refactor: drop the dead _vision_supported flag

Image rejections are now tracked per (provider, model) in
agent._image_rejecting_models, and recover_before_classification gates
on that set. Nothing reads agent._vision_supported any more, so the
write in turn_recovery and the per-turn reset entry in turn_context are
dead state. Remove both so the next reader doesn't assume a turn-global
vision gate still exists; update the two tests that asserted or seeded
the attribute.

agent/turn_context.py

@@ -558,7 +558,7 @@ _PER_TURN_RESET_STATE: Tuple[Tuple[str, Any], ...] = (
     ("_thinking_prefill_retries", 0), ("_post_tool_empty_retried", False),
     ("_last_content_with_tools", None), ("_last_content_tools_all_housekeeping", False),
     ("_mute_post_response", False), ("_unicode_sanitization_passes", 0),
-    ("_tool_guardrail_halt_decision", None), ("_vision_supported", True),
+    ("_tool_guardrail_halt_decision", None),
     ("_iteration_budget_warning_injected", False),
     ("_run_budget_wrapup_injected", False), ("_verification_stop_nudges", 0),
     ("_pre_verify_nudges", 0),

agent/turn_recovery.py

@@ -244,14 +244,13 @@ def recover_before_classification(
     _err_status = getattr(api_error, "status_code", None)
     # 4xx-only gate: 5xx/timeouts are transient and take the retry path.
     _status_ok = _err_status is None or (400 <= int(_err_status) < 500)
-    # Guarded PER MODEL, not by the turn-global ``_vision_supported``: in a fallback chain the next
-    # model can reject images too, and a turn-wide flag would skip its recovery and fail the turn.
+    # Guarded PER MODEL, not by a turn-global flag: in a fallback chain the next model can reject
+    # images too, and a turn-wide flag would skip its recovery and fail the turn.
     _model_key = image_model_key(agent)
     _rejected = getattr(agent, "_image_rejecting_models", None)
     if not isinstance(_rejected, set):
         _rejected = agent._image_rejecting_models = set()
     if _model_key not in _rejected and _looks_like_image_content_rejection(_err_body) and _status_ok:
-        agent._vision_supported = False
         # Send-path only. A rejection says what THIS model accepts, not what the conversation
         # holds: stripping ``messages`` (canonical history) and forcing a flush deleted every
         # image — and every image-only message — from state.db for good, so a later switch to a

tests/agent/test_image_rejection_fallback.py

@@ -227,8 +227,8 @@ class TestStripImagesDropsStaleApiContent:
     previously sent for a message, and the next turn substitutes it back into
     ``content``. Leaving it in place on a message this function rewrote would
     replay the images the strip just removed — and the recovery cannot re-fire,
-    because it sets ``_vision_supported = False`` and gates itself on that. The
-    session would then send rejected images on every subsequent turn.
+    because it records the model in ``_image_rejecting_models`` and gates itself
+    on that. The session would then send rejected images on every subsequent turn.
 
     Same contract the other content-rewrite paths follow (stale-confirmation
     redaction in ``replay_cleanup``, compression rewrites, merge-into-tail):
@@ -317,7 +317,7 @@ class TestRejectionNeverReachesPersistedHistory:
         from types import SimpleNamespace
 
         return SimpleNamespace(
-            provider=provider, model=model, _vision_supported=True, _force_ascii_payload=False,
+            provider=provider, model=model, _force_ascii_payload=False,
             _image_rejecting_models=set(), _db_flush_scan_prefix=7, log_prefix="",
             _vprint=lambda *a, **k: None,
         )

tests/agent/test_turn_context.py

@@ -90,7 +90,6 @@ class _FakeAgent:
         self._stream_think_scrubber = None
         # Attributes the prologue assigns; recorded for assertions.
         self._invalid_tool_retries = -1
-        self._vision_supported = None
         self._persist_calls = 0
         self._session_messages = []
         self._pending_cli_user_message = None
@@ -383,10 +382,9 @@ def test_turn_start_replaces_stale_parent_history_with_compression_child():
 def test_applies_agent_side_effects():
     agent = _FakeAgent()
     _build(agent)
-    # Retry counters reset, guardrails reset, vision re-armed, turn counted.
+    # Retry counters reset, guardrails reset, turn counted.
     assert agent._invalid_tool_retries == 0
     assert agent._tool_guardrails.reset_called is True
-    assert agent._vision_supported is True
     assert agent._user_turn_count == 1
     # Crash-resilience persistence fired once.
     assert agent._persist_calls == 1