fix(bedrock): inherit modalities from foundation models for inference profiles, and surface them in /model

agent/bedrock_adapter.py

@@ -1637,6 +1637,17 @@ def discover_bedrock_models(
     except Exception as e:
         logger.warning("Failed to list Bedrock foundation models: %s", e)
 
+    # Build a lookup map from foundation models so inference profiles can
+    # inherit modalities.  Without this, profiles default to TEXT-only and
+    # models like Claude (which support IMAGE input) lose that capability,
+    # excluding them from vision-capable model filtering in the /model picker.
+    _foundation_modalities: Dict[str, Dict[str, List[str]]] = {}
+    for m in models:
+        _foundation_modalities[m["id"].lower()] = {
+            "input": m["input_modalities"],
+            "output": m["output_modalities"],
+        }
+
     # 2. Discover inference profiles (cross-region, better capacity)
     try:
         profiles = []
@@ -1661,22 +1672,45 @@ def discover_bedrock_models(
             if profile_id.lower() in seen_ids:
                 continue
 
+            profile_models = profile.get("models", [])
+
             # Apply provider filter to underlying models
             if filter_set:
-                profile_models = profile.get("models", [])
                 matches = any(
                     _extract_provider_from_arn(m.get("modelArn", "")).lower() in filter_set
                     for m in profile_models
                 )
                 if not matches:
                     continue
 
+            # Inherit modalities from the underlying foundation model.
+            # Inference profiles wrap foundation models but ListInferenceProfiles
+            # doesn't expose modalities — resolve the base model from the ARN
+            # (arn:aws:bedrock:*::foundation-model/<model-id>) and look it up.
+            # Falls back to stripping the regional prefix
+            # ("us.anthropic.claude-v2" → "anthropic.claude-v2").
+            base_model_id = None
+            for pm in profile_models:
+                arn = pm.get("modelArn", "")
+                match = re.search(r"foundation-model/(.+)$", arn)
+                if match:
+                    base_model_id = match.group(1)
+                    break
+            if base_model_id is None and "." in profile_id:
+                parts = profile_id.split(".", 1)
+                if len(parts) == 2 and parts[0] in ("us", "eu", "ap", "global", "jp", "apac"):
+                    base_model_id = parts[1]
+
+            base_mods = _foundation_modalities.get((base_model_id or "").lower(), {})
+            input_mods = base_mods.get("input", ["TEXT"])
+            output_mods = base_mods.get("output", ["TEXT"])
+
             models.append({
                 "id": profile_id,
                 "name": (profile.get("inferenceProfileName") or profile_id).strip(),
                 "provider": "inference-profile",
-                "input_modalities": ["TEXT"],
-                "output_modalities": ["TEXT"],
+                "input_modalities": input_mods,
+                "output_modalities": output_mods,
                 "streaming": True,
             })
             seen_ids.add(profile_id.lower())
@@ -1704,6 +1738,30 @@ def _extract_provider_from_arn(arn: str) -> str:
     """
     match = re.search(r"foundation-model/([^.]+)", arn)
     return match.group(1) if match else ""
+
+
+def vision_capable_bedrock_model_ids(models: List[Dict[str, Any]]) -> List[str]:
+    """Return the IDs of discovered Bedrock models that accept image input.
+
+    A model qualifies when ``"IMAGE"`` appears in its ``input_modalities``.
+    This is the production consumer of the inference-profile modality
+    inheritance above: ``ListFoundationModels`` reports modalities, but
+    ``ListInferenceProfiles`` does not, so without that inheritance every
+    profile — which is what the ``/model`` picker actually offers, since bare
+    foundation IDs are deduplicated away in favour of profiles — would arrive
+    stamped ``TEXT``-only and every Claude profile would be misreported here as
+    text-only.
+
+    Order follows *models*, so callers keep whatever sort they applied.
+    """
+    ids: List[str] = []
+    for m in models:
+        modalities = [str(x).upper() for x in (m.get("input_modalities") or [])]
+        if "IMAGE" in modalities:
+            ids.append(m["id"])
+    return ids
+
+
 # ---------------------------------------------------------------------------
 # Error classification — Bedrock-specific exceptions
 # ---------------------------------------------------------------------------

