zeph_llm/

provider.rs

// SPDX-FileCopyrightText: 2026 Andrei G <bug-ops>
// SPDX-License-Identifier: MIT OR Apache-2.0

use std::future::Future;
use std::pin::Pin;
use std::{
    any::TypeId,
    collections::HashMap,
    sync::{LazyLock, Mutex},
};

use futures_core::Stream;
use serde::{Deserialize, Serialize};

use zeph_common::ToolName;

pub use zeph_common::ToolDefinition;

use crate::embed::owned_strs;
use crate::error::LlmError;

static SCHEMA_CACHE: LazyLock<Mutex<HashMap<TypeId, (serde_json::Value, String)>>> =
    LazyLock::new(|| Mutex::new(HashMap::new()));

/// Return the JSON schema value and pretty-printed string for type `T`, cached by `TypeId`.
///
/// # Errors
///
/// Returns an error if schema serialization fails.
pub(crate) fn cached_schema<T: schemars::JsonSchema + 'static>()
-> Result<(serde_json::Value, String), crate::LlmError> {
    let type_id = TypeId::of::<T>();
    if let Ok(cache) = SCHEMA_CACHE.lock()
        && let Some(entry) = cache.get(&type_id)
    {
        return Ok(entry.clone());
    }
    let schema = schemars::schema_for!(T);
    let value = serde_json::to_value(&schema)
        .map_err(|e| crate::LlmError::StructuredParse(e.to_string()))?;
    let pretty = serde_json::to_string_pretty(&schema)
        .map_err(|e| crate::LlmError::StructuredParse(e.to_string()))?;
    if let Ok(mut cache) = SCHEMA_CACHE.lock() {
        cache.insert(type_id, (value.clone(), pretty.clone()));
    }
    Ok((value, pretty))
}

/// Extract the short (unqualified) type name for schema prompts and tool names.
///
/// Returns the last `::` segment of [`std::any::type_name::<T>()`], which is always
/// non-empty. The `"Output"` fallback is unreachable in practice (`type_name` never returns
/// an empty string and `rsplit` on a non-empty string always yields at least one element),
/// but is kept for defensive clarity.
///
/// # Examples
///
/// ```
/// struct MyOutput;
/// // short_type_name::<MyOutput>() returns "MyOutput"
/// ```
pub(crate) fn short_type_name<T: ?Sized>() -> &'static str {
    std::any::type_name::<T>()
        .rsplit("::")
        .next()
        .unwrap_or("Output")
}

/// Per-call extras returned alongside the chat response by [`LlmProvider::chat_with_extras`].
///
/// Always paired 1:1 with a single response — no shared state, no races possible.
/// All optional fields default to `None` so providers that do not expose the
/// underlying API (e.g. Claude, Gemini) can simply return the default.
///
/// Marked `#[non_exhaustive]` so future fields (e.g. `cached_tokens`) can be added
/// without breaking match sites.
#[non_exhaustive]
#[derive(Debug, Clone, Default)]
pub struct ChatExtras {
    /// Mean negative log-probability of the generated tokens, when the provider
    /// was configured to request `logprobs` and the API supplied them.
    ///
    /// Lower = more confident. Typical range: `[0.0, ~6.0]` for natural-language tokens.
    pub entropy: Option<f64>,
}

impl ChatExtras {
    /// Return a `ChatExtras` with the given entropy value.
    ///
    /// Used by `MockProvider` (test-only, enabled via `testing` feature) and OpenAI/Ollama providers.
    ///
    /// # Examples
    ///
    /// ```
    /// use zeph_llm::provider::ChatExtras;
    ///
    /// let extras = ChatExtras::with_entropy(0.9);
    /// assert_eq!(extras.entropy, Some(0.9));
    /// ```
    #[must_use]
    pub fn with_entropy(entropy: f64) -> Self {
        Self {
            entropy: Some(entropy),
        }
    }
}

/// A chunk from an LLM streaming response.
///
/// Consumers should match all variants: future providers may emit non-`Content` chunks
/// that callers must not silently drop (e.g. thinking blocks that must be echoed back).
#[non_exhaustive]
#[derive(Debug, Clone)]
pub enum StreamChunk {
    /// Regular response text.
    Content(String),
    /// Internal reasoning/thinking token (e.g. Claude extended thinking, `OpenAI` reasoning).
    Thinking(String),
    /// Server-side compaction summary (Claude compact-2026-01-12 beta).
    /// Delivered when the Claude API automatically summarizes conversation history.
    Compaction(String),
    /// One or more tool calls from the model received during streaming.
    ToolUse(Vec<ToolUseRequest>),
}

/// Boxed stream of typed chunks from an LLM provider.
///
/// Obtain via [`LlmProvider::chat_stream`]. Drive the stream with
/// `futures::StreamExt::next` or `tokio_stream::StreamExt::next`.
pub type ChatStream = Pin<Box<dyn Stream<Item = Result<StreamChunk, LlmError>> + Send>>;

/// Structured tool invocation request from the model.
///
/// Returned by [`LlmProvider::chat_with_tools`] when the model decides to call one or
/// more tools. The caller is responsible for executing the tool and returning results
/// via a [`MessagePart::ToolResult`] in the next turn.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ToolUseRequest {
    /// Opaque call identifier assigned by the model; must be echoed in `ToolResult.tool_use_id`.
    pub id: String,
    /// Name of the tool to invoke, matching a [`ToolDefinition::name`].
    pub name: ToolName,
    /// JSON arguments the model wants to pass to the tool.
    pub input: serde_json::Value,
}

/// Thinking block returned by Claude when extended or adaptive thinking is enabled.
///
/// Both variants must be echoed verbatim in the next turn's `assistant` message so
/// the API can correctly attribute reasoning across turns. Never modify or discard
/// these blocks between turns.
#[non_exhaustive]
#[derive(Debug, Clone)]
pub enum ThinkingBlock {
    /// Visible reasoning token with its cryptographic signature.
    Thinking { thinking: String, signature: String },
    /// Redacted reasoning block (API-side privacy redaction). Preserved as opaque data.
    Redacted { data: String },
}

/// Marker injected into `ChatResponse::Text` when the LLM response was cut off by the
/// token limit. Consumers can detect this substring to signal `MaxTokens` stop reason.
pub const MAX_TOKENS_TRUNCATION_MARKER: &str = "max_tokens limit reached";

/// Response from [`LlmProvider::chat_with_tools`].
///
/// When the model returns `ToolUse`, the caller must:
/// 1. Execute each tool in `tool_calls`.
/// 2. Append an `assistant` message with the original `tool_calls` and any `thinking_blocks`.
/// 3. Append a `user` message containing [`MessagePart::ToolResult`] entries.
/// 4. Call `chat_with_tools` again to continue the conversation.
#[non_exhaustive]
#[derive(Debug, Clone)]
pub enum ChatResponse {
    /// Model produced text output only.
    Text(String),
    /// Model requests one or more tool invocations.
    ToolUse {
        /// Any text the model emitted before/alongside tool calls.
        text: Option<String>,
        tool_calls: Vec<ToolUseRequest>,
        /// Thinking blocks from the model (empty when thinking is disabled).
        /// Must be preserved verbatim in multi-turn requests.
        thinking_blocks: Vec<ThinkingBlock>,
    },
}

/// Boxed future returning an embedding vector, returned by [`EmbedFn`].
pub type EmbedFuture = Pin<Box<dyn Future<Output = Result<Vec<f32>, LlmError>> + Send>>;

/// A Send + Sync closure that embeds a text slice into a vector.
///
/// Obtain a provider-backed `EmbedFn` via [`crate::any::AnyProvider::embed_fn`].
/// The closure captures an `Arc`-wrapped provider clone, so it is cheap to clone.
pub type EmbedFn = Box<dyn Fn(&str) -> EmbedFuture + Send + Sync>;

/// Sender for emitting human-readable status events (retries, fallbacks) to the UI layer.
///
/// When set on a provider, the provider sends short strings such as
/// `"Retrying after rate limit…"` or `"Falling back to secondary provider"`.
/// The TUI consumes these to show real-time activity spinners.
pub type StatusTx = tokio::sync::mpsc::UnboundedSender<String>;

/// Best-effort fallback for debug dump request payloads when a provider does not expose
/// its concrete API request body.
#[must_use]
pub fn default_debug_request_json(
    messages: &[Message],
    tools: &[ToolDefinition],
) -> serde_json::Value {
    serde_json::json!({
        "model": serde_json::Value::Null,
        "max_tokens": serde_json::Value::Null,
        "messages": serde_json::to_value(messages).unwrap_or(serde_json::Value::Array(vec![])),
        "tools": serde_json::to_value(tools).unwrap_or(serde_json::Value::Array(vec![])),
        "temperature": serde_json::Value::Null,
        "cache_control": serde_json::Value::Null,
    })
}

