feat: Add _vision_sidecar plugin

extensions/python/system_prompt/_11_tools_prompt.py

@@ -51,10 +51,13 @@ async def build_prompt(agent: Agent) -> str:
     prompt = agent.read_prompt("agent.system.tools.md", tools=tools_str)
 
     # vision support
-    from plugins._model_config.helpers.model_config import get_chat_model_config
+    from plugins._model_config.helpers.model_config import (
+        get_chat_model_config,
+        get_vision_model_config,
+    )
 
     chat_cfg = get_chat_model_config(agent)
-    if chat_cfg.get("vision", False):
+    if get_vision_model_config(agent) or chat_cfg.get("vision", False):
         prompt += "\n\n" + agent.read_prompt("agent.system.tools_vision.md")
 
     return prompt

helpers/parallel_tools.py

@@ -477,10 +477,15 @@ async def _run_direct_tool_job(parent_context_id: str, job: ParallelJob) -> str:
         worker_context.set_data(PARALLEL_WORKER_PARENT_CONTEXT_KEY, parent_context_id)
         worker_context.set_data(PARALLEL_WORKER_JOB_KEY, job.id)
         worker_context.set_data(PARALLEL_WORKER_KIND_KEY, job.kind)
+        worker_context.set_data(
+            "chat_model_override",
+            parent_context.get_data("chat_model_override"),
+        )
         job.worker_context_id = worker_context.id
         _copy_project(parent_context, worker_context)
 
         worker_agent = worker_context.agent0
+        worker_agent.last_user_message = parent_context.agent0.last_user_message
         worker_agent.loop_data = LoopData()
         return await execute_tool_call(
             worker_agent,

helpers/parallel_tools.py.dox.md

@@ -32,6 +32,7 @@
 - Subordinate child chats are tagged with job metadata, remain outside the scheduler task list, and may use normal child-chat tools including `parallel`.
 - Nested parallel jobs started by a parallel subordinate are registered as child `DeferredTask` instances so stopping the ancestor also stops its descendants.
 - Direct tool jobs run in isolated background contexts and are blocked from recursively invoking `parallel`.
+- Direct tool jobs inherit the parent's active per-chat model override and current user message.
 - Direct tool background context cleanup removes both the in-memory context and any transient chat folder left on disk.
 - Parent-visible child log items are created for each wrapped call so the WebUI can inspect concurrent children separately while the wrapper result remains model-history-only.
 - Child tool logs mirror normal tool-call visible args; job ids remain available through wrapper results and prompt extras rather than visible process-step args.

helpers/responses_tools.py

@@ -114,9 +114,15 @@ def _local_tool_prompts(agent: Any) -> list[tuple[str, str]]:
 
 def _vision_tool_prompt(agent: Any) -> str:
     try:
-        from plugins._model_config.helpers.model_config import get_chat_model_config
+        from plugins._model_config.helpers.model_config import (
+            get_chat_model_config,
+            get_vision_model_config,
+        )
 
-        if not get_chat_model_config(agent).get("vision", False):
+        if not (
+            get_vision_model_config(agent)
+            or get_chat_model_config(agent).get("vision", False)
+        ):
             return ""
         return agent.read_prompt("agent.system.tools_vision.md")
     except Exception:

helpers/responses_tools.py.dox.md

@@ -12,7 +12,7 @@
 
 ## Local Contracts
 
-- Build local function tools from enabled `agent.system.tool.*.md` prompt files and include `vision_load` only when the active chat model enables the matching vision prompt.
+- Build local function tools from enabled `agent.system.tool.*.md` prompt files and include `vision_load` when either Main native vision or the effective preset's Vision Model enables the canonical vision prompt.
 - Discover local prompt files through `helpers.subagents.get_paths`; this module
   owns the Responses-specific prompt-name compatibility rules.
 - Local prompt-derived function names use existing bullet declarations that pair a backticked name with `arg` or `args` for multi-tool prompt files, otherwise prefer explicit `"tool_name"` examples, then the first prompt heading, and finally the prompt filename.
@@ -25,8 +25,8 @@
 - Preserve original Agent Zero tool names through the native Responses name map.
 - Keep MCP tool schemas merged after local prompt-derived tools.
 - Apply `helpers.tool_policy` before emitting local or MCP schemas; a blocked
-  capability is absent from provider-native tool definitions. Vision remains
-  controlled solely by the active chat model configuration.
+  capability is absent from provider-native tool definitions. Vision routing
+  is controlled by the effective model preset rather than Agent Editor.
 - Connector remote tools are advertised only when `_a0_connector` runtime metadata says the matching connected CLI capability is currently available.
 
