codex-a2a 1.3.5__tar.gz → 1.4.0__tar.gz

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/.github/workflows/publish.yml

@@ -55,8 +55,11 @@ jobs:
         run: |
           audit_cache_dir="$(mktemp -d)"
           trap 'rm -rf "${audit_cache_dir}"' EXIT
+          # The export is complete and hash-pinned; avoid a second pip resolver pass.
           XDG_CACHE_HOME="${audit_cache_dir}" \
-            uv run pip-audit --requirement /tmp/runtime-requirements.txt
+            uv run pip-audit \
+            --disable-pip \
+            --requirement /tmp/runtime-requirements.txt
 
       - name: Resolve release tag
         id: release_tag

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/CONTRIBUTING.md

@@ -72,6 +72,8 @@ bash -n scripts/smoke_test_built_cli.sh
 
 If `pre-commit` rewrites files, review the rewritten output and re-run the checks until the working tree is clean.
 
+Do not silence signals with global suppressions. `pytest` `filterwarnings` `ignore:` entries and `mypy` `follow_imports = silent` hide real failures or drift; scope the fix instead (for example `[mypy-<package>.*] ignore_missing_imports = True`) so the gate keeps reporting everything else.
+
 Ruff's version is declared in the `dev` extra in `pyproject.toml` and resolved in `uv.lock`. The local pre-commit hooks run `uv run --locked --extra dev ruff`, so direct commands, hooks, and CI use the same project dependency. To upgrade Ruff, update that dependency and regenerate `uv.lock`; there is no separate Ruff hook revision to maintain.
 
 ## Compatibility Expectations

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codex-a2a
-Version: 1.3.5
+Version: 1.4.0
 Summary: Codex A2A runtime adapter
 Author: liujuanjuan1984@Intelligent-Internet
 License-Expression: Apache-2.0
@@ -23,7 +23,7 @@ Classifier: Topic :: Internet :: WWW/HTTP
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: a2a-sdk==1.1.5
+Requires-Dist: a2a-sdk==1.2.3
 Requires-Dist: aiosqlite<1.0,>=0.20
 Requires-Dist: fastapi<0.142,>=0.121
 Requires-Dist: httpx<0.29,>=0.28.1
@@ -41,8 +41,8 @@ Requires-Dist: pytest>=8.0; extra == "dev"
 Requires-Dist: pytest-asyncio>=0.23; extra == "dev"
 Requires-Dist: pytest-cov<8.0,>=5.0; extra == "dev"
 Requires-Dist: diff-cover<11,>=9; extra == "dev"
-Requires-Dist: ruff==0.16.9; extra == "dev"
-Requires-Dist: mypy<2.4,>=1.11; extra == "dev"
+Requires-Dist: ruff==0.17.2; extra == "dev"
+Requires-Dist: mypy<2.5,>=1.11; extra == "dev"
 Dynamic: license-file
 
 # codex-a2a
@@ -222,7 +222,7 @@ The normative compatibility split and deployment model live in [Compatibility Gu
 ### Normative Upstreams
 
 - **[A2A Protocol](https://github.com/a2aproject/A2A)**: The Linux Foundation open protocol implemented by this adapter at the declared A2A `1.0` line.
-- **[a2aproject/a2a-python](https://github.com/a2aproject/a2a-python)**: The official Apache-2.0 Python SDK; this repository pins the `a2a-sdk` package to `1.1.2`.
+- **[a2aproject/a2a-python](https://github.com/a2aproject/a2a-python)**: The official Apache-2.0 Python SDK; this repository pins the `a2a-sdk` package to `1.2.3`.
 - **[openai/codex](https://github.com/openai/codex)**: The Apache-2.0 Codex CLI repository containing the local App Server used by this adapter. OpenAI's [open-source component list](https://developers.openai.com/codex/open-source) identifies the IDE extension and Codex cloud as not open source. Desktop and web product behavior is also separate from this local runtime and outside this repository's compatibility claim.
 