/// Partial LLM generation parameter overrides for experiment variation injection.
///
/// Applied by the experiment engine to clone-and-patch a provider before evaluation,
/// so each variation is scored with its specific generation parameters.
///
/// Only `Some` fields are applied; `None` fields leave the provider's configured
/// defaults unchanged. Not all providers support all fields — unsupported fields
/// are silently ignored by each backend.
#[derive(Debug, Clone, Default)]
pub struct GenerationOverrides {
    /// Sampling temperature in `[0.0, 2.0]`. Lower = more deterministic.
    pub temperature: Option<f64>,
    /// Nucleus sampling probability in `[0.0, 1.0]`.
    pub top_p: Option<f64>,
    /// Top-K sampling cutoff (number of top tokens to consider).
    pub top_k: Option<usize>,
    /// Penalty for tokens that have already appeared (OpenAI-compatible providers).
    pub frequency_penalty: Option<f64>,
    /// Penalty for topics the model has already covered (OpenAI-compatible providers).
    pub presence_penalty: Option<f64>,
}

/// Message role in a conversation.
///
/// Determines how each message is presented to the model:
/// - `System` — global instructions prepended before the conversation
/// - `User` — human turn input
/// - `Assistant` — previous model output (used for multi-turn context)
#[non_exhaustive]
#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "lowercase")]
pub enum Role {
    System,
    User,
    Assistant,
}

/// A typed content part within a [`Message`].
///
/// Messages may contain zero or more parts that represent heterogeneous content:
/// plain text, tool invocations, memory recall fragments, images, and internal
/// protocol blocks (thinking, compaction). Most providers flatten these into a single
/// string before sending; Claude encodes them as structured content blocks.
///
/// # Ordering invariants
///
/// - `ToolUse` parts must precede their corresponding `ToolResult` parts.
/// - `ThinkingBlock` / `RedactedThinkingBlock` parts must be preserved verbatim in
///   multi-turn requests so the API can correctly attribute reasoning.
/// - `Compaction` parts must be preserved verbatim; the API uses them to prune
///   prior history on subsequent turns (Claude compact-2026-01-12 beta).
#[non_exhaustive]
#[derive(Clone, Debug, Serialize, Deserialize)]
#[serde(tag = "kind", rename_all = "snake_case")]
pub enum MessagePart {
    /// Plain assistant or user text.
    Text { text: String },
    /// Output from a tool execution, optionally compacted.
    ToolOutput {
        tool_name: zeph_common::ToolName,
        body: String,
        #[serde(default, skip_serializing_if = "Option::is_none")]
        compacted_at: Option<i64>,
    },
    /// Memory recall fragment injected by the agent's semantic memory layer.
    Recall { text: String },
    /// Repository or file code context injected by the code indexing layer.
    CodeContext { text: String },
    /// Compaction summary replacing pruned conversation history.
    Summary { text: String },
    /// Cross-session memory fragment carried over from a previous conversation.
    CrossSession { text: String },
    /// Model-initiated tool invocation. Pairs with a subsequent [`MessagePart::ToolResult`].
    ToolUse {
        id: String,
        name: String,
        input: serde_json::Value,
    },
    /// Tool execution result returned to the model after a [`MessagePart::ToolUse`].
    ToolResult {
        tool_use_id: String,
        content: String,
        #[serde(default)]
        is_error: bool,
    },
    /// Inline image payload (vision input).
    Image(Box<ImageData>),
    /// Claude thinking block — must be preserved verbatim in multi-turn requests.
    ThinkingBlock { thinking: String, signature: String },
    /// Claude redacted thinking block — preserved as-is in multi-turn requests.
    RedactedThinkingBlock { data: String },
    /// Claude server-side compaction block — must be preserved verbatim in multi-turn requests
    /// so the API can correctly prune prior history on the next turn.
    Compaction { summary: String },
}

impl MessagePart {
    /// Return the plain text content if this part is a text-like variant (`Text`, `Recall`,
    /// `CodeContext`, `Summary`, `CrossSession`), `None` otherwise.
    #[must_use]
    pub fn as_plain_text(&self) -> Option<&str> {
        match self {
            Self::Text { text }
            | Self::Recall { text }
            | Self::CodeContext { text }
            | Self::Summary { text }
            | Self::CrossSession { text } => Some(text.as_str()),
            _ => None,
        }
    }

    /// Return the image data if this part is an `Image` variant, `None` otherwise.
    #[must_use]
    pub fn as_image(&self) -> Option<&ImageData> {
        if let Self::Image(img) = self {
            Some(img)
        } else {
            None
        }
    }

    /// Return a cloned copy of `parts` with every `Image` entry removed.
    ///
    /// `Image` parts are ephemeral, current-turn-only vision input (spec-072 §4, C1) and must
    /// never reach a persistence sink — `SQLite` `parts_json`, the Qdrant embed path, the durable
    /// JSONL session log, or a sub-agent transcript file. Callers apply this once, immediately
    /// above the persistence write, leaving the in-memory slice used for the current turn's
    /// provider request untouched.
    ///
    /// # Examples
    ///
    /// ```
    /// use zeph_llm::provider::{ImageData, MessagePart};
    ///
    /// let parts = vec![
    ///     MessagePart::Text {
    ///         text: "hello".to_owned(),
    ///     },
    ///     MessagePart::Image(Box::new(ImageData {
    ///         data: vec![0xFF, 0xD8, 0xFF, 0xE0],
    ///         mime_type: "image/jpeg".to_owned(),
    ///     })),
    /// ];
    /// let stripped = MessagePart::strip_images(&parts);
    /// assert_eq!(stripped.len(), 1);
    /// assert!(!stripped.iter().any(|p| matches!(p, MessagePart::Image(_))));
    /// ```
    #[must_use]
    pub fn strip_images(parts: &[MessagePart]) -> Vec<MessagePart> {
        parts
            .iter()
            .filter(|p| !matches!(p, MessagePart::Image(_)))
            .cloned()
            .collect()
    }
}

#[derive(Clone, Serialize, Deserialize)]
/// Raw image payload for vision-capable providers.
///
/// The `data` field is serialized as a Base64 string. `mime_type` must be a valid
/// image MIME type supported by the target provider (e.g. `"image/png"`, `"image/jpeg"`).
pub struct ImageData {
    #[serde(with = "serde_bytes_base64")]
    pub data: Vec<u8>,
    pub mime_type: String,
}

impl std::fmt::Debug for ImageData {
    /// Redacts the raw image bytes — only the MIME type and byte count are printed.
    ///
    /// `data` can carry arbitrary externally-sourced bytes (MCP tool results, user uploads);
    /// a derived `Debug` would dump the full payload into logs/panics.
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(f, "[image: {}, {} bytes]", self.mime_type, self.data.len())
    }
}

mod serde_bytes_base64 {
    use base64::{Engine, engine::general_purpose::STANDARD};
    use serde::{Deserialize, Deserializer, Serializer};

    pub fn serialize<S>(bytes: &[u8], s: S) -> Result<S::Ok, S::Error>
    where
        S: Serializer,
    {
        s.serialize_str(&STANDARD.encode(bytes))
    }

    pub fn deserialize<'de, D>(d: D) -> Result<Vec<u8>, D::Error>
    where
        D: Deserializer<'de>,
    {
        let s = String::deserialize(d)?;
        STANDARD.decode(&s).map_err(serde::de::Error::custom)
    }
}

/// Visibility of a message to agent and user.
///
/// Replaces the former `(agent_visible: bool, user_visible: bool)` pair, which
/// allowed the semantically invalid `(false, false)` combination. Every variant
/// guarantees at least one consumer can see the message.
///
/// # Examples
///
/// ```
/// use zeph_llm::provider::MessageVisibility;
///
/// let v = MessageVisibility::AgentOnly;
/// assert!(v.is_agent_visible());
/// assert!(!v.is_user_visible());
/// ```
#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
#[non_exhaustive]
pub enum MessageVisibility {
    /// Visible to both the agent (LLM context) and the user (conversation log).
    Both,
    /// Visible to the agent only (e.g. compaction summaries, internal context).
    AgentOnly,
    /// Visible to the user only (e.g. compacted originals shown in history).
    UserOnly,
}

impl MessageVisibility {
    /// Returns `true` if this message should be included in the LLM request context.
    #[must_use]
    pub fn is_agent_visible(self) -> bool {
        matches!(self, MessageVisibility::Both | MessageVisibility::AgentOnly)
    }