hermes_cli/auth.py

@@ -7733,6 +7733,7 @@ def _prompt_model_selection(
     confirm_provider: str = "",
     confirm_base_url: str = "",
     confirm_api_key: str = "",
+    vision_model_ids: Optional[List[str]] = None,
 ) -> Optional[str]:
     """Interactive model selection. Puts current_model first with a marker. Returns chosen model ID or None.
 
@@ -7741,6 +7742,11 @@ def _prompt_model_selection(
 
     If *unavailable_models* is provided, those models are shown grayed out
     and unselectable, with an upgrade link to *portal_url*.
+
+    If *vision_model_ids* is provided, those models get a ``(vision)`` marker
+    in their displayed label so image-capable models are distinguishable at the
+    point of choosing one. Display-only — the returned value is always the
+    plain model ID.
     """
     from hermes_cli.cli_output import line_input
     from hermes_cli.models import (
@@ -7749,6 +7755,7 @@ def _prompt_model_selection(
     )
 
     _unavailable = unavailable_models or []
+    _vision = set(vision_model_ids or [])
     # Sale chrome (★ / -N% / was) is Nous Portal-only — never for OpenRouter
     # or other providers even if pricing.original is somehow present.
     sale_chrome = (confirm_provider or "").strip().lower() == "nous"
@@ -7853,6 +7860,8 @@ def _label_segments(mid):
         """Build a rich radiolist row: yellow ★/% , dim was, plain prices."""
         if not has_pricing:
             segs: list[tuple[str, str | None]] = [(mid, None)]
+            if mid in _vision:
+                segs.append(("  (vision)", "dim"))
             if mid == current_model:
                 segs.append(("  ← currently in use", None))
             return segs
@@ -7879,6 +7888,8 @@ def _label_segments(mid):
             segs.append((f"  -{pct}%", "yellow"))
             if was_inp or was_out:
                 segs.append((f"  was {was_inp}/{was_out}", "dim"))
+        if mid in _vision:
+            segs.append(("  (vision)", "dim"))
         if mid == current_model:
             segs.append(("  ← currently in use", None))
         return segs

hermes_cli/model_setup_flows.py

@@ -2583,10 +2583,25 @@ def _sort_key(m):
 
         deduped.sort(key=_sort_key)
         model_list = [m["id"] for m in deduped]
-        print(
-            f"  Found {len(model_list)} text model(s) (filtered from {len(live_models)} total)"
-        )
+
+        # Which of these accept image input. The picker offers inference
+        # profiles (bare foundation IDs are deduplicated away just above), and
+        # ListInferenceProfiles reports no modalities, so this is only accurate
+        # because profiles inherit them from their foundation model.
+        from agent.bedrock_adapter import vision_capable_bedrock_model_ids
+
+        vision_ids = vision_capable_bedrock_model_ids(deduped)
+        if vision_ids:
+            print(
+                f"  Found {len(model_list)} text model(s), {len(vision_ids)} also "
+                f"accept image input (filtered from {len(live_models)} total)"
+            )
+        else:
+            print(
+                f"  Found {len(model_list)} text model(s) (filtered from {len(live_models)} total)"
+            )
     else:
+        vision_ids = []
         model_list = _PROVIDER_MODELS.get("bedrock", [])
         if model_list:
             print(
@@ -2605,6 +2620,7 @@ def _sort_key(m):
             current_model=current_model,
             confirm_provider="bedrock",
             confirm_base_url=f"https://bedrock-runtime.{region}.amazonaws.com",
+            vision_model_ids=vision_ids,
         )
     else:
         try: