Implement Venice.ai provider

2026-06-21 01:40:36 +01:00
parent cc3888a4cf
commit de980f0165
20 changed files with 1800 additions and 4 deletions
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,3 +1,8 @@
+# (2026-06-21) Version 1.22.0
+
+- (**Feature**) Add a native [Venice](https://venice.ai) provider with [🖌️ image-generation](./docs/features.md#️-image-creation) (incl. editing), [💬 text-generation](./docs/features.md#-text-generation) (incl. vision), [🗣️ text-to-speech](./docs/features.md#️-text-to-speech), [🦻 speech-to-text](./docs/features.md#-speech-to-text), and Venice's native web search via the full `venice_parameters` knob set. Unlike the [OpenAI-compatible](./docs/providers.md#openai-compatible) path (which drops images and can't reach Venice's audio or native image endpoints), it talks to Venice's API directly, using the knob-rich native `/image/generate` and `/image/edit` endpoints. See the [Venice provider docs](./docs/providers.md#venice).
+
+
 # (2026-06-05) Version 1.21.1

 - (**Security**) Update the [anthropic](https://github.com/etkecc/anthropic-rs) dependency to use [reqwest](https://crates.io/crates/reqwest) 0.12 / [rustls](https://crates.io/crates/rustls) 0.23, replacing the vulnerable `rustls-webpki` 0.101 line with 0.103.13. This resolves [`GHSA-82j2-j2ch-gfr8`](https://github.com/advisories/GHSA-82j2-j2ch-gfr8) (high — denial of service via panic on a malformed CRL), [`GHSA-xgp8-3hg3-c2mh`](https://github.com/advisories/GHSA-xgp8-3hg3-c2mh) and [`GHSA-965h-392x-2mh5`](https://github.com/advisories/GHSA-965h-392x-2mh5) (name-constraint validation issues).
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -315,7 +315,7 @@ dependencies = [

 [[package]]
 name = "baibot"
-version = "1.21.1"
+version = "1.22.0"
 dependencies = [
 "anthropic",
 "anyhow",
@@ -329,6 +329,7 @@ dependencies = [
 "mxlink",
 "quick_cache",
 "regex",
+ "reqwest 0.12.28",
 "serde",
 "serde_json",
 "serde_yaml_ng",
@@ -1590,6 +1591,7 @@ dependencies = [
 "tokio",
 "tokio-rustls",
 "tower-service",
+ "webpki-roots 1.0.7",
 ]

 [[package]]
@@ -3074,6 +3076,7 @@ dependencies = [
 "hyper-util",
 "js-sys",
 "log",
+ "mime_guess",
 "percent-encoding",
 "pin-project-lite",
 "quinn",
@@ -3095,6 +3098,7 @@ dependencies = [
 "wasm-bindgen-futures",
 "wasm-streams 0.4.2",
 "web-sys",
+ "webpki-roots 1.0.7",
 ]

 [[package]]
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -7,7 +7,7 @@ license = "AGPL-3.0-or-later"
 readme = "README.md"
 keywords = ["matrix", "chat", "bot", "AI", "LLM"]
 include = ["/etc/assets/baibot-torso-768.png", "/src", "/README.md", "/CHANGELOG.md", "/LICENSE"]
-version = "1.21.1"
+version = "1.22.0"
 edition = "2024"

 [lib]
@@ -28,6 +28,10 @@ mxlink = ">=1.15.0"
 etke_openai_api_rust = "0.1.*"
 quick_cache = "0.6.*"
 regex = "1.12.*"
+# Direct dep for the native `venice` provider's HTTP client. Pinned to 0.12 (the version
+# async-openai 0.41 already resolves) with rustls only and default-features off, so we ride
+# the existing reqwest+rustls copy instead of pulling a second TLS stack (native-tls/openssl).
+reqwest = { version = "0.12.*", default-features = false, features = ["json", "multipart", "rustls-tls"] }
 serde = { version = "1.0.*", features = ["derive"], default-features = false }
 serde_json = "1.0.*"
 serde_yaml_ng = "0.10.*"
--- a/README.md
+++ b/README.md
@@ -13,7 +13,7 @@ It's influenced by [chaz](https://github.com/arcuru/chaz), but does **not** use

 ## 🌟 Features

- 🎨 Encourages **[provider](./docs/providers.md) choice** ([Anthropic](./docs/providers.md#anthropic), [Groq](./docs/providers.md#groq), [LocalAI](./docs/providers.md#localai), [OpenAI](./docs/providers.md#openai) and [☁️ many more](./docs/providers.md#️-providers)) as well as **[mixing & matching models](./docs/features.md#-mixing--matching-models)**:
+- 🎨 Encourages **[provider](./docs/providers.md) choice** ([Anthropic](./docs/providers.md#anthropic), [Groq](./docs/providers.md#groq), [LocalAI](./docs/providers.md#localai), [OpenAI](./docs/providers.md#openai), [Venice](./docs/providers.md#venice) and [☁️ many more](./docs/providers.md#️-providers)) as well as **[mixing & matching models](./docs/features.md#-mixing--matching-models)**:

 - Supports **different use purposes** (depending on the [☁️ provider](./docs/providers.md) & model):

--- a/docs/configuration/text-to-speech.md
+++ b/docs/configuration/text-to-speech.md
@@ -56,7 +56,7 @@ Example: `!bai config room text-to-speech set-speed-override 1.5` (this can also

 ### 👫 Voice override

-The voice override setting lets you change the voice being used by the text-to-speech model configured at the [🤖 agent](../agents.md) level (usually `onyx` when using [OpenAI](../providers.md#openai)).
+The voice override setting lets you change the voice being used by the text-to-speech model configured at the [🤖 agent](../agents.md) level (e.g. `onyx` when using [OpenAI](../providers.md#openai), or `af_sky` when using [Venice](../providers.md#venice)).

 Possible values (e.g. `onyx`) depend on the model you're using. For example, for [OpenAI](../providers.md#openai)'s Whisper model, [these voices](https://platform.openai.com/docs/guides/text-to-speech/voice-options) are available.

--- a/docs/providers.md
+++ b/docs/providers.md
@@ -19,6 +19,7 @@ The list of supported providers is below.
  - [OpenAI Compatible](#openai-compatible)
  - [OpenRouter](#openrouter)
  - [Together AI](#together-ai)
+  - [Venice](#venice)


 ### How to choose a provider
@@ -171,3 +172,82 @@ This provider is just as featureful as the [OpenAI](#openai) provider, but is mo
  - create a global agent: `!bai agent create-global together-ai my-together-ai-agent`