    /// Returns `true` if this message should appear in the user-facing conversation log.
    #[must_use]
    pub fn is_user_visible(self) -> bool {
        matches!(self, MessageVisibility::Both | MessageVisibility::UserOnly)
    }
}

impl Default for MessageVisibility {
    /// Defaults to [`Both`](MessageVisibility::Both) — visible to agent and user.
    fn default() -> Self {
        MessageVisibility::Both
    }
}

impl MessageVisibility {
    /// Serialize to the SQLite/PostgreSQL text value stored in the `visibility` column.
    #[must_use]
    pub fn as_db_str(self) -> &'static str {
        match self {
            MessageVisibility::Both => "both",
            MessageVisibility::AgentOnly => "agent_only",
            MessageVisibility::UserOnly => "user_only",
        }
    }

    /// Deserialize from the SQLite/PostgreSQL text value stored in the `visibility` column.
    ///
    /// Unknown values (e.g. from a future migration) default to `Both` for safety.
    #[must_use]
    pub fn from_db_str(s: &str) -> Self {
        match s {
            "agent_only" => MessageVisibility::AgentOnly,
            "user_only" => MessageVisibility::UserOnly,
            _ => MessageVisibility::Both,
        }
    }
}

/// Per-message visibility and metadata controlling agent context and user display.
///
/// Constructors [`agent_only`](Self::agent_only), [`user_only`](Self::user_only),
/// and [`focus_pinned`](Self::focus_pinned) cover the most common combinations.
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct MessageMetadata {
    /// Who can see this message.
    pub visibility: MessageVisibility,
    /// Unix timestamp (seconds) when this message was compacted, if applicable.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub compacted_at: Option<i64>,
    /// Pre-computed tool pair summary, applied lazily when context pressure rises.
    /// Stored on the tool response message; cleared after application.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub deferred_summary: Option<String>,
    /// When true, this message is excluded from all compaction passes (soft pruning,
    /// hard summarization, sidequest eviction). Used for the Focus Knowledge block (#1850).
    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
    pub focus_pinned: bool,
    /// Unique marker UUID set when `start_focus` begins a session. Used by `complete_focus`
    /// to locate the checkpoint without relying on a fragile raw index.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub focus_marker_id: Option<uuid::Uuid>,
    /// `SQLite` row ID for this message. Populated when loading from DB or after persisting.
    /// Never serialized — always re-populated from the database on load.
    #[serde(skip)]
    pub db_id: Option<i64>,
    /// Fidelity level assigned by `FidelityScorer` during context assembly.
    ///
    /// `None` when fidelity scoring is disabled or the message has not yet been scored.
    /// Used for debug tracing and compaction input filtering (INV-02).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub fidelity_tag: Option<zeph_common::ContextFidelity>,
    /// Cached embedding vector for semantic fidelity scoring.
    ///
    /// In-memory only — not serialized or persisted to the database.
    #[serde(skip)]
    pub embedding: Option<Vec<f32>>,
    /// Write-time memory-consent gate content-trust tier (issue #6490 `MemGhost`, #6558 fix).
    ///
    /// Raw `u8` discriminant of `zeph_sanitizer::ContentTrustLevel` (kept as a plain integer
    /// here, not the enum itself, since `zeph-llm` does not depend on `zeph-sanitizer` — mirrors
    /// `zeph_core::memory_tools::MemoryConsentTrustSlot`'s same representation choice). Set on
    /// tool-result batch messages by `Agent::process_tool_result_batch` to the batch's
    /// worst-case trust tier. Scanned by `Agent::context_max_trust_level` so untrusted content
    /// is still recognized by the memory-consent gate for as long as it remains in the live
    /// conversation context — including across a user-turn boundary — rather than only for the
    /// single turn in which it was fetched.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub trust_level: Option<u8>,
}

impl Default for MessageMetadata {
    fn default() -> Self {
        Self {
            visibility: MessageVisibility::Both,
            compacted_at: None,
            deferred_summary: None,
            focus_pinned: false,
            focus_marker_id: None,
            db_id: None,
            fidelity_tag: None,
            embedding: None,
            trust_level: None,
        }
    }
}

impl MessageMetadata {
    /// Message visible only to the agent (e.g. compaction summary).
    #[must_use]
    pub fn agent_only() -> Self {
        Self {
            visibility: MessageVisibility::AgentOnly,
            compacted_at: None,
            deferred_summary: None,
            focus_pinned: false,
            focus_marker_id: None,
            db_id: None,
            fidelity_tag: None,
            embedding: None,
            trust_level: None,
        }
    }

    /// Message visible only to the user (e.g. compacted original).
    #[must_use]
    pub fn user_only() -> Self {
        Self {
            visibility: MessageVisibility::UserOnly,
            compacted_at: None,
            deferred_summary: None,
            focus_pinned: false,
            focus_marker_id: None,
            db_id: None,
            fidelity_tag: None,
            embedding: None,
            trust_level: None,
        }
    }

    /// Pinned Knowledge block — excluded from all compaction passes.
    #[must_use]
    pub fn focus_pinned() -> Self {
        Self {
            visibility: MessageVisibility::AgentOnly,
            compacted_at: None,
            deferred_summary: None,
            focus_pinned: true,
            focus_marker_id: None,
            db_id: None,
            fidelity_tag: None,
            embedding: None,
            trust_level: None,
        }
    }
}

/// A single message in a conversation.
///
/// Each message has a [`Role`], a flat `content` string (used when sending to providers
/// that do not support structured parts), and an optional list of [`MessagePart`]s for
/// providers that accept heterogeneous content blocks (e.g. Claude).
///
/// The `content` field is kept in sync with `parts` via [`Message::rebuild_content`].
/// When building messages from structured parts, always use [`Message::from_parts`] —
/// it populates both `parts` and `content`.
///
/// # Examples
///
/// ```
/// use zeph_llm::provider::{Message, MessagePart, Role};
///
/// // Simple text-only message
/// let msg = Message::from_legacy(Role::User, "What is Rust?");
/// assert_eq!(msg.to_llm_content(), "What is Rust?");
///
/// // Structured message with parts
/// let parts = vec![
///     MessagePart::Text { text: "Explain this code.".into() },
/// ];
/// let msg = Message::from_parts(Role::User, parts);
/// assert!(!msg.parts.is_empty());
/// ```
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Message {
    pub role: Role,
    /// Flat text representation of this message, derived from `parts` when structured.
    pub content: String,
    #[serde(default)]
    pub parts: Vec<MessagePart>,
    #[serde(default)]
    pub metadata: MessageMetadata,
}

impl Default for Message {
    fn default() -> Self {
        Self {
            role: Role::User,
            content: String::new(),
            parts: vec![],
            metadata: MessageMetadata::default(),
        }
    }
}

impl Message {
    /// Create a simple text-only message without structured parts.
    ///
    /// Use this constructor for system prompts, plain user turns, and assistant
    /// messages produced by providers that return a raw string.
    #[must_use]
    pub fn from_legacy(role: Role, content: impl Into<String>) -> Self {
        Self {
            role,
            content: content.into(),
            parts: vec![],
            metadata: MessageMetadata::default(),
        }
    }

    /// Create a message from structured parts, deriving the flat `content` automatically.
    ///
    /// Prefer this constructor when the message contains tool invocations, images,
    /// or other non-text content that providers need to render as separate content blocks.
    #[must_use]
    pub fn from_parts(role: Role, parts: Vec<MessagePart>) -> Self {
        let content = Self::flatten_parts(&parts);
        Self {
            role,
            content,
            parts,
            metadata: MessageMetadata::default(),
        }
    }

    /// Return the flat text content of this message, suitable for providers that do
    /// not support structured content blocks.
    #[must_use]
    pub fn to_llm_content(&self) -> &str {
        &self.content
    }

