molo_agent/agent/

react.rs

//! The ReAct reasoning loop: the classic Agent assembly shipped with the
//! framework.
//!
//! The generic loop shape: record the user input → start a conversation →
//! execute tool requests from the model and feed the results back → until
//! the model answers directly; tool execution failures are fed back to the
//! model as text, and the model decides what to do next.
//!
//! Getting started needs just three required parameters: provider
//! ([`Provider`]) + tools ([`ToolRegistry`]) + system_prompt; Memory
//! defaults to a bounded window ([`WindowMemory`], 128k token budget, long
//! conversations auto-trim the oldest rounds); replace it with
//! [`with_memory`](ReActAgent::with_memory) when you need something custom.
//!
//! Pre-call approval at the loop level is not built in; a custom
//! [`ToolRoundExecutor`] injected via
//! [`with_tool_round_executor`](ReActAgent::with_tool_round_executor) is
//! the application-side hook for round-level gates.

use super::config::AgentConfig;
#[cfg(feature = "structured")]
use super::structured::{StructuredOutcome, StructuredValidator};
#[cfg(feature = "structured")]
use crate::agent::TypedAgent;
use crate::agent::events::ReActEvent;
use crate::agent::{
    Agent, AgentAction, AgentError, AgentEvent, AgentKernel, MessageChunk, ModelObservation,
    ModelRequest, Observation, RunSummary,
};
use crate::effect::{EffectObservation, EffectRequest};
use crate::event_channel::EventChannel;
use crate::memory::{Memory, WindowMemory};
use crate::message::{Message, ToolCall};
use crate::provider::FakeProvider;
use crate::provider::{
    ChatRequest, FinishReason, ModelOptions, Provider, ProviderError, ProviderRequestContext,
    StreamEvent, Usage,
};
#[cfg(feature = "structured")]
use crate::run::TypedRunOutput;
use crate::run::{Artifact, RunContext, RunMetadata, RunOutput, RunRequest};
use crate::tool::{SharedState, Tool, ToolMemoryPolicy, ToolOutput, ToolRegistry, ToolResult};
use futures::StreamExt;
use futures::stream::BoxStream;
#[cfg(feature = "structured")]
use schemars::JsonSchema;
#[cfg(feature = "structured")]
use serde::de::DeserializeOwned;
use std::collections::HashMap;
use std::collections::HashSet;
use std::collections::VecDeque;
use std::fmt;
use std::future::Future;
use std::pin::Pin;
use std::sync::Arc;
use std::time::Instant;
#[cfg(feature = "tracing")]
use tracing::Instrument;

/// Convenience assembly macro: registers a list of tools (possibly
/// heterogeneous) with automatic boxing, creating a
/// [`ToolRegistry`](crate::tool::ToolRegistry) internally. The system prompt
/// is **optional** (omitted = no system prompt). Six arms:
///
/// - `react_agent!(provider)` — no tools, no system prompt;
/// - `react_agent!(provider, "system prompt")` — no tools, with a system
///   prompt (the string must be a **literal**, distinguishing it from the
///   "existing registry" arm; to use a variable, write
///   `react_agent!(provider, [], system_var)` or the three-arg arm);
/// - `react_agent!(provider, [tool_1, tool_2, ...])` — a heterogeneous tool
///   list (types need not match), auto-registered, no system prompt;
/// - `react_agent!(provider, [tool_1, ...], "system prompt")` — list +
///   system prompt;
/// - `react_agent!(provider, registry)` — an existing registry (e.g. a
///   sub-agent trimmed via `subset`), no system prompt;
/// - `react_agent!(provider, registry, "system prompt")` — registry +
///   system prompt.
///
/// Returns [`ReActAgent`](crate::agent::ReActAgent); chained
/// [`with_memory`](ReActAgent::with_memory) / [`with_config`](ReActAgent::with_config) /
/// [`with_state`](ReActAgent::with_state) / [`with_event_channel`](ReActAgent::with_event_channel) /
/// [`with_tool_round_executor`](ReActAgent::with_tool_round_executor) work as usual.
/// `ReActAgent::new` keeps a single signature; the macro handles the
/// multiple shapes.
///
/// ```
/// # extern crate molo_agent as molo;
/// # #[tokio::main]
/// # async fn main() -> Result<(), molo::AgentError> {
/// use molo::{react_agent, Agent, FakeProvider, FakeReply};
///
/// let mut agent = react_agent!(
///     FakeProvider::new([FakeReply::Text("Hello".into())]),
///     "You are a helpful assistant",
/// );
/// assert_eq!(agent.run("Are you there").await?, "Hello");
/// # Ok(())
/// # }
/// ```
#[macro_export]
macro_rules! react_agent {
    // No tools: bare / with a system prompt (a string literal).
    ($provider:expr $(,)?) => {
        $crate::agent::ReActAgent::new(
            $provider,
            $crate::tool::ToolRegistry::new(),
            "",
        )
    };
    ($provider:expr, $system:literal $(,)?) => {
        $crate::agent::ReActAgent::new(
            $provider,
            $crate::tool::ToolRegistry::new(),
            $system,
        )
    };
    // Tool list: auto-registered, without / with a system prompt.
    ($provider:expr, [$($tool:expr),* $(,)?] $(,)?) => {{
        let mut __molo_registry = $crate::tool::ToolRegistry::new();
        $(__molo_registry.register($tool);)*
        $crate::agent::ReActAgent::new($provider, __molo_registry, "")
    }};
    ($provider:expr, [$($tool:expr),* $(,)?], $system:expr $(,)?) => {{
        let mut __molo_registry = $crate::tool::ToolRegistry::new();
        $(__molo_registry.register($tool);)*
        $crate::agent::ReActAgent::new($provider, __molo_registry, $system)
    }};
    // Existing registry: without / with a system prompt.
    ($provider:expr, $registry:expr $(,)?) => {
        $crate::agent::ReActAgent::new($provider, $registry, "")
    };
    ($provider:expr, $registry:expr, $system:expr $(,)?) => {
        $crate::agent::ReActAgent::new($provider, $registry, $system)
    };
}

/// The classic ReAct reasoning loop: conversation → tool execution fed
/// back → until the model answers directly.
///
/// Assembly: three required parameters (Provider / ToolRegistry /
/// system_prompt) + chained optional
/// ([`with_memory`](ReActAgent::with_memory) / [`with_config`](ReActAgent::with_config) /
/// [`with_state`](ReActAgent::with_state) / [`with_event_channel`](ReActAgent::with_event_channel) /
/// [`with_tool_round_executor`](ReActAgent::with_tool_round_executor));
/// for simpler assembly use the macro [`react_agent!`](crate::react_agent).
///
/// # Examples
///
/// ```rust
/// # extern crate molo_agent as molo;
/// # #[tokio::main]
/// # async fn main() -> Result<(), molo::AgentError> {
/// use molo::agent::{Agent, ReActAgent};
/// use molo::provider::{FakeProvider, FakeReply};
/// use molo::tool::ToolRegistry;
///
/// let mut agent = ReActAgent::new(
///     FakeProvider::new([FakeReply::Text("Hello".into())]),
///     ToolRegistry::new(),
///     "You are a helpful assistant",
/// );
/// let answer = agent.run("Are you there").await?;
/// assert_eq!(answer, "Hello");
/// # Ok(())
/// # }
/// ```
///
/// # Cancellation semantics
///
/// ReActAgent observes cancellation and deadlines carried by [`RunContext`].
/// Cancellation is cooperative and only checked at safe points:
///
/// - Cancelled mid-conversation: the in-flight request is dropped, and
///   nothing is recorded for this round;
/// - Cancelled during a tool round: the running tool call is not
///   interrupted (tools are user code; a partial execution risks side
///   effects); the Assistant message already recorded this round and the
///   ToolResult fed back right after remain complete;
/// - Streaming path: already-dispatched `Delta` chunks are kept, then
///   the stream concludes with a [`MessageChunk::Cancelled`] terminal
///   chunk — no `Done`;
/// - After cancellation, already-recorded messages are kept, not rolled
///   back; subsequent runs continue from the retained memory.
///
/// Callers that don't need cancellation can just use
/// [`run`](Agent::run) / [`run_stream`](Agent::run_stream) — they
/// internally use a token that is never cancelled.
///
/// # Errors
///
/// Tool execution failures are **not** [`AgentError`] — the error text
/// is fed back to the model via `ToolResult`, the model decides what to
/// do next, and the loop continues. [`AgentError`] only represents
/// run-level failures (Memory / Provider failures, exceeding the
/// tool-round limit, cancellation).
///
/// # Structured output
///
/// The ability to require the model's answer to conform to a given JSON
/// Schema; both entry points share the same validation-and-retry loop
/// (see [`run_typed`](ReActAgent::run_typed) and
/// [`with_structured_output`](ReActAgent::with_structured_output)):
///
/// - [`run_typed`](ReActAgent::run_typed): **typed output** —
///   `run_typed`'s return type directly declares the target type; this
///   run auto-generates a JSON Schema from the type
///   ([`schemars`](https://docs.rs/schemars)-derived, the same pipeline
///   as tool parameter schemas), and deserializes after validation;
/// - [`with_structured_output`](ReActAgent::with_structured_output): a
///   hand-written Schema (or the serialized result of
///   `schemars::schema_for!(T)`), paired with [`Agent::run`](Agent::run)
///   returning JSON text that you parse yourself.
///
/// Both entry points validate the final answer with framework-side
/// jsonschema; on failure the error is fed back to the model for retry
/// (budget [`AgentConfig::max_structured_retries`]); compatible
/// endpoints are additionally constrained as best effort via
/// `response_format` (see
/// [`ModelOptions::structured`](crate::provider::ModelOptions::structured)).
pub struct ReActAgent {
    provider: Box<dyn Provider>,
    memory: Box<dyn Memory>,
    registry: ToolRegistry,
    system_prompt: String,
    config: AgentConfig,
    /// Shared state: the application reads and writes this field directly
    /// across runs; it is injected via [`Tool::call`](crate::Tool::call) on
    /// every tool call, so multiple tools / Agents can share one instance.
    pub state: SharedState,
    /// Observation channel (optional, not attached by default): the loop
    /// pushes process events here, and the host side subscribes.
    events: Option<Arc<dyn EventChannel>>,
    /// Tool-round execution policy (default: [`SerialToolRoundExecutor`] —
    /// calls run one by one, in request order). The application injects a
    /// custom policy (order / concurrency / approval gates) via
    /// [`with_tool_round_executor`](ReActAgent::with_tool_round_executor).
    executor: Box<dyn ToolRoundExecutor>,
    /// Step-wise kernel state. `None` means no `AgentKernel` run is active.
    kernel_state: Option<ReActKernelState>,
}

/// Per-round reply text accumulation limit (4 MiB, shared by streaming and
/// non-streaming): exceeding it is treated as a malicious / abnormal
/// endpoint and terminates the round with `AgentError::Provider`.
const MAX_ROUND_TEXT: usize = 4 << 20;

/// Token budget of the default Memory (128k): `ReActAgent::new` defaults
/// to [`WindowMemory`](crate::memory::WindowMemory), which auto-trims the
/// oldest rounds in long conversations — an unbounded default would let
/// memory grow indefinitely with the conversation; callers needing full
/// history explicitly replace it with
/// [`with_memory`](ReActAgent::with_memory).
const DEFAULT_MEMORY_TOKENS: usize = 128_000;

/// Builder for assembling a [`ReActAgent`] from the same components used by
/// [`ReActAgent::new`], plus optional runtime configuration.
///
/// The builder is only an assembly helper: provider calls still go through
/// the agent loop, immediate tools still run through the tool registry, and
/// governed side effects still belong to the outer harness/kernel path.
/// Skill assembly is handled by the `molo-skills` layer so `ReActAgent`
/// does not own skill policy.
///
/// # Examples
///
/// ```
/// # extern crate molo_agent as molo;
/// # #[tokio::main]
/// # async fn main() -> Result<(), molo::AgentError> {
/// use molo::{Agent, FakeProvider, FakeReply, ReActAgent};
///
/// let mut agent = ReActAgent::builder(FakeProvider::new([
///     FakeReply::Text("Hello".into()),
/// ]))
/// .with_system_prompt("You are a helpful assistant")
/// .build();
///
/// assert_eq!(agent.run("Are you there").await?, "Hello");
/// # Ok(())
/// # }
/// ```
pub struct ReActAgentBuilder {
    provider: Box<dyn Provider>,
    memory: Box<dyn Memory>,
    tools: ToolRegistry,
    system_prompt: String,
    config: AgentConfig,
    state: SharedState,
    events: Option<Arc<dyn EventChannel>>,
    executor: Box<dyn ToolRoundExecutor>,
}

impl ReActAgentBuilder {
    /// Starts a builder with a provider and default optional components.
    pub fn new(provider: impl Provider + 'static) -> Self {
        Self {
            provider: Box::new(provider),
            memory: Box::new(WindowMemory::new(DEFAULT_MEMORY_TOKENS)),
            tools: ToolRegistry::new(),
            system_prompt: String::new(),
            config: AgentConfig::default(),
            state: SharedState::default(),
            events: None,
            executor: Box::new(SerialToolRoundExecutor),
        }
    }

    /// Replaces the provider.
    pub fn with_provider(mut self, provider: impl Provider + 'static) -> Self {
        self.provider = Box::new(provider);
        self
    }

    /// Replaces the tool registry.
    pub fn with_tools(mut self, tools: ToolRegistry) -> Self {
        self.tools = tools;
        self
    }

    /// Registers one tool into the builder's registry.
    pub fn with_tool(mut self, tool: impl Tool + 'static) -> Self {
        self.tools.register(tool);
        self
    }

    /// Sets the system prompt. Empty means no system message is assembled.
    pub fn with_system_prompt(mut self, system_prompt: impl Into<String>) -> Self {
        self.system_prompt = system_prompt.into();
        self
    }

    /// Replaces the default Memory.
    pub fn with_memory(mut self, memory: impl Memory + 'static) -> Self {
        self.memory = Box::new(memory);
        self
    }

