fix(xai): OAuth Responses native web_search, incomplete guard, grok-composer context

- model_metadata: grok-composer-2.5-fast → 262144 (OAuth slug not in /v1/models)
- codex transport: inject native {"type":"web_search"} for is_xai_responses;
  drop client web_search to avoid duplicate-name 400s
- codex adapter: do not treat in-progress server-side *_call items as incomplete
- tests: adapter, transport build_kwargs, model_metadata, oauth recovery

tests/agent/test_codex_responses_adapter.py

@@ -5,6 +5,7 @@ import pytest
 from agent.codex_responses_adapter import (
     _format_responses_error,
     _normalize_codex_response,
+    _preflight_codex_api_kwargs,
 )
 
 
@@ -68,6 +69,115 @@ def test_normalize_codex_response_treats_summary_only_reasoning_as_incomplete():
     assert assistant_message.codex_reasoning_items is None
 
 
+# ---------------------------------------------------------------------------
+# Server-side built-in tool calls (xAI native web_search, code interpreter,
+# etc.) come back as discrete ``*_call`` output items that xAI's
+# /v1/responses surface routinely leaves at ``status="in_progress"`` even
+# when the overall ``response.status == "completed"``.  These must NOT mark
+# the turn incomplete — otherwise grok-composer-2.5-fast research queries
+# (which invoke server-side web_search) get misclassified as
+# ``finish_reason="incomplete"`` and burn 3 fruitless continuation retries
+# before failing with "Codex response remained incomplete after 3
+# continuation attempts".  Observed live against grok-composer-2.5-fast on
+# SuperGrok OAuth (2026-06).
+# ---------------------------------------------------------------------------
+
+
+def test_normalize_codex_response_ignores_in_progress_server_side_tool_calls():
+    """A completed response with a final message + lingering in_progress
+    server-side web_search_call items resolves to 'stop', not 'incomplete'."""
+    response = SimpleNamespace(
+        status="completed",
+        incomplete_details=None,
+        output=[
+            SimpleNamespace(
+                type="reasoning",
+                id="rs_1",
+                encrypted_content="opaque",
+                summary=[SimpleNamespace(text="researching blades")],
+            ),
+            SimpleNamespace(
+                type="message",
+                role="assistant",
+                status="completed",
+                content=[SimpleNamespace(
+                    type="output_text",
+                    text="Milwaukee M18 blade 49-16-2734, ~$30 OEM.",
+                )],
+            ),
+            SimpleNamespace(type="web_search_call", status="in_progress"),
+            SimpleNamespace(type="web_search_call", status="in_progress"),
+            SimpleNamespace(type="web_search_call", status="in_progress"),
+        ],
+    )
+
+    assistant_message, finish_reason = _normalize_codex_response(response)
+
+    assert finish_reason == "stop"
+    assert assistant_message.content == "Milwaukee M18 blade 49-16-2734, ~$30 OEM."
+
+
+def test_normalize_codex_response_in_progress_message_still_incomplete():
+    """Guard scope: an in_progress *message* item (genuine model output that
+    is still streaming) must still mark the turn incomplete — only
+    server-side ``*_call`` items are exempted."""
+    response = SimpleNamespace(
+        status="completed",
+        incomplete_details=None,
+        output=[
+            SimpleNamespace(
+                type="message",
+                role="assistant",
+                status="in_progress",
+                content=[SimpleNamespace(type="output_text", text="partial...")],
+            ),
+        ],
+    )
+
+    _assistant_message, finish_reason = _normalize_codex_response(response)
+
+    assert finish_reason == "incomplete"
+
+
+# ---------------------------------------------------------------------------
+# _preflight_codex_api_kwargs — built-in (provider-executed) tools must pass
+# through validation.  Regression guard for the xAI native web_search
+# injection: the preflight validator previously rejected any tool whose
+# ``type != "function"`` with "unsupported type", which would 400 every xAI
+# turn once the native web_search tool is declared.
+# ---------------------------------------------------------------------------
+
+
+def test_preflight_passes_native_web_search_tool_through():
+    kwargs = {
+        "model": "grok-composer-2.5-fast",
+        "instructions": "You are helpful.",
+        "input": [{"role": "user", "content": [{"type": "input_text", "text": "hi"}]}],
+        "store": False,
+        "tools": [
+            {"type": "function", "name": "read_file", "description": "Read.",
+             "parameters": {"type": "object", "properties": {}}},
+            {"type": "web_search"},
+        ],
+    }
+    out = _preflight_codex_api_kwargs(kwargs, allow_stream=True)
+    tools = out["tools"]
+    assert {"type": "web_search"} in tools
+    assert any(t.get("type") == "function" and t.get("name") == "read_file" for t in tools)
+
+
+def test_preflight_still_rejects_unknown_tool_type():
+    kwargs = {
+        "model": "grok-composer-2.5-fast",
+        "instructions": "You are helpful.",
+        "input": [{"role": "user", "content": [{"type": "input_text", "text": "hi"}]}],
+        "store": False,
+        "tools": [{"type": "totally_made_up_tool"}],
+    }
+    with pytest.raises(ValueError, match="unsupported type"):
+        _preflight_codex_api_kwargs(kwargs, allow_stream=True)
+
+
 # ---------------------------------------------------------------------------
 # _format_responses_error — adapted from anomalyco/opencode#28757.
 # Provider failures should surface BOTH the code (rate_limit_exceeded /