    /// Re-synchronize `content` from `parts` after in-place mutation.
    pub fn rebuild_content(&mut self) {
        if !self.parts.is_empty() {
            self.content = Self::flatten_parts(&self.parts);
        }
    }

    fn flatten_parts(parts: &[MessagePart]) -> String {
        use std::fmt::Write;
        let mut out = String::new();
        for part in parts {
            match part {
                MessagePart::Text { text }
                | MessagePart::Recall { text }
                | MessagePart::CodeContext { text }
                | MessagePart::Summary { text }
                | MessagePart::CrossSession { text } => out.push_str(text),
                MessagePart::ToolOutput {
                    tool_name,
                    body,
                    compacted_at,
                } => {
                    if compacted_at.is_some() {
                        if body.is_empty() {
                            let _ = write!(out, "[tool output: {tool_name}] (pruned)");
                        } else {
                            let _ = write!(out, "[tool output: {tool_name}] {body}");
                        }
                    } else {
                        let _ = write!(out, "[tool output: {tool_name}]\n```\n{body}\n```");
                    }
                }
                MessagePart::ToolUse { id, name, .. } => {
                    let _ = write!(out, "[tool_use: {name}({id})]");
                }
                MessagePart::ToolResult {
                    tool_use_id,
                    content,
                    ..
                } => {
                    let _ = write!(out, "[tool_result: {tool_use_id}]\n{content}");
                }
                MessagePart::Image(img) => {
                    let _ = write!(out, "[image: {}, {} bytes]", img.mime_type, img.data.len());
                }
                // Thinking and compaction blocks are internal API metadata — not rendered in text.
                MessagePart::ThinkingBlock { .. }
                | MessagePart::RedactedThinkingBlock { .. }
                | MessagePart::Compaction { .. } => {}
            }
        }
        out
    }
}

/// Core abstraction for all LLM inference backends.
///
/// Every backend — `Ollama`, `Claude`, `OpenAI`, `Gemini`, `Candle` — implements this trait.
/// The [`crate::any::AnyProvider`] enum erases the concrete type so callers can
/// hold any backend behind a single type, and [`crate::router::RouterProvider`]
/// implements this trait to multiplex across multiple backends.
///
/// # Object safety
///
/// This trait is **not** object-safe: 6 methods return `impl Future + Send` (RPIT),
/// and [`chat_typed`](Self::chat_typed) carries a generic type parameter `T`.
/// The `where Self: Sized` bound on `chat_typed` is a mitigation — it excludes that
/// method from the vtable — but the RPIT methods remain and prevent `dyn LlmProvider`.
/// Attempting `Box<dyn LlmProvider>` will produce a compile error.
///
/// For dynamic dispatch, use [`Arc<dyn LlmProviderDyn>`](crate::provider_dyn::LlmProviderDyn)
/// instead — a blanket impl wires every `LlmProvider` implementor automatically.
/// See the [`provider_dyn`](crate::provider_dyn) module for details.
///
/// # Required methods
///
/// Implementors must provide: [`chat`](Self::chat), [`chat_stream`](Self::chat_stream),
/// [`supports_streaming`](Self::supports_streaming), [`embed`](Self::embed),
/// [`supports_embeddings`](Self::supports_embeddings), and [`name`](Self::name).
///
/// # Optional methods
///
/// All other methods have default implementations that are safe to accept:
/// - [`context_window`](Self::context_window) — returns `None`
/// - [`embed_batch`](Self::embed_batch) — sequential fallback via [`embed`](Self::embed)
/// - [`chat_with_tools`](Self::chat_with_tools) — falls back to [`chat`](Self::chat)
/// - [`chat_typed`](Self::chat_typed) — schema-prompt injection + retry
///   (requires `Self: Sized`; use [`chat_typed_dyn`](crate::provider_dyn::chat_typed_dyn)
///   for trait objects)
/// - [`supports_vision`](Self::supports_vision) — returns `false`
/// - [`supports_tool_use`](Self::supports_tool_use) — returns `false`, matching the
///   [`chat_with_tools`](Self::chat_with_tools) default fallback which silently drops
///   tool definitions; providers must override both together to opt into tool use
///
/// # Examples
///
/// ```rust,no_run
/// use zeph_llm::provider::{LlmProvider, Message, Role, ChatStream};
/// use zeph_llm::LlmError;
///
/// struct EchoProvider;
///
/// impl LlmProvider for EchoProvider {
///     async fn chat(&self, messages: &[Message]) -> Result<String, LlmError> {
///         Ok(messages.last().map(|m| m.content.clone()).unwrap_or_default())
///     }
///
///     async fn chat_stream(&self, messages: &[Message]) -> Result<ChatStream, LlmError> {
///         use zeph_llm::provider::StreamChunk;
///         let text = self.chat(messages).await?;
///         Ok(Box::pin(tokio_stream::once(Ok(StreamChunk::Content(text)))))
///     }
///
///     fn supports_streaming(&self) -> bool { true }
///
///     async fn embed(&self, _text: &str) -> Result<Vec<f32>, LlmError> {
///         Err(LlmError::EmbedUnsupported { provider: "echo".into() })
///     }
///
///     fn supports_embeddings(&self) -> bool { false }
///
///     fn name(&self) -> &str { "echo" }
/// }
/// ```
pub trait LlmProvider: Send + Sync {
    /// Report the model's context window size in tokens.
    ///
    /// Returns `None` if unknown. Used for auto-budget calculation.
    fn context_window(&self) -> Option<usize> {
        None
    }

    /// Send messages to the LLM and return the assistant response.
    ///
    /// # Errors
    ///
    /// Returns an error if the provider fails to communicate or the response is invalid.
    fn chat(&self, messages: &[Message]) -> impl Future<Output = Result<String, LlmError>> + Send;

    /// Send messages and return a stream of response chunks.
    ///
    /// # Errors
    ///
    /// Returns an error if the provider fails to communicate or the response is invalid.
    fn chat_stream(
        &self,
        messages: &[Message],
    ) -> impl Future<Output = Result<ChatStream, LlmError>> + Send;

    /// Whether this provider supports native streaming.
    fn supports_streaming(&self) -> bool;

    /// Generate an embedding vector from text.
    ///
    /// # Errors
    ///
    /// Returns an error if the provider does not support embeddings or the request fails.
    fn embed(&self, text: &str) -> impl Future<Output = Result<Vec<f32>, LlmError>> + Send;

    /// Embed multiple texts in a single API call.
    ///
    /// Default implementation calls [`embed`][Self::embed] sequentially for each input.
    /// Providers with native batch APIs should override this.
    ///
    /// # Errors
    ///
    /// Returns an error if any embedding fails. On native batch backends the entire batch
    /// fails atomically; on the sequential fallback the first error aborts.
    fn embed_batch(
        &self,
        texts: &[&str],
    ) -> impl Future<Output = Result<Vec<Vec<f32>>, LlmError>> + Send {
        let owned = owned_strs(texts);
        async move {
            let mut results = Vec::with_capacity(owned.len());
            for text in &owned {
                results.push(self.embed(text).await?);
            }
            Ok(results)
        }
    }

    /// Whether this provider supports embedding generation.
    fn supports_embeddings(&self) -> bool;

    /// Provider name for logging and identification.
    fn name(&self) -> &str;

    /// Model identifier string (e.g. `gpt-4o-mini`, `claude-sonnet-5`).
    /// Used by cost-estimation heuristics. Returns `""` when not applicable.
    #[allow(clippy::unnecessary_literal_bound)]
    fn model_identifier(&self) -> &str {
        ""
    }

    /// Model identifier that actually served the most recent dispatch.
    ///
    /// For a concrete single-model provider (Claude, `OpenAI`, Candle, ...) this is
    /// identical to [`model_identifier`](Self::model_identifier) — the default
    /// implementation simply forwards to it. Routing providers (`Router`,
    /// `TriageRouter`) override it to resolve the sub-provider that served the last
    /// call, since their own `model_identifier()` returns a stable routing-policy
    /// label (e.g. `"router"`) rather than a real model id. Callers that need to
    /// reason about the model that produced a specific response (for example,
    /// `is_reasoning_model` detection) should call this method instead of
    /// `model_identifier()`.
    fn effective_model_identifier(&self) -> &str {
        self.model_identifier()
    }

    /// Whether this provider supports image input (vision).
    fn supports_vision(&self) -> bool {
        false
    }

    /// Whether this provider supports native `tool_use` / function calling.
    ///
    /// Defaults to `false` because [`chat_with_tools`](Self::chat_with_tools) defaults
    /// to falling back on [`chat`](Self::chat), which silently discards tool
    /// definitions. Providers implementing real tool calling must override both
    /// this method and `chat_with_tools` together.
    fn supports_tool_use(&self) -> bool {
        false
    }

    /// Send messages with tool definitions, returning a structured response.
    ///
    /// Default: falls back to `chat()` and wraps the result in `ChatResponse::Text`.
    ///
    /// # Errors
    ///
    /// Returns an error if the provider fails to communicate or the response is invalid.
    fn chat_with_tools(
        &self,
        messages: &[Message],
        _tools: &[ToolDefinition],
    ) -> impl std::future::Future<Output = Result<ChatResponse, LlmError>> + Send {
        let msgs = messages.to_vec();
        async move { Ok(ChatResponse::Text(self.chat(&msgs).await?)) }
    }

