fix(media): OpenRouter image/video generation via TinyInference; use_skill live parent; media agents off vision-v1

crates/openhuman-cli/Cargo.toml

@@ -251,6 +251,11 @@ path = "../../tests/mcp_registry_multi_server.rs"
 name = "mcp_stdio_integration"
 path = "../../tests/mcp_stdio_integration.rs"
 
+[[test]]
+name = "media_generation_e2e"
+path = "../../tests/media_generation_e2e.rs"
+required-features = ["media"]
+
 [[test]]
 name = "memory_roundtrip_e2e"
 path = "../../tests/memory_roundtrip_e2e.rs"

crates/openhuman-core/Cargo.toml

@@ -999,16 +999,17 @@ runtime-node = []
 # and removing it from one without the other fails that lane.
 contacts = []
 # Media-generation + image domains: the `media_generate_*` agent tools
-# (image/video via GMI through the backend) and the `openhuman::image` tool
+# (image/video via OpenRouter through the backend's
+# `/agent-integrations/openrouter` proxy) and the `openhuman::image` tool
 # contracts scaffold. Default-ON. Slim builds opt out via
 # `--no-default-features --features "<explicit list without media>"`.
 # Composes with the runtime `DomainSet::media` flag (#4796).
-# NOTE: this gate sheds no exclusive dependencies — media generation is
-# backend-proxied (reqwest, shared). It is a surface-only gate (drops the tool
-# code + module from the compile), not a dependency-shedding one. There are no
-# controllers / stores / subscribers tagged `Media` (agent tools only), and
-# `openhuman::image` is currently unwired scaffold (added #2997).
-media = []
+# Enables `tinyagents-harness/media`, which pulls in `tinyinference-image` and
+# `tinyinference-video` (the wire contract, job loop and generic tools); both
+# are light (serde + the already-shared reqwest), so this remains mostly a
+# surface gate. There are no controllers / stores / subscribers tagged `Media`
+# (agent tools only), and `openhuman::image` is currently unwired scaffold.
+media = ["tinyagents-harness/media"]
 # Flows domains: the `flows::` automation surface (saved tinyflows graphs —
 # create/run/schedule + the workflow_builder / flow_discovery agents), the
 # `tinyflows::` adapter seam, and the `rhai_workflows::` language-workflow tool.

crates/openhuman-core/src/agent/registry/agents/image_agent/agent.toml

@@ -11,10 +11,16 @@ omit_safety_preamble = false
 omit_profile = true
 omit_memory_md = true
 
-# Multimodal tier so the agent can review the images it generates (and any
-# reference images) via the image_info / inline-image path, then iterate.
+# Pinned to a dedicated OpenRouter passthrough model rather than the
+# deprecated `hint:vision` (regression R4). This exact id is in
+# `MANAGED_MULTIMODAL_MODELS`, so it keeps the multimodal tier needed to
+# review the images it generates (and any reference images) via the
+# image_info / inline-image path, then iterate. Actual image generation goes
+# through `media_generate_image`, which defaults to
+# `bytedance-seed/seedream-5-0-lite` on OpenRouter — a separate model from
+# the one this agent's own turns run on.
 [model]
-hint = "vision"
+exact = "openrouter/qwen/qwen3.7-flash"
 
 [tools]
 # media_generate_image submits the generation and returns a saved local path;

crates/openhuman-core/src/agent/registry/agents/image_agent/prompt.md

@@ -1,28 +1,32 @@
 # Image-generation specialist
 
 You are a focused **image-creation** sub-agent. You turn a delegating agent's
-request into one or more finished image files using the hosted GMI image models
-(Seedream for text-to-image, SeedEdit for edits). You run on a multimodal model,
-so you can look at reference images and at the images you generate.
+request into one or more finished image files using a hosted image-generation
+model — the default is `bytedance-seed/seedream-5-0-lite` via OpenRouter for
+text-to-image and edits, but the catalog offers other supported models too.
+You run on a multimodal model, so you can look at reference images and at the
+images you generate.
 
 ## Your job
 
 - **Create** images from a text prompt (`media_generate_image`).
-- **Edit / restyle** a supplied image by passing its URL(s) as `input_images`.
-- **Pick the right model** when it matters — call `media_list_models` to see the
-  catalog (defaults are fine for most requests; `include_upstream` exposes the
-  full GMI list).
+- **Edit / restyle** a supplied image by passing it in `references` (https
+  URLs, `data:` URLs, or workspace file paths).
+- **Pick the right model** when it matters — call `media_list_models`
+  (`kind: "image"`, optional `search`) to see the catalog. The default suits
+  most requests.
 