 💡 When creating an agent, the bot will show you an up-to-date sample configuration for this provider which looks [like this](./sample-provider-configs/together-ai.yml).
+
+
+### Venice
+
+[Venice AI](https://venice.ai) runs inference on Venice-controlled GPUs or zero-data-retention partner infrastructure and stores no prompts or responses, so your conversations don't linger anywhere. It serves both frontier proprietary models and the latest open-source ones.
+
+- 🆔 Identifier: `venice`
+- 🔗 Links: [🏠 Home page](https://venice.ai), [👤 Sign up](https://venice.ai), [📋 Models list](https://api.venice.ai/api/v1/models)
+- 🌟 Capabilities: [🖌️ image-generation](./features.md#️-image-creation) (incl. editing, via the native knob-rich `/image/generate` and `/image/edit` endpoints), [💬 text-generation](./features.md#-text-generation) (incl. vision; native web search via the `venice_parameters` config), [🗣️ text-to-speech](./features.md#️-text-to-speech), [🦻 speech-to-text](./features.md#-speech-to-text)
+- 🗲 Quick start:
+  - create a room-local agent: `!bai agent create-room-local venice my-venice-agent`
+  - create a global agent: `!bai agent create-global venice my-venice-agent`
+
+💡 When creating an agent, the bot will show you an up-to-date sample configuration for this provider which looks [like this](./sample-provider-configs/venice.yml).
+
+Unlike the [OpenAI Compatible](#openai-compatible) provider (which can talk to Venice but drops images and can't reach its audio or native image endpoints), this is a first-class Venice integration that exposes Venice's full parameter set. Image generation uses the native `/image/generate` endpoint rather than the OpenAI-compatible `/images/generations` shim, so every Venice-specific knob below is available.
+
+#### Configuration reference
+
+Every parameter below is optional unless marked otherwise. Omitting a knob lets Venice apply its own server-side default; this is **not** the same as setting it to `false`, which actively sends `false`.
+
+**`text_generation.venice_parameters`** — Venice-specific request knobs sent in the `venice_parameters` bag (alongside the standard `model_id`, `prompt`, `temperature`, `max_response_tokens`, and `max_context_tokens` fields). Set any of them to override Venice's behavior. The `Default` column shows the value baibot's sample config ships; a `—` means the knob is left unset, so Venice's own default applies.
+
+| Knob | What it does | Default |
+|------|--------------|---------|
+| `enable_web_search` | Web search mode: `auto` (model decides), `on` (always), or `off`. | `auto` |
+| `enable_web_citations` | Append source citations to web-search answers. | — |
+| `enable_web_scraping` | Allow the model to scrape page contents during web search. | — |
+| `enable_x_search` | Include X (Twitter) in web search. | — |
+| `include_search_results_in_stream` | Stream search results back as they arrive. | — |
+| `return_search_results_as_documents` | Return search results as structured documents. | — |
+| `include_venice_system_prompt` | Prepend Venice's own system prompt alongside yours. | — |
+| `character_slug` | Use a public Venice character by its slug. | — |
+| `strip_thinking_response` | Strip `<think></think>` blocks from reasoning models so the user sees only the answer. | `true` |
+| `disable_thinking` | Disable the model's reasoning step entirely. | — |
+| `enable_e2ee` | Run in end-to-end-encrypted mode rather than the default TEE-only mode. | `false` |
+
+**`text_to_speech`**:
+
+| Knob | What it does | Default |
+|------|--------------|---------|
+| `model_id` | The Venice TTS model (e.g. `tts-kokoro`, `tts-qwen3-1-7b`, `tts-xai-v1`). | `tts-kokoro` |
+| `voice` | The voice to synthesize with. Model-specific (Kokoro: `af_*`/`am_*`/`bf_*`/`bm_*`); a cloned-voice handle (`vv_<id>`) also works. | `af_sky` |
+| `response_format` | Audio format: `mp3`, `opus`, `aac`, `flac`, `wav`, or `pcm`. | `mp3` |
+| `speed` | Playback speed, `0.25`–`4.0`. | `1.0` |
+| `prompt` | A style prompt steering emotion/delivery. Only Qwen 3 TTS honors it. | — |
+| `temperature` | Sampling temperature, `0.0`–`2.0`. Only Qwen 3 / Orpheus / Chatterbox HD honor it. | — |
+| `top_p` | Nucleus sampling, `0.0`–`1.0`. Only Qwen 3 TTS honors it. | — |
+
+**`image_generation`**:
+
+| Knob | What it does | Default |
+|------|--------------|---------|
+| `model_id` | The image-generation model. | `chroma` |
+| `negative_prompt` | A description of what should **not** appear in the image. | — |
+| `cfg_scale` | CFG scale, `0`–`20`. Higher values adhere more closely to the prompt. | — |
+| `steps` | Number of inference steps. Model-specific; some models ignore it. | — |
+| `style_preset` | A named style to apply (e.g. `3D Model`). | — |
+| `seed` | Random seed, `-999999999`–`999999999`. Fix it for reproducible results. | random |
+| `safe_mode` | Blur images classified as adult content. | `true` |
+| `hide_watermark` | Hide the Venice watermark (may be ignored for some content). | `false` |
+| `format` | Output format: `jpeg`, `png`, or `webp`. | `webp` |
+| `width` / `height` | Image dimensions in pixels, each `1`–`1280`. | `1024` |
+| `aspect_ratio` | Aspect ratio for models that support it (e.g. `1:1`, `16:9`). Alternative to `width`/`height`. | — |
+| `resolution` | Resolution tier for models that support it (`1K`, `2K`, `4K`). | — |
+| `quality` | Output quality for supported models: `low`, `medium`, `high`. Higher can cost more. | — |
+| `lora_strength` | Lora strength, `0`–`100`. Only applies if the model uses additional Loras. | — |
+| `embed_exif_metadata` | Embed the generation prompt into the image's EXIF metadata. | `false` |
+| `enable_web_search` | Let the model pull the latest info from the web. Model-specific; costs extra credits. | — |
+
+**`image_generation.edit`** — image editing reuses the `image_generation` block; only the model and a few output knobs differ:
+
+| Knob | What it does | Default |
+|------|--------------|---------|
+| `model_id` | The image-edit model. | `firered-image-edit` |
+| `output_format` | Output format: `jpeg`, `png`, or `webp`. When omitted, Venice infers it (PNG at 1K, JPEG at 2K/4K). | inferred |
+| `aspect_ratio` | Aspect ratio of the result: `auto`, `1:1`, `3:2`, `16:9`, `21:9`, `9:16`, `2:3`, `3:4`, `4:5` (model-specific). | — |
+| `resolution` | Resolution tier: `1K`, `2K`, `4K` (model-specific). | `1K` |
+| `safe_mode` | Blur images classified as adult content. | `true` |
--- a/docs/sample-provider-configs/venice.yml
+++ b/docs/sample-provider-configs/venice.yml
@@ -0,0 +1,97 @@
+base_url: https://api.venice.ai/api/v1
+api_key: YOUR_API_KEY_HERE
+text_generation:
+  model_id: kimi-k2-5
+  prompt: "You are a brief, but helpful bot called {{ baibot_name }} powered by the {{ baibot_model_id }} model. The date/time of this conversation's start is: {{ baibot_conversation_start_time_utc }}."
+  temperature: 1.0
+  max_response_tokens: 4096
+  max_context_tokens: 128000
+  # Venice-specific request parameters. Only the keys present below are sent to Venice; omit a
+  # key to fall back to Venice's own default. Omitting a knob is NOT the same as setting it to
+  # `false` — `false` actively sends `false`.
+  venice_parameters:
+    # Web search: "auto" (model decides), "on" (always), or "off".
+    enable_web_search: "auto"
+    # Strip <think></think> blocks from reasoning models so the user sees only the answer.
+    strip_thinking_response: true
+    # Run in TEE-only mode instead of end-to-end encryption (works across all models).
+    enable_e2ee: false
+    # Other available knobs — uncomment to override Venice's default:
+    # enable_web_citations: true
+    # enable_web_scraping: true
+    # include_venice_system_prompt: false
+    # include_search_results_in_stream: true
+    # return_search_results_as_documents: true
+    # enable_x_search: true
+    # disable_thinking: true
+    # character_slug: public-character-id
+speech_to_text:
+  model_id: nvidia/parakeet-tdt-0.6b-v3
+text_to_speech:
+  # The Venice TTS model. Others include tts-qwen3-1-7b, tts-xai-v1,
+  # tts-elevenlabs-turbo-v2-5, tts-minimax-speech-02-hd. See the models list endpoint.
+  model_id: tts-kokoro
+  # The voice to synthesize with. Voices are model-specific: Kokoro uses af_*/am_*/bf_*/bm_*
+  # (e.g. af_sky, am_adam), other models have their own sets. You can also pass a cloned-voice
+  # handle (vv_<id>) created via Venice's voice-cloning API. An incompatible voice returns an error.
+  voice: af_sky
+  # Output audio format: mp3, opus, aac, flac, wav, or pcm. mp3 is the broadest Matrix-client fit.
+  response_format: mp3
+  # Other available knobs — uncomment to override Venice's default:
+  # Playback speed, 0.25–4.0 (1.0 is normal).
+  # speed: 1.0
+  # A style prompt steering emotion/delivery (e.g. "Excited and energetic."). Only Qwen 3 TTS uses it.
+  # prompt: "Calm and warm."
+  # Sampling temperature, 0.0–2.0 (higher = more varied). Only Qwen 3 / Orpheus / Chatterbox HD use it.
+  # temperature: 0.9
+  # Nucleus sampling, 0.0–1.0. Only Qwen 3 TTS uses it.
+  # top_p: 1.0
+image_generation:
+  # The image-generation model. See the models list endpoint for the full set.
+  model_id: chroma
+  # The image-edit model, used when editing an existing image rather than generating a new one.
+  # Editing shares this same image_generation config block; only the model differs.
+  edit:
+    model_id: firered-image-edit
+    # Other edit knobs — uncomment to override Venice's default:
+    # Output format: jpeg, png, or webp. When omitted, Venice infers it (PNG at 1K, JPEG at 2K/4K).
+    # output_format: png
+    # Aspect ratio of the result: auto, 1:1, 3:2, 16:9, 21:9, 9:16, 2:3, 3:4, 4:5 (model-specific).
+    # aspect_ratio: auto
+    # Resolution tier: 1K, 2K, 4K (model-specific). Defaults to 1K.
+    # resolution: 1K
+    # Blur images classified as adult content. Defaults to true.
+    # safe_mode: true
+  # Other generation knobs — uncomment to override Venice's default. Omitting a knob is NOT the same
+  # as setting it: an omitted knob lets Venice apply its own default, a set value is sent verbatim.
+  # A description of what should NOT appear in the image.
+  # negative_prompt: "blurry, watermark, text"
+  # CFG scale, 0–20. Higher values make the image adhere more closely to the prompt.
+  # cfg_scale: 7.5
+  # Number of inference steps. Model-specific; some models ignore it.
+  # steps: 8
+  # A named style to apply (e.g. "3D Model"). See Venice's image-styles reference.
+  # style_preset: "3D Model"
+  # Random seed, -999999999–999999999. Fix it for reproducible results; omit for a random seed.
+  # seed: 123456789
+  # Blur images classified as adult content. Defaults to true.
+  # safe_mode: true
+  # Hide the Venice watermark. Venice may ignore this for certain generated content. Defaults to false.
+  # hide_watermark: false
+  # Output format: jpeg, png, or webp. webp is smallest; png is highest-quality. Defaults to webp.
+  # format: webp
+  # Image dimensions in pixels, each 1–1280. Default 1024×1024.
+  # width: 1024
+  # height: 1024
+  # Aspect ratio (used by certain models, e.g. Nano Banana): "1:1", "16:9". An alternative to width/height.
+  # aspect_ratio: "1:1"
+  # Resolution tier (used by certain models): "1K", "2K", "4K".
+  # resolution: "1K"
+  # Output quality for supported models (e.g. GPT Image 2): low, medium, high. Higher can cost more.
+  # quality: high
+  # Lora strength, 0–100. Only applies if the model uses additional Loras.
+  # lora_strength: 50
+  # Embed the generation prompt into the image's EXIF metadata. Defaults to false.
+  # embed_exif_metadata: false
+  # Let the model pull the latest info from the web for the image. Model-specific; costs extra credits.
+  # enable_web_search: false
--- a/src/agent/instantiation.rs
+++ b/src/agent/instantiation.rs
@@ -109,6 +109,9 @@ fn create_controller_from_provider_and_json_value_config(
        AgentProvider::TogetherAI => {
            provider::openai_compat::create_controller_from_yaml_value_config(agent_id, config)
        }
+        AgentProvider::Venice => {
+            provider::venice::create_controller_from_yaml_value_config(agent_id, config)
+        }
    }
 }