    /// Return the cache usage from the last API call, if available.
    /// Returns `(cache_creation_tokens, cache_read_tokens)`.
    fn last_cache_usage(&self) -> Option<(u64, u64)> {
        None
    }

    /// Return token counts from the last API call, if available.
    /// Returns `(input_tokens, output_tokens)`.
    fn last_usage(&self) -> Option<(u64, u64)> {
        None
    }

    /// Return reasoning tokens from the last API call, if the provider reports them.
    ///
    /// Reasoning tokens are a **subset** of completion tokens (`OpenAI` o-series only).
    /// Returns `None` for providers that do not expose reasoning token counts.
    fn last_reasoning_tokens(&self) -> Option<u64> {
        None
    }

    /// Return the time-to-first-byte (milliseconds) of the last API call, if available.
    ///
    /// This trait method always returns a TTFB proxy (measured around the HTTP request send,
    /// `retry::send_with_retry` or the equivalent per-attempt retry loop for providers that
    /// don't use that shared helper) — it never reflects true time-to-first-token, even for
    /// calls that streamed. `None` for providers with no HTTP round-trip (Candle) or that
    /// have not made a call yet.
    ///
    /// The one production streaming path (speculative decoding) captures true
    /// time-to-first-token separately, at its own stream-consumption point, and injects it
    /// alongside (overriding) this method's TTFB value when building a `usage_records` row —
    /// see `zeph_core::agent::speculative::stream_drainer::SpeculativeStreamDrainer::drive`.
    /// This trait method itself is not aware of that override; callers that need the true
    /// streaming value must go through that call site, not this one.
    fn last_ttft_ms(&self) -> Option<u64> {
        None
    }

    /// Return the compaction summary from the most recent API call, if a server-side
    /// compaction occurred (Claude compact-2026-01-12 beta). Clears the stored value.
    fn take_compaction_summary(&self) -> Option<String> {
        None
    }

    /// Send messages and return the assistant response together with per-call extras.
    ///
    /// Default implementation calls [`chat`][Self::chat] and returns [`ChatExtras::default()`],
    /// keeping every existing implementor source-compatible at zero cost.
    ///
    /// Providers that support logprobs (`OpenAI`, `Compatible`, `Ollama`) override this to
    /// populate [`ChatExtras::entropy`] with the mean negative log-probability.
    ///
    /// `CoE` is the only caller of this method; the canonical entry point for the agent
    /// loop remains [`chat`][Self::chat].
    ///
    /// # Errors
    ///
    /// Same as [`chat`][Self::chat].
    fn chat_with_extras(
        &self,
        messages: &[Message],
    ) -> impl Future<Output = Result<(String, ChatExtras), LlmError>> + Send {
        let msgs = messages.to_vec();
        async move { Ok((self.chat(&msgs).await?, ChatExtras::default())) }
    }

    /// Return the request payload that will be sent to the provider, for debug dumps.
    ///
    /// Implementations should mirror the provider's request body as closely as practical.
    #[must_use]
    fn debug_request_json(
        &self,
        messages: &[Message],
        tools: &[ToolDefinition],
        _stream: bool,
    ) -> serde_json::Value {
        default_debug_request_json(messages, tools)
    }

    /// Return the list of model identifiers this provider can serve.
    /// Default: empty (provider does not advertise models).
    fn list_models(&self) -> Vec<String> {
        vec![]
    }

    /// Whether this provider supports native structured output.
    fn supports_structured_output(&self) -> bool {
        false
    }

    /// Send messages and parse the response into a typed value `T`.
    ///
    /// Default implementation injects JSON schema into the system prompt and retries once
    /// on parse failure. Providers with native structured output should override this.
    ///
    /// # Object safety
    ///
    /// This method requires `Self: Sized` and is therefore unavailable on trait objects.
    /// Use [`chat_typed_dyn`](crate::provider_dyn::chat_typed_dyn) when working with
    /// `Arc<dyn LlmProviderDyn>`.
    #[allow(async_fn_in_trait)]
    async fn chat_typed<T>(&self, messages: &[Message]) -> Result<T, LlmError>
    where
        T: serde::de::DeserializeOwned + schemars::JsonSchema + 'static,
        Self: Sized,
    {
        let (_, schema_json) = cached_schema::<T>()?;
        let type_name = short_type_name::<T>();

        let mut augmented = messages.to_vec();
        let instruction = format!(
            "Respond with a valid JSON object matching this schema. \
             Output ONLY the JSON, no markdown fences or extra text.\n\n\
             Type: {type_name}\nSchema:\n```json\n{schema_json}\n```"
        );
        augmented.insert(0, Message::from_legacy(Role::System, instruction));

        let raw = self.chat(&augmented).await?;
        let cleaned = strip_json_fences(&raw);
        match serde_json::from_str::<T>(cleaned) {
            Ok(val) => Ok(val),
            Err(first_err) => {
                augmented.push(Message::from_legacy(Role::Assistant, &raw));
                augmented.push(Message::from_legacy(
                    Role::User,
                    format!(
                        "Your response was not valid JSON. Error: {first_err}. \
                         Please output ONLY valid JSON matching the schema."
                    ),
                ));
                let retry_raw = self.chat(&augmented).await?;
                let retry_cleaned = strip_json_fences(&retry_raw);
                serde_json::from_str::<T>(retry_cleaned).map_err(|e| {
                    LlmError::StructuredParse(format!("parse failed after retry: {e}"))
                })
            }
        }
    }
}

/// Strip markdown code fences from LLM output. Only handles outer fences;
/// JSON containing trailing triple backticks in string values may be
/// incorrectly trimmed (acceptable for MVP — see review R2).
fn strip_json_fences(s: &str) -> &str {
    s.trim()
        .trim_start_matches("```json")
        .trim_start_matches("```")
        .trim_end_matches("```")
        .trim()
}

#[cfg(test)]
mod tests {
    use std::assert_matches;
    use tokio_stream::StreamExt;

    use super::*;

    struct StubProvider {
        response: String,
    }

    impl LlmProvider for StubProvider {
        async fn chat(&self, _messages: &[Message]) -> Result<String, LlmError> {
            Ok(self.response.clone())
        }

        async fn chat_stream(&self, messages: &[Message]) -> Result<ChatStream, LlmError> {
            let response = self.chat(messages).await?;
            Ok(Box::pin(tokio_stream::once(Ok(StreamChunk::Content(
                response,
            )))))
        }

        fn supports_streaming(&self) -> bool {
            false
        }

        async fn embed(&self, _text: &str) -> Result<Vec<f32>, LlmError> {
            Ok(vec![0.1, 0.2, 0.3])
        }

        fn supports_embeddings(&self) -> bool {
            false
        }

        fn name(&self) -> &'static str {
            "stub"
        }
    }

    #[test]
    fn test_image_data_debug_redacts_bytes() {
        let img = ImageData {
            data: vec![0xAB, 0xCD, 0xEF],
            mime_type: "image/png".to_owned(),
        };
        let debug = format!("{img:?}");
        assert_eq!(debug, "[image: image/png, 3 bytes]");
        assert!(!debug.contains("171") && !debug.contains("205") && !debug.contains("239"));
    }

    #[test]
    fn context_window_default_returns_none() {
        let provider = StubProvider {
            response: String::new(),
        };
        assert!(provider.context_window().is_none());
    }

    #[test]
    fn supports_streaming_default_returns_false() {
        let provider = StubProvider {
            response: String::new(),
        };
        assert!(!provider.supports_streaming());
    }

    #[test]
    fn supports_tool_use_default_returns_false() {
        // StubProvider overrides neither `supports_tool_use` nor `chat_with_tools`,
        // mirroring CandleProvider (crates/zeph-llm/src/candle_provider/mod.rs). The
        // default must report `false` so callers gating on this method skip providers
        // that would otherwise silently drop tool definitions via the `chat_with_tools`
        // fallback (issue #5687).
        let provider = StubProvider {
            response: String::new(),
        };
        assert!(!provider.supports_tool_use());
    }