tests/agent/test_model_metadata.py

@@ -142,6 +142,7 @@ class TestDefaultContextLengths:
                 ("grok-4", 256000),
                 ("grok-4-0709", 256000),
                 ("grok-build-0.1", 256000),
+                ("grok-composer-2.5-fast", 262144),
                 ("grok-code-fast-1", 256000),
                 ("grok-3", 131072),
                 ("grok-3-mini", 131072),

tests/agent/transports/test_codex_transport.py

@@ -263,6 +263,72 @@ class TestCodexBuildKwargs:
         # full history.
         assert "reasoning.encrypted_content" in kw.get("include", [])
 
+    def test_xai_injects_native_web_search_tool(self, transport):
+        """xAI path declares xAI's native server-side web_search built-in so
+        grok server-side search runs to completion (otherwise the turn stalls
+        as reasoning-with-no-answer -> false 'incomplete' -> 3 retries -> fail).
+        """
+        messages = [{"role": "user", "content": "Find current prices."}]
+        kw = transport.build_kwargs(
+            model="grok-composer-2.5-fast", messages=messages,
+            tools=[{"type": "function", "function": {
+                "name": "read_file", "description": "Read a file.",
+                "parameters": {"type": "object",
+                               "properties": {"path": {"type": "string"}}}}}],
+            is_xai_responses=True,
+        )
+        tool_types = [t.get("type") for t in kw.get("tools", [])]
+        assert "web_search" in tool_types, kw.get("tools")
+        # Non-conflicting client-side tools are preserved.
+        names = [t.get("name") for t in kw.get("tools", []) if t.get("type") == "function"]
+        assert "read_file" in names
+
+    def test_xai_drops_clientside_web_search_to_avoid_duplicate(self, transport):
+        """When the client registers its own 'web_search' function, the xAI
+        path must drop it and rely on the native built-in — otherwise xAI
+        returns HTTP 400 'Duplicate tool names: web_search'."""
+        messages = [{"role": "user", "content": "Search the web."}]
+        kw = transport.build_kwargs(
+            model="grok-composer-2.5-fast", messages=messages,
+            tools=[{"type": "function", "function": {
+                "name": "web_search", "description": "Search the web.",
+                "parameters": {"type": "object",
+                               "properties": {"query": {"type": "string"}}}}}],
+            is_xai_responses=True,
+        )
+        tools = kw.get("tools", [])
+        # Exactly one tool named/typed web_search, and it is the native built-in.
+        web_search_entries = [
+            t for t in tools
+            if t.get("name") == "web_search" or t.get("type") == "web_search"
+        ]
+        assert len(web_search_entries) == 1
+        assert web_search_entries[0] == {"type": "web_search"}
+        # No client-side function form of web_search survives.
+        assert not any(
+            t.get("type") == "function" and t.get("name") == "web_search"
+            for t in tools
+        )
+
+    def test_non_xai_path_does_not_inject_native_web_search(self, transport):
+        """Native web_search injection is scoped to xAI — Codex/GitHub paths
+        keep the client-side web_search function untouched."""
+        messages = [{"role": "user", "content": "Search."}]
+        kw = transport.build_kwargs(
+            model="gpt-5.4", messages=messages,
+            tools=[{"type": "function", "function": {
+                "name": "web_search", "description": "Search the web.",
+                "parameters": {"type": "object",
+                               "properties": {"query": {"type": "string"}}}}}],
+            is_xai_responses=False,
+        )
+        tools = kw.get("tools", [])
+        assert not any(t.get("type") == "web_search" for t in tools)
+        assert any(
+            t.get("type") == "function" and t.get("name") == "web_search"
+            for t in tools
+        )
+
     def test_xai_reasoning_disabled_no_reasoning_key(self, transport):
         messages = [{"role": "user", "content": "Hi"}]
         kw = transport.build_kwargs(

tests/run_agent/test_codex_xai_oauth_recovery.py

@@ -949,6 +949,26 @@ def test_grok_4_still_resolves_to_256k():
         assert DEFAULT_CONTEXT_LENGTHS[matched_key] == 256_000
 
 
+def test_grok_composer_context_length_is_262k():
+    """grok-composer-2.5-fast is OAuth-only and missing from /v1/models.
+
+    Without a specific entry it fell through to the generic ``grok`` 131k
+    catch-all, under-reporting ~262k enforced on /v1/responses.
+    """
+    from agent.model_metadata import DEFAULT_CONTEXT_LENGTHS
+
+    assert DEFAULT_CONTEXT_LENGTHS["grok-composer"] == 262_144
+    slug = "grok-composer-2.5-fast"
+    matched_key = max(
+        (k for k in DEFAULT_CONTEXT_LENGTHS if k in slug.lower()),
+        key=len,
+    )
+    assert matched_key == "grok-composer", (
+        f"Expected longest-first match on grok-composer for {slug}, got {matched_key}"
+    )
+    assert DEFAULT_CONTEXT_LENGTHS[matched_key] == 262_144
+
+
 # ---------------------------------------------------------------------------
 # Cross-issuer reasoning replay guard
 #