.11/3.12.
+- `release.yml`: extracts release notes from `CHANGELOG_SDK.md` and creates a GitHub Release for `vX.Y.Z`.
+- `publish.yml`: builds (`uv build --no-sources`) and publishes to PyPI on release publish (or successful release workflow).
+- `docs-deploy.yml`: optional docs deployment (gated by `DOCS_DEPLOY_ENABLED`).
+- `uv-lock-update.yml`: scheduled weekly lockfile refresh PR.
+
+`ci.yml` test jobs install Node.js and run:
+
+```bash
+python scripts/setup_binary.py
+```
+
+before `pytest`, so test runs do not depend on committed vendor binary updates.
 
 <a id="operations"></a>
 ## ![Operations](https://img.shields.io/badge/Operations-Health%20%26%20Sessions-10b981?style=for-the-badge&logo=serverless&logoColor=white)

{codex_sdk_python-0.104.0 → codex_sdk_python-0.105.0}/README.md

@@ -8,7 +8,7 @@ Embed the Codex agent in Python workflows. This SDK wraps the bundled `codex` CL
       <td><strong>Lifecycle</strong></td>
       <td>
         <a href="#ci-cd"><img src="https://img.shields.io/badge/CI%2FCD-Active-16a34a?style=flat&logo=githubactions&logoColor=white" alt="CI/CD badge" /></a>
-        <img src="https://img.shields.io/badge/Release-0.104.0-6b7280?style=flat&logo=pypi&logoColor=white" alt="Release 0.104.0 badge" />
+        <img src="https://img.shields.io/badge/Release-0.105.0-6b7280?style=flat&logo=pypi&logoColor=white" alt="Release 0.105.0 badge" />
         <a href="#license"><img src="https://img.shields.io/badge/License-Apache--2.0-0f766e?style=flat&logo=apache&logoColor=white" alt="License badge" /></a>
       </td>
     </tr>
@@ -37,7 +37,7 @@ Embed the Codex agent in Python workflows. This SDK wraps the bundled `codex` CL
 </div>
 
 - Runtime dependency-free: uses only the Python standard library.
-- Codex CLI binaries are downloaded separately; use `scripts/setup_binary.py` from the repo or install the Codex CLI and set `codex_path_override`.
+- Codex CLI binaries can be sourced either from the vendored `src/codex_sdk/vendor/` tree (when present) or from a system-installed `codex` binary (`codex_path_override` supported).
 - Async-first API with sync helpers, streaming events, and structured output.
 - Python 3.8/3.9 support is deprecated and will be removed in a future release; use Python 3.10+.
 
@@ -132,6 +132,8 @@ python examples/app_server_turn_session.py
 python examples/config_overrides.py
 python examples/hooks_streaming.py
 python examples/notify_hook.py
+python examples/pydantic_ai_run_compat.py
+python examples/pydantic_ai_run_stream_compat.py
 ```
 
 <a id="features"></a>
@@ -149,6 +151,15 @@ python examples/notify_hook.py
 | ![Abort](https://img.shields.io/badge/Abort-Signals-ef4444?style=flat&logo=gnubash&logoColor=white)                | Cancel running turns via `AbortController` and `AbortSignal`.           |
 | ![Telemetry](https://img.shields.io/badge/Telemetry-Logfire%20Spans-f97316?style=flat&logo=simpleicons&logoColor=white) | Optional spans if Logfire is installed and initialized.                 |
 
+### Execution surfaces and transports
+
+| Surface | Primary transport | Typical use |
+| ------- | ----------------- | ----------- |
+| `Codex` + `Thread` | `codex exec --experimental-json` (JSONL over stdio) | Conversational turns, streaming item events, structured output (`run_json` / `run_pydantic`). |
+| `AppServerClient` | `codex app-server` (JSON-RPC over stdio) | Full app-server protocol: thread management, turn sessions, approvals, account/config/skills APIs. |
+| `CodexModel` (PydanticAI provider) | `codex app-server` (JSON-RPC over stdio) | Tool-call planning + text generation through app-server turn sessions with incremental stream events. |
+| `codex_handoff_tool` (PydanticAI tool) | `Thread.run(...)` on top of `codex exec` | Delegate repository-aware tasks from another model to Codex. |
+
 <a id="configuration"></a>
 ## ![Configuration](https://img.shields.io/badge/Configuration-Options%20%26%20Env-0ea5e9?style=for-the-badge&logo=json&logoColor=white)
 
@@ -276,6 +287,10 @@ For richer integrations (thread fork, requirements, explicit skill input), use t
 protocol. The client handles the initialize/initialized handshake and gives you access to
 JSON-RPC notifications.
 
+`AppServerClient` starts `codex app-server` and multiplexes notifications + server-initiated
+requests (for example approval requests in `turn_session(...)`) on top of the same stdio
+connection.
+
 ```python
 import asyncio
 from codex_sdk import AppServerClient, AppServerOptions
@@ -391,6 +406,11 @@ If you are working from source and the vendor directory is missing, run
 `python scripts/setup_binary.py` to fetch and assemble the platform `@openai/codex`
 artifacts into `src/codex_sdk/vendor/`.
 