    #[tokio::test]
    async fn chat_stream_default_yields_single_chunk() {
        let provider = StubProvider {
            response: "hello world".into(),
        };
        let messages = vec![Message {
            role: Role::User,
            content: "test".into(),
            parts: vec![],
            metadata: MessageMetadata::default(),
        }];

        let mut stream = provider.chat_stream(&messages).await.unwrap();
        let chunk = stream.next().await.unwrap().unwrap();
        assert_matches!(chunk, StreamChunk::Content(s) if s == "hello world");
        assert!(stream.next().await.is_none());
    }

    #[tokio::test]
    async fn chat_stream_default_propagates_chat_error() {
        struct FailProvider;

        impl LlmProvider for FailProvider {
            async fn chat(&self, _messages: &[Message]) -> Result<String, LlmError> {
                Err(LlmError::Unavailable)
            }

            async fn chat_stream(&self, messages: &[Message]) -> Result<ChatStream, LlmError> {
                let response = self.chat(messages).await?;
                Ok(Box::pin(tokio_stream::once(Ok(StreamChunk::Content(
                    response,
                )))))
            }

            fn supports_streaming(&self) -> bool {
                false
            }

            async fn embed(&self, _text: &str) -> Result<Vec<f32>, LlmError> {
                Err(LlmError::Unavailable)
            }

            fn supports_embeddings(&self) -> bool {
                false
            }

            fn name(&self) -> &'static str {
                "fail"
            }
        }

        let provider = FailProvider;
        let messages = vec![Message {
            role: Role::User,
            content: "test".into(),
            parts: vec![],
            metadata: MessageMetadata::default(),
        }];

        let result = provider.chat_stream(&messages).await;
        assert!(result.is_err());
        if let Err(e) = result {
            assert!(e.to_string().contains("provider unavailable"));
        }
    }

    #[tokio::test]
    async fn stub_provider_embed_returns_vector() {
        let provider = StubProvider {
            response: String::new(),
        };
        let embedding = provider.embed("test").await.unwrap();
        assert_eq!(embedding, vec![0.1, 0.2, 0.3]);
    }

    #[tokio::test]
    async fn fail_provider_embed_propagates_error() {
        struct FailProvider;

        impl LlmProvider for FailProvider {
            async fn chat(&self, _messages: &[Message]) -> Result<String, LlmError> {
                Err(LlmError::Unavailable)
            }

            async fn chat_stream(&self, messages: &[Message]) -> Result<ChatStream, LlmError> {
                let response = self.chat(messages).await?;
                Ok(Box::pin(tokio_stream::once(Ok(StreamChunk::Content(
                    response,
                )))))
            }

            fn supports_streaming(&self) -> bool {
                false
            }

            async fn embed(&self, _text: &str) -> Result<Vec<f32>, LlmError> {
                Err(LlmError::EmbedUnsupported {
                    provider: "fail".into(),
                })
            }

            fn supports_embeddings(&self) -> bool {
                false
            }

            fn name(&self) -> &'static str {
                "fail"
            }
        }

        let provider = FailProvider;
        let result = provider.embed("test").await;
        assert!(result.is_err());
        assert!(
            result
                .unwrap_err()
                .to_string()
                .contains("embedding not supported")
        );
    }

    #[test]
    fn role_serialization() {
        let system = Role::System;
        let user = Role::User;
        let assistant = Role::Assistant;

        assert_eq!(serde_json::to_string(&system).unwrap(), "\"system\"");
        assert_eq!(serde_json::to_string(&user).unwrap(), "\"user\"");
        assert_eq!(serde_json::to_string(&assistant).unwrap(), "\"assistant\"");
    }

    #[test]
    fn role_deserialization() {
        let system: Role = serde_json::from_str("\"system\"").unwrap();
        let user: Role = serde_json::from_str("\"user\"").unwrap();
        let assistant: Role = serde_json::from_str("\"assistant\"").unwrap();

        assert_eq!(system, Role::System);
        assert_eq!(user, Role::User);
        assert_eq!(assistant, Role::Assistant);
    }

    #[test]
    fn message_clone() {
        let msg = Message {
            role: Role::User,
            content: "test".into(),
            parts: vec![],
            metadata: MessageMetadata::default(),
        };
        let cloned = msg.clone();
        assert_eq!(cloned.role, msg.role);
        assert_eq!(cloned.content, msg.content);
    }

    #[test]
    fn message_debug() {
        let msg = Message {
            role: Role::Assistant,
            content: "response".into(),
            parts: vec![],
            metadata: MessageMetadata::default(),
        };
        let debug = format!("{msg:?}");
        assert!(debug.contains("Assistant"));
        assert!(debug.contains("response"));
    }

    #[test]
    fn message_serialization() {
        let msg = Message {
            role: Role::User,
            content: "hello".into(),
            parts: vec![],
            metadata: MessageMetadata::default(),
        };
        let json = serde_json::to_string(&msg).unwrap();
        assert!(json.contains("\"role\":\"user\""));
        assert!(json.contains("\"content\":\"hello\""));
    }

    #[test]
    fn message_part_serde_round_trip() {
        let parts = vec![
            MessagePart::Text {
                text: "hello".into(),
            },
            MessagePart::ToolOutput {
                tool_name: "bash".into(),
                body: "output".into(),
                compacted_at: None,
            },
            MessagePart::Recall {
                text: "recall".into(),
            },
            MessagePart::CodeContext {
                text: "code".into(),
            },
            MessagePart::Summary {
                text: "summary".into(),
            },
        ];
        let json = serde_json::to_string(&parts).unwrap();
        let deserialized: Vec<MessagePart> = serde_json::from_str(&json).unwrap();
        assert_eq!(deserialized.len(), 5);
    }

    #[test]
    fn from_legacy_creates_empty_parts() {
        let msg = Message::from_legacy(Role::User, "hello");
        assert_eq!(msg.role, Role::User);
        assert_eq!(msg.content, "hello");
        assert!(msg.parts.is_empty());
        assert_eq!(msg.to_llm_content(), "hello");
    }

    #[test]
    fn from_parts_flattens_content() {
        let msg = Message::from_parts(
            Role::System,
            vec![MessagePart::Recall {
                text: "recalled data".into(),
            }],
        );
        assert_eq!(msg.content, "recalled data");
        assert_eq!(msg.to_llm_content(), "recalled data");
        assert_eq!(msg.parts.len(), 1);
    }

    #[test]
    fn from_parts_tool_output_format() {
        let msg = Message::from_parts(
            Role::User,
            vec![MessagePart::ToolOutput {
                tool_name: "bash".into(),
                body: "hello world".into(),
                compacted_at: None,
            }],
        );
        assert!(msg.content.contains("[tool output: bash]"));
        assert!(msg.content.contains("hello world"));
    }

    #[test]
    fn message_deserializes_without_parts() {
        let json = r#"{"role":"user","content":"hello"}"#;
        let msg: Message = serde_json::from_str(json).unwrap();
        assert_eq!(msg.content, "hello");
        assert!(msg.parts.is_empty());
    }

    #[test]
    fn flatten_skips_compacted_tool_output_empty_body() {
        // When compacted_at is set and body is empty, renders "(pruned)".
        let msg = Message::from_parts(
            Role::User,
            vec![
                MessagePart::Text {
                    text: "prefix ".into(),
                },
                MessagePart::ToolOutput {
                    tool_name: "bash".into(),
                    body: String::new(),
                    compacted_at: Some(1234),
                },
                MessagePart::Text {
                    text: " suffix".into(),
                },
            ],
        );
        assert!(msg.content.contains("(pruned)"));
        assert!(msg.content.contains("prefix "));
        assert!(msg.content.contains(" suffix"));
    }

    #[test]
    fn flatten_compacted_tool_output_with_reference_renders_body() {
        // When compacted_at is set and body contains a reference notice, renders the body.
        let ref_notice = "[tool output pruned; full content at /tmp/overflow/big.txt]";
        let msg = Message::from_parts(
            Role::User,
            vec![MessagePart::ToolOutput {
                tool_name: "bash".into(),
                body: ref_notice.into(),
                compacted_at: Some(1234),
            }],
        );
        assert!(msg.content.contains(ref_notice));
        assert!(!msg.content.contains("(pruned)"));
    }

    #[test]
    fn rebuild_content_syncs_after_mutation() {
        let mut msg = Message::from_parts(
            Role::User,
            vec![MessagePart::ToolOutput {
                tool_name: "bash".into(),
                body: "original".into(),
                compacted_at: None,
            }],
        );
        assert!(msg.content.contains("original"));

        if let MessagePart::ToolOutput {
            ref mut compacted_at,
            ref mut body,
            ..
        } = msg.parts[0]
        {
            *compacted_at = Some(999);
            body.clear(); // simulate pruning: body cleared, no overflow notice
        }
        msg.rebuild_content();

        assert!(msg.content.contains("(pruned)"));
        assert!(!msg.content.contains("original"));
    }