    /// Replaces the agent config.
    pub fn with_config(mut self, config: AgentConfig) -> Self {
        self.config = config;
        self
    }

    /// Enables structured output for [`Agent::run`](crate::agent::Agent::run).
    #[cfg(feature = "structured")]
    pub fn with_structured_output(mut self, schema: serde_json::Value) -> Self {
        self.config.options.structured = Some(schema);
        self
    }

    /// Attaches shared state for tool calls.
    pub fn with_state(mut self, state: SharedState) -> Self {
        self.state = state;
        self
    }

    /// Attaches an observation channel.
    pub fn with_event_channel(mut self, channel: impl EventChannel + 'static) -> Self {
        self.events = Some(Arc::new(channel));
        self
    }

    /// Replaces the tool-round executor.
    pub fn with_tool_round_executor(mut self, executor: impl ToolRoundExecutor + 'static) -> Self {
        self.executor = Box::new(executor);
        self
    }

    /// Builds the agent.
    pub fn build(self) -> ReActAgent {
        ReActAgent {
            provider: self.provider,
            memory: self.memory,
            registry: self.tools,
            system_prompt: self.system_prompt,
            config: self.config,
            state: self.state,
            events: self.events,
            executor: self.executor,
            kernel_state: None,
        }
    }
}

impl fmt::Debug for ReActAgentBuilder {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        let mut debug = f.debug_struct("ReActAgentBuilder");
        debug
            .field("provider", &"Box<dyn Provider>")
            .field("memory", &"Box<dyn Memory>")
            .field("tools", &self.tools)
            .field("system_prompt", &self.system_prompt)
            .field("config", &self.config)
            .field("state", &self.state)
            .field(
                "events",
                &match &self.events {
                    Some(_) => "Some<dyn EventChannel>",
                    None => "None",
                },
            );
        debug.finish()
    }
}

impl ReActAgent {
    /// Starts a [`ReActAgentBuilder`] with a provider.
    pub fn builder(provider: impl Provider + 'static) -> ReActAgentBuilder {
        ReActAgentBuilder::new(provider)
    }

    /// Simple construction: three required parameters — provider
    /// ([`Provider`]) + tools ([`ToolRegistry`]) + system_prompt (empty
    /// string = no system prompt); `run` returns the model's answer text
    /// (consistent with [`Agent::run`]).
    ///
    /// Memory defaults to a bounded window
    /// ([`WindowMemory`](crate::memory::WindowMemory), 128k token budget,
    /// long conversations auto-trim the oldest rounds); when you need
    /// something custom (full history / persistence, etc.), chain
    /// [`with_memory`](ReActAgent::with_memory) to replace it; other
    /// optional behaviors (such as the max tool rounds) go through
    /// [`with_config`](ReActAgent::with_config).
    ///
    /// Structured output: typed output goes straight to
    /// [`run_typed`](ReActAgent::run_typed) (this run generates the Schema
    /// from the target type, no constructor configuration needed); for
    /// hand-written Schemas use
    /// [`with_structured_output`](ReActAgent::with_structured_output).
    ///
    /// For simpler assembly see the macro
    /// [`react_agent!`](crate::react_agent).
    pub fn new(
        provider: impl Provider + 'static,
        tools: ToolRegistry,
        system_prompt: impl Into<String>,
    ) -> Self {
        Self {
            provider: Box::new(provider),
            memory: Box::new(WindowMemory::new(DEFAULT_MEMORY_TOKENS)),
            registry: tools,
            system_prompt: system_prompt.into(),
            config: AgentConfig::default(),
            state: SharedState::default(),
            events: None,
            executor: Box::new(SerialToolRoundExecutor),
            kernel_state: None,
        }
    }

    /// Constructs a provider-free ReAct kernel for use with an outer
    /// `HarnessRuntime`.
    ///
    /// The runtime owns provider execution; this kernel only assembles
    /// model requests, runs immediate tools, requests governed effects, and
    /// consumes observations.
    pub fn kernel(tools: ToolRegistry, system_prompt: impl Into<String>) -> Self {
        Self::new(FakeProvider::new([]), tools, system_prompt)
    }

    /// Replace the default Memory (default
    /// [`WindowMemory`](crate::memory::WindowMemory), 128k token budget,
    /// long conversations auto-trim the oldest rounds).
    ///
    /// The Agent owns the Memory: `Memory::record` needs `&mut self`, which
    /// a shared (Arc) form can't write through; sharing the same
    /// conversation history across Agents is a Workflow-orchestration
    /// concern, out of this method's scope.
    ///
    /// # Examples
    ///
    /// Replace it with [`InMemoryMemory`](crate::memory::InMemoryMemory),
    /// which keeps all history verbatim:
    ///
    /// ```
    /// # extern crate molo_agent as molo;
    /// # #[tokio::main]
    /// # async fn main() {
    /// use molo::agent::{Agent, ReActAgent};
    /// use molo::memory::InMemoryMemory;
    /// use molo::provider::{FakeProvider, FakeReply};
    /// use molo::tool::ToolRegistry;
    ///
    /// let mut agent = ReActAgent::new(
    ///     FakeProvider::new([FakeReply::Text("Hello".into())]),
    ///     ToolRegistry::new(),
    ///     "",
    /// )
    /// .with_memory(InMemoryMemory::default());
    ///
    /// assert_eq!(agent.run("Are you there").await.unwrap(), "Hello");
    /// # }
    /// ```
    pub fn with_memory(mut self, memory: impl Memory + 'static) -> Self {
        self.memory = Box::new(memory);
        self
    }

    /// Replace the default config (default [`AgentConfig::default`]): the
    /// `max_tool_rounds` round limit and the `options` model parameters
    /// (temperature / max_tokens / extra parameters).
    ///
    /// See [`AgentConfig`](crate::agent::AgentConfig) for how to write a
    /// config.
    pub fn with_config(mut self, config: AgentConfig) -> Self {
        self.config = config;
        self
    }

    /// Enable structured output: the final answer must be JSON conforming
    /// to this **JSON Schema document**.
    ///
    /// `schema` is usually the serialized `RootSchema` produced by
    /// `schemars::schema_for!(T)` (the same pipeline as tool parameter
    /// schemas); it can also be hand-written. Semantics:
    ///
    /// - Compatible endpoints constrain the model as best effort via
    ///   `response_format` once they receive it;
    /// - **Framework-side fallback validation**: when the final answer
    ///   doesn't conform, the validation error is fed back to the model for
    ///   retry (independent budget
    ///   [`AgentConfig::max_structured_retries`]; exceeding it fails);
    /// - Once validation passes, the JSON text is returned as-is as the
    ///   answer.
    ///
    /// Typed output goes through [`run_typed`](ReActAgent::run_typed): the
    /// schema is auto-generated from the target type and deserialized,
    /// ignoring the hand-written schema set here — the two paths don't
    /// combine (the hand-written schema serves `Agent::run`'s text form;
    /// `run_typed` is always "the type is the schema").
    ///
    /// # Examples
    ///
    /// ```rust
    /// # extern crate molo_agent as molo;
    /// # #[tokio::main]
    /// # async fn main() -> Result<(), molo::AgentError> {
    /// use molo::{react_agent, Agent, FakeProvider, FakeReply};
    /// use serde_json::json;
    ///
    /// let mut agent = react_agent!(
    ///     FakeProvider::new([FakeReply::Text(r#"{"city":"Beijing"}"#.into())]),
    ///     "You are a structured-output assistant",
    /// )
    /// .with_structured_output(json!({
    ///     "type": "object",
    ///     "properties": { "city": { "type": "string" } },
    ///     "required": ["city"],
    /// }));
    ///
    /// let answer = agent.run("How's the weather in Beijing").await?;
    /// assert_eq!(answer, r#"{"city":"Beijing"}"#);
    /// # Ok(())
    /// # }
    /// ```
    #[cfg(feature = "structured")]
    pub fn with_structured_output(mut self, schema: serde_json::Value) -> Self {
        self.config.options.structured = Some(schema);
        self
    }

    /// Attach shared state (empty by default). The application reads and
    /// writes via [`state`](ReActAgent::state) across runs, and tools
    /// access the same instance through the `state` parameter at call time.
    pub fn with_state(mut self, state: SharedState) -> Self {
        self.state = state;
        self
    }

    /// Attach an observation channel (not attached by default, zero cost):
    /// the loop pushes process events into the pipeline, and the host side
    /// `subscribe`s.
    ///
    /// The pipeline is **long-lived**: attached once, events from many runs
    /// flow into the same pipeline (the Agent is the event source); each
    /// run is one segment of the pipeline, delimited by
    /// [`ReActEvent::RunStarted`] / [`ReActEvent::RunEnded`]; both the
    /// streaming and non-streaming paths publish. See
    /// [`ReActEvent`](crate::agent::ReActEvent) for the event set.
    pub fn with_event_channel(mut self, channel: impl EventChannel + 'static) -> Self {
        self.events = Some(Arc::new(channel));
        self
    }

    /// Replace the default tool-round executor (default:
    /// [`SerialToolRoundExecutor`] — the round's calls run one by one, in
    /// request order) with a custom round policy: order / concurrency /
    /// approval gates.
    ///
    /// The executor decides *how* the round's tool calls are executed; the
    /// per-call mechanics (tool span, ToolStarted / ToolCompleted events,
    /// registry dispatch) stay framework-owned via
    /// [`ToolRoundCtx::run`]. Each outcome is recorded (and, on the
    /// streaming path, dispatched) by the agent as the executor's stream
    /// yields it. The round stays atomic by contract — the executor owns
    /// that responsibility.
    ///
    /// # Examples
    ///
    /// A round policy that runs the round's calls concurrently; outcomes
    /// flow back as each completes:
    ///
    /// ```
    /// # extern crate molo_agent as molo;
    /// # #[tokio::main]
    /// # async fn main() -> Result<(), molo::AgentError> {
    /// use molo::agent::{
    ///     Agent, ReActAgent, ToolCallOutcome, ToolRoundCtx, ToolRoundExecutor,
    /// };
    /// use molo::message::ToolCall;
    /// use molo::provider::{FakeProvider, FakeReply};
    /// use molo::tool::ToolRegistry;
    /// use futures::stream::{BoxStream, FuturesUnordered};
    /// use futures::StreamExt;
    ///
    /// #[derive(Default)]
    /// struct ParallelToolRoundExecutor;
    ///
    /// #[molo::async_trait]
    /// impl ToolRoundExecutor for ParallelToolRoundExecutor {
    ///     async fn execute_round<'a>(
    ///         &'a mut self,
    ///         ctx: ToolRoundCtx<'a>,
    ///         calls: Vec<ToolCall>,
    ///     ) -> BoxStream<'a, ToolCallOutcome> {
    ///         // The run futures borrow ctx, so they materialize inside the
    ///         // generator (which owns ctx), not across the return.
    ///         Box::pin(async_stream::stream! {
    ///             let mut tasks = FuturesUnordered::new();
    ///             for call in calls {
    ///                 tasks.push(ctx.run(call));
    ///             }
    ///             while let Some(outcome) = tasks.next().await {
    ///                 yield outcome;
    ///             }
    ///         })
    ///     }
    /// }
    ///
    /// let mut agent = ReActAgent::new(
    ///     FakeProvider::new([FakeReply::Text("Hello".into())]),
    ///     ToolRegistry::new(),
    ///     "",
    /// )
    /// .with_tool_round_executor(ParallelToolRoundExecutor);
    ///
    /// assert_eq!(agent.run("Are you there").await?, "Hello");
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_tool_round_executor(mut self, executor: impl ToolRoundExecutor + 'static) -> Self {
        self.executor = Box::new(executor);
        self
    }

    /// Publish an event (no-op when no channel is attached). Takes a
    /// **construction closure** rather than the event itself: with no
    /// observation channel attached (the default), construction cost is
    /// zero — the hot path (every Delta chunk) skips an unconditional heap
    /// allocation and copy.
    fn publish<E: AgentEvent + 'static>(&self, make_event: impl FnOnce() -> Arc<E>) {
        if let Some(pipe) = &self.events {
            pipe.publish(make_event());
        }
    }

    /// Non-streaming round loop (same semantics as the streaming path).
    async fn run_rounds_with_context(
        &mut self,
        context: &RunContext,
        run_id: &str,
        counters: &mut RunCounters,
        options: &ModelOptions,
        schema: Option<&serde_json::Value>,
    ) -> Result<FinalAnswer, AgentError> {
        let schemas = self.registry.schemas();
        // Structured validator: built when this run has a schema; the retry
        // budget lives in the component (the count accumulates across
        // rounds and doesn't consume the tool-round budget).
        #[cfg(feature = "structured")]
        let mut validator = schema.map(|schema| {
            StructuredValidator::new(schema.clone(), self.config.max_structured_retries)
        });
        #[cfg(not(feature = "structured"))]
        let _ = schema;
        // Tool-round counter: the limit check only counts rounds where the
        // model requested tools — conversation rounds and structured
        // validation-retry rounds don't consume the tool-round budget
        // (structured retries have their own `max_structured_retries`
        // budget); errors no longer report "tool round limit exceeded"
        // because of structured retries.
        let mut tool_rounds = 0usize;
        loop {
            if tool_rounds >= self.config.max_tool_rounds {
                return Err(AgentError::TooManyToolRounds(self.config.max_tool_rounds));
            }
            counters.rounds += 1;
            // Between-round check: cancellation takes effect immediately —
            // this round hasn't started a conversation, so the memory is
            // clean (nothing from this round beyond the user message).
            check_run_context(context)?;