+`scripts/setup_binary.py` behavior:
+- Resolves `CODEX_NPM_VERSION` when set; otherwise resolves latest `@openai/codex`.
+- Tries legacy `@openai/codex-sdk` package first, then falls back to per-target `@openai/codex@<version>-<suffix>` artifacts.
+- Rebuilds `src/codex_sdk/vendor/` for all supported targets and validates current-platform binary presence.
+
 <a id="auth"></a>
 ## ![Auth](https://img.shields.io/badge/Auth%20%26%20Credentials-Access-2563eb?style=for-the-badge&logo=gnubash&logoColor=white)
 
@@ -597,6 +617,8 @@ Example scripts under `examples/`:
 - `notify_hook.py`: notify script for CLI callbacks.
 - `pydantic_ai_model_provider.py`: Codex as a PydanticAI model provider.
 - `pydantic_ai_handoff.py`: Codex as a PydanticAI tool.
+- `pydantic_ai_run_compat.py`: minimal compatibility harness for `Agent.run(...)`.
+- `pydantic_ai_run_stream_compat.py`: minimal compatibility harness for `Agent.run_stream(...)`.
 
 <a id="sandbox"></a>
 ## ![Sandbox](https://img.shields.io/badge/Sandbox-Permissions%20%26%20Safety-1f2937?style=for-the-badge&logo=gnubash&logoColor=white)
@@ -619,6 +641,9 @@ Additional controls:
 
 This SDK offers two ways to integrate with PydanticAI:
 
+Note: starting in `0.105.0`, `CodexModel` is app-server-only and no longer supports
+the legacy `CodexModel(codex=...)` fallback path.
+
 ### 1) Codex as a PydanticAI model provider
 
 Use `CodexModel` to delegate tool-call planning and text generation to Codex, while PydanticAI executes tools and validates outputs.
@@ -646,11 +671,20 @@ print(result.output)
 ```
 
 How it works:
+- `CodexModel` keeps a persistent app-server session and starts/reuses threads based on
+  `thread_reuse_mode` (`"run"` by default, `"always"` optional).
 - `CodexModel` builds a JSON schema envelope with `tool_calls` and `final`.
 - Codex emits tool calls as JSON strings; PydanticAI runs them.
 - If `allow_text_output` is true, Codex can place final text in `final`.
-- Streaming APIs (`Agent.run_stream_events()`, `Agent.run_stream_sync()`) are supported; Codex
-  emits streamed responses as a single chunk once the turn completes.
+- Streaming APIs emit incremental events while app-server turn notifications arrive.
+
+Provider options:
+- `performance_profile`: `"balanced"` (default) or `"max"`.
+- `thread_reuse_mode`: `"run"` (default) or `"always"`.
+
+Lifecycle:
+- `CodexModel` owns an app-server client by default; call `await model.close()` in
+  long-lived processes to release resources.
 
 Safety defaults (you can override with your own `ThreadOptions`):
 - `sandbox_mode="read-only"`
@@ -713,38 +747,49 @@ If `logfire` is installed and initialized, the SDK emits spans:
 If Logfire is missing or not initialized, the span context manager is a no-op.
 
 <a id="architecture"></a>
-<a id="acheature"></a>
 ## ![Architecture](https://img.shields.io/badge/Architecture-Stack%20map-1f2937?style=for-the-badge&logo=serverless&logoColor=white)
 
-### System components
+### System components and transport paths
 