@@ -150,5 +153,9 @@ pub fn default_config_for_provider(provider: &AgentProvider) -> serde_yaml_ng::V
            let config = super::provider::togetherai::default_config();
            serde_yaml_ng::to_value(config).expect("Failed to serialize config")
        }
+        AgentProvider::Venice => {
+            let config = super::provider::venice::default_config();
+            serde_yaml_ng::to_value(config).expect("Failed to serialize config")
+        }
    }
 }
--- a/src/agent/provider/controller.rs
+++ b/src/agent/provider/controller.rs
@@ -61,6 +61,7 @@ pub enum ControllerType {
    OpenAI(Box<super::openai::Controller>),
    OpenAICompat(Box<super::openai_compat::Controller>),
    Anthropic(Box<super::anthropic::Controller>),
+    Venice(Box<super::venice::Controller>),
 }

 impl ControllerTrait for ControllerType {
@@ -69,6 +70,7 @@ impl ControllerTrait for ControllerType {
            ControllerType::OpenAI(controller) => controller.supports_purpose(purpose),
            ControllerType::OpenAICompat(controller) => controller.supports_purpose(purpose),
            ControllerType::Anthropic(controller) => controller.supports_purpose(purpose),
+            ControllerType::Venice(controller) => controller.supports_purpose(purpose),
        }
    }

@@ -77,6 +79,7 @@ impl ControllerTrait for ControllerType {
            ControllerType::OpenAI(controller) => controller.text_generation_model_id(),
            ControllerType::OpenAICompat(controller) => controller.text_generation_model_id(),
            ControllerType::Anthropic(controller) => controller.text_generation_model_id(),
+            ControllerType::Venice(controller) => controller.text_generation_model_id(),
        }
    }

@@ -85,6 +88,7 @@ impl ControllerTrait for ControllerType {
            ControllerType::OpenAI(controller) => controller.text_generation_prompt(),
            ControllerType::OpenAICompat(controller) => controller.text_generation_prompt(),
            ControllerType::Anthropic(controller) => controller.text_generation_prompt(),
+            ControllerType::Venice(controller) => controller.text_generation_prompt(),
        }
    }

@@ -93,6 +97,7 @@ impl ControllerTrait for ControllerType {
            ControllerType::OpenAI(controller) => controller.text_to_speech_voice(),
            ControllerType::OpenAICompat(controller) => controller.text_to_speech_voice(),
            ControllerType::Anthropic(controller) => controller.text_to_speech_voice(),
+            ControllerType::Venice(controller) => controller.text_to_speech_voice(),
        }
    }

@@ -101,6 +106,7 @@ impl ControllerTrait for ControllerType {
            ControllerType::OpenAI(controller) => controller.text_to_speech_speed(),
            ControllerType::OpenAICompat(controller) => controller.text_to_speech_speed(),
            ControllerType::Anthropic(controller) => controller.text_to_speech_speed(),
+            ControllerType::Venice(controller) => controller.text_to_speech_speed(),
        }
    }

@@ -109,6 +115,7 @@ impl ControllerTrait for ControllerType {
            ControllerType::OpenAI(controller) => controller.text_generation_temperature(),
            ControllerType::OpenAICompat(controller) => controller.text_generation_temperature(),
            ControllerType::Anthropic(controller) => controller.text_generation_temperature(),
+            ControllerType::Venice(controller) => controller.text_generation_temperature(),
        }
    }

@@ -117,6 +124,7 @@ impl ControllerTrait for ControllerType {
            ControllerType::OpenAI(controller) => controller.ping().await,
            ControllerType::OpenAICompat(controller) => controller.ping().await,
            ControllerType::Anthropic(controller) => controller.ping().await,
+            ControllerType::Venice(controller) => controller.ping().await,
        }
    }

@@ -135,6 +143,9 @@ impl ControllerTrait for ControllerType {
            ControllerType::Anthropic(controller) => {
                controller.generate_text(conversation, params).await
            }
+            ControllerType::Venice(controller) => {
+                controller.generate_text(conversation, params).await
+            }
        }
    }

@@ -154,6 +165,9 @@ impl ControllerTrait for ControllerType {
            ControllerType::Anthropic(controller) => {
                controller.speech_to_text(mime_type, media, params).await
            }
+            ControllerType::Venice(controller) => {
+                controller.speech_to_text(mime_type, media, params).await
+            }
        }
    }

@@ -170,6 +184,9 @@ impl ControllerTrait for ControllerType {
            ControllerType::Anthropic(controller) => {
                controller.generate_image(prompt, params).await
            }
+            ControllerType::Venice(controller) => {
+                controller.generate_image(prompt, params).await
+            }
        }
    }

@@ -189,6 +206,9 @@ impl ControllerTrait for ControllerType {
            ControllerType::Anthropic(controller) => {
                controller.create_image_edit(prompt, images, params).await
            }
+            ControllerType::Venice(controller) => {
+                controller.create_image_edit(prompt, images, params).await
+            }
        }
    }

@@ -203,6 +223,7 @@ impl ControllerTrait for ControllerType {
                controller.text_to_speech(text, params).await
            }
            ControllerType::Anthropic(controller) => controller.text_to_speech(text, params).await,
+            ControllerType::Venice(controller) => controller.text_to_speech(text, params).await,
        }
    }
 }
--- a/src/agent/provider/entity/agent_provider.rs
+++ b/src/agent/provider/entity/agent_provider.rs
@@ -11,6 +11,7 @@ pub enum AgentProvider {
    OpenAICompat,
    OpenRouter,
    TogetherAI,
+    Venice,
 }

 impl AgentProvider {
@@ -25,6 +26,7 @@ impl AgentProvider {
            &Self::OpenAICompat,
            &Self::OpenRouter,
            &Self::TogetherAI,
+            &Self::Venice,
        ]
    }

@@ -39,6 +41,7 @@ impl AgentProvider {
            Self::OpenAICompat => "openai-compatible",
            Self::OpenRouter => "openrouter",
            Self::TogetherAI => "together-ai",
+            Self::Venice => "venice",
        }
    }

@@ -53,6 +56,7 @@ impl AgentProvider {
            "openai-compatible" => Ok(Self::OpenAICompat),
            "openrouter" => Ok(Self::OpenRouter),
            "together-ai" => Ok(Self::TogetherAI),
+            "venice" => Ok(Self::Venice),
            _ => Err("Unexpected string value"),
        }
    }
@@ -181,6 +185,25 @@ impl AgentProvider {
                text_generation_supports_vision: false,
                text_generation_supports_tools: false,
            },
+            Self::Venice => AgentProviderInfo {
+                id: Self::Venice.to_static_str(),
+                name: "Venice",
+                description: "Venice AI runs inference on Venice-controlled GPUs or zero-data-retention partner infrastructure and stores no prompts or responses. It serves frontier proprietary and open-source models with text-generation (including vision), speech-to-text, text-to-speech, native image generation and editing, and native web search.",
+                homepage_url: Some("https://venice.ai"),
+                wiki_url: None,
+                sign_up_url: Some("https://venice.ai"),
+                models_list_url: Some("https://api.venice.ai/api/v1/models"),
+                supported_purposes: vec![
+                    AgentPurpose::ImageGeneration,
+                    AgentPurpose::TextGeneration,
+                    AgentPurpose::TextToSpeech,
+                    AgentPurpose::SpeechToText,
+                ],
+                text_generation_supports_vision: true,
+                // Venice does native web search via `venice_parameters`, NOT baibot's built-in
+                // tools mechanism (the OpenAI web_search/code_interpreter block), so this is false.
+                text_generation_supports_tools: false,
+            },
        }
    }
 }
--- a/src/agent/provider/mod.rs
+++ b/src/agent/provider/mod.rs
@@ -10,6 +10,7 @@ pub mod openai;
 pub mod openai_compat;
 pub(super) mod openrouter;
 pub(super) mod togetherai;
+pub mod venice;