            // Provider-call span for this round: duration comes from span
            // timing automatically; usage goes out through both channels
            // (the event stream's RunSummary as usual, and observability
            // records it on the span fields at wrap-up); the span is created
            // while the run is on the stack (the outer run block is
            // instrumented), so the hierarchy is established by the ambient
            // parent automatically. The round body runs directly inside the
            // run scope.
            // The block returns Option<String> = whether this round already
            // answered directly; internal errors propagate via `?` (the two
            // Ok arms pin the error type with turbofish — inference would
            // otherwise fail under multiple From impls).
            let answer: Option<FinalAnswer> = async {
                let llm_span = span_llm(run_id, counters.rounds);
                // Cancelled mid-conversation: run_until_cancelled drops the
                // in-flight request and returns immediately (nothing is
                // recorded for this round). When cancellation races with
                // completion, completion wins (primitive semantics).
                // Provider errors are recorded on the llm span here
                // (observability pinpoints the failed call); cancellations
                // are not (cancellation is a run-level outcome, covered by
                // the run span and the RunEnded event).
                // Model parameters on the request: the schema passed for
                // this run (typed path) takes precedence over the
                // hand-written schema in the config — endpoint-side
                // constraint and framework-side validation use the same
                // one.
                let model_request_id = format!("{run_id}-model-{}", counters.rounds);
                let provider_context =
                    ProviderRequestContext::from_run_context(model_request_id, context);
                let chat = self.provider.chat_with_context(
                    ChatRequest {
                        messages: self.assemble_messages(self.memory.context().await?),
                        tools: schemas.clone(),
                        options: options.clone(),
                    },
                    &provider_context,
                );
                let response =
                    match run_until_context(context, instrument(chat, llm_span.clone())).await {
                        Ok(Ok(response)) => response,
                        Ok(Err(e)) => {
                            #[cfg(feature = "tracing")]
                            llm_span.record("error", e.to_string());
                            return Err(AgentError::Provider(e));
                        }
                        Err(e) => return Err(e),
                    };
                // Usage may be absent (the endpoint didn't report it);
                // reported parts are summed and omitted turns are tracked so
                // the summary's sum is not misread as exact.
                if let Some(usage) = response.usage {
                    #[cfg(feature = "tracing")]
                    {
                        llm_span.record("usage.prompt_tokens", usage.prompt_tokens);
                        llm_span.record("usage.completion_tokens", usage.completion_tokens);
                    }
                    counters.usage_total += usage;
                } else {
                    counters.usage_omitted = true;
                }
                let finish_reason = response.finish_reason.clone();

                // Per the Provider contract, this round's reply is exactly
                // one Assistant message; text, reasoning, and tool requests
                // stay in the same message (wire constraint: multiple tool
                // requests in the same round are not split apart).
                let Message::Assistant {
                    content,
                    reasoning,
                    tool_calls,
                } = response.message
                else {
                    // Defensive handling: a custom Provider may return a
                    // non-Assistant message; the library boundary responds
                    // to expected inputs with an error (same rigor as tool
                    // panic catching, see the registry).
                    return Err(AgentError::Provider(ProviderError::Protocol {
                        message: "provider returned a non-assistant message".into(),
                    }));
                };
                // Round text limit: a malicious endpoint emitting unbounded
                // text in one round could blow up memory (the Provider layer
                // only limits single lines; this is the per-round backstop,
                // sharing the same constant as the streaming path; reasoning
                // counts toward it too).
                if content.len() + reasoning.as_deref().map_or(0, str::len) > MAX_ROUND_TEXT {
                    return Err(AgentError::Provider(ProviderError::ResponseTooLarge {
                        limit_bytes: MAX_ROUND_TEXT,
                    }));
                }

                // Empty replies are not recorded: Assistant messages with no
                // content, no reasoning, and no tool requests don't go into
                // Memory; reasoning and tool requests are saved along with
                // the content.
                let final_message = Message::Assistant {
                    content: content.clone(),
                    reasoning: reasoning.clone(),
                    tool_calls: Vec::new(),
                };

                if !content.is_empty() || reasoning.is_some() || !tool_calls.is_empty() {
                    self.memory
                        .record(Message::Assistant {
                            content: content.clone(),
                            reasoning,
                            tool_calls: tool_calls.clone(),
                        })
                        .await?;
                }

                if tool_calls.is_empty() {
                    // The model answered directly. Structured output: the
                    // component validates the final answer — on failure,
                    // feed it back to the model for retry (budget is built
                    // into the component, decoupled from the tool-round
                    // limit); exceeding it fails the run.
                    #[cfg(feature = "structured")]
                    {
                        if let Some(validator) = &mut validator {
                            match validator.validate(&content) {
                                StructuredOutcome::Passed => {}
                                StructuredOutcome::Retry { message } => {
                                    self.memory.record(message).await?;
                                    return Ok::<Option<FinalAnswer>, AgentError>(None);
                                }
                                StructuredOutcome::Exhausted { max_retries } => {
                                    return Err(AgentError::StructuredRetriesExhausted(
                                        max_retries,
                                    ));
                                }
                            }
                        }
                    }
                    return Ok::<Option<FinalAnswer>, AgentError>(Some(FinalAnswer {
                        answer: content,
                        final_message,
                        finish_reason: Some(finish_reason),
                    }));
                }

                counters.tool_calls_total += tool_calls.len();
                tool_rounds += 1;
                // Tool round: executed atomically and never interrupted —
                // tools are user code with no cancellation interface, and a
                // partial execution risks side effects; interrupting
                // mid-way would also leave the Assistant recorded without
                // its ToolResult, breaking the next round's message
                // sequence. Cancellation takes effect naturally after the
                // tool round ends, before the next round's conversation.
                // The round policy (default: one by one, in request order)
                // is injected via the ToolRoundExecutor; each outcome is
                // recorded as it arrives.
                let ctx = ToolRoundCtx {
                    context,
                    round: counters.rounds,
                    registry: &self.registry,
                    state: &self.state,
                    events: &self.events,
                };
                let mut outcomes = self.executor.execute_round(ctx, tool_calls).await;
                while let Some(outcome) = outcomes.next().await {
                    if let Some(effect) = outcome.effect_request() {
                        return Err(AgentError::EffectRequiresHarness(format!(
                            "{} ({})",
                            effect.description, effect.id
                        )));
                    }
                    // Protected results (skill bodies, etc.) are recorded
                    // via record_protected, exempt from window trimming;
                    // when recording fails, the error text is recorded as a
                    // fallback (memory integrity, see record_tool_result).
                    record_tool_result(&mut self.memory, &outcome).await?;
                }
                Ok::<Option<FinalAnswer>, AgentError>(None)
            }
            .await?;
            if let Some(answer) = answer {
                return Ok(answer);
            }
        }
    }

    /// Assemble the full prompt for each request: System first, followed by
    /// Memory's conversation history.
    ///
    /// Assembly happens when the conversation starts and is not written to
    /// Memory — the system prompt is static configuration while Memory is
    /// the dynamic conversation record; the two are managed separately. The
    /// skill part reads the registry fresh on every request: hot-swapped
    /// additions/removals take effect on the next request.
    fn assemble_messages(&self, context: Vec<Message>) -> Vec<Message> {
        let mut messages = Vec::with_capacity(context.len() + 1);
        let system = self.assemble_system_prompt();
        if !system.is_empty() {
            messages.push(Message::system(&system));
        }
        messages.extend(context);
        messages
    }

    /// Assemble the system prompt; an empty result = no system prompt.
    fn assemble_system_prompt(&self) -> String {
        self.system_prompt.clone()
    }
}

/// What a round policy gets handed to execute one tool round.
///
/// The agent constructs the context and passes it by value to your
/// [`ToolRoundExecutor::execute_round`]; you only read it, and execute
/// calls through [`run`](ToolRoundCtx::run) — the framework-owned per-call
/// mechanics (span, events, registry dispatch) stay out of your hands, so
/// a custom policy reorders / parallelizes / gates calls without
/// reimplementing any of that. See [`ToolRoundExecutor`] for a working
/// example.
pub struct ToolRoundCtx<'a> {
    /// Run context (the run id is the correlation key carried by spans and
    /// events).
    pub context: &'a RunContext,
    /// The current round number (carried into the tool span).
    pub round: usize,
    /// The tool registry (call dispatch + protected declarations).
    pub registry: &'a ToolRegistry,
    /// Shared state, injected into every tool call.
    pub state: &'a SharedState,
    /// The observation channel (None when not attached).
    pub events: &'a Option<Arc<dyn EventChannel>>,
}

impl fmt::Debug for ToolRoundCtx<'_> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        // The observation channel is opaque (no Debug on the trait
        // object); finish_non_exhaustive marks the omission.
        f.debug_struct("ToolRoundCtx")
            .field("run_id", &self.context.run_id)
            .field("round", &self.round)
            .field("registry", &self.registry)
            .field("state", &self.state)
            .finish_non_exhaustive()
    }
}

impl ToolRoundCtx<'_> {
    /// Dispatch one tool call with the framework's standard mechanics and
    /// return its outcome.
    ///
    /// Call this for each call your policy decides to run: the call is
    /// dispatched to the registry (tool span + ToolStarted / ToolCompleted
    /// events included), and the outcome carries immediate output text, a
    /// registry error text, or an effect request. Tool failures are **not**
    /// errors here; they ride along in [`content`](ToolCallOutcome::content)
    /// and are fed back to the model.
    ///
    /// Takes `&self` rather than `&mut self`: the mechanics are all shared
    /// borrows, so a policy can hold multiple `run` futures concurrently
    /// (see the parallel example on
    /// [`with_tool_round_executor`](ReActAgent::with_tool_round_executor)).
    pub async fn run(&self, call: ToolCall) -> ToolCallOutcome {
        // Tool-call span: carries duration; records error on failure. The
        // span is created while the run is on the stack (ambient parent on
        // both paths, see the span-construction comments), so the hierarchy
        // is correct automatically.
        let tool_span = span_tool(&self.context.run_id, self.round, &call.name);
        // Tool-started event (carries id / arguments; subscribers pair it
        // with ToolCompleted by id).
        self.publish(|| {
            Arc::new(ReActEvent::ToolStarted {
                id: call.id.clone(),
                name: call.name.clone(),
                arguments: call.arguments.clone(),
            })
        });
        // Execute; the Result status rides along with the event (Ok/Err is
        // classified by the registry), and the text (Ok result / Err's
        // Display) is recorded and fed back to the model.
        let call_future = self.registry.call(&call, self.context, self.state);
        let result = instrument(call_future, tool_span.clone()).await;
        #[cfg(feature = "tracing")]
        if let Err(e) = &result {
            tool_span.record("error", e.to_string());
        }
        let outcome = match &result {
            Ok(ToolResult::Output(output)) => ToolCallOutcome::output(call.clone(), output.clone()),
            Ok(ToolResult::Effect(request)) => {
                ToolCallOutcome::effect(call.clone(), request.clone())
            }
            Ok(other) => ToolCallOutcome::text(call.clone(), other.to_string()),
            Err(e) => ToolCallOutcome::text(call.clone(), e.to_string()),
        };
        let publish_tool_completed = {
            let id = call.id.clone();
            let name = call.name.clone();
            move || Arc::new(ReActEvent::ToolCompleted { id, name, result })
        };
        self.publish(publish_tool_completed);
        outcome
    }

    /// Publish an event (no-op when no channel is attached). Takes a
    /// construction closure rather than the event itself: with no
    /// observation channel attached (the default), construction cost is
    /// zero.
    fn publish<E: AgentEvent + 'static>(&self, make_event: impl FnOnce() -> Arc<E>) {
        if let Some(pipe) = self.events {
            pipe.publish(make_event());
        }
    }
}

/// Round-level tool-call policy: decides how one round's tool calls are
/// executed — order, concurrency, approval gates. Inject it with
/// [`with_tool_round_executor`](ReActAgent::with_tool_round_executor); the
/// default [`SerialToolRoundExecutor`] runs calls one by one in request
/// order, the classic serial loop with identical behavior.
///
/// The executor owns *how* calls run; the framework owns *that* they run
/// correctly: every call goes through [`ToolRoundCtx::run`], which carries
/// the standard mechanics (tool span, ToolStarted / ToolCompleted events,
/// registry dispatch, protected-result declaration) — a custom executor
/// can reorder, parallelize, or gate calls without reimplementing any of
/// that.
///
/// # Contract
///
/// - **Atomicity**: tool rounds are executed atomically and never
///   interrupted — tools are user code with no cancellation interface, and
///   a partial execution risks side effects; cancellation takes effect
///   after the round ends, before the next round's conversation. An
///   executor that interrupts mid-round accepts that risk itself.
/// - **Outcomes stream back as they complete**: the agent records (and,
///   on the streaming path, dispatches) each outcome as the stream yields
///   it. Dropping the stream early aborts the rest of the round —
///   already-produced outcomes stay recorded, and the model adapts to the
///   partial results.
/// - **Decisions feed back as text**: to deny a call (approval gate, rate
///   limit, ...), yield a synthetic [`ToolCallOutcome`] carrying the
///   decision text — it reaches the model through the normal ToolResult
///   channel without the tool ever running.
///
/// # Examples
///
/// An approval gate: calls to "dangerous" are denied without running,
/// everything else executes as usual.
///
/// ```
/// # extern crate molo_agent as molo;
/// # #[tokio::main]
/// # async fn main() -> Result<(), molo::AgentError> {
/// use molo::agent::{
///     Agent, ReActAgent, ToolCallOutcome, ToolRoundCtx, ToolRoundExecutor,
/// };
/// use molo::message::ToolCall;
/// use molo::provider::{FakeProvider, FakeReply};
/// use molo::tool::ToolRegistry;
/// use futures::stream::BoxStream;
///
/// #[derive(Default)]
/// struct ApprovalGateToolRoundExecutor;
///
/// #[molo::async_trait]
/// impl ToolRoundExecutor for ApprovalGateToolRoundExecutor {
///     async fn execute_round<'a>(
///         &'a mut self,
///         ctx: ToolRoundCtx<'a>,
///         calls: Vec<ToolCall>,
///     ) -> BoxStream<'a, ToolCallOutcome> {
///         Box::pin(async_stream::stream! {
///             for call in calls {
///                 if call.name == "dangerous" {
///                     yield ToolCallOutcome::text(call, "denied by approval gate");
///                 } else {
///                     yield ctx.run(call).await;
///                 }
///             }
///         })
///     }
/// }
///
/// let mut agent = ReActAgent::new(
///     FakeProvider::new([FakeReply::Text("Hello".into())]),
///     ToolRegistry::new(),
///     "",
/// )
/// .with_tool_round_executor(ApprovalGateToolRoundExecutor);
///
/// assert_eq!(agent.run("Are you there").await?, "Hello");
/// # Ok(())
/// # }
/// ```
#[async_trait::async_trait]
pub trait ToolRoundExecutor: Send + Sync {
    /// Execute one tool round: decide the policy over `calls` — which to
    /// run, in what order, with what concurrency — and yield each
    /// [`ToolCallOutcome`] as it completes.
    ///
    /// Use [`ToolRoundCtx::run`] to execute a call; its futures borrow
    /// `ctx`, so when the policy runs calls concurrently, the futures must
    /// materialize inside the returned stream (which owns `ctx`) rather
    /// than across this method's return — see the parallel example on
    /// [`with_tool_round_executor`](ReActAgent::with_tool_round_executor).
    ///
    /// # Errors
    ///
    /// None at the round level: a tool's failure is not an error here — it
    /// rides along in the outcome's text, which the agent records and
    /// feeds back to the model.
    async fn execute_round<'a>(
        &'a mut self,
        ctx: ToolRoundCtx<'a>,
        calls: Vec<ToolCall>,
    ) -> BoxStream<'a, ToolCallOutcome>;
}