 ```mermaid
 flowchart LR
   subgraph App[Your Python App]
     U[User Code]
-    T[Thread API]
+    T[Codex + Thread API]
+    A[AppServerClient API]
+    PM[PydanticAI CodexModel]
+    PH[PydanticAI codex_handoff_tool]
   end
 
   subgraph SDK[Codex SDK]
     C[Codex]
     E[CodexExec]
-    P[Event Parser]
+    P[Thread Event Parser]
+    R[JSON-RPC Client]
   end
 
-  subgraph CLI[Bundled Codex CLI]
+  subgraph CLI[Codex CLI]
     X["codex exec --experimental-json"]
+    S["codex app-server"]
   end
 
   FS[(Filesystem)]
   NET[(Network)]
 
   U --> T --> C --> E --> X
+  U --> A --> R --> S
+  U --> PM --> R
+  U --> PH --> C
   X -->|JSONL events| P --> T
+  S -->|JSON-RPC messages| R --> A
+  S -->|JSON-RPC messages| R --> PM
   X --> FS
   X --> NET
+  S --> FS
+  S --> NET
 ```
 
-### Streaming event lifecycle
+### Thread API streaming lifecycle (`codex exec`)
 
 ```mermaid
 sequenceDiagram
@@ -767,21 +812,41 @@ sequenceDiagram
   Thread-->>Dev: turn.completed / turn.failed
 ```
 
-### PydanticAI model-provider loop
+### App-server lifecycle (`codex app-server`)
+
+```mermaid
+sequenceDiagram
+  participant Dev as Developer
+  participant App as AppServerClient
+  participant CLI as codex app-server
+
+  Dev->>App: start()
+  App->>CLI: spawn process
+  App->>CLI: initialize
+  CLI-->>App: initialize result
+  App->>CLI: initialized (notification)
+  Dev->>App: turn_session(thread_id, input)
+  App->>CLI: turn/start
+  CLI-->>App: thread/turn/item notifications
+  CLI-->>App: request (approval needed)
+  App-->>CLI: response (approve/reject)
+  CLI-->>App: turn completed notification
+  App-->>Dev: final turn payload
+```
+
+### PydanticAI model-provider loop (current implementation)
 
 ```mermaid
 sequenceDiagram
   participant Agent as PydanticAI Agent
   participant Model as CodexModel
-  participant SDK as Codex SDK
-  participant CLI as codex exec
+  participant App as Codex app-server
   participant Tools as User Tools
 
   Agent->>Model: request(messages, tools)
-  Model->>SDK: start_thread + run_json(prompt, output_schema)
-  SDK->>CLI: codex exec --output-schema
-  CLI-->>SDK: JSON envelope {tool_calls, final}
-  SDK-->>Model: ParsedTurn
+  Model->>App: thread/start (if needed)
+  Model->>App: turn/session(input + output_schema envelope)
+  App-->>Model: item/updated + turn/completed notifications
   alt tool_calls present
     Model-->>Agent: ToolCallPart(s)
     Agent->>Tools: execute tool(s)
@@ -797,12 +862,18 @@ sequenceDiagram
 flowchart LR
   Agent[PydanticAI Agent] --> Tool[codex_handoff_tool]
   Tool --> SDK[Codex SDK Thread]
-  SDK --> CLI[Codex CLI]
+  SDK --> CLI[codex exec]
   CLI --> SDK
   SDK --> Tool
   Tool --> Agent
 ```
 
+### Streaming semantics summary
+
+- `Thread.run_streamed()` and `Thread.run_streamed_events()` stream incremental `ThreadEvent` values as JSONL lines arrive from `codex exec`.
+- `AppServerClient.notifications()` streams incremental JSON-RPC notifications from `codex app-server`.
+- `CodexModel.request_stream(...)` emits incremental events from app-server turn notifications, with compatibility handling for current and legacy PydanticAI stream interfaces.
+
 <a id="testing"></a>
 ## ![Testing](https://img.shields.io/badge/Testing-Pytest%20%26%20Coverage-2563eb?style=for-the-badge&logo=pytest&logoColor=white)
 
@@ -811,9 +882,12 @@ This repo uses unit tests with mocked CLI processes to keep the test suite fast
 Test focus areas:
 - `tests/test_exec.py`: CLI invocation, environment handling, config flags, abort behavior.
 - `tests/test_thread.py`: parsing, streaming, JSON schema, Pydantic validation, input normalization.
+- `tests/test_app_server.py` and `tests/test_app_server_session.py`: JSON-RPC lifecycle, method wrappers, turn-session approvals/notifications.
 - `tests/test_codex.py`: resume helpers and option wiring.
+- `tests/test_config_overrides.py`: `--config` serialization and merge behavior.
 - `tests/test_abort.py`: abort signal semantics.
 - `tests/test_telemetry.py`: Logfire span behavior.
+- `tests/test_tool_envelope.py`: tool envelope parsing and schema normalization helpers.
 - `tests/test_pydantic_ai_*`: PydanticAI model provider and handoff integration.
 
 ### Run tests
@@ -855,10 +929,21 @@ uv run mypy src
 ## ![CI/CD](https://img.shields.io/badge/CI%2FCD-Overview-1F4B99?style=for-the-badge&logo=gnubash&logoColor=white)
 
 This repository includes GitHub Actions workflows under `.github/workflows/`.
-The CI pipeline runs linting, type checks, and `pytest --cov=codex_sdk`.
-Release automation creates GitHub releases from `CHANGELOG_SDK.md` when you push a
-`vX.Y.Z` tag or manually dispatch the workflow, then the publish workflow uploads
-the package to PyPI on release publish.
+
+Current workflows:
+- `ci.yml`: lint (`black`, `isort`, `flake8`), type-check (`mypy`), and tests (`pytest --cov=codex_sdk`) on Python 3.10/3.11/3.12.
+- `release.yml`: extracts release notes from `CHANGELOG_SDK.md` and creates a GitHub Release for `vX.Y.Z`.
+- `publish.yml`: builds (`uv build --no-sources`) and publishes to PyPI on release publish (or successful release workflow).
+- `docs-deploy.yml`: optional docs deployment (gated by `DOCS_DEPLOY_ENABLED`).
+- `uv-lock-update.yml`: scheduled weekly lockfile refresh PR.
+
+`ci.yml` test jobs install Node.js and run:
+
+```bash
+python scripts/setup_binary.py
+```
+
+before `pytest`, so test runs do not depend on committed vendor binary updates.
 
 <a id="operations"></a>
 ## ![Operations](https://img.shields.io/badge/Operations-Health%20%26%20Sessions-10b981?style=for-the-badge&logo=serverless&logoColor=white)

{codex_sdk_python-0.104.0 → codex_sdk_python-0.105.0}/pyproject.toml

@@ -1,10 +1,10 @@
 [build-system]
-requires = ["uv_build>=0.9.21,<0.11.0"]
+requires = ["uv_build>=0.9.21,<0.10.0"]
 build-backend = "uv_build"
 
 [project]
 name = "codex-sdk-python"
-version = "0.104.0"
+version = "0.105.0"
 description = "Python SDK for the Codex CLI agent with async threads, streaming events, and structured outputs"
 readme = "README.md"
 license = {text = "Apache-2.0"}

{codex_sdk_python-0.104.0 → codex_sdk_python-0.105.0}/src/codex_sdk/__init__.py

@@ -79,7 +79,7 @@ from .thread import (
     Turn,
 )
 
-__version__ = "0.104.0"
+__version__ = "0.105.0"
 
 __all__ = [
     "AbortController",