 fn default_temperature() -> f32 {
    1.0
--- a/src/agent/provider/venice/audio.rs
+++ b/src/agent/provider/venice/audio.rs
@@ -0,0 +1,156 @@
+use crate::agent::AgentPurpose;
+use crate::agent::provider::entity::{TextToSpeechParams, TextToSpeechResult};
+use crate::agent::provider::{SpeechToTextParams, SpeechToTextResult};
+use crate::strings;
+
+use super::config::Config;
+use super::wire::{SpeechRequest, TranscriptionResponse};
+
+pub async fn speech_to_text(
+    config: &Config,
+    http: &reqwest::Client,
+    mime_type: &mxlink::mime::Mime,
+    media: Vec<u8>,
+    params: SpeechToTextParams,
+) -> anyhow::Result<SpeechToTextResult> {
+    let Some(speech_to_text_config) = &config.speech_to_text else {
+        return Err(anyhow::anyhow!(
+            strings::agent::no_configuration_for_purpose_so_cannot_be_used(
+                &AgentPurpose::SpeechToText
+            ),
+        ));
+    };
+
+    // Unlike the openai_compat path (which writes the audio to a temp file because its library
+    // can't take bytes), reqwest's multipart takes the bytes directly.
+    let part = reqwest::multipart::Part::bytes(media)
+        .file_name("audio")
+        .mime_str(mime_type.as_ref())?;
+
+    let mut form = reqwest::multipart::Form::new()
+        .part("file", part)
+        .text("model", speech_to_text_config.model_id.clone())
+        .text("response_format", "json");
+
+    if let Some(language) = &params.language_override {
+        form = form.text("language", language.clone());
+    }
+
+    let url = format!(
+        "{}/audio/transcriptions",
+        config.base_url.trim_end_matches('/')
+    );
+
+    tracing::trace!(
+        model_id = speech_to_text_config.model_id,
+        language = ?params.language_override,
+        "Sending Venice audio transcription API request"
+    );
+
+    let response = http
+        .post(&url)
+        .bearer_auth(&config.api_key)
+        .multipart(form)
+        .send()
+        .await?;
+
+    let status = response.status();
+    if !status.is_success() {
+        // Body to the server log only, not into the returned error (which reaches the Matrix room).
+        let body = response.text().await.unwrap_or_default();
+        tracing::warn!(%status, body, "Venice audio transcription request failed");
+        return Err(anyhow::anyhow!(
+            "Venice audio transcription request failed with status {status}"
+        ));
+    }
+
+    let response: TranscriptionResponse = response.json().await?;
+
+    Ok(SpeechToTextResult {
+        text: response.text,
+    })
+}
+
+pub async fn text_to_speech(
+    config: &Config,
+    http: &reqwest::Client,
+    input: &str,
+    params: TextToSpeechParams,
+) -> anyhow::Result<TextToSpeechResult> {
+    let Some(text_to_speech_config) = &config.text_to_speech else {
+        return Err(anyhow::anyhow!(
+            strings::agent::no_configuration_for_purpose_so_cannot_be_used(
+                &AgentPurpose::TextToSpeech
+            ),
+        ));
+    };
+
+    // Per-call overrides win over the configured defaults.
+    let voice = params
+        .voice_override
+        .or_else(|| text_to_speech_config.voice.clone());
+    let speed = params.speed_override.or(text_to_speech_config.speed);
+
+    let response_format = text_to_speech_config.response_format.clone();
+    let mime_type = response_format_to_mime_type(response_format.as_deref());
+
+    let request = SpeechRequest {
+        model: text_to_speech_config.model_id.clone(),
+        input: input.to_owned(),
+        voice,
+        speed,
+        response_format,
+        prompt: text_to_speech_config.prompt.clone(),
+        temperature: text_to_speech_config.temperature,
+        top_p: text_to_speech_config.top_p,
+    };
+
+    let url = format!("{}/audio/speech", config.base_url.trim_end_matches('/'));
+
+    tracing::trace!(
+        model_id = text_to_speech_config.model_id,
+        voice = ?request.voice,
+        "Sending Venice text-to-speech API request"
+    );
+
+    let response = http
+        .post(&url)
+        .bearer_auth(&config.api_key)
+        .json(&request)
+        .send()
+        .await?;
+
+    let status = response.status();
+    if !status.is_success() {
+        // Body to the server log only, not into the returned error (which reaches the Matrix room).
+        let body = response.text().await.unwrap_or_default();
+        tracing::warn!(%status, body, "Venice text-to-speech request failed");
+        return Err(anyhow::anyhow!(
+            "Venice text-to-speech request failed with status {status}"
+        ));
+    }
+
+    // The speech endpoint answers with raw binary audio; read the body directly.
+    let bytes = response.bytes().await?.to_vec();
+
+    Ok(TextToSpeechResult { bytes, mime_type })
+}
+
+/// Map a Venice TTS `response_format` to its MIME type. Defaults to `audio/mpeg` (the
+/// IANA-registered MP3 type, RFC 3003) when the format is unset, matching Venice's own `mp3`
+/// default. This deliberately uses `audio/mpeg` rather than the `audio/mp3` alias the openai
+/// provider emits; baibot's downstream audio-filename mapping treats both as `.mp3`.
+fn response_format_to_mime_type(response_format: Option<&str>) -> mxlink::mime::Mime {
+    let raw = match response_format.unwrap_or("mp3") {
+        "mp3" => "audio/mpeg",
+        "opus" => "audio/ogg",
+        "aac" => "audio/aac",
+        "flac" => "audio/flac",
+        "wav" => "audio/wav",
+        "pcm" => "audio/L8",
+        _ => "audio/mpeg",
+    };
+
+    raw.parse()
+        .unwrap_or(mxlink::mime::APPLICATION_OCTET_STREAM)
+}
--- a/src/agent/provider/venice/chat.rs
+++ b/src/agent/provider/venice/chat.rs
@@ -0,0 +1,122 @@
+use crate::agent::AgentPurpose;
+use crate::agent::provider::entity::{TextGenerationParams, TextGenerationResult};
+use crate::conversation::llm::{
+    Author as LLMAuthor, Conversation as LLMConversation, Message as LLMMessage,
+    MessageContent as LLMMessageContent, shorten_messages_list_to_context_size,
+};
+use crate::strings;
+
+use super::config::Config;
+use super::utils::convert_llm_messages_to_venice;
+use super::wire::{ChatCompletionRequest, ChatCompletionResponse};
+
+pub async fn generate_text(
+    config: &Config,
+    http: &reqwest::Client,
+    conversation: LLMConversation,
+    params: TextGenerationParams,
+) -> anyhow::Result<TextGenerationResult> {
+    let Some(text_generation_config) = &config.text_generation else {
+        return Err(anyhow::anyhow!(
+            strings::agent::no_configuration_for_purpose_so_cannot_be_used(
+                &AgentPurpose::TextGeneration
+            ),
+        ));
+    };
+
+    let prompt_text = params.prompt_variables.format(
+        params
+            .prompt_override
+            .unwrap_or(text_generation_config.prompt.clone().unwrap_or_default())
+            .trim(),
+    );
+
+    let prompt_message = if prompt_text.is_empty() {
+        None
+    } else {
+        Some(LLMMessage {
+            author: LLMAuthor::Prompt,
+            sender_id: None,
+            content: LLMMessageContent::Text(prompt_text),
+            timestamp: chrono::Utc::now(),
+        })
+    };
+
+    let mut conversation_messages = conversation.messages;
+
+    if params.context_management_enabled {
+        conversation_messages = shorten_messages_list_to_context_size(
+            &text_generation_config.model_id,
+            &prompt_message,
+            conversation_messages,
+            text_generation_config.max_response_tokens,
+            text_generation_config.max_context_tokens,
+        );
+    }
+
+    if let Some(prompt_message) = prompt_message {
+        conversation_messages.insert(0, prompt_message);
+    }
+
+    let messages = convert_llm_messages_to_venice(conversation_messages);
+
+    let temperature = params
+        .temperature_override
+        .unwrap_or(text_generation_config.temperature);
+
+    let request = ChatCompletionRequest {
+        model: text_generation_config.model_id.clone(),
+        messages,
+        temperature: Some(temperature),
+        // Web search rides entirely inside `venice_parameters`; there is no `tools` array here.
+        // `max_tokens` is deprecated on Venice in favor of `max_completion_tokens`.
+        max_completion_tokens: text_generation_config.max_response_tokens,
+        venice_parameters: text_generation_config.venice_parameters.clone(),
+    };
+
+    let url = format!(
+        "{}/chat/completions",
+        config.base_url.trim_end_matches('/')
+    );
+
+    tracing::trace!(
+        model = text_generation_config.model_id,
+        messages_count = request.messages.len(),
+        "Sending Venice chat completion API request"
+    );
+
+    let response = http
+        .post(&url)
+        .bearer_auth(&config.api_key)
+        .json(&request)
+        .send()
+        .await?;
+
+    let status = response.status();
+    if !status.is_success() {
+        // Log the body server-side for debugging (Venice explains a rejected strict body there),
+        // but keep it OUT of the returned error: that error surfaces in the Matrix room, and the
+        // body can carry account / rate-limit details that shouldn't reach room members.
+        let body = response.text().await.unwrap_or_default();
+        tracing::warn!(%status, body, "Venice chat completion request failed");
+        return Err(anyhow::anyhow!(
+            "Venice chat completion request failed with status {status}"
+        ));
+    }
+
+    let response: ChatCompletionResponse = response.json().await?;
+
+    let Some(choice) = response.choices.into_iter().next() else {
+        return Err(anyhow::anyhow!(
+            "No choices were returned from the Venice chat completion API"
+        ));
+    };
+
+    let Some(content) = choice.message.content else {
+        return Err(anyhow::anyhow!(
+            "No message content was returned from the Venice chat completion API"
+        ));
+    };
+
+    Ok(TextGenerationResult { text: content })
+}
--- a/src/agent/provider/venice/config.rs
+++ b/src/agent/provider/venice/config.rs
@@ -0,0 +1,355 @@
+use serde::{Deserialize, Serialize};
+
+use crate::agent::{default_prompt, provider::ConfigTrait};
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct Config {
+    pub base_url: String,
+
+    pub api_key: String,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub text_generation: Option<TextGenerationConfig>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub speech_to_text: Option<SpeechToTextConfig>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub text_to_speech: Option<TextToSpeechConfig>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_generation: Option<ImageGenerationConfig>,
+}
+