 ### Community Gateways and Hubs

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/README.md

@@ -175,7 +175,7 @@ The normative compatibility split and deployment model live in [Compatibility Gu
 ### Normative Upstreams
 
 - **[A2A Protocol](https://github.com/a2aproject/A2A)**: The Linux Foundation open protocol implemented by this adapter at the declared A2A `1.0` line.
-- **[a2aproject/a2a-python](https://github.com/a2aproject/a2a-python)**: The official Apache-2.0 Python SDK; this repository pins the `a2a-sdk` package to `1.1.2`.
+- **[a2aproject/a2a-python](https://github.com/a2aproject/a2a-python)**: The official Apache-2.0 Python SDK; this repository pins the `a2a-sdk` package to `1.2.3`.
 - **[openai/codex](https://github.com/openai/codex)**: The Apache-2.0 Codex CLI repository containing the local App Server used by this adapter. OpenAI's [open-source component list](https://developers.openai.com/codex/open-source) identifies the IDE extension and Codex cloud as not open source. Desktop and web product behavior is also separate from this local runtime and outside this repository's compatibility claim.
 
 ### Community Gateways and Hubs

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/docs/compatibility.md

@@ -7,8 +7,8 @@ This document explains the compatibility promises this repository currently trie
 | Layer | Supported contract | Evidence and boundary |
 | --- | --- | --- |
 | Python | 3.11–3.14 | Package classifiers cover all four versions; CI runs 3.13 as the quality gate and 3.11, 3.12, and 3.14 in the runtime matrix. |
-| A2A Python SDK | Exactly `a2a-sdk==1.1.5` | Pinned in `pyproject.toml`; this is the [official A2A Python SDK](https://github.com/a2aproject/a2a-python), not a fork. |
-| Protocol Buffers | `protobuf>=6.33.5,<8.0` | The lockfile selects 7.36.2. SDK 1.1.5 supports protobuf 7; Google API Core, common protos, and proto-plus must resolve within the same range. Task retrieval delegates validation and history trimming to the SDK. |
+| A2A Python SDK | Exactly `a2a-sdk==1.2.3` | Pinned in `pyproject.toml`; this is the [official A2A Python SDK](https://github.com/a2aproject/a2a-python), not a fork. |
+| Protocol Buffers | `protobuf>=6.33.5,<8.0` | The lockfile selects 7.36.2. SDK 1.2.3 supports protobuf 7; Google API Core, common protos, and proto-plus must resolve within the same range. Task retrieval delegates validation and history trimming to the SDK. |
 | SQLAlchemy | `sqlalchemy[asyncio]>=2.0,<3.0` | The `asyncio` extra explicitly installs `greenlet` for the async persistence engine, including standalone wheel installations. The lockfile selects the development version. |
 | A2A protocol | `1.0` wire line | The released upstream A2A specification is normative for portable core semantics; the current compatible release is `v1.0.1`. SDK compatibility with other protocol versions does not extend this adapter's claim. |
 | Inbound A2A transports | JSON-RPC and HTTP+JSON | Both are repository-owned runtime surfaces. gRPC is not exposed by this adapter. |
@@ -74,13 +74,19 @@ Service-level behavior layered on top of those core methods should also be decla
 
 Task-store resilience is also service-level behavior in this deployment:
 
-- Once a task reaches a terminal state, later conflicting persistence writes are dropped on a first-terminal-state-wins basis.
-- In the default SQLite-backed deployment, terminal-task persistence is guarded with an atomic database upsert instead of a process-local read-before-write check.
+- A `SendMessage` that references a task already in a terminal state follows the SDK contract and raises `UnsupportedOperationError`; the task is not restarted in place.
+- Database-backed deployments use the SDK `VersionedDatabaseTaskStore`, `DatabaseTaskEventStream`, and their per-task compare-and-swap versioning; stale writes are rejected rather than overwriting the current task snapshot.
+- Memory-backed deployments use the SDK `CopyingTaskStoreAdapter` and `LegacyTaskStoreAdapter`; they are process-local and do not provide cross-replica version coordination.
 - Task-store I/O failures are surfaced as stable service errors instead of leaking backend-specific exceptions through request handlers.
 
