SMOODEV-1472: Structured output (JSON-schema constrained responses) in the LLM client

.changeset/structured-output.md

@@ -0,0 +1,19 @@
+---
+'@smooai/smooth-operator-core-monorepo': minor
+---
+
+Add structured output (schema-constrained JSON responses) to the LLM client (SMOODEV-1472).
+
+New public API:
+
+- `ResponseFormat` enum (`JsonSchema { name, schema, strict }`) with a `ResponseFormat::json_schema(name, schema)` constructor (defaults `strict = true`).
+- `LlmClient::chat_structured(messages, &ResponseFormat)` and the lower-level `LlmClient::chat_with_format(messages, tools, Option<&ResponseFormat>)`. `chat` now delegates to `chat_with_format(.., None)`.
+- `LlmResponse::structured_json() -> serde_json::Value` and `LlmResponse::deserialize_json::<T>()` — both surface a clear error (never a silent empty value) when the model returned empty or non-JSON content.
+- `LlmProvider::chat_structured` trait method; `MockLlmClient` records the requested `ResponseFormat` on `RecordedCall.response_format` for assertions.
+
+Provider handling:
+
+- **OpenAI-compatible** (LiteLLM gateway, etc.): serialized on `/chat/completions` as `response_format: { type: "json_schema", json_schema: { name, schema, strict } }`.
+- **Anthropic-native** (`/v1/messages`): achieved via a forced single tool call — a synthetic tool whose `input_schema` is the requested schema, forced with `tool_choice: { type: "tool", name }`; the tool's `input` is surfaced back as the JSON content string.
+
+Streaming structured output and agent-level (`Agent::run`) wiring are deliberately deferred as follow-ups.

rust/smooth-operator-core/src/lib.rs

@@ -32,7 +32,7 @@ pub use conversation::{CompactionResult, CompactionStrategy, Conversation, Messa
 pub use cost::{BudgetExceeded, CostBudget, CostEntry, CostTracker, ModelPricing};
 pub use human::{human_channel, ConfirmationHook, HumanChannelPair, HumanRequest, HumanResponse};
 pub use knowledge::{Document, DocumentType, InMemoryKnowledge, KnowledgeBase, KnowledgeResult};
-pub use llm::{accumulate_stream_events, LlmClient, LlmConfig, LlmResponse, StreamEvent};
+pub use llm::{accumulate_stream_events, LlmClient, LlmConfig, LlmResponse, ResponseFormat, StreamEvent};
 pub use memory::{InMemoryMemory, Memory, MemoryEntry, MemoryType};
 pub use providers::{Activity, ModelRouting, ModelSlot, ProviderConfig, ProviderRegistry};
 pub use tool::{Tool, ToolCall, ToolRegistry, ToolResult, ToolSchema};