 ## How to work
 
 - Write a vivid, specific prompt. Translate a terse request into concrete visual
   detail — subject, composition, lighting, style, mood, colour — but stay true
   to what was asked. Don't invent requirements the user didn't state.
-- Default the model and size unless the task calls for something specific. Use a
-  `size` like `1024x1024` (square), `1536x1024` (landscape), or `1024x1536`
-  (portrait) when the aspect ratio matters.
-- For edits, pass the source image URL(s) in `input_images` and describe the
-  change precisely.
+- Default the model and shape unless the task calls for something specific. Set
+  `aspect_ratio` (`1:1`, `16:9`, `9:16`, `4:3`, …) and optionally `resolution`
+  (`1K`, `2K`, `4K`); use `size` (e.g. `1536x1024`) only when exact pixels
+  matter. `n` asks for several variants; `seed` makes a result reproducible.
+- For edits, pass the source image(s) in `references` and describe the change
+  precisely.
 - Each generation **saves the image to the workspace and returns a local file
   path**. Always report that path back so the deck/answer can reference the
   concrete artifact. Do not paste raw base64 or invent URLs.
@@ -34,5 +38,6 @@ so you can look at reference images and at the images you generate.
 - Report results to the delegating agent — you are not talking to the end user.
 - If a request is unsafe or disallowed, decline rather than attempting a
   work-around.
-- If generation fails or times out, say so plainly and surface the request id;
-  don't fabricate a path or claim success.
+- If generation fails, say so plainly and surface the request id; don't
+  fabricate a path or claim success. When the error says the call was billed,
+  do **not** call again — report it.

crates/openhuman-core/src/agent/registry/agents/loader_tests_builtin_registration_tests.rs

@@ -224,12 +224,17 @@ fn every_builtin_is_stamped_builtin_source() {
 }
 
 #[test]
-fn vision_agent_loads_on_vision_hint() {
-    // The vision sub-agent rides the multimodal `vision-v1` tier (via the
-    // `vision` hint) so its model is image-capable, and it must be reachable
-    // from the orchestrator's subagent allowlist.
+fn vision_agent_loads_on_its_pinned_multimodal_model() {
+    // The vision sub-agent used to ride the multimodal `vision-v1` tier (via
+    // the `vision` hint), which is now deprecated — `vision-v1` silently
+    // falls back to the chat default on managed routes (regression R4). It
+    // is pinned to a dedicated OpenRouter passthrough model instead, and
+    // must remain reachable from the orchestrator's subagent allowlist.
     let def = find("vision_agent");
-    assert!(matches!(def.model, ModelSpec::Hint(ref h) if h == "vision"));
+    assert!(matches!(
+        def.model,
+        ModelSpec::Exact(ref m) if m == crate::config::MODEL_MEDIA_UNDERSTANDING
+    ));
 
     let orchestrator = find("orchestrator");
     assert!(

crates/openhuman-core/src/agent/registry/agents/loader_tests_specialist_agents_tests.rs

@@ -528,6 +528,44 @@ fn code_executor_has_curl_for_artifact_downloads() {
     }
 }
 
+/// R4 regression: `hint:vision` is deprecated (`vision-v1` silently falls
+/// back to the chat default on managed routes, with no error), so no
+/// built-in agent may still declare `ModelSpec::Hint("vision")`.
+#[test]
+fn no_builtin_agent_declares_the_deprecated_vision_hint() {
+    for def in load_builtins().expect("built-ins load") {
+        assert!(
+            !matches!(&def.model, ModelSpec::Hint(h) if h == "vision"),
+            "`{}` still declares the deprecated `hint:vision` — pin an exact model instead",
+            def.id
+        );
+    }
+}
+
+/// The three media agents are pinned to their dedicated OpenRouter
+/// passthrough models (regression R4), not left on `Inherit` or a `Hint`.
+#[test]
+fn media_agents_are_pinned_to_their_exact_models() {
+    use crate::config::{
+        MODEL_IMAGE_GENERATION_AGENT, MODEL_MEDIA_UNDERSTANDING, MODEL_VIDEO_GENERATION_AGENT,
+    };
+
+    for (agent_id, expected_model) in [
+        ("vision_agent", MODEL_MEDIA_UNDERSTANDING),
+        ("image_agent", MODEL_IMAGE_GENERATION_AGENT),
+        ("video_agent", MODEL_VIDEO_GENERATION_AGENT),
+    ] {
+        let def = find(agent_id);
+        match &def.model {
+            ModelSpec::Exact(model) => assert_eq!(
+                model, expected_model,
+                "{agent_id} must be pinned to `{expected_model}`, got `{model}`"
+            ),
+            other => panic!("{agent_id} must use ModelSpec::Exact, got {other:?}"),
+        }
+    }
+}
+
 #[test]
 fn orchestrator_does_not_get_curl() {
     // Per design: curl is a `Write` permission tool that writes

crates/openhuman-core/src/agent/registry/agents/orchestrator/agent.toml

@@ -121,8 +121,9 @@ allowlist = [
     # the memory tree only when a message needs it, not before every turn.
     "agent_memory",
     # Image-understanding specialist. Route anything that hinges on the content
-    # of an attached or on-disk user-provided image file here — it rides the
-    # multimodal `hint:vision` tier, so it can actually see the image.
+    # of an attached or on-disk user-provided image file here — it is pinned to
+    # a dedicated multimodal model (`MODEL_MEDIA_UNDERSTANDING`), so it can
+    # actually see the image.
     "vision_agent",
     # Image-generation specialist. Synthesised into a `delegate_create_image`
     # tool. Route make/generate/edit an image requests here — it owns prompt

crates/openhuman-core/src/agent/registry/agents/video_agent/agent.toml

@@ -11,10 +11,15 @@ omit_safety_preamble = false
 omit_profile = true
 omit_memory_md = true
 