+HTTP+JSON message endpoints accept the A2A 1.0 `message.parts` shape. A JSON-RPC envelope, legacy `message.content`, lowercase role values, or Part wrappers using `kind`, `type`, or `file` are rejected before SDK parsing rather than being partially accepted through unknown-field handling. JSON-RPC send methods apply the same message-shape guard and return `-32602` for invalid message fields.
+
+Successful HTTP+JSON responses use the SDK's `application/a2a+json` media type. REST error envelopes remain AIP-193 `application/json` payloads, and OpenAPI response declarations follow the same split.
+
 Task durability is deployment-dependent:
 
 - `A2A_DATABASE_URL=<sqlalchemy-async-url>` preserves task lookup/cancel/resubscribe state across process restarts.
+- SDK keyset page tokens remain usable after a task from the preceding page is deleted; memory and database regressions cover this pagination behavior.
 - `A2A_DATABASE_URL` now defaults to a workspace-local SQLite database at `${CODEX_WORKSPACE_ROOT}/.codex-a2a/codex-a2a.db` when `CODEX_WORKSPACE_ROOT` is configured; otherwise it falls back to `sqlite+aiosqlite:///./codex-a2a.db`. Persistence therefore remains the default runtime behavior.
 - The same database-backed mode also preserves session-binding ownership state and pending interrupt callback requests that still fall within their TTL. Session-binding and ownership persistence are independent from the in-memory session cache TTL.
 - The default SQLite + WAL profile is intended for a stable local filesystem path, not for network filesystems or sync-managed folders with weaker locking behavior.

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/docs/conformance-triage.md

@@ -5,7 +5,7 @@ This document is the standing triage template for local `./scripts/conformance.s
 ## Authorities Used For Triage
 
 - The latest compatible released upstream A2A specification is authoritative; currently this is `v1.0.1` for the repository's advertised `1.0` wire line.
-- `a2a-sdk==1.1.5` as installed in this repository.
+- `a2a-sdk==1.2.3` as installed in this repository.
 - The default A2A protocol version advertised by this repository: `1.0`.
 - TCK assertions are non-normative observations and may lag the released specification.
 - Repository compatibility policy:

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/docs/conformance.md

@@ -74,7 +74,7 @@ When a TCK run fails, inspect the raw report before changing the runtime:
 
 - Some failures may point to real runtime gaps.
 - Some failures may reflect an outdated embedded specification or a TCK-specific interpretation. Resolve those disagreements against the latest compatible released A2A specification.
-- Some failures may come from TCK assumptions that do not match `a2a-sdk==1.1.5`.
+- Some failures may come from TCK assumptions that do not match `a2a-sdk==1.2.3`.
 - Some failures may come from older A2A naming or schema expectations that no longer match the repository's `1.0` contract.
 - Some failures may be local experiment artifacts from the dummy-backed runtime.
 

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/docs/guide.md

@@ -9,6 +9,7 @@ This guide covers runtime configuration, transport contracts, streaming/session/
   - Shared JSON-RPC (`POST /`) for both core A2A methods and provider-private extension methods
 - Agent Card publishes both HTTP+JSON and JSON-RPC endpoints through `supported_interfaces[]`.
 - Both supported interfaces publish the same `public_url`. For HTTP+JSON, that URL is the REST service root: per A2A 1.0 the concrete REST methods (`/message:send`, `/tasks/...`) live at that root, and the protocol version is negotiated via the `A2A-Version` header rather than a URL prefix.
+- Successful HTTP+JSON responses use `application/a2a+json`; REST errors remain AIP-193 `application/json` envelopes.
 - The public Agent Card at `/.well-known/agent-card.json` is intentionally slimmed to the minimum discovery surface.
 - The authenticated extended card exposes the authenticated provider-private skill inventory and deployment-aware examples:
   - preferred: JSON-RPC `GetExtendedAgentCard`
@@ -541,7 +542,7 @@ On the current npm global install layout for Linux x64, the command above resolv
 - Successful terminal chat results are artifact-first. Non-streaming terminal `Task` objects carry the completed answer in `Task.artifacts`; streaming terminal status updates carry lifecycle state and metadata but do not repeat the final answer in `status.message`.
 - `status.message` remains reserved for user-facing status, interrupt, or error details. Clients should not treat it as the canonical completed-result carrier on successful terminal surfaces.
 - `SubscribeToTask` remains part of the core A2A method baseline, but this deployment's terminal-task replay-once policy is a declared service-level behavior rather than a generic A2A guarantee.
-- Task persistence also applies a service-level resilience policy: once a task has been durably stored in a terminal state, later conflicting writes are dropped instead of overwriting that terminal snapshot.
+- Database-backed task persistence uses the SDK's versioned compare-and-swap store so stale writers cannot overwrite a newer task snapshot.
 