/// The default round policy: execute the round's calls one by one, in
/// request order — the classic serial loop, identical behavior.
///
/// This is the executor used when none is injected, so you usually don't
/// need to name it; reach for it when replacing the default with a custom
/// policy and keeping the serial behavior handy as a fallback:
///
/// ```
/// # extern crate molo_agent as molo;
/// use molo::agent::SerialToolRoundExecutor;
/// use molo::agent::ToolRoundExecutor;
///
/// let serial: Box<dyn ToolRoundExecutor> = Box::new(SerialToolRoundExecutor::default());
/// ```
#[derive(Debug, Default, Clone, Copy)]
pub struct SerialToolRoundExecutor;

#[async_trait::async_trait]
impl ToolRoundExecutor for SerialToolRoundExecutor {
    async fn execute_round<'a>(
        &'a mut self,
        ctx: ToolRoundCtx<'a>,
        calls: Vec<ToolCall>,
    ) -> BoxStream<'a, ToolCallOutcome> {
        Box::pin(async_stream::stream! {
            for call in calls {
                yield ctx.run(call).await;
            }
        })
    }
}

/// The outcome of one tool call dispatch: the call info, the text fed back
/// to the model when immediately available, recording metadata, and an
/// optional effect request.
///
/// `run` produces outcomes; policies can also construct **synthetic**
/// outcomes to feed a decision back to the model without running the tool
/// (see the approval-gate example on [`ToolRoundExecutor`]) — the agent
/// records the text through the normal ToolResult channel:
///
/// ```
/// # extern crate molo_agent as molo;
/// use molo::agent::ToolCallOutcome;
/// use molo::message::ToolCall;
///
/// let denied = ToolCallOutcome {
///     call: ToolCall {
///         id: "call_1".into(),
///         name: "dangerous".into(),
///         arguments: "{}".into(),
///     },
///     content: "denied by approval gate".into(),
///     memory_policy: molo::ToolMemoryPolicy::Normal,
///     effect: None,
/// };
/// ```
#[derive(Debug, Clone, PartialEq)]
pub struct ToolCallOutcome {
    /// The call as-is (with id / name / arguments, used for event pairing
    /// and locating the record).
    pub call: ToolCall,
    /// The text fed back (Ok result / Err's Display, visible to the model).
    pub content: String,
    /// Memory policy for the model-visible content.
    pub memory_policy: ToolMemoryPolicy,
    /// Effect request produced by the tool, when the tool requested a
    /// governed side effect instead of immediate output.
    pub effect: Option<EffectRequest>,
}

impl ToolCallOutcome {
    /// Constructs an immediate text outcome.
    pub fn text(call: ToolCall, content: impl Into<String>) -> Self {
        Self::output(call, ToolOutput::text(content))
    }

    /// Constructs an immediate output outcome.
    pub fn output(call: ToolCall, output: ToolOutput) -> Self {
        Self {
            call,
            content: output.content,
            memory_policy: output.memory_policy,
            effect: None,
        }
    }

    /// Constructs an effect outcome.
    pub fn effect(call: ToolCall, request: EffectRequest) -> Self {
        Self {
            call,
            content: String::new(),
            memory_policy: ToolMemoryPolicy::Normal,
            effect: Some(request),
        }
    }

    /// Whether this outcome requests a governed effect.
    pub fn effect_request(&self) -> Option<&EffectRequest> {
        self.effect.as_ref()
    }
}

#[cfg(feature = "structured")]
impl ReActAgent {
    /// Typed run: same semantics as [`Agent::run`](Agent::run) (records
    /// input, drives the loop), but the final answer is deserialized into
    /// the type parameter `U` after validation — this run auto-generates a
    /// JSON Schema from `U`
    /// ([`schemars`](https://docs.rs/schemars)-derived, the same pipeline
    /// as tool parameter schemas), with no constructor configuration
    /// needed.
    ///
    /// Relation to [`Agent::run`](Agent::run): run returns the model's
    /// answer text (free text, or the JSON text of a hand-written Schema
    /// from
    /// [`with_structured_output`](ReActAgent::with_structured_output));
    /// run_typed generates the Schema from the type and returns the type.
    /// The return type is declared in the let annotation (see `# Examples`);
    /// no turbofish needed.
    ///
    /// # Errors
    ///
    /// - [`AgentError::StructuredRetriesExhausted`][]: validation failures
    ///   are fed back to the model for retry, but the
    ///   [`AgentConfig::max_structured_retries`] budget is exhausted
    ///   without success;
    /// - [`AgentError::StructuredParse`][]: validation passed but
    ///   deserialization failed (when a derived schema customized with
    ///   `#[schemars(...)]` disagrees with the serde representation);
    /// - otherwise the same as [`Agent::run`](Agent::run) (Memory /
    ///   Provider / round limit / cancellation).
    ///
    /// # Examples
    ///
    /// ```
    /// # extern crate molo_agent as molo;
    /// # #[tokio::main]
    /// # async fn main() -> Result<(), molo::AgentError> {
    /// use molo::{FakeProvider, FakeReply, ReActAgent};
    /// use schemars::JsonSchema;
    /// use serde::Deserialize;
    ///
    /// #[derive(Deserialize, JsonSchema)]
    /// struct Weather {
    ///     city: String,
    /// }
    ///
    /// let mut agent = ReActAgent::new(
    ///     FakeProvider::new([FakeReply::Text(r#"{"city":"Beijing"}"#.into())]),
    ///     molo::tool::ToolRegistry::new(),
    ///     "You are a weather assistant",
    /// );
    /// let weather: Weather = agent.run_typed("How's the weather in Beijing").await?;
    /// assert_eq!(weather.city, "Beijing");
    /// # Ok(())
    /// # }
    /// ```
    pub async fn run_typed<U>(&mut self, input: &str) -> Result<U, AgentError>
    where
        U: DeserializeOwned + JsonSchema + Send + Sync,
    {
        // Implementation lives in [`TypedAgent`]; the inherent delegate
        // lets call sites avoid importing the trait.
        TypedAgent::run_typed(self, input).await
    }
}

#[cfg(feature = "structured")]
#[async_trait::async_trait]
impl TypedAgent for ReActAgent {
    async fn run_typed_request_with_context<U>(
        &mut self,
        request: RunRequest,
        context: RunContext,
    ) -> Result<TypedRunOutput<U>, AgentError>
    where
        U: DeserializeOwned + JsonSchema + Send + Sync,
    {
        let schema = serde_json::to_value(schemars::schema_for!(U))
            .expect("schemars-generated schema always serializes (pure JSON value structure)");
        let output = self
            .run_request_inner(request, context, Some(&schema))
            .await?;
        let value = serde_json::from_str(&output.answer)
            .map_err(|e| AgentError::StructuredParse(e.to_string()))?;
        Ok(TypedRunOutput { value, output })
    }
}

/// Cumulative counters for a single run (rounds / tools / usage), for the
/// `RunEnded` summary at wrap-up; the non-streaming path accumulates
/// through a mutable reference (the streaming path accumulates in local
/// variables inside the generator, same semantics).
/// Structured validation-retry counts don't live here — they're held by the
/// [`StructuredValidator`] component.
#[derive(Default)]
struct RunCounters {
    /// Number of conversation rounds (same semantics as
    /// `RunSummary.rounds`).
    rounds: usize,
    /// Total number of tool executions.
    tool_calls_total: usize,
    /// Sum of token usage across rounds (reported parts only).
    usage_total: Usage,
    /// `true` = at least one round's provider did not report usage, so
    /// [`usage_total`](Self::usage_total) is a lower bound.
    usage_omitted: bool,
}

struct FinalAnswer {
    answer: String,
    final_message: Message,
    finish_reason: Option<FinishReason>,
}

struct RunExecution {
    answer: String,
    final_message: Message,
    summary: RunSummary,
    artifacts: Vec<Artifact>,
    metadata: RunMetadata,
}

struct ReActKernelState {
    run_id: String,
    started_at: Instant,
    provider_model: Option<String>,
    counters: RunCounters,
    options: ModelOptions,
    schemas: Vec<crate::tool::ToolSchema>,
    #[cfg(feature = "structured")]
    validator: Option<StructuredValidator>,
    tool_rounds: usize,
    pending_tools: VecDeque<ToolCall>,
    pending_tool_results: VecDeque<PendingToolResult>,
    next_model_request: u64,
}

#[derive(Debug)]
enum PendingToolResult {
    Outcome(ToolCallOutcome),
    Effect {
        effect_id: String,
        call: ToolCall,
        observation: Option<EffectObservation>,
    },
}

impl PendingToolResult {
    fn effect_id(&self) -> Option<&str> {
        match self {
            Self::Outcome(_) => None,
            Self::Effect { effect_id, .. } => Some(effect_id),
        }
    }
}

impl ReActKernelState {
    fn next_model_request(&mut self, messages: Vec<Message>) -> AgentAction {
        self.next_model_request += 1;
        AgentAction::RequestModel {
            request: ModelRequest::new(
                format!("{}-model-{}", self.run_id, self.next_model_request),
                ChatRequest {
                    messages,
                    tools: self.schemas.clone(),
                    options: self.options.clone(),
                },
            ),
        }
    }

    fn summary(&self, finish_reason: Option<FinishReason>) -> RunSummary {
        run_summary(
            &self.counters,
            finish_reason,
            self.started_at,
            self.provider_model.clone(),
        )
    }
}

impl fmt::Debug for ReActAgent {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        // Provider / Memory are trait objects and can't be Debug'd; print
        // the composition and behavior config.
        let mut debug = f.debug_struct("ReActAgent");
        debug
            .field("provider", &"Box<dyn Provider>")
            .field("memory", &"Box<dyn Memory>")
            .field("tools", &self.registry)
            .field("system_prompt", &self.system_prompt)
            .field("config", &self.config)
            .field("state", &self.state)
            .field(
                "events",
                &match &self.events {
                    Some(_) => "Some<dyn EventChannel>",
                    None => "None",
                },
            );
        debug.finish()
    }
}

#[async_trait::async_trait]
impl Agent for ReActAgent {
    async fn run_request_with_context(
        &mut self,
        request: RunRequest,
        context: RunContext,
    ) -> Result<RunOutput, AgentError> {
        self.run_request_inner(request, context, None).await
    }

    async fn run_stream_request_with_context<'a>(
        &'a mut self,
        request: RunRequest,
        context: RunContext,
    ) -> Result<BoxStream<'a, Result<MessageChunk, AgentError>>, AgentError> {
        self.run_stream_request_inner(request, context).await
    }
}