-# Multimodal tier so the agent can inspect a reference/first-frame image or the
-# returned thumbnail when shaping an image-to-video request.
+# Pinned to a dedicated OpenRouter passthrough model rather than the
+# deprecated `hint:vision` (regression R4). This exact id is in
+# `MANAGED_MULTIMODAL_MODELS`, so it keeps the multimodal tier needed to
+# inspect a reference/first-frame image or the returned thumbnail when
+# shaping an image-to-video request. Actual video generation goes through
+# `media_generate_video`, which defaults to `bytedance/seedance-2.0-mini` on
+# OpenRouter — a separate model from the one this agent's own turns run on.
 [model]
-hint = "vision"
+exact = "openrouter/qwen/qwen3.7-flash"
 
 [tools]
 # media_generate_video submits the generation and returns a saved local path;

crates/openhuman-core/src/agent/registry/agents/video_agent/prompt.md

@@ -1,27 +1,33 @@
 # Video-generation specialist
 
 You are a focused **video-creation** sub-agent. You turn a delegating agent's
-request into a finished video clip using the hosted GMI video models (Seedance
-for fast clips, Veo for premium-tier output). You can do text-to-video or
-animate a supplied first-frame/reference image (image-to-video).
+request into a finished video clip using a hosted video-generation model — the
+default is `bytedance/seedance-2.0-mini` via OpenRouter for fast clips, with
+premium-tier models available in the catalog for higher-quality output. You
+can do text-to-video or animate a supplied first-frame/reference image
+(image-to-video).
 
 ## Your job
 
 - **Create** a clip from a text prompt (`media_generate_video`).
-- **Animate** a supplied image by passing its URL as `input_image`.
-- **Pick the right model** when it matters — call `media_list_models` to see the
-  catalog (the fast Seedance default suits most requests; `include_upstream`
-  exposes the full GMI list, including premium tiers).
+- **Animate** a supplied image by passing it as `first_frame` (and optionally
+  `last_frame`) — an https URL, `data:` URL, or workspace file path.
+- **Pick the right model** when it matters — call `media_list_models`
+  (`kind: "video"`, optional `search`) to see the catalog. The fast default
+  suits most requests.
 
 ## How to work
 
 - Write a concrete prompt describing the motion, subject, and scene — what
   happens over the clip, not just a static description. Mention camera movement,
   pacing, and style when relevant.
-- Use `duration_seconds` and `aspect_ratio` (e.g. `16:9`, `9:16`, `1:1`) when the
-  task specifies them; otherwise let the model default.
-- For image-to-video, pass the source image URL in `input_image` and describe
-  the motion you want applied to it.
+- Use `duration` (seconds; the default model accepts 4–15), `aspect_ratio`
+  (e.g. `16:9`, `9:16`, `1:1`) and `resolution` (`480p`, `720p`) when the task
+  specifies them; otherwise let the model default. `generate_audio` adds a
+  soundtrack where supported.
+- For image-to-video, pass the source image in `first_frame` and describe the
+  motion you want applied to it. `references` guide subject or style without
+  fixing a frame.
 - Generation is **asynchronous and can take minutes** — the tool blocks until the
   clip is ready, saves it to the workspace, and returns a local file path. Report
   that path back. Set expectations: tell the delegating agent it may take a
@@ -34,5 +40,7 @@ animate a supplied first-frame/reference image (image-to-video).
 - Report results to the delegating agent — you are not talking to the end user.
 - If a request is unsafe or disallowed, decline rather than attempting a
   work-around.
-- If generation fails or times out, say so plainly and surface the request id;
-  don't fabricate a path or claim success.
+- If generation fails, say so plainly and surface the job id; don't fabricate a
+  path or claim success. If it **times out**, call the tool again with
+  `resume_job_id` set to that job id to collect the clip — never submit a new
+  job for the same request, since each submit is billed.

crates/openhuman-core/src/agent/registry/agents/vision_agent/agent.toml

@@ -10,12 +10,14 @@ omit_identity = true
 omit_memory_context = true
 omit_safety_preamble = true
 
-# Multimodal tier. `ModelSpec::Hint("vision")` resolves to `hint:vision`, which
-# `oh_tier_supports_vision` reports as vision-capable — so this sub-agent's
-# model is always treated as image-enabled (managed or BYOK), and the turn
-# engine never strips the attached image at the vision gate.
+# Pinned to a dedicated OpenRouter passthrough model rather than the
+# deprecated `hint:vision` (`vision-v1` silently falls back to the chat
+# default on managed routes, with no error — regression R4). This exact id
+# is in `MANAGED_MULTIMODAL_MODELS`, so `oh_tier_supports_vision` still
+# reports it as image-enabled and the turn engine never strips the attached
+# image at the vision gate.
 [model]
-hint = "vision"
+exact = "openrouter/qwen/qwen3.7-flash"
 
 [tools]
 # Attached images arrive inline in the sub-agent's context via the multimodal