    #[test]
    fn message_part_tool_use_serde_round_trip() {
        let part = MessagePart::ToolUse {
            id: "toolu_123".into(),
            name: "bash".into(),
            input: serde_json::json!({"command": "ls"}),
        };
        let json = serde_json::to_string(&part).unwrap();
        let deserialized: MessagePart = serde_json::from_str(&json).unwrap();
        if let MessagePart::ToolUse { id, name, input } = deserialized {
            assert_eq!(id, "toolu_123");
            assert_eq!(name, "bash");
            assert_eq!(input["command"], "ls");
        } else {
            panic!("expected ToolUse");
        }
    }

    #[test]
    fn message_part_tool_result_serde_round_trip() {
        let part = MessagePart::ToolResult {
            tool_use_id: "toolu_123".into(),
            content: "file1.rs\nfile2.rs".into(),
            is_error: false,
        };
        let json = serde_json::to_string(&part).unwrap();
        let deserialized: MessagePart = serde_json::from_str(&json).unwrap();
        if let MessagePart::ToolResult {
            tool_use_id,
            content,
            is_error,
        } = deserialized
        {
            assert_eq!(tool_use_id, "toolu_123");
            assert_eq!(content, "file1.rs\nfile2.rs");
            assert!(!is_error);
        } else {
            panic!("expected ToolResult");
        }
    }

    #[test]
    fn message_part_tool_result_is_error_default() {
        let json = r#"{"kind":"tool_result","tool_use_id":"id","content":"err"}"#;
        let part: MessagePart = serde_json::from_str(json).unwrap();
        if let MessagePart::ToolResult { is_error, .. } = part {
            assert!(!is_error);
        } else {
            panic!("expected ToolResult");
        }
    }

    #[test]
    fn chat_response_construction() {
        let text = ChatResponse::Text("hello".into());
        assert_matches!(text, ChatResponse::Text(s) if s == "hello");

        let tool_use = ChatResponse::ToolUse {
            text: Some("I'll run that".into()),
            tool_calls: vec![ToolUseRequest {
                id: "1".into(),
                name: "bash".into(),
                input: serde_json::json!({}),
            }],
            thinking_blocks: vec![],
        };
        assert_matches!(tool_use, ChatResponse::ToolUse { .. });
    }

    #[test]
    fn flatten_parts_tool_use() {
        let msg = Message::from_parts(
            Role::Assistant,
            vec![MessagePart::ToolUse {
                id: "t1".into(),
                name: "bash".into(),
                input: serde_json::json!({"command": "ls"}),
            }],
        );
        assert!(msg.content.contains("[tool_use: bash(t1)]"));
    }

    #[test]
    fn flatten_parts_tool_result() {
        let msg = Message::from_parts(
            Role::User,
            vec![MessagePart::ToolResult {
                tool_use_id: "t1".into(),
                content: "output here".into(),
                is_error: false,
            }],
        );
        assert!(msg.content.contains("[tool_result: t1]"));
        assert!(msg.content.contains("output here"));
    }

    #[test]
    fn tool_definition_serde_round_trip() {
        let def = ToolDefinition {
            name: "bash".into(),
            description: "Execute a shell command".into(),
            parameters: serde_json::json!({"type": "object"}),
            output_schema: None,
        };
        let json = serde_json::to_string(&def).unwrap();
        let deserialized: ToolDefinition = serde_json::from_str(&json).unwrap();
        assert_eq!(deserialized.name, "bash");
        assert_eq!(deserialized.description, "Execute a shell command");
    }

    #[tokio::test]
    async fn chat_with_tools_default_delegates_to_chat() {
        let provider = StubProvider {
            response: "hello".into(),
        };
        let messages = vec![Message::from_legacy(Role::User, "test")];
        let result = provider.chat_with_tools(&messages, &[]).await.unwrap();
        assert_matches!(result, ChatResponse::Text(s) if s == "hello");
    }

    #[test]
    fn tool_output_compacted_at_serde_default() {
        let json = r#"{"kind":"tool_output","tool_name":"bash","body":"out"}"#;
        let part: MessagePart = serde_json::from_str(json).unwrap();
        if let MessagePart::ToolOutput { compacted_at, .. } = part {
            assert!(compacted_at.is_none());
        } else {
            panic!("expected ToolOutput");
        }
    }

    // --- M27: strip_json_fences tests ---

    #[test]
    fn strip_json_fences_plain_json() {
        assert_eq!(strip_json_fences(r#"{"a": 1}"#), r#"{"a": 1}"#);
    }

    #[test]
    fn strip_json_fences_with_json_fence() {
        assert_eq!(strip_json_fences("```json\n{\"a\": 1}\n```"), r#"{"a": 1}"#);
    }

    #[test]
    fn strip_json_fences_with_plain_fence() {
        assert_eq!(strip_json_fences("```\n{\"a\": 1}\n```"), r#"{"a": 1}"#);
    }

    #[test]
    fn strip_json_fences_whitespace() {
        assert_eq!(strip_json_fences("  \n  "), "");
    }

    #[test]
    fn strip_json_fences_empty() {
        assert_eq!(strip_json_fences(""), "");
    }

    #[test]
    fn strip_json_fences_outer_whitespace() {
        assert_eq!(
            strip_json_fences("  ```json\n{\"a\": 1}\n```  "),
            r#"{"a": 1}"#
        );
    }

    #[test]
    fn strip_json_fences_only_opening_fence() {
        assert_eq!(strip_json_fences("```json\n{\"a\": 1}"), r#"{"a": 1}"#);
    }

    // --- M27: chat_typed tests ---

    #[derive(Debug, serde::Deserialize, schemars::JsonSchema, PartialEq)]
    struct TestOutput {
        value: String,
    }

    struct SequentialStub {
        responses: std::sync::Mutex<Vec<Result<String, LlmError>>>,
    }

    impl SequentialStub {
        fn new(responses: Vec<Result<String, LlmError>>) -> Self {
            Self {
                responses: std::sync::Mutex::new(responses),
            }
        }
    }

    impl LlmProvider for SequentialStub {
        async fn chat(&self, _messages: &[Message]) -> Result<String, LlmError> {
            let mut responses = self.responses.lock().unwrap();
            if responses.is_empty() {
                return Err(LlmError::Other("no more responses".into()));
            }
            responses.remove(0)
        }

        async fn chat_stream(&self, messages: &[Message]) -> Result<ChatStream, LlmError> {
            let response = self.chat(messages).await?;
            Ok(Box::pin(tokio_stream::once(Ok(StreamChunk::Content(
                response,
            )))))
        }

        fn supports_streaming(&self) -> bool {
            false
        }

        async fn embed(&self, _text: &str) -> Result<Vec<f32>, LlmError> {
            Err(LlmError::EmbedUnsupported {
                provider: "sequential-stub".into(),
            })
        }

        fn supports_embeddings(&self) -> bool {
            false
        }

        fn name(&self) -> &'static str {
            "sequential-stub"
        }
    }

    #[tokio::test]
    async fn chat_typed_happy_path() {
        let provider = StubProvider {
            response: r#"{"value": "hello"}"#.into(),
        };
        let messages = vec![Message::from_legacy(Role::User, "test")];
        let result: TestOutput = provider.chat_typed(&messages).await.unwrap();
        assert_eq!(
            result,
            TestOutput {
                value: "hello".into()
            }
        );
    }

    #[tokio::test]
    async fn chat_typed_retry_succeeds() {
        let provider = SequentialStub::new(vec![
            Ok("not valid json".into()),
            Ok(r#"{"value": "ok"}"#.into()),
        ]);
        let messages = vec![Message::from_legacy(Role::User, "test")];
        let result: TestOutput = provider.chat_typed(&messages).await.unwrap();
        assert_eq!(result, TestOutput { value: "ok".into() });
    }

    #[tokio::test]
    async fn chat_typed_both_fail() {
        let provider = SequentialStub::new(vec![Ok("bad json".into()), Ok("still bad".into())]);
        let messages = vec![Message::from_legacy(Role::User, "test")];
        let result = provider.chat_typed::<TestOutput>(&messages).await;
        let err = result.unwrap_err();
        assert!(err.to_string().contains("parse failed after retry"));
    }

    #[tokio::test]
    async fn chat_typed_chat_error_propagates() {
        let provider = SequentialStub::new(vec![Err(LlmError::Unavailable)]);
        let messages = vec![Message::from_legacy(Role::User, "test")];
        let result = provider.chat_typed::<TestOutput>(&messages).await;
        assert_matches!(result, Err(LlmError::Unavailable));
    }