#[async_trait::async_trait]
impl AgentKernel for ReActAgent {
    async fn start(
        &mut self,
        request: RunRequest,
        context: &RunContext,
    ) -> Result<AgentAction, AgentError> {
        check_run_context(context)?;
        let run_id = context.run_id.clone();
        let input = request.input;
        self.memory.record(input.clone().into_message()).await?;
        self.publish(|| {
            Arc::new(ReActEvent::RunStarted {
                run_id: run_id.clone(),
                input,
            })
        });

        let options = request
            .options
            .clone()
            .unwrap_or_else(|| self.config.options.clone());
        #[cfg(feature = "structured")]
        let validator = options.structured.as_ref().map(|schema| {
            StructuredValidator::new(schema.clone(), self.config.max_structured_retries)
        });
        let mut state = ReActKernelState {
            run_id,
            started_at: Instant::now(),
            provider_model: self.provider.model().map(str::to_string),
            counters: RunCounters::default(),
            options,
            schemas: self.registry.schemas(),
            #[cfg(feature = "structured")]
            validator,
            tool_rounds: 0,
            pending_tools: VecDeque::new(),
            pending_tool_results: VecDeque::new(),
            next_model_request: 0,
        };
        state.counters.rounds += 1;
        let messages = self.assemble_messages(self.memory.context().await?);
        let action = state.next_model_request(messages);
        self.kernel_state = Some(state);
        Ok(action)
    }

    async fn observe(
        &mut self,
        observation: Observation,
        context: &RunContext,
    ) -> Result<AgentAction, AgentError> {
        check_run_context(context)?;
        match observation {
            Observation::Model(observation) => self.observe_model(observation, context).await,
            Observation::Effect(observation) => self.observe_effect(observation, context).await,
            Observation::Effects(observations) => self.observe_effects(observations, context).await,
            _ => Err(AgentError::InvalidStep("unsupported observation".into())),
        }
    }
}

impl ReActAgent {
    async fn observe_model(
        &mut self,
        observation: ModelObservation,
        context: &RunContext,
    ) -> Result<AgentAction, AgentError> {
        let mut state = self.kernel_state.take().ok_or_else(|| {
            AgentError::InvalidStep("model observation without active run".into())
        })?;
        if !state.pending_tool_results.is_empty() || !state.pending_tools.is_empty() {
            self.kernel_state = Some(state);
            return Err(AgentError::InvalidStep(
                "model observation received while tool calls are pending".into(),
            ));
        }

        let response = observation.response;
        if let Some(usage) = response.usage {
            state.counters.usage_total += usage;
        } else {
            state.counters.usage_omitted = true;
        }
        let finish_reason = response.finish_reason.clone();
        let Message::Assistant {
            content,
            reasoning,
            tool_calls,
        } = response.message
        else {
            self.kernel_state = Some(state);
            return Err(AgentError::Provider(ProviderError::Protocol {
                message: "provider returned a non-assistant message".into(),
            }));
        };
        if content.len() + reasoning.as_deref().map_or(0, str::len) > MAX_ROUND_TEXT {
            self.kernel_state = Some(state);
            return Err(AgentError::Provider(ProviderError::ResponseTooLarge {
                limit_bytes: MAX_ROUND_TEXT,
            }));
        }

        let final_message = Message::Assistant {
            content: content.clone(),
            reasoning: reasoning.clone(),
            tool_calls: Vec::new(),
        };
        if !content.is_empty() || reasoning.is_some() || !tool_calls.is_empty() {
            self.memory
                .record(Message::Assistant {
                    content: content.clone(),
                    reasoning,
                    tool_calls: tool_calls.clone(),
                })
                .await?;
        }

        if tool_calls.is_empty() {
            #[cfg(feature = "structured")]
            {
                if let Some(validator) = &mut state.validator {
                    match validator.validate(&content) {
                        StructuredOutcome::Passed => {}
                        StructuredOutcome::Retry { message } => {
                            self.memory.record(message).await?;
                            state.counters.rounds += 1;
                            let messages = self.assemble_messages(self.memory.context().await?);
                            let action = state.next_model_request(messages);
                            self.kernel_state = Some(state);
                            return Ok(action);
                        }
                        StructuredOutcome::Exhausted { max_retries } => {
                            self.kernel_state = None;
                            return Err(AgentError::StructuredRetriesExhausted(max_retries));
                        }
                    }
                }
            }
            let summary = state.summary(Some(finish_reason));
            let output = RunOutput {
                run_id: state.run_id.clone(),
                answer: content,
                summary: summary.clone(),
                final_message,
                artifacts: Vec::new(),
                metadata: RunMetadata::new(),
            };
            publish_ended(&self.events, summary, None);
            self.kernel_state = None;
            return Ok(AgentAction::Respond { output });
        }

        if state.tool_rounds >= self.config.max_tool_rounds {
            self.kernel_state = None;
            return Err(AgentError::TooManyToolRounds(self.config.max_tool_rounds));
        }
        state.tool_rounds += 1;
        state.counters.tool_calls_total += tool_calls.len();
        state.pending_tools = VecDeque::from(tool_calls);
        let action = self.process_kernel_pending_tools(state, context).await?;
        Ok(action)
    }

    async fn observe_effect(
        &mut self,
        observation: EffectObservation,
        context: &RunContext,
    ) -> Result<AgentAction, AgentError> {
        let mut state = self.kernel_state.take().ok_or_else(|| {
            AgentError::InvalidStep("effect observation without active run".into())
        })?;
        let pending_effect_count = state
            .pending_tool_results
            .iter()
            .filter(|result| result.effect_id().is_some())
            .count();
        if pending_effect_count > 1 {
            self.kernel_state = Some(state);
            return Err(AgentError::InvalidStep(
                "single effect observation received while a batch is pending".into(),
            ));
        }
        let Some(expected_effect_id) = state
            .pending_tool_results
            .iter()
            .find_map(PendingToolResult::effect_id)
            .map(str::to_string)
        else {
            self.kernel_state = Some(state);
            return Err(AgentError::InvalidStep(
                "effect observation received with no pending effect".into(),
            ));
        };
        if observation.effect_id != expected_effect_id {
            self.kernel_state = Some(state);
            return Err(AgentError::InvalidStep(format!(
                "effect observation id mismatch: expected {expected_effect_id}, got {}",
                observation.effect_id
            )));
        }

        if observation.output.observation_for_model.len() > MAX_ROUND_TEXT {
            self.kernel_state = Some(state);
            return Err(AgentError::Provider(ProviderError::ResponseTooLarge {
                limit_bytes: MAX_ROUND_TEXT,
            }));
        }
        let recorded = Self::mark_pending_effect_observed(
            &mut state.pending_tool_results,
            &expected_effect_id,
            observation,
        );
        debug_assert!(recorded);
        if let Err(error) = self.record_pending_tool_results(&mut state).await {
            self.kernel_state = Some(state);
            return Err(error);
        }
        self.process_kernel_pending_tools(state, context).await
    }

    async fn observe_effects(
        &mut self,
        observations: Vec<EffectObservation>,
        context: &RunContext,
    ) -> Result<AgentAction, AgentError> {
        let mut state = self.kernel_state.take().ok_or_else(|| {
            AgentError::InvalidStep("effect observations without active run".into())
        })?;
        let pending_effect_ids = state
            .pending_tool_results
            .iter()
            .filter_map(PendingToolResult::effect_id)
            .map(str::to_string)
            .collect::<Vec<_>>();
        if pending_effect_ids.is_empty() {
            self.kernel_state = Some(state);
            return Err(AgentError::InvalidStep(
                "effect observations received with no pending effects".into(),
            ));
        }
        if observations.len() != pending_effect_ids.len() {
            let expected = pending_effect_ids.len();
            self.kernel_state = Some(state);
            return Err(AgentError::InvalidStep(format!(
                "effect observation count mismatch: expected {expected}, got {}",
                observations.len()
            )));
        }

        let expected_ids = pending_effect_ids.iter().cloned().collect::<HashSet<_>>();
        let mut observations_by_id = HashMap::with_capacity(observations.len());
        for observation in observations {
            if observation.output.observation_for_model.len() > MAX_ROUND_TEXT {
                self.kernel_state = Some(state);
                return Err(AgentError::Provider(ProviderError::ResponseTooLarge {
                    limit_bytes: MAX_ROUND_TEXT,
                }));
            }
            if !expected_ids.contains(&observation.effect_id) {
                let effect_id = observation.effect_id;
                self.kernel_state = Some(state);
                return Err(AgentError::InvalidStep(format!(
                    "unexpected effect observation for {effect_id}"
                )));
            }

            let effect_id = observation.effect_id.clone();
            if observations_by_id
                .insert(effect_id.clone(), observation)
                .is_some()
            {
                self.kernel_state = Some(state);
                return Err(AgentError::InvalidStep(format!(
                    "duplicate effect observation for {effect_id}"
                )));
            }
        }
        for effect_id in &pending_effect_ids {
            if !observations_by_id.contains_key(effect_id) {
                self.kernel_state = Some(state);
                return Err(AgentError::InvalidStep(format!(
                    "missing effect observation for {effect_id}"
                )));
            }
        }

        for effect_id in pending_effect_ids {
            let observation = observations_by_id
                .remove(&effect_id)
                .expect("effect observation was prevalidated");
            let recorded = Self::mark_pending_effect_observed(
                &mut state.pending_tool_results,
                &effect_id,
                observation,
            );
            debug_assert!(recorded);
        }
        if let Err(error) = self.record_pending_tool_results(&mut state).await {
            self.kernel_state = Some(state);
            return Err(error);
        }
        self.process_kernel_pending_tools(state, context).await
    }

    fn mark_pending_effect_observed(
        pending_tool_results: &mut VecDeque<PendingToolResult>,
        expected_effect_id: &str,
        observation: EffectObservation,
    ) -> bool {
        for result in pending_tool_results {
            let PendingToolResult::Effect {
                effect_id,
                observation: pending_observation,
                ..
            } = result
            else {
                continue;
            };
            if effect_id == expected_effect_id {
                *pending_observation = Some(observation);
                return true;
            }
        }
        false
    }

    async fn record_pending_tool_results(
        &mut self,
        state: &mut ReActKernelState,
    ) -> Result<(), AgentError> {
        while let Some(outcome) =
            state
                .pending_tool_results
                .front()
                .and_then(|pending| match pending {
                    PendingToolResult::Outcome(outcome) => Some(outcome.clone()),
                    PendingToolResult::Effect {
                        call,
                        observation: Some(observation),
                        ..
                    } => Some(Self::effect_observation_outcome(
                        call.clone(),
                        observation.clone(),
                    )),
                    PendingToolResult::Effect {
                        observation: None, ..
                    } => None,
                })
        {
            record_tool_result(&mut self.memory, &outcome).await?;
            state.pending_tool_results.pop_front();
        }
        Ok(())
    }

    fn effect_observation_outcome(
        call: ToolCall,
        observation: EffectObservation,
    ) -> ToolCallOutcome {
        ToolCallOutcome {
            call,
            content: observation.output.observation_for_model,
            memory_policy: observation.output.memory_policy,
            effect: None,
        }
    }

    async fn process_kernel_pending_tools(
        &mut self,
        mut state: ReActKernelState,
        context: &RunContext,
    ) -> Result<AgentAction, AgentError> {
        let mut effects = Vec::new();
        let mut effect_ids = HashSet::new();
        while let Some(call) = state.pending_tools.pop_front() {
            let ctx = ToolRoundCtx {
                context,
                round: state.counters.rounds,
                registry: &self.registry,
                state: &self.state,
                events: &self.events,
            };
            let outcome = ctx.run(call).await;
            if let Some(effect) = outcome.effect_request().cloned() {
                if !effect_ids.insert(effect.id.clone()) {
                    return Err(AgentError::InvalidStep(format!(
                        "duplicate effect request id: {}",
                        effect.id
                    )));
                }
                state
                    .pending_tool_results
                    .push_back(PendingToolResult::Effect {
                        effect_id: effect.id.clone(),
                        call: outcome.call.clone(),
                        observation: None,
                    });
                effects.push(effect);
                continue;
            }
            state
                .pending_tool_results
                .push_back(PendingToolResult::Outcome(outcome));
        }
        if let Err(error) = self.record_pending_tool_results(&mut state).await {
            self.kernel_state = Some(state);
            return Err(error);
        }
        if !effects.is_empty() {
            self.kernel_state = Some(state);
            return if effects.len() == 1 {
                let request = effects
                    .pop()
                    .expect("single effect request must be present");
                Ok(AgentAction::RequestEffect { request })
            } else {
                Ok(AgentAction::RequestEffects { requests: effects })
            };
        }

        check_run_context(context)?;
        state.counters.rounds += 1;
        let messages = self.assemble_messages(self.memory.context().await?);
        let action = state.next_model_request(messages);
        self.kernel_state = Some(state);
        Ok(action)
    }
}

// Span construction is unified — both run / run_stream paths share the
// same field set. round is passed in by the caller (already incremented on
// the non-streaming path; current round + 1 on the streaming path). A
// single hierarchy mechanism: the ambient parent at creation time — llm/tool
// spans are created inside the round body, and while the round body runs,
// `agent.run` is always on the stack (non-streaming: the run block
// instruments the whole span; streaming: SpanStream enters on every poll),
// so no explicit parent and no double instrumenting are needed.

/// The `agent.run` root span: the observability identifier of a whole run
/// (including the error recorded on failure).
#[cfg(feature = "tracing")]
type TraceSpan = tracing::Span;

#[cfg(not(feature = "tracing"))]
#[derive(Debug, Clone)]
struct TraceSpan;

#[cfg(feature = "tracing")]
fn instrument<F>(future: F, span: TraceSpan) -> tracing::instrument::Instrumented<F>
where
    F: Future,
{
    future.instrument(span)
}

#[cfg(not(feature = "tracing"))]
fn instrument<F>(future: F, _span: TraceSpan) -> F
where
    F: Future,
{
    future
}

fn span_run(run_id: &str) -> TraceSpan {
    #[cfg(feature = "tracing")]
    {
        tracing::info_span!("agent.run", "run.id" = %run_id, error = tracing::field::Empty)
    }
    #[cfg(not(feature = "tracing"))]
    {
        let _ = run_id;
        TraceSpan
    }
}

/// Provider-call span: duration comes from span timing automatically; usage
/// goes out through both channels (the business-side RunSummary as usual,
/// observability records it on the span fields at wrap-up); errors are
/// recorded when the Provider fails (both paths' error branches record).
fn span_llm(run_id: &str, round: usize) -> TraceSpan {
    #[cfg(feature = "tracing")]
    {
        tracing::debug_span!(
            "llm_request",
            "run.id" = %run_id,
            round = round,
            usage.prompt_tokens = tracing::field::Empty,
            usage.completion_tokens = tracing::field::Empty,
            error = tracing::field::Empty,
        )
    }
    #[cfg(not(feature = "tracing"))]
    {
        let _ = (run_id, round);
        TraceSpan
    }
}

/// Tool-call span: carries duration; records error on failure (see
/// run_tool_call).
fn span_tool(run_id: &str, round: usize, name: &str) -> TraceSpan {
    #[cfg(feature = "tracing")]
    {
        tracing::debug_span!(
            "tool",
            "run.id" = %run_id,
            round = round,
            name = %name,
            error = tracing::field::Empty,
        )
    }
    #[cfg(not(feature = "tracing"))]
    {
        let _ = (run_id, round, name);
        TraceSpan
    }
}

/// Generator terminal wrap-up: publish `RunEnded` (with the accumulated
/// summary and error state; error=None on normal completion, error set on
/// cancellation / failure; no-op when no channel is attached).
///
/// A free function rather than a method: the generator borrows both
/// `self.events` (immutably) and other `&mut self` calls (record, etc.)
/// simultaneously; passing references pointwise avoids closure-capture
/// borrow tangles.
fn publish_ended(
    events: &Option<Arc<dyn EventChannel>>,
    summary: RunSummary,
    error: Option<AgentError>,
) {
    if let Some(pipe) = events {
        pipe.publish(Arc::new(ReActEvent::RunEnded { summary, error }));
    }
}