 ### Streaming and Interrupt Contract
 
@@ -1280,7 +1281,7 @@ Terminal-task note:
 - If the task is already terminal, this service replays one final task snapshot and then closes the stream.
 - That replay-once terminal behavior is a service-level contract for this deployment. It is published through the compatibility profile and wire contract so clients do not mistake it for a generic A2A baseline rule.
 - If the task store is unavailable while loading subscribe state, the service returns a controlled failure instead of exposing backend exception details.
-- In the default SQLite-backed deployment, terminal-task persistence is also database-guarded so late conflicting writes are dropped without relying on a process-local stale snapshot check.
+- In the default SQLite-backed deployment, the SDK version table rejects stale writes without relying on a process-local snapshot check.
 
 ## Development Setup
 

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/mypy.ini

@@ -2,7 +2,6 @@
 python_version = 3.11
 files = src/codex_a2a, tests/client/test_agent_card.py, tests/client/test_errors.py, tests/client/test_lifecycle.py, tests/client/test_payload_text.py, tests/client/test_request_context.py, tests/config/test_settings.py, tests/contracts/test_extension_contract_consistency.py, tests/contracts/test_extension_registry.py, tests/execution/test_agent_errors.py, tests/execution/test_cancellation.py, tests/execution/test_directory_validation.py, tests/execution/test_metrics.py, tests/execution/test_request_metadata.py, tests/execution/test_session_persistence.py, tests/execution/test_session_runtime.py, tests/execution/test_stream_chunks.py, tests/execution/test_stream_interrupts.py, tests/execution/test_tool_call_payloads.py, tests/jsonrpc/test_codex_discovery_extension.py, tests/jsonrpc/test_codex_exec_extension.py, tests/jsonrpc/test_codex_session_extension.py, tests/jsonrpc/test_codex_thread_lifecycle_extension.py, tests/jsonrpc/test_codex_turn_review_extension.py, tests/jsonrpc/test_dispatch.py, tests/jsonrpc/test_guard_hooks.py, tests/jsonrpc/test_interrupt_params.py, tests/jsonrpc/test_jsonrpc_models.py, tests/jsonrpc/test_protocol_payload_mapping.py, tests/package/test_logging.py, tests/package/test_publish_github_release_script.py, tests/package/test_release_distribution_contract.py, tests/package/test_typing_contract.py, tests/package/test_version.py, tests/profile/test_profile.py, tests/scripts/test_audit_low_call_sites.py, tests/scripts/test_script_health_contract.py, tests/scripts/test_tck_auth_plugin.py, tests/server/test_agent_card.py, tests/server/test_auth.py, tests/server/test_call_context_builder.py, tests/server/test_cli.py, tests/server/test_database.py, tests/server/test_database_app_persistence.py, tests/server/test_migrations.py, tests/server/test_output_negotiation.py, tests/server/test_push_config_store.py, tests/server/test_request_handler.py, tests/server/test_runtime_limits.py, tests/server/test_runtime_state.py, tests/support/context.py, tests/support/dummy_clients.py, tests/support/fixtures.py, tests/support/http_auth.py, tests/support/jsonrpc_errors.py, tests/support/settings.py, tests/test_a2a_proto.py, tests/test_input_mapping.py, tests/test_protocol_versions.py, tests/upstream/test_interrupt_bridge.py, tests/upstream/test_interrupt_persistence.py, tests/upstream/test_request_mapping.py
 mypy_path = src
-follow_imports = silent
 warn_unused_ignores = True
 warn_redundant_casts = True
 show_error_codes = True
@@ -10,5 +9,8 @@ show_error_codes = True
 [mypy-a2a.*]
 ignore_missing_imports = True
 
+[mypy-tck.*]
+ignore_missing_imports = True
+
 [mypy-setuptools_scm]
 ignore_missing_imports = True