    #[tokio::test]
    async fn chat_typed_strips_fences() {
        let provider = StubProvider {
            response: "```json\n{\"value\": \"fenced\"}\n```".into(),
        };
        let messages = vec![Message::from_legacy(Role::User, "test")];
        let result: TestOutput = provider.chat_typed(&messages).await.unwrap();
        assert_eq!(
            result,
            TestOutput {
                value: "fenced".into()
            }
        );
    }

    #[test]
    fn supports_structured_output_default_false() {
        let provider = StubProvider {
            response: String::new(),
        };
        assert!(!provider.supports_structured_output());
    }

    #[test]
    fn structured_parse_error_display() {
        let err = LlmError::StructuredParse("test error".into());
        assert_eq!(
            err.to_string(),
            "structured output parse failed: test error"
        );
    }

    #[test]
    fn message_part_image_roundtrip_json() {
        let part = MessagePart::Image(Box::new(ImageData {
            data: vec![1, 2, 3, 4],
            mime_type: "image/jpeg".into(),
        }));
        let json = serde_json::to_string(&part).unwrap();
        let decoded: MessagePart = serde_json::from_str(&json).unwrap();
        match decoded {
            MessagePart::Image(img) => {
                assert_eq!(img.data, vec![1, 2, 3, 4]);
                assert_eq!(img.mime_type, "image/jpeg");
            }
            _ => panic!("expected Image variant"),
        }
    }

    #[test]
    fn flatten_parts_includes_image_placeholder() {
        let msg = Message::from_parts(
            Role::User,
            vec![
                MessagePart::Text {
                    text: "see this".into(),
                },
                MessagePart::Image(Box::new(ImageData {
                    data: vec![0u8; 100],
                    mime_type: "image/png".into(),
                })),
            ],
        );
        let content = msg.to_llm_content();
        assert!(content.contains("see this"));
        assert!(content.contains("[image: image/png"));
    }

    #[test]
    fn supports_vision_default_false() {
        let provider = StubProvider {
            response: String::new(),
        };
        assert!(!provider.supports_vision());
    }

    #[test]
    fn message_metadata_default_both_visible() {
        let m = MessageMetadata::default();
        assert!(m.visibility.is_agent_visible());
        assert!(m.visibility.is_user_visible());
        assert_eq!(m.visibility, MessageVisibility::Both);
        assert!(m.compacted_at.is_none());
    }

    #[test]
    fn message_metadata_agent_only() {
        let m = MessageMetadata::agent_only();
        assert!(m.visibility.is_agent_visible());
        assert!(!m.visibility.is_user_visible());
        assert_eq!(m.visibility, MessageVisibility::AgentOnly);
    }

    #[test]
    fn message_metadata_user_only() {
        let m = MessageMetadata::user_only();
        assert!(!m.visibility.is_agent_visible());
        assert!(m.visibility.is_user_visible());
        assert_eq!(m.visibility, MessageVisibility::UserOnly);
    }

    #[test]
    fn message_metadata_serde_default() {
        let json = r#"{"role":"user","content":"hello"}"#;
        let msg: Message = serde_json::from_str(json).unwrap();
        assert!(msg.metadata.visibility.is_agent_visible());
        assert!(msg.metadata.visibility.is_user_visible());
    }

    #[test]
    fn message_metadata_round_trip() {
        let msg = Message {
            role: Role::User,
            content: "test".into(),
            parts: vec![],
            metadata: MessageMetadata::agent_only(),
        };
        let json = serde_json::to_string(&msg).unwrap();
        let decoded: Message = serde_json::from_str(&json).unwrap();
        assert!(decoded.metadata.visibility.is_agent_visible());
        assert!(!decoded.metadata.visibility.is_user_visible());
        assert_eq!(decoded.metadata.visibility, MessageVisibility::AgentOnly);
    }

    #[test]
    fn message_part_compaction_round_trip() {
        let part = MessagePart::Compaction {
            summary: "Context was summarized.".to_owned(),
        };
        let json = serde_json::to_string(&part).unwrap();
        let decoded: MessagePart = serde_json::from_str(&json).unwrap();
        assert!(
            matches!(decoded, MessagePart::Compaction { summary } if summary == "Context was summarized.")
        );
    }

    #[test]
    fn flatten_parts_compaction_contributes_no_text() {
        // MessagePart::Compaction must not appear in the flattened content string
        // (it's metadata-only; the summary is stored on the Message separately).
        let parts = vec![
            MessagePart::Text {
                text: "Hello".to_owned(),
            },
            MessagePart::Compaction {
                summary: "Summary".to_owned(),
            },
        ];
        let msg = Message::from_parts(Role::Assistant, parts);
        // Only the Text part should appear in content.
        assert_eq!(msg.content.trim(), "Hello");
    }

    #[test]
    fn stream_chunk_compaction_variant() {
        let chunk = StreamChunk::Compaction("A summary".to_owned());
        assert_matches!(chunk, StreamChunk::Compaction(s) if s == "A summary");
    }

    #[test]
    fn short_type_name_extracts_last_segment() {
        struct MyOutput;
        assert_eq!(short_type_name::<MyOutput>(), "MyOutput");
    }

    #[test]
    fn short_type_name_primitive_returns_full_name() {
        // Primitives have no "::" in their type_name — rsplit returns the full name.
        assert_eq!(short_type_name::<u32>(), "u32");
        assert_eq!(short_type_name::<bool>(), "bool");
    }

    #[test]
    fn short_type_name_nested_path_returns_last() {
        // Use a type whose path contains "::" segments.
        assert_eq!(
            short_type_name::<std::collections::HashMap<u32, u32>>(),
            "HashMap<u32, u32>"
        );
    }

    // Regression test for #2257: `MessagePart::Summary` must serialize to the
    // internally-tagged format `{"kind":"summary","text":"..."}` and round-trip correctly.
    #[test]
    fn summary_roundtrip() {
        let part = MessagePart::Summary {
            text: "hello".to_string(),
        };
        let json = serde_json::to_string(&part).expect("serialization must not fail");
        assert!(
            json.contains("\"kind\":\"summary\""),
            "must use internally-tagged format, got: {json}"
        );
        assert!(
            !json.contains("\"Summary\""),
            "must not use externally-tagged format, got: {json}"
        );
        let decoded: MessagePart =
            serde_json::from_str(&json).expect("deserialization must not fail");
        match decoded {
            MessagePart::Summary { text } => assert_eq!(text, "hello"),
            other => panic!("expected MessagePart::Summary, got {other:?}"),
        }
    }

    #[tokio::test]
    async fn embed_batch_default_empty_returns_empty() {
        let provider = StubProvider {
            response: String::new(),
        };
        let result = provider.embed_batch(&[]).await.unwrap();
        assert!(result.is_empty());
    }

    #[tokio::test]
    async fn embed_batch_default_calls_embed_sequentially() {
        let provider = StubProvider {
            response: String::new(),
        };
        let texts = ["hello", "world", "foo"];
        let result = provider.embed_batch(&texts).await.unwrap();
        assert_eq!(result.len(), 3);
        // StubProvider::embed always returns [0.1, 0.2, 0.3]
        for vec in &result {
            assert_eq!(vec, &[0.1_f32, 0.2, 0.3]);
        }
    }

    #[test]
    fn message_visibility_db_roundtrip_both() {
        assert_eq!(MessageVisibility::Both.as_db_str(), "both");
        assert_eq!(
            MessageVisibility::from_db_str("both"),
            MessageVisibility::Both
        );
    }

    #[test]
    fn message_visibility_db_roundtrip_agent_only() {
        assert_eq!(MessageVisibility::AgentOnly.as_db_str(), "agent_only");
        assert_eq!(
            MessageVisibility::from_db_str("agent_only"),
            MessageVisibility::AgentOnly
        );
    }

    #[test]
    fn message_visibility_db_roundtrip_user_only() {
        assert_eq!(MessageVisibility::UserOnly.as_db_str(), "user_only");
        assert_eq!(
            MessageVisibility::from_db_str("user_only"),
            MessageVisibility::UserOnly
        );
    }

    #[test]
    fn message_visibility_from_db_str_unknown_defaults_to_both() {
        assert_eq!(
            MessageVisibility::from_db_str("unknown_future_value"),
            MessageVisibility::Both
        );
        assert_eq!(MessageVisibility::from_db_str(""), MessageVisibility::Both);
    }
}