/// Streaming termination helper: observability wrap-up (span error +
/// RunEnded event) plus the terminal chunk — cancellation →
/// `Ok(Cancelled)`, other errors → `Err(e)`; the call site `yield`s and
/// `break`s. Consolidates the fixed four-step sequence of every failure
/// branch in the generator (record → publish_ended → yield → break).
///
/// A free function rather than a method: the generator borrows both
/// `self.events` and `&mut self` calls (record, etc.) simultaneously;
/// passing references pointwise avoids closure-capture borrow tangles (same
/// as publish_ended).
fn stream_end(
    events: &Option<Arc<dyn EventChannel>>,
    #[cfg_attr(not(feature = "tracing"), allow(unused_variables))] run_span: &TraceSpan,
    summary: RunSummary,
    error: AgentError,
) -> Result<MessageChunk, AgentError> {
    // Cancellation is a normal outcome (terminating with the Cancelled
    // terminal chunk), so no span error is recorded — otherwise the
    // observability dashboard would mark a user-initiated stop as a
    // failure.
    #[cfg(feature = "tracing")]
    if !matches!(error, AgentError::Cancelled) {
        run_span.record("error", error.to_string());
    }
    publish_ended(events, summary, Some(error.clone()));
    match error {
        AgentError::Cancelled => Ok(MessageChunk::Cancelled),
        e => Err(e),
    }
}

fn run_summary(
    counters: &RunCounters,
    finish_reason: Option<FinishReason>,
    started_at: Instant,
    provider_model: Option<String>,
) -> RunSummary {
    run_summary_from_parts(
        counters.rounds,
        counters.tool_calls_total,
        counters.usage_total,
        counters.usage_omitted,
        finish_reason,
        started_at,
        provider_model,
    )
}

fn run_summary_from_parts(
    rounds: usize,
    tool_calls: usize,
    usage: Usage,
    usage_omitted: bool,
    finish_reason: Option<FinishReason>,
    started_at: Instant,
    provider_model: Option<String>,
) -> RunSummary {
    RunSummary {
        rounds,
        tool_calls,
        usage,
        usage_omitted,
        finish_reason,
        latency: started_at.elapsed(),
        provider_model,
    }
}

fn check_run_context(context: &RunContext) -> Result<(), AgentError> {
    if context.is_cancelled() {
        Err(AgentError::Cancelled)
    } else if context.is_expired() {
        Err(AgentError::DeadlineExceeded)
    } else {
        Ok(())
    }
}

async fn run_until_context<F>(context: &RunContext, future: F) -> Result<F::Output, AgentError>
where
    F: Future,
{
    check_run_context(context)?;
    match context.remaining() {
        Some(remaining) if remaining.is_zero() => Err(AgentError::DeadlineExceeded),
        Some(remaining) => {
            tokio::select! {
                _ = context.cancellation.cancelled() => Err(AgentError::Cancelled),
                _ = tokio::time::sleep(remaining) => Err(AgentError::DeadlineExceeded),
                output = future => Ok(output),
            }
        }
        None => context
            .cancellation
            .run_until_cancelled(future)
            .await
            .ok_or(AgentError::Cancelled),
    }
}

/// Record a tool result, falling back to recording the error text before
/// re-raising on failure.
///
/// The tool has already run (side effects happened), so its result text
/// must not be lost: an Assistant message recorded without its ToolResult
/// leaves the memory incomplete, the next request would carry an unpaired
/// assistant/tool sequence, real endpoints reject it (400), and the
/// reported cause would be disconnected from the real failure. If the
/// fallback recording itself fails, the error is swallowed — the original
/// error takes priority, and there's no point retrying when the same
/// Memory keeps failing.
///
/// A free function rather than a method: while the round's outcome stream
/// is consumed, the agent's registry / state / events are borrowed (via
/// [`ToolRoundCtx`]); recording goes through the memory field directly —
/// a disjoint borrow (same rationale as publish_ended / stream_end).
async fn record_tool_result(
    memory: &mut Box<dyn Memory>,
    outcome: &ToolCallOutcome,
) -> Result<(), AgentError> {
    let message = Message::tool_result(outcome.call.id.clone(), outcome.content.clone());
    let record_result = if outcome.memory_policy.is_protected() {
        memory.record_protected(message).await
    } else {
        memory.record(message).await
    };
    match record_result {
        Ok(()) => Ok(()),
        Err(e) => {
            let fallback = Message::tool_result(
                outcome.call.id.clone(),
                format!("memory record failed: {e}"),
            );
            let _ = if outcome.memory_policy.is_protected() {
                memory.record_protected(fallback).await
            } else {
                memory.record(fallback).await
            };
            Err(e.into())
        }
    }
}

impl ReActAgent {
    async fn run_request_inner(
        &mut self,
        request: RunRequest,
        context: RunContext,
        schema: Option<&serde_json::Value>,
    ) -> Result<RunOutput, AgentError> {
        // Run id and root span — one agent.run span for the whole run (from
        // recording the input to the wrap-up event), sharing the same
        // run.id as the event stream's RunStarted (the correlation key
        // between the two channels); on Err, the error is recorded at
        // wrap-up (observability spots the failed run at a glance).
        let run_id = context.run_id.clone();
        let run_span = span_run(&run_id);
        let provider_model = self.provider.model().map(str::to_string);
        let started_at = Instant::now();
        let run_future = async {
            // Record first, publish after (consistent with the streaming
            // path): RunStarted's claim that "the user input is recorded"
            // holds; when recording fails, neither RunStarted nor RunEnded
            // is published, leaving no dangling segment in the event
            // stream.
            let input = request.input;
            self.memory.record(input.clone().into_message()).await?;
            self.publish(|| {
                Arc::new(ReActEvent::RunStarted {
                    run_id: run_id.clone(),
                    input,
                })
            });

            // Round / tool / structured-retry / usage counters: for the
            // RunEnded summary at wrap-up (also accumulated on the
            // non-streaming path).
            let mut counters = RunCounters::default();
            let mut options = request
                .options
                .clone()
                .unwrap_or_else(|| self.config.options.clone());
            if let Some(schema) = schema {
                options.structured = Some(schema.clone());
            }
            let validation_schema = options.structured.as_ref();
            let result: Result<FinalAnswer, AgentError> = self
                .run_rounds_with_context(
                    &context,
                    &run_id,
                    &mut counters,
                    &options,
                    validation_schema,
                )
                .await;
            let output_result = result.map(|final_answer| {
                let summary = run_summary(
                    &counters,
                    final_answer.finish_reason.clone(),
                    started_at,
                    provider_model.clone(),
                );
                RunExecution {
                    answer: final_answer.answer,
                    final_message: final_answer.final_message,
                    summary,
                    artifacts: Vec::new(),
                    metadata: RunMetadata::new(),
                }
            });
            let summary = match &output_result {
                Ok(execution) => execution.summary.clone(),
                Err(_) => run_summary(&counters, None, started_at, provider_model.clone()),
            };
            // Wrap-up event: error=None on success; error set on
            // cancellation / failure (bystanders can observe the outcome of
            // a non-streaming run); built with publish_ended, shared with
            // the streaming path.
            publish_ended(&self.events, summary, output_result.as_ref().err().cloned());
            output_result
        };
        let result = instrument(run_future, run_span.clone()).await;
        #[cfg(feature = "tracing")]
        if let Err(e) = &result {
            run_span.record("error", e.to_string());
        }
        result.map(|execution| RunOutput {
            run_id,
            answer: execution.answer,
            summary: execution.summary,
            final_message: execution.final_message,
            artifacts: execution.artifacts,
            metadata: execution.metadata,
        })
    }

    async fn run_stream_request_inner<'a>(
        &'a mut self,
        request: RunRequest,
        context: RunContext,
    ) -> Result<BoxStream<'a, Result<MessageChunk, AgentError>>, AgentError> {
        // Run id and root span — the run span covers the consumption period
        // from "stream returned" to "stream dropped" (SpanStream enters on
        // every poll); the error field is recorded by the generator's error
        // branches; the same run.id is shared with the event stream's
        // RunStarted (the correlation key between the two channels).
        //
        // Recording and RunStarted are published before the stream is
        // returned (failing fast, so callers see the error before holding
        // the stream): if a caller drops the returned stream without
        // polling, the event stream is left with a dangling segment without
        // RunEnded — event consumers pair by run.id and wrap up such cases
        // themselves.
        let run_id = context.run_id.clone();
        let run_span = span_run(&run_id);
        let stream_span = run_span.clone();
        let input = request.input;
        self.memory.record(input.clone().into_message()).await?;
        self.publish(|| {
            Arc::new(ReActEvent::RunStarted {
                run_id: run_id.clone(),
                input,
            })
        });
        let schemas = self.registry.schemas();
        let max_rounds = self.config.max_tool_rounds;
        let provider_model = self.provider.model().map(str::to_string);
        let started_at = Instant::now();
        let options = request
            .options
            .clone()
            .unwrap_or_else(|| self.config.options.clone());
        let validation_schema = options.structured.clone();

        // Streaming state machine: an async_stream generator — sequential
        // awaits in the loop body, syntax isomorphic to a synchronous loop
        // (replacing a hand-written unfold state machine).
        let context = context.clone();
        let stream = async_stream::stream! {
            let mut rounds = 0usize;
            // Execution summary (carried with Done): rounds is the `rounds`
            // above; tool counts and usage accumulate per round; rounds whose
            // provider omitted usage are tracked so the sum is not misread
            // as exact.
            let mut tool_calls_total = 0usize;
            let mut usage_total = Usage::default();
            let mut usage_omitted = false;
            // Structured validator: built when this run has a hand-written
            // schema from the request or config; the retry budget lives in
            // the component.
            #[cfg(feature = "structured")]
            let mut validator = validation_schema.as_ref().map(|schema| {
                StructuredValidator::new(schema.clone(), self.config.max_structured_retries)
            });
            #[cfg(not(feature = "structured"))]
            let _ = &validation_schema;
            // Tool-round counter (same semantics as the non-streaming
            // path): the limit only counts rounds where the model requested
            // tools; structured validation-retry rounds don't consume it
            // (independent max_structured_retries budget).
            let mut tool_rounds = 0usize;
            // Labeled loop: in-stream errors must terminate the whole
            // generator — a plain break only exits the current loop;
            // break 'rounds terminates the entire generator.
            'rounds: loop {
                // Same semantics as run: the model still requests tools
                // past the limit → error event.
                if tool_rounds >= max_rounds {
                    let summary = run_summary_from_parts(
                        rounds,
                        tool_calls_total,
                        usage_total,
                        usage_omitted,
                        None,
                        started_at,
                        provider_model.clone(),
                    );
                    yield stream_end(&self.events, &run_span, summary,
                        AgentError::TooManyToolRounds(max_rounds));
                    break;
                }

                // Increment at the start of the round (same semantics as
                // the non-streaming path): RunEnded reports "rounds
                // started" — both paths report 1 for a pre-cancelled token,
                // no drift.
                rounds += 1;

                // Between-round check: cancellation takes effect
                // immediately — this round hasn't started a conversation,
                // so the memory is clean.
                if let Err(e) = check_run_context(&context) {
                    // This counted round got no provider usage: the sum is a
                    // lower bound from here on.
                    usage_omitted = true;
                    let summary = run_summary_from_parts(
                        rounds,
                        tool_calls_total,
                        usage_total,
                        usage_omitted,
                        None,
                        started_at,
                        provider_model.clone(),
                    );
                    yield stream_end(&self.events, &run_span, summary, e);
                    break;
                }

                // Start a streaming conversation round (cancelled during
                // establishment → nothing recorded for this round, memory
                // clean).
                let messages = match self.memory.context().await {
                    Ok(messages) => messages,
                    Err(e) => {
                        // This counted round got no provider usage: the sum
                        // is a lower bound from here on.
                        usage_omitted = true;
                        let summary = run_summary_from_parts(
                            rounds,
                            tool_calls_total,
                            usage_total,
                            usage_omitted,
                            None,
                            started_at,
                            provider_model.clone(),
                        );
                        yield stream_end(&self.events, &run_span, summary,
                            AgentError::Memory(e));
                        break;
                    }
                };
                // Provider-call span (streaming: establishment and
                // per-event consumption both belong to llm_request, the span
                // is shared across the round's consumption loop; usage is
                // recorded at wrap-up when the Done event arrives — the
                // business-side RunSummary as usual); the span is created
                // while the run is on the stack (SpanStream enters on every
                // poll), so the hierarchy is correct automatically.
                let llm_span = span_llm(&run_id, rounds);
                let model_request_id = format!("{run_id}-model-{rounds}");
                let provider_context =
                    ProviderRequestContext::from_run_context(model_request_id, &context);
                let stream_chat = self.provider.stream_chat_with_context(
                    ChatRequest {
                    messages: self.assemble_messages(messages),
                    tools: schemas.clone(),
                    options: options.clone(),
                    },
                    &provider_context,
                );
                let mut provider_stream = match run_until_context(
                    &context,
                    instrument(stream_chat, llm_span.clone()),
                )
                .await
                {
                    Ok(Ok(stream)) => stream,
                    Ok(Err(e)) => {
                        // Provider errors are recorded on the llm span
                        // (observability pinpoints the failed call).
                        #[cfg(feature = "tracing")]
                        llm_span.record("error", e.to_string());
                        // This counted round got no provider usage: the sum
                        // is a lower bound from here on.
                        usage_omitted = true;
                        let summary = run_summary_from_parts(
                            rounds,
                            tool_calls_total,
                            usage_total,
                            usage_omitted,
                            None,
                            started_at,
                            provider_model.clone(),
                        );
                        yield stream_end(&self.events, &run_span, summary,
                            AgentError::Provider(e));
                        break;
                    }
                    Err(e) => {
                        // This counted round got no provider usage: the sum
                        // is a lower bound from here on.
                        usage_omitted = true;
                        let summary = run_summary_from_parts(
                            rounds,
                            tool_calls_total,
                            usage_total,
                            usage_omitted,
                            None,
                            started_at,
                            provider_model.clone(),
                        );
                        yield stream_end(&self.events, &run_span, summary, e);
                        break;
                    }
                };