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/pyproject.toml

@@ -25,7 +25,7 @@ classifiers = [
     "Topic :: Internet :: WWW/HTTP",
 ]
 dependencies = [
-    "a2a-sdk==1.1.5",
+    "a2a-sdk==1.2.3",
     "aiosqlite>=0.20,<1.0",
     "fastapi>=0.121,<0.142",
     "httpx>=0.28.1,<0.29",
@@ -46,8 +46,8 @@ dev = [
     "pytest-asyncio>=0.23",
     "pytest-cov>=5.0,<8.0",
     "diff-cover>=9,<11",
-    "ruff==0.16.9",
-    "mypy>=1.11,<2.4",
+    "ruff==0.17.2",
+    "mypy>=1.11,<2.5",
 ]
 
 [project.scripts]
@@ -90,7 +90,6 @@ tag_regex = "^v?(?P<version>.+)$"
 [tool.pytest.ini_options]
 addopts = "--cov=src/codex_a2a --cov-branch --cov-report=term-missing:skip-covered --cov-report=xml"
 filterwarnings = [
-    "ignore:'HTTP_413_REQUEST_ENTITY_TOO_LARGE' is deprecated. Use 'HTTP_413_CONTENT_TOO_LARGE' instead.:DeprecationWarning:a2a\\.server\\.apps\\.jsonrpc\\.fastapi_app",
     "error::pytest.PytestUnhandledThreadExceptionWarning",
     "error::pytest.PytestUnraisableExceptionWarning",
     "always:.*was deleted before being closed.*:ResourceWarning:aiosqlite.core",

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/scripts/README.md

@@ -38,6 +38,7 @@ The `Publish` workflow now separates build, PyPI publish, and GitHub Release syn
 - `validate_baseline.sh` now includes a conservative private dead-code check before type-checking and test execution.
 - `audit_low_call_sites.py` is intentionally advisory only. Use it to surface low-call-count candidates, then inspect the actual consumer chain before removing an abstraction.
 - `validate_baseline.sh` now blocks on runtime dependency vulnerability audit, while `dependency_health.sh` remains focused on broader dependency review (`outdated` + dev audit).
+- Vulnerability audits consume complete hashed `uv export` requirements with `pip-audit --disable-pip`, avoiding a second pip resolver pass while retaining hash verification.
 - The baseline makes up to three runtime audit attempts (the initial attempt plus at most two retries), with a fresh cache per attempt. If every attempt fails, it returns the last audit exit code and stops before package build and wheel smoke testing. A successful retry allows validation to continue.
 - [`.github/dependabot.yml`](../.github/dependabot.yml) prefers a single weekly grouped Dependabot PR for `uv`, while the repository scripts remain the explicit audit and validation entrypoints.
 - End-user runtime startup does not use repository scripts. Prefer the published CLI command documented in [README.md](../README.md) and [docs/guide.md](../docs/guide.md).

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/scripts/dependency_health.sh

@@ -24,4 +24,7 @@ uv export \
   --output-file "${dev_requirements}" >/dev/null
 
 echo "[dependency-health] run dev dependency vulnerability audit"
-XDG_CACHE_HOME="${audit_cache_dir}" uv run pip-audit --requirement "${dev_requirements}"
+# uv export is complete and hash-pinned; do not run a second pip resolver pass.
+XDG_CACHE_HOME="${audit_cache_dir}" uv run pip-audit \
+  --disable-pip \
+  --requirement "${dev_requirements}"

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/scripts/validate_baseline.sh

@@ -16,7 +16,10 @@ run_pip_audit() {
 
   while true; do
     audit_cache_dir="$(mktemp -d)"
-    if XDG_CACHE_HOME="${audit_cache_dir}" uv run pip-audit --requirement "${requirement_file}"; then
+    # uv export is complete and hash-pinned; do not run a second pip resolver pass.
+    if XDG_CACHE_HOME="${audit_cache_dir}" uv run pip-audit \
+      --disable-pip \
+      --requirement "${requirement_file}"; then
       rm -rf "${audit_cache_dir}"
       return 0
     else

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/src/codex_a2a/a2a_proto.py

@@ -53,34 +53,6 @@ def new_data_part(data: Any) -> Part:
     return Part(data=ParseDict(data, Value()))
 