+impl Default for Config {
+    fn default() -> Self {
+        Self {
+            base_url: "https://api.venice.ai/api/v1".to_owned(),
+            api_key: "YOUR_API_KEY_HERE".to_owned(),
+            text_generation: Some(TextGenerationConfig::default()),
+            speech_to_text: Some(SpeechToTextConfig::default()),
+            text_to_speech: Some(TextToSpeechConfig::default()),
+            image_generation: Some(ImageGenerationConfig::default()),
+        }
+    }
+}
+
+impl ConfigTrait for Config {
+    fn validate(&self) -> Result<(), String> {
+        if self.base_url.is_empty() {
+            return Err("The base URL must not be empty.".to_owned());
+        }
+
+        Ok(())
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct TextGenerationConfig {
+    #[serde(default = "default_text_model_id")]
+    pub model_id: String,
+
+    #[serde(default)]
+    pub prompt: Option<String>,
+
+    #[serde(default = "super::super::default_temperature")]
+    pub temperature: f32,
+
+    #[serde(default)]
+    pub max_response_tokens: Option<u32>,
+
+    #[serde(default)]
+    pub max_context_tokens: u32,
+
+    /// Venice-specific request knobs, serialized 1:1 into the `venice_parameters` bag on the
+    /// wire. Any unset field is omitted, so Venice applies its own server-side default.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub venice_parameters: Option<VeniceParameters>,
+}
+
+impl Default for TextGenerationConfig {
+    fn default() -> Self {
+        Self {
+            model_id: default_text_model_id(),
+            prompt: Some(default_prompt().to_owned()),
+            temperature: super::super::default_temperature(),
+            // Reserved output budget: sent as the response cap AND subtracted from the context
+            // window when trimming history. Mirrors the openai_compat sibling's default.
+            max_response_tokens: Some(4096),
+            // Matches Venice's own `availableContextTokens` (131072) and the non-OpenAI sibling
+            // providers (ollama/localai/mistral all default to 128_000).
+            max_context_tokens: 128_000,
+            // A usable starting point, not an everything-set dump: only these three are sent;
+            // every other knob stays None so Venice applies its own default (omitting != false).
+            venice_parameters: Some(VeniceParameters {
+                enable_web_search: Some(WebSearchMode::Auto),
+                strip_thinking_response: Some(true),
+                enable_e2ee: Some(false),
+                ..Default::default()
+            }),
+        }
+    }
+}
+
+fn default_text_model_id() -> String {
+    "kimi-k2-5".to_owned()
+}
+
+/// The full `venice_parameters` knob set, mirroring Venice's `ChatCompletionRequest`
+/// schema field-for-field. Every field is optional with `skip_serializing_if`, so the
+/// request never carries a knob the user didn't set (the body is `additionalProperties: false`,
+/// and an unset knob simply omits rather than sending `null`).
+#[derive(Debug, Clone, Serialize, Deserialize, Default)]
+pub struct VeniceParameters {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub enable_web_search: Option<WebSearchMode>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub enable_web_citations: Option<bool>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub enable_web_scraping: Option<bool>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub include_venice_system_prompt: Option<bool>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub include_search_results_in_stream: Option<bool>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub return_search_results_as_documents: Option<bool>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub enable_x_search: Option<bool>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub enable_e2ee: Option<bool>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub character_slug: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub strip_thinking_response: Option<bool>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub disable_thinking: Option<bool>,
+}
+
+#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
+#[serde(rename_all = "lowercase")]
+pub enum WebSearchMode {
+    Auto,
+    On,
+    Off,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct SpeechToTextConfig {
+    #[serde(default = "default_speech_to_text_model_id")]
+    pub model_id: String,
+}
+
+impl Default for SpeechToTextConfig {
+    fn default() -> Self {
+        Self {
+            model_id: default_speech_to_text_model_id(),
+        }
+    }
+}
+
+fn default_speech_to_text_model_id() -> String {
+    "nvidia/parakeet-tdt-0.6b-v3".to_owned()
+}
+
+/// `/audio/speech` (`CreateSpeechRequestSchema`) request knobs. Only `model_id` is required on
+/// the wire; everything else is optional with `skip_serializing_if` so an unset knob is omitted
+/// rather than sent as `null` (the body is `additionalProperties: false`). `voice` is a free
+/// `Option<String>`, not a closed enum: Venice's voice set spans dozens of model-specific names
+/// plus arbitrary cloned-voice handles (`vv_<id>`), so an enum would reject valid handles.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct TextToSpeechConfig {
+    #[serde(default = "default_text_to_speech_model_id")]
+    pub model_id: String,
+
+    #[serde(
+        default = "default_text_to_speech_voice",
+        skip_serializing_if = "Option::is_none"
+    )]
+    pub voice: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub speed: Option<f32>,
+
+    #[serde(
+        default = "default_text_to_speech_response_format",
+        skip_serializing_if = "Option::is_none"
+    )]
+    pub response_format: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub prompt: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub temperature: Option<f32>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub top_p: Option<f32>,
+}
+
+impl Default for TextToSpeechConfig {
+    fn default() -> Self {
+        Self {
+            model_id: default_text_to_speech_model_id(),
+            voice: default_text_to_speech_voice(),
+            speed: None,
+            response_format: default_text_to_speech_response_format(),
+            prompt: None,
+            temperature: None,
+            top_p: None,
+        }
+    }
+}
+
+fn default_text_to_speech_model_id() -> String {
+    "tts-kokoro".to_owned()
+}
+
+fn default_text_to_speech_voice() -> Option<String> {
+    Some("af_sky".to_owned())
+}
+
+fn default_text_to_speech_response_format() -> Option<String> {
+    Some("mp3".to_owned())
+}
+
+/// `/image/generate` (`GenerateImageRequest`) request knobs, mirroring Venice's schema
+/// field-for-field. Only `model_id` is required; every other knob is optional with
+/// `skip_serializing_if` so unset knobs are omitted (the body is `additionalProperties: false`).
+/// The full knob set is deliberate: the native `/image/generate` endpoint is the flagship's
+/// reason to exist over the knob-dropping OpenAI-compat path, so the knobs ARE the feature.
+/// The deprecated `inpaint` knob is intentionally absent.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ImageGenerationConfig {
+    #[serde(default = "default_image_generation_model_id")]
+    pub model_id: String,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub negative_prompt: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub cfg_scale: Option<f32>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub steps: Option<u32>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub style_preset: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub seed: Option<i64>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub safe_mode: Option<bool>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub hide_watermark: Option<bool>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub format: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub width: Option<u32>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub height: Option<u32>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub aspect_ratio: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub resolution: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub quality: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub lora_strength: Option<u32>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub embed_exif_metadata: Option<bool>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub enable_web_search: Option<bool>,
+
+    /// Image-edit settings, nested here because baibot has a single `ImageGeneration` purpose
+    /// and edit shares its config gate. The gen and edit model sets are disjoint, so edit
+    /// carries its own model field.
+    #[serde(default)]
+    pub edit: ImageEditSettings,
+}
+
+impl Default for ImageGenerationConfig {
+    fn default() -> Self {
+        Self {
+            model_id: default_image_generation_model_id(),
+            negative_prompt: None,
+            cfg_scale: None,
+            steps: None,
+            style_preset: None,
+            seed: None,
+            safe_mode: None,
+            hide_watermark: None,
+            format: None,
+            width: None,
+            height: None,
+            aspect_ratio: None,
+            resolution: None,
+            quality: None,
+            lora_strength: None,
+            embed_exif_metadata: None,
+            enable_web_search: None,
+            edit: ImageEditSettings::default(),
+        }
+    }
+}
+
+fn default_image_generation_model_id() -> String {
+    "chroma".to_owned()
+}
+
+/// `/image/edit` (`EditImageRequest`) request knobs, mirroring Venice's schema. The source image
+/// and prompt are supplied per-call (not config), so only the model and the output-shaping knobs
+/// live here. Each knob is optional with `skip_serializing_if`.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ImageEditSettings {
+    #[serde(default = "default_image_edit_model_id")]
+    pub model_id: String,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub output_format: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub aspect_ratio: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub resolution: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub safe_mode: Option<bool>,
+}
+
+impl Default for ImageEditSettings {
+    fn default() -> Self {
+        Self {
+            model_id: default_image_edit_model_id(),
+            output_format: None,
+            aspect_ratio: None,
+            resolution: None,
+            safe_mode: None,
+        }
+    }
+}
+
+fn default_image_edit_model_id() -> String {
+    "firered-image-edit".to_owned()
+}
--- a/src/agent/provider/venice/controller.rs
+++ b/src/agent/provider/venice/controller.rs
@@ -0,0 +1,146 @@