                // Consume this round's full event stream: text dispatches
                // character by character in real time; tool requests are
                // collected whole at the end of the round.
                // Every next is wrapped in run_until_cancelled —
                // cancellation stops per-character dispatch immediately;
                // when cancellation races with event arrival, completion
                // wins (same trade-off as the non-streaming path): keep
                // consuming as long as events keep flowing, and only void
                // the round when next itself is preempted by cancellation
                // (None) (record happens at the end of the round, so the
                // memory is clean when voided).
                let mut text = String::new();
                let mut reasoning = String::new();
                let mut calls = Vec::new();
                let mut round_finish_reason = None::<FinishReason>;
                // True when this round's Done carried reported usage; a
                // round that ends without it marks the run's usage sum as a
                // lower bound (see the wrap-up after the loop below).
                let mut round_usage_reported = false;
                loop {
                    // next()'s Output is itself an Option, and
                    // run_until_cancelled wraps it in another (returns None
                    // on cancellation): Some(Some(event)) / Some(None) =
                    // stream ended naturally / None = cancelled during next
                    // (this round is voided).
                    let next = run_until_context(
                        &context,
                        instrument(provider_stream.next(), llm_span.clone()),
                    )
                    .await;
                    let Some(event) = (match next {
                        Ok(event) => event,
                        Err(e) => {
                            // The round was aborted before its Done: its
                            // usage is unknown, the sum is a lower bound.
                            usage_omitted = true;
                            let summary = run_summary_from_parts(
                                rounds,
                                tool_calls_total,
                                usage_total,
                                usage_omitted,
                                None,
                                started_at,
                                provider_model.clone(),
                            );
                            yield stream_end(&self.events, &run_span, summary, e);
                            break 'rounds;
                        }
                    }) else {
                        // The event stream ended naturally; move to
                        // end-of-round wrap-up.
                        break;
                    };
                    match event {
                        Ok(StreamEvent::Delta(delta)) => {
                            // Round text accumulation limit: a malicious
                            // endpoint can keep sending Deltas without
                            // bound (the Provider layer only limits single
                            // lines; this is the per-round backstop).
                            if text.len() + delta.len() > MAX_ROUND_TEXT {
                                // The round was aborted before its Done:
                                // its usage is unknown, the sum is a lower
                                // bound.
                                usage_omitted = true;
                                let summary = run_summary_from_parts(
                                    rounds,
                                    tool_calls_total,
                                    usage_total,
                                    usage_omitted,
                                    None,
                                    started_at,
                                    provider_model.clone(),
                                );
                                yield stream_end(&self.events, &run_span, summary,
                                    AgentError::Provider(ProviderError::ResponseTooLarge {
                                        limit_bytes: MAX_ROUND_TEXT,
                                    }));
                                break 'rounds;
                            }
                            text.push_str(&delta);
                            self.publish(|| Arc::new(ReActEvent::Delta { text: delta.clone() }));
                            yield Ok(MessageChunk::Delta(delta));
                        }
                        Ok(StreamEvent::Reasoning(chunk)) => {
                            // Same limit as text: a custom Provider may
                            // send reasoning deltas without bound.
                            if reasoning.len() + chunk.len() > MAX_ROUND_TEXT {
                                // The round was aborted before its Done:
                                // its usage is unknown, the sum is a lower
                                // bound.
                                usage_omitted = true;
                                let summary = run_summary_from_parts(
                                    rounds,
                                    tool_calls_total,
                                    usage_total,
                                    usage_omitted,
                                    None,
                                    started_at,
                                    provider_model.clone(),
                                );
                                yield stream_end(&self.events, &run_span, summary,
                                    AgentError::Provider(ProviderError::ResponseTooLarge {
                                        limit_bytes: MAX_ROUND_TEXT,
                                    }));
                                break 'rounds;
                            }
                            reasoning.push_str(&chunk);
                            self.publish(move || Arc::new(ReActEvent::Reasoning { text: chunk }));
                        }
                        Ok(StreamEvent::ToolCall { id, name, arguments }) => {
                            calls.push(ToolCall {
                                id: id.clone(),
                                name: name.clone(),
                                arguments: arguments.clone(),
                            });
                            yield Ok(MessageChunk::ToolCall { id, name, arguments });
                        }
                        Ok(StreamEvent::Done { reason, usage }) => {
                            // Streaming usage may be absent (the endpoint
                            // didn't return it); missing rounds are tracked
                            // so the summary's sum is not misread as exact.
                            // usage goes through both channels:
                            // observability records it on the llm_request
                            // span at wrap-up, and the business-side
                            // RunSummary accumulates as usual.
                            if let Some(usage) = usage {
                                #[cfg(feature = "tracing")]
                                {
                                llm_span.record("usage.prompt_tokens", usage.prompt_tokens);
                                llm_span.record("usage.completion_tokens", usage.completion_tokens);
                                }
                                usage_total += usage;
                                round_usage_reported = true;
                            } else {
                                usage_omitted = true;
                            }
                            round_finish_reason = Some(reason);
                            // Done = this round is complete: move to
                            // end-of-round wrap-up immediately, consuming no
                            // further events — the Provider's Done is always
                            // the stream-final event, and anything after it
                            // was already discarded by the Provider; this
                            // break is belt and braces.
                            break;
                        }
                        Err(e) => {
                            // Errors are produced as Err events and
                            // terminate the stream (no Done afterwards);
                            // recorded on the llm span (observability
                            // pinpoints the failed call).
                            #[cfg(feature = "tracing")]
                            llm_span.record("error", e.to_string());
                            // The round was aborted before its Done: its
                            // usage is unknown, the sum is a lower bound.
                            usage_omitted = true;
                            let summary = run_summary_from_parts(
                                rounds,
                                tool_calls_total,
                                usage_total,
                                usage_omitted,
                                None,
                                started_at,
                                provider_model.clone(),
                            );
                            yield stream_end(&self.events, &run_span, summary,
                                AgentError::Provider(e));
                            break 'rounds;
                        }
                        Ok(_) => {}
                    }
                }

                // Round ended without reported usage (stream ended without
                // Done, or Done carried None): the accumulated sum is a
                // lower bound from here on.
                usage_omitted |= !round_usage_reported;

                // Empty Assistant messages are not recorded (consistent
                // with run).
                if !text.is_empty() || !reasoning.is_empty() || !calls.is_empty() {
                    let message = Message::Assistant {
                        content: text.clone(),
                        reasoning: (!reasoning.is_empty()).then_some(reasoning),
                        tool_calls: calls.clone(),
                    };
                    match self.memory.record(message).await {
                        Ok(()) => {}
                        Err(e) => {
                            let summary = run_summary_from_parts(
                                rounds,
                                tool_calls_total,
                                usage_total,
                                usage_omitted,
                                None,
                                started_at,
                                provider_model.clone(),
                            );
                            yield stream_end(&self.events, &run_span, summary,
                                AgentError::Memory(e));
                            break;
                        }
                    }
                }

                if calls.is_empty() {
                    // The model answered directly. Structured output: the
                    // component validates the final answer — on failure,
                    // feed it back to the model for retry (no Done
                    // dispatched; the loop continues; the budget is built
                    // into the component, decoupled from the tool-round
                    // limit), and exceeding it terminates the stream with
                    // an error; only a pass wraps up.
                    #[cfg(feature = "structured")]
                    {
                        if let Some(validator) = &mut validator {
                            match validator.validate(&text) {
                                StructuredOutcome::Passed => {}
                                StructuredOutcome::Retry { message } => {
                                    // Record the feedback and continue to the
                                    // next round (no Done dispatched).
                                    match self.memory.record(message).await {
                                        Ok(()) => continue 'rounds,
                                        Err(e) => {
                                            let summary = run_summary_from_parts(
                                                rounds,
                                                tool_calls_total,
                                                usage_total,
                                                usage_omitted,
                                                None,
                                                started_at,
                                                provider_model.clone(),
                                            );
                                            yield stream_end(&self.events, &run_span, summary,
                                                AgentError::Memory(e));
                                            break;
                                        }
                                    }
                                }
                                StructuredOutcome::Exhausted { max_retries } => {
                                    let summary = run_summary_from_parts(
                                        rounds,
                                        tool_calls_total,
                                        usage_total,
                                        usage_omitted,
                                        None,
                                        started_at,
                                        provider_model.clone(),
                                    );
                                    yield stream_end(&self.events, &run_span, summary,
                                        AgentError::StructuredRetriesExhausted(max_retries));
                                    break 'rounds;
                                }
                            }
                        }
                    }
                    // Validation passed (or no structured constraint): this
                    // round ran to completion; wrap up directly.
                    let summary = run_summary_from_parts(
                        rounds,
                        tool_calls_total,
                        usage_total,
                        usage_omitted,
                        round_finish_reason,
                        started_at,
                        provider_model.clone(),
                    );
                    publish_ended(&self.events, summary.clone(), None);
                    yield Ok(MessageChunk::Done(summary));
                    break;
                }

                // Tool round: the round policy (default: execute one by
                // one, feeding results back right after each) is injected
                // via the ToolRoundExecutor; each outcome is dispatched
                // and recorded as it arrives.
                tool_calls_total += calls.len();
                tool_rounds += 1;
                let ctx = ToolRoundCtx {
                    context: &context,
                    round: rounds,
                    registry: &self.registry,
                    state: &self.state,
                    events: &self.events,
                };
                let mut outcomes = self.executor.execute_round(ctx, calls).await;
                while let Some(outcome) = outcomes.next().await {
                    if let Some(effect) = outcome.effect_request() {
                        let summary = run_summary_from_parts(
                            rounds,
                            tool_calls_total,
                            usage_total,
                            usage_omitted,
                            None,
                            started_at,
                            provider_model.clone(),
                        );
                        yield stream_end(
                            &self.events,
                            &run_span,
                            summary,
                            AgentError::EffectRequiresHarness(format!(
                                "{} ({})",
                                effect.description, effect.id
                            )),
                        );
                        break 'rounds;
                    }
                    yield Ok(MessageChunk::ToolResult {
                        id: outcome.call.id.clone(),
                        name: outcome.call.name.clone(),
                        content: outcome.content.clone(),
                    });
                    // Protected results (skill bodies, etc.) are recorded
                    // via record_protected, exempt from window trimming;
                    // when recording fails, the error text is recorded as a
                    // fallback (memory integrity, see record_tool_result).
                    if let Err(e) = record_tool_result(&mut self.memory, &outcome).await {
                        let summary = run_summary_from_parts(
                            rounds,
                            tool_calls_total,
                            usage_total,
                            usage_omitted,
                            None,
                            started_at,
                            provider_model.clone(),
                        );
                        yield stream_end(&self.events, &run_span, summary, e);
                        // Errors are produced as Err items and terminate
                        // the stream — the break must exit the whole round
                        // loop, not just the outcome loop (otherwise the
                        // loop would continue to the next round after an
                        // Err).
                        break 'rounds;
                    }
                }
            }
        };
        Ok(Box::pin(SpanStream {
            stream: Box::pin(stream),
            span: stream_span,
        }))
    }
}

/// Wrap a span around a stream: enters on each poll and exits on return;
/// the span's lifetime = the stream's consumption period (creation to drop,
/// including time spent waiting while the consumer hasn't polled yet).
///
/// tracing's `Instrument` only applies to Futures; Streams need this thin
/// wrapper (per-await instrumenting inside the generator is already done
/// explicitly at each await point; this covers the `agent.run` root span
/// across the whole consumption). The inner stream is boxed and pinned at
/// construction (the generator is not Unpin — async_stream uses internal
/// pin_project; once pinned it never moves), and both `Pin<Box<S>>` and
/// `Span` are Unpin ⇒ the wrapper is always Unpin, so projection is safe,
/// no unsafe.
struct SpanStream<S> {
    stream: Pin<Box<S>>,
    #[cfg_attr(not(feature = "tracing"), allow(dead_code))]
    span: TraceSpan,
}

impl<S: futures::Stream> futures::Stream for SpanStream<S> {
    type Item = S::Item;

    fn poll_next(
        mut self: Pin<&mut Self>,
        cx: &mut std::task::Context<'_>,
    ) -> std::task::Poll<Option<Self::Item>> {
        // Clone the span handle before entering: the guard borrows the
        // local handle, not self (we need &mut self below; Span is a shared
        // handle, so the clone refers to the same span).
        #[cfg(feature = "tracing")]
        let span = self.span.clone();
        #[cfg(feature = "tracing")]
        let _enter = span.enter();
        // SpanStream is Unpin (both Pin<Box<S>> and Span are Unpin) ⇒
        // DerefMut is safe; poll the pinned inner stream directly, no
        // unsafe.
        self.stream.as_mut().poll_next(cx)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::CancellationToken;
    use crate::effect::{EffectKind, EffectObservation, EffectRequest};
    use crate::memory::MemoryError;
    use crate::message::ContentBlock;
    use crate::provider::{
        ChatResponse, FakeProvider, FakeReply, FinishReason, ModelOptions, ProviderError,
        StreamEvent, TimeoutStage,
    };
    use crate::tool::{Tool, ToolContext, ToolError, ToolOutput, ToolResult, ToolSchema};
    use futures::StreamExt;
    #[cfg(feature = "structured")]
    use serde::Deserialize;
    use std::sync::Arc;
    use std::sync::atomic::{AtomicUsize, Ordering};
    use std::time::Duration;