 ## Work Guidance

plugins/_model_config/AGENTS.md

@@ -14,15 +14,20 @@
 
 ## Local Contracts
 
-- `Default` is the first global preset and cannot be deleted or renamed. It owns the complete main, utility, and embedding baseline.
+- `Default` is the first global preset and cannot be deleted or renamed. It owns the complete main, utility, and embedding baseline; its Vision Model slot is optional.
 - Preset definitions are global. Global, project, agent-profile, and project/profile plugin configs persist only `model_preset`; chats may persist a preset reference as their explicit override.
 - Preserve scoped plugin resolution order and fall back invalid or missing scope/chat references to `Default`.
 - Project Settings `llm` payloads are owned here through the generic `helpers.projects` project extension-data hooks; keep project helper code agnostic to `_model_config` paths, presets, and inheritance rules.
 - Keep provider metadata and API-key checks safe around secrets.
 - Check API-key readiness only for the effective model configuration; unused global presets must not produce Welcome-screen warnings.
 - Coordinate OAuth-backed providers with `_oauth` instead of hardcoding provider-specific auth here.
 - `model_config_get` exposes `model_configured` as a derived chat-model readiness flag from provider, model name, and API-key availability.
-- Non-default presets may inherit omitted slots or durable tuning from `Default`, but must replace or clear per-slot `kwargs` so provider-specific extra params never leak across model providers.
+- Non-default presets may inherit omitted main, utility, or embedding slots and durable tuning from `Default`, but must replace or clear per-slot `kwargs` so provider-specific extra params never leak across model providers.
+- The optional `vision` slot is strictly per preset and never inherited from `Default`; an empty slot disables the separate Vision Model for that preset.
+- Main native vision wins by default. A configured Vision Model handles `vision_load` when Main lacks vision, or when that preset explicitly enables `override_main`.
+- Keep the optional Vision provider/model selector inside the Main Model card and flush with Main's field alignment, without a nested left inset. Show it only while Main vision is disabled or `override_main` is enabled; do not render a standalone Vision Model card.
+- Show `Use separate Vision Model` immediately below `Supports Vision` while Main vision is enabled, not inside Advanced Settings; describe the disabled state as using Main's native vision.
+- In model overviews, render the effective Vision Model as a text-only `Vision override / Provider / Model` child aligned with Main's provider column, not as an icon-bearing peer row.
 - Changing a model provider in the settings UI must clear `api_base` and `kwargs` because both may be provider-specific.
 - Repair provider-specific model-config aliases at the model-config read/build boundary; keep provider-specific repairs out of provider-agnostic core wrappers such as `models.py`.
 - `modelConfig.createPresetEditor()` owns local preset drafts, row actions, and stable UI-only row keys so deletion or renaming cannot rebind nested model fields.

plugins/_model_config/extensions/webui/chat-input-progress-start/model-switcher.html

@@ -65,6 +65,16 @@
                         </span>
                       </div>
                     </template>
+                    <template x-if="preset.vision?.name && (!preset.chat?.vision || preset.vision?.override_main)">
+                      <div class="model-switcher-model-row">
+                        <span class="model-switcher-model-label">Vision</span>
+                        <span class="model-switcher-model-value">
+                          <span x-text="preset.vision.provider" style="opacity:0.5;"></span>
+                          <span style="opacity:0.3; margin:0 3px;">/</span>
+                          <span x-text="preset.vision.name"></span>
+                        </span>
+                      </div>
+                    </template>
                     <template x-if="preset.utility?.name">
                       <div class="model-switcher-model-row">
                         <span class="model-switcher-model-label">Utility</span>

plugins/_model_config/helpers/model_config.py

@@ -16,11 +16,21 @@
 PRESET_SCOPE_PROJECT = "project"
 PRESET_SLOT_CONFIG_SECTIONS = {
     "chat": "chat_model",
+    "vision": "vision_model",
     "utility": "utility_model",
     "embedding": "embedding_model",
 }
 MODEL_SLOT_PRESET_REPLACE_FIELDS = {"kwargs"}
 IMPLICIT_PRESET_SLOT_DEFAULTS = {
+    "vision": {
+        "vision": True,
+        "max_embeds": 10,
+        "override_main": False,
+        "rl_requests": 0,
+        "rl_input": 0,
+        "rl_output": 0,
+        "kwargs": {},
+    },
     "utility": {
         "ctx_length": 128000,
         "ctx_input": 0.7,
@@ -339,7 +349,12 @@ def validate_presets(presets: list, *, require_default: bool = True) -> list:
 def normalize_config_for_save(config: dict) -> dict:
     """Remove UI-only fields and inline API keys before storing scoped config."""
     cleaned = deepcopy(config or {})
-    for section_name in ("chat_model", "utility_model", "embedding_model"):
+    for section_name in (
+        "chat_model",
+        "vision_model",
+        "utility_model",
+        "embedding_model",
+    ):
         section = cleaned.get(section_name)
         if isinstance(section, dict):
             cleaned[section_name] = _strip_ui_fields(section, strip_api_key=True)
@@ -659,10 +674,12 @@ def build_config_from_preset(
             continue
         slot_config = _get_preset_slot_config(preset, slot)
         if not _should_apply_preset_slot(slot, slot_config):
+            if slot == "vision":
+                config[section] = {}
             continue
         config[section] = _merge_model_slot(
             slot,
-            config.get(section, {}),
+            {} if slot == "vision" else config.get(section, {}),
             slot_config,
             strip_api_key=strip_api_key,
         )
@@ -738,6 +755,22 @@ def get_chat_model_config(agent=None) -> dict:
     return get_effective_config(agent).get("chat_model", {})
 