+use crate::agent::AgentPurpose;
+use crate::agent::provider::entity::{
+    ImageEditResult, ImageGenerationResult, ImageSource, PingResult, TextGenerationParams,
+    TextGenerationResult, TextToSpeechParams, TextToSpeechResult,
+};
+use crate::agent::provider::{
+    ImageEditParams, ImageGenerationParams, SpeechToTextParams, SpeechToTextResult,
+};
+use crate::conversation::llm::{
+    Author as LLMAuthor, Conversation as LLMConversation, Message as LLMMessage,
+    MessageContent as LLMMessageContent,
+};
+
+use super::super::ControllerTrait;
+use super::config::Config;
+
+#[derive(Debug, Clone)]
+pub struct Controller {
+    config: Config,
+    http: reqwest::Client,
+}
+
+impl Controller {
+    pub fn new(config: Config) -> Self {
+        // Image generation and text-to-speech can run long, so give the client a generous timeout
+        // instead of reqwest's default (none). `build` only fails on TLS/system init; fall back to
+        // the infallible `Client::new()` so this constructor stays infallible.
+        let http = reqwest::Client::builder()
+            .timeout(std::time::Duration::from_secs(120))
+            .build()
+            .unwrap_or_else(|_| reqwest::Client::new());
+
+        Self { config, http }
+    }
+}
+
+impl ControllerTrait for Controller {
+    async fn ping(&self) -> anyhow::Result<PingResult> {
+        if !self.supports_purpose(AgentPurpose::TextGeneration) {
+            return Ok(PingResult::Inconclusive);
+        }
+
+        // Mirror the openai/openai_compat ping: a real "Hello!" round-trip exercises the strict
+        // /chat/completions body and auth, so a successful ping proves text generation works.
+        let messages = vec![LLMMessage {
+            author: LLMAuthor::User,
+            sender_id: None,
+            content: LLMMessageContent::Text("Hello!".to_string()),
+            timestamp: chrono::Utc::now(),
+        }];
+
+        let conversation = LLMConversation { messages };
+
+        self.generate_text(conversation, TextGenerationParams::default())
+            .await?;
+
+        Ok(PingResult::Successful)
+    }
+
+    async fn generate_text(
+        &self,
+        conversation: LLMConversation,
+        params: TextGenerationParams,
+    ) -> anyhow::Result<TextGenerationResult> {
+        super::chat::generate_text(&self.config, &self.http, conversation, params).await
+    }
+
+    async fn speech_to_text(
+        &self,
+        mime_type: &mxlink::mime::Mime,
+        media: Vec<u8>,
+        params: SpeechToTextParams,
+    ) -> anyhow::Result<SpeechToTextResult> {
+        super::audio::speech_to_text(&self.config, &self.http, mime_type, media, params).await
+    }
+
+    async fn generate_image(
+        &self,
+        prompt: &str,
+        params: ImageGenerationParams,
+    ) -> anyhow::Result<ImageGenerationResult> {
+        super::images::generate_image(&self.config, &self.http, prompt, params).await
+    }
+
+    async fn create_image_edit(
+        &self,
+        prompt: &str,
+        images: Vec<ImageSource>,
+        params: ImageEditParams,
+    ) -> anyhow::Result<ImageEditResult> {
+        super::images::create_image_edit(&self.config, &self.http, prompt, images, params).await
+    }
+
+    async fn text_to_speech(
+        &self,
+        input: &str,
+        params: TextToSpeechParams,
+    ) -> anyhow::Result<TextToSpeechResult> {
+        super::audio::text_to_speech(&self.config, &self.http, input, params).await
+    }
+
+    fn supports_purpose(&self, purpose: AgentPurpose) -> bool {
+        match purpose {
+            AgentPurpose::TextGeneration => self.config.text_generation.is_some(),
+            AgentPurpose::SpeechToText => self.config.speech_to_text.is_some(),
+            AgentPurpose::TextToSpeech => self.config.text_to_speech.is_some(),
+            AgentPurpose::ImageGeneration => self.config.image_generation.is_some(),
+            AgentPurpose::CatchAll => true,
+        }
+    }
+
+    fn text_generation_model_id(&self) -> Option<String> {
+        self.config
+            .text_generation
+            .as_ref()
+            .map(|config| config.model_id.to_owned())
+    }
+
+    fn text_generation_prompt(&self) -> Option<String> {
+        self.config
+            .text_generation
+            .as_ref()
+            .and_then(|config| config.prompt.clone())
+    }
+
+    fn text_generation_temperature(&self) -> Option<f32> {
+        self.config
+            .text_generation
+            .as_ref()
+            .map(|config| config.temperature)
+    }
+
+    fn text_to_speech_voice(&self) -> Option<String> {
+        self.config
+            .text_to_speech
+            .as_ref()
+            .and_then(|config| config.voice.clone())
+    }
+
+    fn text_to_speech_speed(&self) -> Option<f32> {
+        self.config
+            .text_to_speech
+            .as_ref()
+            .and_then(|config| config.speed)
+    }
+}
--- a/src/agent/provider/venice/images.rs
+++ b/src/agent/provider/venice/images.rs
@@ -0,0 +1,192 @@
+use crate::agent::AgentPurpose;
+use crate::agent::provider::entity::{ImageEditResult, ImageGenerationResult, ImageSource};
+use crate::agent::provider::{ImageEditParams, ImageGenerationParams};
+use crate::strings;
+use crate::utils::base64::{base64_decode, base64_encode};
+
+use super::config::Config;
+use super::wire::{EditImageRequest, GenerateImageRequest, GenerateImageResponse};
+
+/// Generate an image via Venice's native `/image/generate` endpoint.
+///
+/// This is the base64-in-JSON path: we pin `return_binary: false` so Venice answers with a JSON
+/// envelope (`GenerateImageResponse`) carrying the image as a base64 string, which we decode. The
+/// sibling `create_image_edit` is the *other* response shape (raw binary); the two must not be
+/// crossed. `params` is advisory only; the Venice config drives the request.
+pub async fn generate_image(
+    config: &Config,
+    http: &reqwest::Client,
+    prompt: &str,
+    _params: ImageGenerationParams,
+) -> anyhow::Result<ImageGenerationResult> {
+    let Some(image_generation_config) = &config.image_generation else {
+        return Err(anyhow::anyhow!(
+            strings::agent::no_configuration_for_purpose_so_cannot_be_used(
+                &AgentPurpose::ImageGeneration
+            ),
+        ));
+    };
+
+    let request = GenerateImageRequest {
+        model: image_generation_config.model_id.clone(),
+        prompt: prompt.to_owned(),
+        // Pinned: baibot wants exactly one image, returned as base64-in-JSON so `GenerateImageResponse`
+        // can decode it. Flipping `return_binary` would make Venice answer with raw binary and break
+        // the JSON decode below, so neither knob is configurable.
+        return_binary: false,
+        variants: 1,
+        negative_prompt: image_generation_config.negative_prompt.clone(),
+        cfg_scale: image_generation_config.cfg_scale,
+        steps: image_generation_config.steps,
+        style_preset: image_generation_config.style_preset.clone(),
+        seed: image_generation_config.seed,
+        safe_mode: image_generation_config.safe_mode,
+        hide_watermark: image_generation_config.hide_watermark,
+        format: image_generation_config.format.clone(),
+        width: image_generation_config.width,
+        height: image_generation_config.height,
+        aspect_ratio: image_generation_config.aspect_ratio.clone(),
+        resolution: image_generation_config.resolution.clone(),
+        quality: image_generation_config.quality.clone(),
+        lora_strength: image_generation_config.lora_strength,
+        embed_exif_metadata: image_generation_config.embed_exif_metadata,
+        enable_web_search: image_generation_config.enable_web_search,
+    };
+
+    let url = format!("{}/image/generate", config.base_url.trim_end_matches('/'));
+
+    // The prompt is user content; keep it out of logs (mirrors the STT/TTS paths).
+    tracing::trace!(
+        model_id = image_generation_config.model_id,
+        "Sending Venice image generation API request"
+    );
+
+    let response = http
+        .post(&url)
+        .bearer_auth(&config.api_key)
+        .json(&request)
+        .send()
+        .await?;
+
+    let status = response.status();
+    if !status.is_success() {
+        // Body to the server log only, not into the returned error (which reaches the Matrix room).
+        let body = response.text().await.unwrap_or_default();
+        tracing::warn!(%status, body, "Venice image generation request failed");
+        return Err(anyhow::anyhow!(
+            "Venice image generation request failed with status {status}"
+        ));
+    }
+
+    let response: GenerateImageResponse = response.json().await?;
+
+    tracing::trace!(request_id = ?response.id, "Venice image generation succeeded");
+
+    let Some(image_base64) = response.images.into_iter().next() else {
+        return Err(anyhow::anyhow!(
+            "The Venice image generation API returned no images"
+        ));
+    };
+
+    // Swallow the decode error's detail (it can echo input bytes/offsets); the returned error
+    // reaches the Matrix room, so it stays generic while the real cause goes to the server log.
+    let bytes = base64_decode(&image_base64).map_err(|decode_err| {
+        tracing::warn!(%decode_err, "Venice image generation returned undecodable base64");
+        anyhow::anyhow!("Venice image generation returned invalid base64 image data")
+    })?;
+
+    Ok(ImageGenerationResult {
+        bytes,
+        mime_type: image_format_to_mime_type(image_generation_config.format.as_deref()),
+        revised_prompt: None,
+    })
+}
+
+/// Edit an image via Venice's native `/image/edit` endpoint.
+///
+/// This is the raw-binary path: the request is JSON carrying the source image as a base64 string
+/// (Venice's `image` field is `anyOf` upload/base64/URL; we send base64, no multipart), and the
+/// response body IS the edited image bytes (no JSON envelope). `params` is advisory only.
+pub async fn create_image_edit(
+    config: &Config,
+    http: &reqwest::Client,
+    prompt: &str,
+    images: Vec<ImageSource>,
+    _params: ImageEditParams,
+) -> anyhow::Result<ImageEditResult> {
+    let Some(image_generation_config) = &config.image_generation else {
+        return Err(anyhow::anyhow!(
+            strings::agent::no_configuration_for_purpose_so_cannot_be_used(
+                &AgentPurpose::ImageGeneration
+            ),
+        ));
+    };
+
+    let edit_config = &image_generation_config.edit;
+
+    let Some(source) = images.into_iter().next() else {
+        return Err(anyhow::anyhow!("No image sources provided"));
+    };
+
+    let request = EditImageRequest {