 
-def new_file_url_part(
-    url: str,
-    *,
-    media_type: str | None = None,
-    filename: str | None = None,
-) -> Part:
-    kwargs: dict[str, Any] = {"url": url}
-    if media_type is not None:
-        kwargs["media_type"] = media_type
-    if filename is not None:
-        kwargs["filename"] = filename
-    return Part(**kwargs)
-
-
-def new_file_bytes_part(
-    raw: bytes,
-    *,
-    media_type: str | None = None,
-    filename: str | None = None,
-) -> Part:
-    kwargs: dict[str, Any] = {"raw": raw}
-    if media_type is not None:
-        kwargs["media_type"] = media_type
-    if filename is not None:
-        kwargs["filename"] = filename
-    return Part(**kwargs)
-
-
 def part_kind(part: Part | None) -> str | None:
     if part is None:
         return None

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/src/codex_a2a/client/errors.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 import httpx
 from a2a.client import A2AClientError as SDKA2AClientError
 from a2a.client.errors import A2AClientTimeoutError, AgentCardResolutionError
+from google.protobuf.json_format import ParseError  # type: ignore[import-untyped]
 
 
 class A2AClientError(RuntimeError):
@@ -127,6 +128,9 @@ def map_a2a_sdk_error(
     if isinstance(exc, A2AClientTimeoutError):
         return A2AAgentUnavailableError(str(exc))
 
+    if isinstance(exc, ParseError):
+        return A2AClientProtocolError(str(exc))
+
     if isinstance(exc, SDKA2AClientError):
         return A2AClientProtocolError(str(exc))
 

{codex_a2a-1.3.5 → codex_a2a-1.4.0}/src/codex_a2a/server/application.py

@@ -2,20 +2,15 @@ from __future__ import annotations
 
 import logging
 from contextlib import AsyncExitStack, asynccontextmanager
-from typing import Any
+from typing import Any, cast
 
 import uvicorn
+from a2a.server.request_handlers.request_handler import RequestHandler
 from a2a.server.routes.agent_card_routes import create_agent_card_routes
-from a2a.server.routes.rest_dispatcher import RestDispatcher
 from a2a.server.routes.rest_routes import create_rest_routes
-from a2a.types import SubscribeToTaskRequest, a2a_pb2
-from a2a.utils import proto_utils
-from a2a.utils.error_handlers import build_rest_error_payload
 from fastapi import FastAPI
-from google.protobuf.json_format import MessageToDict, Parse  # type: ignore[import-untyped]
-from starlette.requests import Request
-from starlette.responses import JSONResponse, PlainTextResponse, Response
-from starlette.routing import BaseRoute, Mount, Route
+from starlette.responses import JSONResponse, PlainTextResponse
+from starlette.routing import BaseRoute, Mount
 
 from codex_a2a.client.manager import A2AClientManager
 from codex_a2a.config import Settings
@@ -29,7 +24,6 @@ from codex_a2a.jsonrpc.application import (
     CodexSessionQueryJSONRPCApplication,
     create_extension_jsonrpc_routes,
 )
-from codex_a2a.jsonrpc.errors import build_http_error_body
 from codex_a2a.jsonrpc.hooks import SessionGuardHooks
 from codex_a2a.logging_context import install_log_record_factory
 from codex_a2a.metrics import get_metrics_registry
@@ -47,7 +41,6 @@ from codex_a2a.server.runtime_limits import (
     OperationCapacity,
     OperationCapacityMiddleware,
     RequestBodyLimitMiddleware,
-    StreamBudgetExceeded,
     apply_stream_budget,
 )
 from codex_a2a.server.runtime_state import build_runtime_state_runtime
@@ -67,151 +60,66 @@ def _is_sdk_tenant_mount(route: BaseRoute) -> bool:
     return isinstance(route, Mount) and route.path == "/{tenant}"
 
 
-_STREAMING_REST_PATHS = frozenset({"/message:stream", "/tasks/{id}:subscribe"})
-_STREAM_BUDGET_REJECT_REASON = "STREAM_BUDGET_EXCEEDED"
-
-
-class BudgetedRestDispatcher(RestDispatcher):
-    """REST dispatcher that applies streaming output budgets to SSE streams."""
+class StreamBudgetedRequestHandler:
+    """Apply stream budgets through the public RequestHandler API."""
 