 
+def get_vision_model_config(agent=None) -> dict:
+    """Get the active Vision Model config after applying Main-first routing."""
+    cfg = get_effective_config(agent)
+    vision_cfg = cfg.get("vision_model", {})
+    if not all(
+        str(vision_cfg.get(key) or "").strip() for key in ("provider", "name")
+    ):
+        return {}
+    chat_cfg = cfg.get("chat_model", {})
+    return (
+        vision_cfg
+        if not chat_cfg.get("vision") or vision_cfg.get("override_main")
+        else {}
+    )
+
+
 def get_utility_model_config(agent=None) -> dict:
     """Get utility model config, with per-chat override if active."""
     return get_effective_config(agent).get("utility_model", {})
@@ -835,6 +868,16 @@ def build_utility_model(agent=None):
     )
 
 
+def build_vision_model(agent=None):
+    """Build the optional Vision Model selected by the effective preset."""
+    cfg = get_vision_model_config(agent)
+    mc = build_model_config(cfg, models.ModelType.CHAT)
+    mc.vision = True
+    return models.get_chat_model(
+        mc.provider, mc.name, model_config=mc, **mc.build_kwargs()
+    )
+
+
 def build_embedding_model(agent=None):
     """Build and return an embedding model wrapper."""
     cfg = get_embedding_model_config(agent)
@@ -881,6 +924,9 @@ def get_missing_api_key_providers(agent=None) -> list[dict]:
         ("Utility Model", cfg.get("utility_model", {})),
         ("Embedding Model", get_embedding_model_config(agent)),
     ]
+    vision_cfg = get_vision_model_config(agent)
+    if vision_cfg:
+        checks.insert(1, ("Vision Model", vision_cfg))
 
     for label, model_cfg in checks:
         provider = model_cfg.get("provider", "")

plugins/_model_config/webui/main.html

@@ -21,7 +21,7 @@
         <div class="preset-editor-intro">
           <div class="field-title">Model Presets</div>
           <div class="field-description">
-            Each preset contains the complete main, utility, and embedding model setup. Default is always available.
+            Each preset contains the main, utility, embedding, and optional vision-sidecar setup. Default is always available.
           </div>
         </div>
 
@@ -67,9 +67,21 @@
                 <div class="section-title">Main Model</div>
                 <div class="section-description">Primary model for conversations, reasoning, and tools.</div>
               </div>
-              <div x-data="{ get model() { return selectedPreset.chat; }, modelType: 'chat', providers: $store.modelConfig.chatProviders, searchType: 'chat', apiKeyMode: 'store' }">
+              <div x-data="{ get model() { return selectedPreset.chat; }, get visionModel() { return selectedPreset.vision; }, modelType: 'chat', providers: $store.modelConfig.chatProviders, searchType: 'chat', apiKeyMode: 'store' }">
                 <x-component path="/plugins/_model_config/webui/model-field.html"></x-component>
               </div>
+              <template x-if="!selectedPreset.chat.vision || selectedPreset.vision.override_main">
+                <div class="vision-sidecar-selector">
+                  <div class="field-title">Vision sidecar</div>
+                  <div class="field-description"
+                       x-text="selectedPreset.chat.vision
+                         ? 'Overrides Main model vision for vision_load.'
+                         : 'Used by vision_load while Main model vision is disabled. Leave empty to disable image analysis.'"></div>
+                  <div x-data="{ get model() { return selectedPreset.vision; }, modelType: 'vision', providers: $store.modelConfig.chatProviders, searchType: 'chat', apiKeyMode: 'store' }">
+                    <x-component path="/plugins/_model_config/webui/model-field.html"></x-component>
+                  </div>
+                </div>
+              </template>
             </section>
 
             <section class="preset-model-section">
@@ -166,6 +178,11 @@
       margin-top: 0;
     }
 
+    .vision-sidecar-selector {
+      margin: 0.75rem 0 0;
+      padding: 0.25rem 0 0;
+    }
+
     .preset-editor-secondary-actions {
       display: flex;
       flex-wrap: wrap;