+        model: edit_config.model_id.clone(),
+        prompt: prompt.to_owned(),
+        image: base64_encode(&source.bytes),
+        output_format: edit_config.output_format.clone(),
+        aspect_ratio: edit_config.aspect_ratio.clone(),
+        resolution: edit_config.resolution.clone(),
+        safe_mode: edit_config.safe_mode,
+    };
+
+    let url = format!("{}/image/edit", config.base_url.trim_end_matches('/'));
+
+    // The prompt is user content; keep it out of logs (mirrors the STT/TTS paths).
+    tracing::trace!(
+        model_id = edit_config.model_id,
+        "Sending Venice image edit API request"
+    );
+
+    let response = http
+        .post(&url)
+        .bearer_auth(&config.api_key)
+        .json(&request)
+        .send()
+        .await?;
+
+    let status = response.status();
+    if !status.is_success() {
+        // Body to the server log only, not into the returned error (which reaches the Matrix room).
+        let body = response.text().await.unwrap_or_default();
+        tracing::warn!(%status, body, "Venice image edit request failed");
+        return Err(anyhow::anyhow!(
+            "Venice image edit request failed with status {status}"
+        ));
+    }
+
+    // The edit endpoint answers with raw binary image bytes, so read the body directly instead of
+    // parsing JSON. The actual format comes from the response Content-Type header; fall back to the
+    // configured `output_format` when the header is missing or unparseable.
+    let mime_type = response
+        .headers()
+        .get(reqwest::header::CONTENT_TYPE)
+        .and_then(|value| value.to_str().ok())
+        .and_then(|value| value.parse::<mxlink::mime::Mime>().ok())
+        .unwrap_or_else(|| image_format_to_mime_type(edit_config.output_format.as_deref()));
+
+    let bytes = response.bytes().await?.to_vec();
+
+    Ok(ImageEditResult { bytes, mime_type })
+}
+
+/// Map a Venice image `format`/`output_format` value (`jpeg`/`png`/`webp`) to its MIME type.
+/// Venice defaults to `webp` when the format is unset, so an absent value maps to `image/webp`.
+fn image_format_to_mime_type(format: Option<&str>) -> mxlink::mime::Mime {
+    match format.unwrap_or("webp") {
+        "jpeg" | "jpg" => mxlink::mime::IMAGE_JPEG,
+        "png" => mxlink::mime::IMAGE_PNG,
+        // No mxlink::mime constant for webp; parse it, falling back to PNG on any surprise value.
+        _ => "image/webp"
+            .parse()
+            .unwrap_or(mxlink::mime::IMAGE_PNG),
+    }
+}
--- a/src/agent/provider/venice/mod.rs
+++ b/src/agent/provider/venice/mod.rs
@@ -0,0 +1,47 @@
+mod audio;
+mod chat;
+mod config;
+mod controller;
+mod images;
+mod utils;
+mod wire;
+
+#[cfg(test)]
+mod tests;
+
+pub use config::Config;
+pub use controller::Controller;
+
+use super::super::AgentInstantiationError;
+use super::super::AgentInstantiationResult;
+use super::ConfigTrait;
+use super::controller::ControllerType;
+
+pub fn create_controller_from_yaml_value_config(
+    agent_id: &str,
+    config: serde_yaml_ng::Value,
+) -> AgentInstantiationResult<ControllerType> {
+    let config = match &config {
+        serde_yaml_ng::Value::Mapping(_) => {
+            let config: Config =
+                serde_yaml_ng::from_value(config).map_err(AgentInstantiationError::Yaml)?;
+
+            config
+                .validate()
+                .map_err(AgentInstantiationError::ConfigFailsValidation)?;
+
+            config
+        }
+        _ => {
+            return Err(AgentInstantiationError::ConfigForAgentIsNotAMapping(
+                agent_id.to_owned(),
+            ));
+        }
+    };
+
+    Ok(ControllerType::Venice(Box::new(Controller::new(config))))
+}
+
+pub fn default_config() -> Config {
+    Config::default()
+}
--- a/src/agent/provider/venice/tests.rs
+++ b/src/agent/provider/venice/tests.rs
@@ -0,0 +1,269 @@
+use mxlink::matrix_sdk::ruma::OwnedMxcUri;
+use mxlink::matrix_sdk::ruma::events::room::message::{
+    FileMessageEventContent, ImageMessageEventContent,
+};
+use mxlink::mime;
+
+use super::super::ControllerTrait;
+use crate::agent::AgentPurpose;
+use crate::conversation::llm::{
+    Author as LLMAuthor, FileDetails, ImageDetails, Message as LLMMessage,
+    MessageContent as LLMMessageContent,
+};
+
+use super::config::{Config, VeniceParameters, WebSearchMode};
+use super::controller::Controller;
+use super::utils::convert_llm_messages_to_venice;
+use super::wire::{
+    ContentPart, EditImageRequest, GenerateImageRequest, MessageContent, SpeechRequest,
+};
+
+#[test]
+fn config_round_trips_with_venice_parameters() {
+    let yaml = r#"
+base_url: https://api.venice.ai/api/v1
+api_key: test-key
+text_generation:
+  model_id: kimi-k2-5
+  temperature: 0.7
+  max_response_tokens: 1024
+  max_context_tokens: 65536
+  venice_parameters:
+    enable_web_search: "auto"
+    enable_web_citations: true
+speech_to_text:
+  model_id: nvidia/parakeet-tdt-0.6b-v3
+"#;
+
+    let config: Config = serde_yaml_ng::from_str(yaml).expect("config should deserialize");
+
+    let tg = config.text_generation.expect("text_generation present");
+    let vp = tg.venice_parameters.expect("venice_parameters present");
+
+    assert!(matches!(vp.enable_web_search, Some(WebSearchMode::Auto)));
+    assert_eq!(vp.enable_web_citations, Some(true));
+    assert_eq!(vp.character_slug, None);
+
+    // The bag must serialize the enum to the exact wire string, and an unset knob must be ABSENT
+    // (not `null`) so the strict `additionalProperties: false` body is honored.
+    let json = serde_json::to_string(&vp).expect("serialize venice_parameters");
+    assert!(
+        json.contains("\"enable_web_search\":\"auto\""),
+        "web search should be the literal \"auto\": {json}"
+    );
+    assert!(
+        !json.contains("character_slug"),
+        "an unset knob must be omitted entirely: {json}"
+    );
+    assert!(!json.contains("null"), "no nulls belong in the body: {json}");
+}
+
+#[test]
+fn converts_image_to_data_uri_and_skips_files() {
+    let messages = vec![
+        LLMMessage {
+            author: LLMAuthor::User,
+            sender_id: None,
+            timestamp: chrono::Utc::now(),
+            content: LLMMessageContent::Text("describe this".to_owned()),
+        },
+        LLMMessage {
+            author: LLMAuthor::User,
+            sender_id: None,
+            timestamp: chrono::Utc::now(),
+            content: LLMMessageContent::Image(ImageDetails::new(
+                ImageMessageEventContent::plain(
+                    "pic.png".to_owned(),
+                    OwnedMxcUri::from("mxc://example.com/abc"),
+                ),
+                mime::IMAGE_PNG,
+                vec![1, 2, 3],
+            )),
+        },
+        LLMMessage {
+            author: LLMAuthor::User,
+            sender_id: None,
+            timestamp: chrono::Utc::now(),
+            content: LLMMessageContent::File(FileDetails::new(
+                FileMessageEventContent::plain(
+                    "doc.pdf".to_owned(),
+                    OwnedMxcUri::from("mxc://example.com/def"),
+                ),
+                mime::APPLICATION_PDF,
+                vec![4, 5, 6],
+            )),
+        },
+    ];
+
+    let converted = convert_llm_messages_to_venice(messages);
+
+    // Text and image survive; the file is warn-skipped.
+    assert_eq!(converted.len(), 2);
+
+    match &converted[0].content {
+        MessageContent::Text(text) => assert_eq!(text, "describe this"),
+        other => panic!("expected bare text, got {other:?}"),
+    }
+
+    match &converted[1].content {
+        MessageContent::Parts(parts) => match &parts[0] {
+            ContentPart::ImageUrl { image_url } => assert!(
+                image_url.url.starts_with("data:image/png;base64,"),
+                "image should be inlined as a data URI: {}",
+                image_url.url
+            ),
+        },
+        other => panic!("expected image parts, got {other:?}"),
+    }
+}
+
+#[test]
+fn supports_purpose_truth_table() {
+    let config: Config = serde_yaml_ng::from_str(
+        r#"
+base_url: https://api.venice.ai/api/v1
+api_key: test-key
+text_generation:
+  model_id: kimi-k2-5
+speech_to_text:
+  model_id: nvidia/parakeet-tdt-0.6b-v3
+"#,
+    )
+    .expect("config should deserialize");
+
+    let controller = Controller::new(config);
+
+    assert!(controller.supports_purpose(AgentPurpose::TextGeneration));
+    assert!(controller.supports_purpose(AgentPurpose::SpeechToText));
+    assert!(controller.supports_purpose(AgentPurpose::CatchAll));
+    assert!(!controller.supports_purpose(AgentPurpose::TextToSpeech));
+    assert!(!controller.supports_purpose(AgentPurpose::ImageGeneration));
+}
+
+#[test]
+fn supports_purpose_true_when_image_and_tts_blocks_present() {
+    let config: Config = serde_yaml_ng::from_str(
+        r#"
+base_url: https://api.venice.ai/api/v1
+api_key: test-key
+text_to_speech:
+  model_id: tts-kokoro
+image_generation:
+  model_id: chroma
+"#,
+    )
+    .expect("config should deserialize");
+
+    let controller = Controller::new(config);
+
+    assert!(controller.supports_purpose(AgentPurpose::TextToSpeech));
+    assert!(controller.supports_purpose(AgentPurpose::ImageGeneration));
+}
+
+#[test]
+fn speech_request_serializes_voice_and_omits_unset() {
+    let request = SpeechRequest {
+        model: "tts-kokoro".to_owned(),
+        input: "hello".to_owned(),
+        voice: Some("af_sky".to_owned()),
+        speed: None,
+        response_format: Some("mp3".to_owned()),
+        prompt: None,
+        temperature: None,
+        top_p: None,
+    };
+
+    let json = serde_json::to_string(&request).expect("serialize SpeechRequest");
+
+    assert!(
+        json.contains("\"voice\":\"af_sky\""),
+        "voice should be present: {json}"
+    );
+    assert!(
+        !json.contains("temperature"),
+        "an unset knob must be omitted (not null): {json}"
+    );
+    assert!(!json.contains("null"), "no nulls belong in the body: {json}");
+}
+
+#[test]
+fn generate_image_request_pins_flags_and_omits_unset() {
+    let request = GenerateImageRequest {
+        model: "chroma".to_owned(),
+        prompt: "a cat".to_owned(),
+        return_binary: false,
+        variants: 1,
+        negative_prompt: None,
+        cfg_scale: None,
+        steps: None,
+        style_preset: None,
+        seed: None,
+        safe_mode: None,
+        hide_watermark: None,
+        format: None,
+        width: None,
+        height: None,
+        aspect_ratio: None,
+        resolution: None,
+        quality: None,
+        lora_strength: None,
+        embed_exif_metadata: None,
+        enable_web_search: None,
+    };
+