     def __init__(
         self,
-        request_handler: Any,
-        context_builder: Any,
+        request_handler: RequestHandler,
         *,
         stream_budget_max_bytes: int,
         stream_budget_max_duration_seconds: float,
         stream_budget_idle_timeout_seconds: float,
     ) -> None:
-        super().__init__(
-            request_handler=request_handler,
-            context_builder=context_builder,
-        )
+        self._request_handler = request_handler
         self._stream_budget_max_bytes = stream_budget_max_bytes
         self._stream_budget_max_duration_seconds = stream_budget_max_duration_seconds
         self._stream_budget_idle_timeout_seconds = stream_budget_idle_timeout_seconds
 
-    async def _handle_streaming(
-        self,
-        request: Request,
-        handler_func: Any,
-    ) -> Any:
-        async def budgeted_handler(context: Any) -> Any:
-            async for item in apply_stream_budget(
-                aiter(handler_func(context)),
-                max_bytes=self._stream_budget_max_bytes,
-                max_duration_seconds=self._stream_budget_max_duration_seconds,
-                idle_timeout_seconds=self._stream_budget_idle_timeout_seconds,
-            ):
-                yield item
-
-        return await super()._handle_streaming(request, budgeted_handler)
-
-
-def _stream_budget_error_response(error: StreamBudgetExceeded) -> JSONResponse:
-    return JSONResponse(
-        build_http_error_body(
-            status_code=429,
-            status="RESOURCE_EXHAUSTED",
-            message=str(error),
-            reason=_STREAM_BUDGET_REJECT_REASON,
-        ),
-        status_code=429,
-    )
+    def __getattr__(self, name: str) -> Any:
+        return getattr(self._request_handler, name)
 
+    def _apply_budget(self, stream: Any) -> Any:
+        return apply_stream_budget(
+            stream,
+            max_bytes=self._stream_budget_max_bytes,
+            max_duration_seconds=self._stream_budget_max_duration_seconds,
+            idle_timeout_seconds=self._stream_budget_idle_timeout_seconds,
+        )
+
+    async def on_message_send_stream(self, params: Any, context: Any) -> Any:
+        async for item in self._apply_budget(
+            self._request_handler.on_message_send_stream(params, context)
+        ):
+            yield item
 
-def _rest_error_response(error: Exception) -> JSONResponse:
-    """Map a pre-SSE error to the same REST error payload the SDK emits."""
-    payload = build_rest_error_payload(error)
-    http_code = payload.get("error", {}).get("code", 500)
-    return JSONResponse(content=payload, status_code=http_code)
+    async def on_subscribe_to_task(self, params: Any, context: Any) -> Any:
+        async for item in self._apply_budget(
+            self._request_handler.on_subscribe_to_task(params, context)
+        ):
+            yield item
 