+    let json = serde_json::to_string(&request).expect("serialize GenerateImageRequest");
+
+    assert!(json.contains("\"model\":\"chroma\""), "{json}");
+    assert!(
+        json.contains("\"return_binary\":false"),
+        "return_binary must be pinned false: {json}"
+    );
+    assert!(
+        json.contains("\"variants\":1"),
+        "variants must be pinned 1: {json}"
+    );
+    assert!(
+        !json.contains("cfg_scale"),
+        "an unset knob must be omitted: {json}"
+    );
+    assert!(!json.contains("null"), "no nulls belong in the body: {json}");
+}
+
+#[test]
+fn edit_image_request_carries_model_and_base64_image() {
+    let request = EditImageRequest {
+        model: "firered-image-edit".to_owned(),
+        prompt: "make it a sunrise".to_owned(),
+        image: "aGVsbG8=".to_owned(),
+        output_format: None,
+        aspect_ratio: None,
+        resolution: None,
+        safe_mode: None,
+    };
+
+    let json = serde_json::to_string(&request).expect("serialize EditImageRequest");
+
+    assert!(
+        json.contains("\"model\":\"firered-image-edit\""),
+        "{json}"
+    );
+    assert!(
+        json.contains("\"image\":\"aGVsbG8=\""),
+        "the base64 image string must be present: {json}"
+    );
+    assert!(
+        !json.contains("output_format"),
+        "an unset knob must be omitted: {json}"
+    );
+}
+
+#[test]
+fn web_search_mode_off_deserializes_from_bare_yaml_off() {
+    // `off` is a YAML-1.1 boolean but a plain string under serde_yaml_ng's YAML-1.2 core schema,
+    // so it deserializes straight into the lowercase `WebSearchMode::Off`. This pins that the
+    // sample config and docs can use the bare, unquoted `off` without it parsing as a boolean.
+    let params: VeniceParameters =
+        serde_yaml_ng::from_str("enable_web_search: off").expect("bare `off` should deserialize");
+
+    assert!(matches!(params.enable_web_search, Some(WebSearchMode::Off)));
+}
--- a/src/agent/provider/venice/utils.rs
+++ b/src/agent/provider/venice/utils.rs
@@ -0,0 +1,55 @@
+use crate::conversation::llm::{
+    Author as LLMAuthor, Message as LLMMessage, MessageContent as LLMMessageContent,
+};
+use crate::utils::base64::base64_encode;
+
+use super::wire::{ChatMessage, ContentPart, ImageUrl, MessageContent};
+
+pub fn convert_llm_messages_to_venice(messages: Vec<LLMMessage>) -> Vec<ChatMessage> {
+    let mut venice_messages: Vec<ChatMessage> = Vec::with_capacity(messages.len());
+
+    for message in messages {
+        if let Some(venice_message) = convert_llm_message_to_venice(message) {
+            venice_messages.push(venice_message);
+        }
+    }
+
+    venice_messages
+}
+
+fn convert_llm_message_to_venice(message: LLMMessage) -> Option<ChatMessage> {
+    let role = match message.author {
+        LLMAuthor::Prompt => "system",
+        LLMAuthor::Assistant => "assistant",
+        LLMAuthor::User => "user",
+    };
+
+    match message.content {
+        LLMMessageContent::Text(text) => Some(ChatMessage {
+            role: role.to_owned(),
+            content: MessageContent::Text(text),
+        }),
+        LLMMessageContent::Image(image_details) => {
+            // Inline the image as a base64 data URI, the same shape the OpenAI vision content
+            // part uses. This is the gap the openai_compat provider can't fill (it drops images).
+            let data_uri = format!(
+                "data:{};base64,{}",
+                image_details.mime,
+                base64_encode(&image_details.data)
+            );
+
+            Some(ChatMessage {
+                role: role.to_owned(),
+                content: MessageContent::Parts(vec![ContentPart::ImageUrl {
+                    image_url: ImageUrl { url: data_uri },
+                }]),
+            })
+        }
+        LLMMessageContent::File(_file_details) => {
+            tracing::warn!(
+                "The Venice provider does not support file content. This file message will be skipped."
+            );
+            None
+        }
+    }
+}
--- a/src/agent/provider/venice/wire.rs
+++ b/src/agent/provider/venice/wire.rs
@@ -0,0 +1,212 @@
+//! Serde structs modeling Venice's `/chat/completions`, `/audio/transcriptions`,
+//! `/audio/speech`, `/image/generate`, and `/image/edit` wire shapes. Request types are
+//! `Serialize`-only (we build them, Venice never sends them back); response types are
+//! `Deserialize`-only. Keeping the split means the untagged request content enum is never on a
+//! deserialize path, so a surprise response shape can't fail to match it.
+//!
+//! Field names match Venice's schema 1:1 (so the config's `model_id` becomes `model` here). Every
+//! request body is `additionalProperties: false`, so optional knobs carry `skip_serializing_if`
+//! to omit rather than send `null`. `/audio/speech` and `/image/edit` return raw binary (no
+//! response struct); only `/image/generate` returns JSON (`GenerateImageResponse`).
+
+use serde::{Deserialize, Serialize};
+
+use super::config::VeniceParameters;
+
+#[derive(Debug, Serialize)]
+pub struct ChatCompletionRequest {
+    pub model: String,
+
+    pub messages: Vec<ChatMessage>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub temperature: Option<f32>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub max_completion_tokens: Option<u32>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub venice_parameters: Option<VeniceParameters>,
+}
+
+#[derive(Debug, Serialize)]
+pub struct ChatMessage {
+    pub role: String,
+
+    pub content: MessageContent,
+}
+
+/// A message body is either a bare string or a list of content parts. Venice accepts both; we
+/// send the parts form only when a message carries an image (baibot keeps text and images in
+/// separate messages, so a parts list only ever holds images in v1).
+#[derive(Debug, Serialize)]
+#[serde(untagged)]
+pub enum MessageContent {
+    Text(String),
+    Parts(Vec<ContentPart>),
+}
+
+#[derive(Debug, Serialize)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum ContentPart {
+    ImageUrl { image_url: ImageUrl },
+}
+
+#[derive(Debug, Serialize)]
+pub struct ImageUrl {
+    /// A `data:<mime>;base64,<data>` URI for inline images.
+    pub url: String,
+}
+
+/// Standard OpenAI-shaped chat completion response. We only read `choices[0].message.content`;
+/// when web search is on, Venice inlines citations as `^n^` superscripts in that content and we
+/// pass it through untouched.
+#[derive(Debug, Deserialize)]
+pub struct ChatCompletionResponse {
+    pub choices: Vec<ChatChoice>,
+}
+
+#[derive(Debug, Deserialize)]
+pub struct ChatChoice {
+    pub message: ResponseMessage,
+}
+
+#[derive(Debug, Deserialize)]
+pub struct ResponseMessage {
+    #[serde(default)]
+    pub content: Option<String>,
+}
+
+/// `/audio/transcriptions` response. We read `text`; the optional `duration`/`timestamps` the
+/// API can return are not used in v1.
+#[derive(Debug, Deserialize)]
+pub struct TranscriptionResponse {
+    pub text: String,
+}
+
+/// `/audio/speech` (`CreateSpeechRequestSchema`) request. `input` and `model` are always sent;
+/// the rest are omitted when unset. The response is raw binary audio, so there is no response
+/// struct.
+#[derive(Debug, Serialize)]
+pub struct SpeechRequest {
+    pub model: String,
+
+    pub input: String,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub voice: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub speed: Option<f32>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub response_format: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub prompt: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub temperature: Option<f32>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub top_p: Option<f32>,
+}
+
+/// `/image/generate` (`GenerateImageRequest`) request. `return_binary` is pinned `false` and
+/// `variants` to `1` by the builder: baibot wants exactly one image returned as base64-in-JSON,
+/// which `GenerateImageResponse` then decodes. Flipping `return_binary` would make Venice answer
+/// with raw binary and break that JSON decode, so it is not configurable.
+#[derive(Debug, Serialize)]
+pub struct GenerateImageRequest {
+    pub model: String,
+
+    pub prompt: String,
+
+    pub return_binary: bool,
+
+    pub variants: u32,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub negative_prompt: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cfg_scale: Option<f32>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub steps: Option<u32>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub style_preset: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub seed: Option<i64>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub safe_mode: Option<bool>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub hide_watermark: Option<bool>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub format: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub width: Option<u32>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub height: Option<u32>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub aspect_ratio: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub resolution: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub quality: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub lora_strength: Option<u32>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub embed_exif_metadata: Option<bool>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub enable_web_search: Option<bool>,
+}
+
+/// `/image/generate` response when `return_binary` is false: a JSON envelope carrying the images
+/// as base64 strings. We read `images[0]`; `request`/`timing` and other fields are ignored. `id`
+/// is telemetry only (logged, never used for correctness), so it is optional: a response that
+/// carries usable `images` must not fail to deserialize just because the telemetry field drifted.
+#[derive(Debug, Deserialize)]
+pub struct GenerateImageResponse {
+    #[serde(default)]
+    pub id: Option<String>,
+    pub images: Vec<String>,
+}
+
+/// `/image/edit` (`EditImageRequest`) request. The source `image` is a base64-encoded string
+/// (Venice's `image` field is `anyOf` upload/base64/URL; we send base64-in-JSON, no multipart).
+/// The response is raw binary, so there is no response struct.
+#[derive(Debug, Serialize)]
+pub struct EditImageRequest {
+    pub model: String,
+
+    pub prompt: String,
+
+    /// Base64-encoded source image bytes.
+    pub image: String,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub output_format: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub aspect_ratio: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub resolution: Option<String>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub safe_mode: Option<bool>,
+}