 
 def _create_single_tenant_rest_routes(
     *,
-    request_handler: Any,
+    request_handler: RequestHandler,
     context_builder: Any = None,
     path_prefix: str = "",
     enable_v0_3_compat: bool = False,
-    stream_budget_max_bytes: int = 0,
-    stream_budget_max_duration_seconds: float = 0.0,
-    stream_budget_idle_timeout_seconds: float = 0.0,
 ) -> list[BaseRoute]:
     # The SDK exposes a tenant-prefixed REST alias by default. This service's
     # supported HTTP+JSON contract is the spec-rooted single-tenant surface
     # (A2A 1.0 resolves REST paths from the advertised interface URL, with no
     # version prefix in the URL), so the application assembly narrows the route
-    # set explicitly and replaces the streaming routes with budgeted ones.
+    # set explicitly. Streaming budgets are applied through the public
+    # RequestHandler wrapper passed by the caller.
     sdk_routes = create_rest_routes(
         request_handler=request_handler,
         context_builder=context_builder,
         enable_v0_3_compat=enable_v0_3_compat,
         path_prefix=path_prefix,
     )
-    base_routes = [route for route in sdk_routes if not _is_sdk_tenant_mount(route)]
-
-    dispatcher = BudgetedRestDispatcher(
-        request_handler=request_handler,
-        context_builder=context_builder,
-        stream_budget_max_bytes=stream_budget_max_bytes,
-        stream_budget_max_duration_seconds=stream_budget_max_duration_seconds,
-        stream_budget_idle_timeout_seconds=stream_budget_idle_timeout_seconds,
-    )
-
-    async def _handle_streaming_route(request: Request, handler_func: Any) -> Response:
-        try:
-            return await dispatcher._handle_streaming(request, handler_func)
-        except StreamBudgetExceeded as error:
-            return _stream_budget_error_response(error)
-        except Exception as error:  # noqa: BLE001 - mirrors SDK rest_stream_error_handler
-            return _rest_error_response(error)
-
-    async def message_stream_route(request: Request) -> Response:
-        async def _handler(context: Any) -> Any:
-            body = await request.body()
-            params = a2a_pb2.SendMessageRequest()
-            Parse(body, params)
-            async for event in request_handler.on_message_send_stream(params, context):
-                yield MessageToDict(proto_utils.to_stream_response(event))
-
-        return await _handle_streaming_route(request, _handler)
-
-    async def subscribe_route(request: Request) -> Response:
-        task_id = request.path_params["id"]
-
-        async def _handler(context: Any) -> Any:
-            async for event in request_handler.on_subscribe_to_task(
-                SubscribeToTaskRequest(id=task_id),
-                context,
-            ):
-                yield MessageToDict(proto_utils.to_stream_response(event))
-
-        return await _handle_streaming_route(request, _handler)
-
-    prefixed_streaming_paths = {f"{path_prefix}{path}" for path in _STREAMING_REST_PATHS}
-    retained = [
-        route
-        for route in base_routes
-        if not (hasattr(route, "path") and route.path in prefixed_streaming_paths)
-    ]
-    streaming_routes = [
-        Route(
-            path=f"{path_prefix}/message:stream",
-            endpoint=message_stream_route,
-            methods=["POST"],
-        ),
-        Route(
-            path=f"{path_prefix}/tasks/{{id}}:subscribe",
-            endpoint=subscribe_route,
-            methods=["GET"],
-        ),
-        Route(
-            path=f"{path_prefix}/tasks/{{id}}:subscribe",
-            endpoint=subscribe_route,
-            methods=["POST"],
-        ),
-    ]
-    return retained + streaming_routes
+    return [route for route in sdk_routes if not _is_sdk_tenant_mount(route)]
 
 
 def create_app(settings: Settings) -> FastAPI:
@@ -257,6 +165,7 @@ def create_app(settings: Settings) -> FastAPI:
         push_config_store=push_config_store_runtime.push_config_store,
         agent_card=agent_card,
         extended_agent_card=extended_agent_card,
+        event_stream=task_store_runtime.event_stream,
     )
     exec_runtime = CodexExecRuntime(
         client=client,
@@ -379,16 +288,22 @@ def create_app(settings: Settings) -> FastAPI:
             stream_budget_idle_timeout_seconds=settings.a2a_stream_idle_timeout_seconds,
         )
     )
+    rest_request_handler = cast(
+        RequestHandler,
+        StreamBudgetedRequestHandler(
+            handler,
+            stream_budget_max_bytes=settings.a2a_stream_max_bytes,
+            stream_budget_max_duration_seconds=settings.a2a_stream_max_duration_seconds,
+            stream_budget_idle_timeout_seconds=settings.a2a_stream_idle_timeout_seconds,
+        ),
+    )
     app.router.routes.extend(
         _create_single_tenant_rest_routes(
-            request_handler=handler,
+            request_handler=rest_request_handler,
             context_builder=context_builder,
             # A2A 1.0 roots the HTTP+JSON surface at the service root; the SDK
             # default prefix is empty, matching the URL the Agent Card advertises.
             path_prefix="",
-            stream_budget_max_bytes=settings.a2a_stream_max_bytes,
-            stream_budget_max_duration_seconds=settings.a2a_stream_max_duration_seconds,
-            stream_budget_idle_timeout_seconds=settings.a2a_stream_idle_timeout_seconds,
         )
     )
     app.state.codex_client = client
@@ -433,7 +348,6 @@ def create_app(settings: Settings) -> FastAPI:
     install_http_middlewares(
         app,
         settings=settings,
-        task_store=task_store,
         agent_card=agent_card,
         extended_agent_card=extended_agent_card,
     )