    /// Shared wrapper: a shared reference to FakeProvider + Provider
    /// delegation, letting tests inspect the request history
    /// ([`SharedFake::requests`]) after the Agent has run to assert
    /// behavioral expectations.
    #[derive(Clone)]
    struct SharedFake(Arc<FakeProvider>);

    impl SharedFake {
        fn new(replies: impl IntoIterator<Item = FakeReply>) -> Self {
            Self(Arc::new(FakeProvider::new(replies)))
        }

        fn requests(&self) -> Vec<ChatRequest> {
            self.0.requests()
        }
    }

    #[async_trait::async_trait]
    impl Provider for SharedFake {
        fn model(&self) -> Option<&str> {
            self.0.model()
        }

        async fn chat(&self, request: ChatRequest) -> Result<ChatResponse, ProviderError> {
            self.0.chat(request).await
        }

        async fn chat_with_context(
            &self,
            request: ChatRequest,
            context: &ProviderRequestContext,
        ) -> Result<ChatResponse, ProviderError> {
            self.0.chat_with_context(request, context).await
        }

        async fn stream_chat(
            &self,
            request: ChatRequest,
        ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError> {
            self.0.stream_chat(request).await
        }

        async fn stream_chat_with_context(
            &self,
            request: ChatRequest,
            context: &ProviderRequestContext,
        ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError> {
            self.0.stream_chat_with_context(request, context).await
        }
    }

    /// Fake tool for tests: returns fixed text and counts calls.
    #[derive(Debug, Clone)]
    struct FakeTool {
        name: &'static str,
        result: &'static str,
        calls: Arc<AtomicUsize>,
    }

    impl FakeTool {
        fn new(name: &'static str, result: &'static str) -> (Self, Arc<AtomicUsize>) {
            let calls = Arc::new(AtomicUsize::new(0));
            (
                Self {
                    name,
                    result,
                    calls: calls.clone(),
                },
                calls,
            )
        }
    }

    #[async_trait::async_trait]
    impl Tool for FakeTool {
        fn schema(&self) -> ToolSchema {
            ToolSchema::new(self.name, "Test tool", serde_json::json!({}))
        }

        async fn call(
            &self,
            _arguments: serde_json::Value,
            _context: ToolContext<'_>,
        ) -> Result<ToolResult, ToolError> {
            self.calls.fetch_add(1, Ordering::Relaxed);
            Ok(ToolOutput::text(self.result).into())
        }
    }

    #[derive(Debug, Clone)]
    struct EffectTool {
        name: &'static str,
        effect_id: &'static str,
        description: &'static str,
    }

    impl EffectTool {
        fn new(name: &'static str, effect_id: &'static str, description: &'static str) -> Self {
            Self {
                name,
                effect_id,
                description,
            }
        }
    }

    #[async_trait::async_trait]
    impl Tool for EffectTool {
        fn schema(&self) -> ToolSchema {
            ToolSchema::new(self.name, "Effect tool", serde_json::json!({}))
        }

        async fn call(
            &self,
            _arguments: serde_json::Value,
            _context: ToolContext<'_>,
        ) -> Result<ToolResult, ToolError> {
            Ok(ToolResult::Effect(
                EffectRequest::new(
                    EffectKind::Custom("test.effect".into()),
                    self.description,
                    serde_json::json!({}),
                )
                .with_id(self.effect_id),
            ))
        }
    }

    fn call(id: &str, name: &str, arguments: &str) -> ToolCall {
        ToolCall {
            id: id.into(),
            name: name.into(),
            arguments: arguments.into(),
        }
    }

    fn done_summary(chunk: &MessageChunk) -> Option<&RunSummary> {
        match chunk {
            MessageChunk::Done(summary) => Some(summary),
            _ => None,
        }
    }

    fn assert_done_summary(
        chunk: &MessageChunk,
        rounds: usize,
        tool_calls: usize,
        usage: Usage,
        usage_omitted: bool,
    ) {
        assert_done_summary_with_finish(
            chunk,
            rounds,
            tool_calls,
            usage,
            usage_omitted,
            Some(FinishReason::Stop),
        );
    }

    fn assert_done_summary_with_finish(
        chunk: &MessageChunk,
        rounds: usize,
        tool_calls: usize,
        usage: Usage,
        usage_omitted: bool,
        finish_reason: Option<FinishReason>,
    ) {
        let summary = done_summary(chunk).expect("expected Done chunk");
        assert_eq!(summary.rounds, rounds);
        assert_eq!(summary.tool_calls, tool_calls);
        assert_eq!(summary.usage, usage);
        assert_eq!(summary.usage_omitted, usage_omitted);
        assert_eq!(summary.finish_reason, finish_reason);
        assert_eq!(summary.provider_model, None);
    }

    /// Simple assembly: default Memory + optional system prompt (empty
    /// string = none).
    fn agent(fake: SharedFake, system_prompt: &str) -> ReActAgent {
        ReActAgent::new(fake, ToolRegistry::new(), system_prompt)
    }

    /// Assembly: custom registry + config (chained).
    fn agent_with_registry(
        fake: SharedFake,
        registry: ToolRegistry,
        config: AgentConfig,
    ) -> ReActAgent {
        ReActAgent::new(fake, registry, "").with_config(config)
    }

    fn cancellation_context(token: &CancellationToken) -> RunContext {
        RunContext::generated().with_cancellation(token.clone())
    }

    #[tokio::test]
    async fn builder_assembles_agent_components() {
        let fake = SharedFake::new([FakeReply::Text("built".into())]);
        let (tool, _calls) = FakeTool::new("builder_tool", "unused");
        let mut agent = ReActAgent::builder(fake.clone())
            .with_tool(tool)
            .with_system_prompt("Builder system")
            .with_config(AgentConfig {
                options: ModelOptions {
                    temperature: Some(0.4),
                    ..Default::default()
                },
                ..Default::default()
            })
            .build();

        assert_eq!(agent.run("hi").await.unwrap(), "built");
        let request = &fake.requests()[0];
        assert_eq!(request.messages[0], Message::system("Builder system"));
        assert_eq!(request.tools.len(), 1);
        assert_eq!(request.tools[0].name, "builder_tool");
        assert_eq!(request.options.temperature, Some(0.4));
    }

    #[tokio::test]
    async fn direct_answer() {
        let fake = SharedFake::new([FakeReply::Text("Hello".into())]);
        let mut agent = agent(fake.clone(), "");

        let answer = agent.run("Are you there").await.unwrap();
        assert_eq!(answer, "Hello");

        let requests = fake.requests();
        assert_eq!(requests.len(), 1);
        assert_eq!(requests[0].messages.len(), 1);
        assert_eq!(requests[0].messages[0], Message::user("Are you there"));
    }

    #[tokio::test]
    async fn run_request_returns_structured_output() {
        let fake = SharedFake::new([FakeReply::text_with_usage("Hello", Usage::new(5, 2))]);
        let mut agent = agent(fake.clone(), "");
        let output = agent
            .run_request_with_context(RunRequest::text("Are you there"), RunContext::new("r1"))
            .await
            .unwrap();

        assert_eq!(output.run_id, "r1");
        assert_eq!(output.answer, "Hello");
        assert_eq!(output.final_message, Message::assistant("Hello"));
        assert!(output.artifacts.is_empty());
        assert!(output.metadata.is_empty());
        assert_eq!(output.summary.rounds, 1);
        assert_eq!(output.summary.tool_calls, 0);
        assert_eq!(output.summary.usage, Usage::new(5, 2));
        assert_eq!(output.summary.finish_reason, Some(FinishReason::Stop));
        assert_eq!(output.summary.provider_model, None);
    }

    #[tokio::test]
    async fn run_request_blocks_are_recorded_as_user_blocks() {
        let blocks = vec![ContentBlock::Text("What is this?".into())];
        let fake = SharedFake::new([FakeReply::Text("A block".into())]);
        let mut agent = agent(fake.clone(), "");

        agent
            .run_request(RunRequest::blocks(blocks.clone()))
            .await
            .unwrap();

        let requests = fake.requests();
        assert_eq!(requests[0].messages[0], Message::user_blocks(blocks));
    }

    #[tokio::test]
    async fn run_summary_carries_provider_model_when_available() {
        #[derive(Clone)]
        struct NamedProvider(SharedFake);

        #[async_trait::async_trait]
        impl Provider for NamedProvider {
            fn model(&self) -> Option<&str> {
                Some("named-model")
            }

            async fn chat(&self, request: ChatRequest) -> Result<ChatResponse, ProviderError> {
                self.0.chat(request).await
            }

            async fn chat_with_context(
                &self,
                request: ChatRequest,
                context: &ProviderRequestContext,
            ) -> Result<ChatResponse, ProviderError> {
                self.0.chat_with_context(request, context).await
            }

            async fn stream_chat(
                &self,
                request: ChatRequest,
            ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError>
            {
                self.0.stream_chat(request).await
            }

            async fn stream_chat_with_context(
                &self,
                request: ChatRequest,
                context: &ProviderRequestContext,
            ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError>
            {
                self.0.stream_chat_with_context(request, context).await
            }
        }

        let mut agent = ReActAgent::new(
            NamedProvider(SharedFake::new([FakeReply::Text("hi".into())])),
            ToolRegistry::new(),
            "",
        );
        let output = agent.run_request(RunRequest::text("hi")).await.unwrap();
        assert_eq!(output.summary.provider_model, Some("named-model".into()));
    }

    #[tokio::test]
    async fn single_tool_round() {
        let (calc, calls) = FakeTool::new("calc", "42");
        let mut registry = ToolRegistry::new();
        registry.register(calc);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "calc", r#"{"a":1}"#)],
            },
            FakeReply::Text("The answer is 42".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());

        let answer = agent.run("Compute 1+1").await.unwrap();
        assert_eq!(answer, "The answer is 42");
        assert_eq!(calls.load(Ordering::Relaxed), 1);

        // Second-round request: the tool result was fed back (the history
        // contains the ToolResult).
        let requests = fake.requests();
        assert_eq!(requests.len(), 2);
        assert!(requests[1].messages.iter().any(|m| matches!(
            m,
            Message::ToolResult { id, content } if id == "c1" && content == "42"
        )));
    }

    #[tokio::test]
    async fn multiple_tools_same_round() {
        let (t1, calls1) = FakeTool::new("t1", "one");
        let (t2, calls2) = FakeTool::new("t2", "two");
        let mut registry = ToolRegistry::new();
        registry.register(t1).register(t2);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "t1", "{}"), call("c2", "t2", "{}")],
            },
            FakeReply::Text("done".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());

        let answer = agent.run("Run them all").await.unwrap();
        assert_eq!(answer, "done");
        assert_eq!(calls1.load(Ordering::Relaxed), 1);
        assert_eq!(calls2.load(Ordering::Relaxed), 1);

        // Two calls in the same round live in the same Assistant message;
        // results follow in order.
        let requests = fake.requests();
        let assistant = requests[1]
            .messages
            .iter()
            .find_map(|m| match m {
                Message::Assistant { tool_calls, .. } => Some(tool_calls),
                _ => None,
            })
            .expect("second round should contain Assistant");
        assert_eq!(assistant.len(), 2);

        let results: Vec<&str> = requests[1]
            .messages
            .iter()
            .filter_map(|m| match m {
                Message::ToolResult { content, .. } => Some(content.as_str()),
                _ => None,
            })
            .collect();
        assert_eq!(results, vec!["one", "two"]);
    }

    /// Empty Assistant messages are not recorded: a purely empty reply is
    /// not recorded, and the request history only contains the user.
    #[tokio::test]
    async fn empty_assistant_not_recorded() {
        let fake = SharedFake::new([FakeReply::Text("".into())]);
        let mut agent = agent(fake.clone(), "");

        let answer = agent.run("hi").await.unwrap();
        assert_eq!(answer, "");

        let requests = fake.requests();
        assert_eq!(requests.len(), 1);
        assert_eq!(requests[0].messages.len(), 1); // user only
    }

    /// Default Memory is a bounded window: over-budget conversations
    /// auto-trim the oldest rounds and stop growing unboundedly; short
    /// conversations don't trigger trimming.
    #[tokio::test]
    async fn default_memory_is_bounded_window() {
        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = agent(fake, "");

        // First round: 600k ASCII chars ≈ 150k tokens (CharTokenCounter,
        // 4 chars / 1 token) > the default 128k budget — triggers trimming;
        // the second round is small and should be fully kept.
        agent
            .memory
            .record(Message::user("x".repeat(600_000)))
            .await
            .unwrap();
        agent
            .memory
            .record(Message::assistant("first-round reply"))
            .await
            .unwrap();
        agent
            .memory
            .record(Message::user("second round"))
            .await
            .unwrap();
        agent
            .memory
            .record(Message::assistant("second-round reply"))
            .await
            .unwrap();

        let ctx = agent.memory.context().await.unwrap();
        assert_eq!(
            ctx,
            vec![
                Message::user("second round"),
                Message::assistant("second-round reply")
            ]
        );
    }

    /// Round limit: the model keeps requesting tools → Err(TooManyToolRounds);
    /// no further conversations are started.
    #[tokio::test]
    async fn too_many_tool_rounds() {
        let (calc, _calls) = FakeTool::new("calc", "42");
        let mut registry = ToolRegistry::new();
        registry.register(calc);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "calc", "{}")],
            },
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c2", "calc", "{}")],
            },
        ]);
        let mut agent = agent_with_registry(
            fake.clone(),
            registry,
            AgentConfig {
                max_tool_rounds: 2,
                ..Default::default()
            },
        );

        let err = agent.run("Keep computing").await.unwrap_err();
        assert!(matches!(err, AgentError::TooManyToolRounds(2)));
        assert_eq!(fake.requests().len(), 2); // only two conversation rounds were sent
    }

    /// system_prompt: assembled on every request, exactly one System at the
    /// front; not written to Memory.
    #[tokio::test]
    async fn system_prompt_assembled_every_request() {
        let fake = SharedFake::new([
            FakeReply::Text("Hello".into()),
            FakeReply::Text("Goodbye".into()),
        ]);
        let mut agent = agent(fake.clone(), "You are an assistant");

        agent.run("Are you there").await.unwrap();
        agent.run("Any more?").await.unwrap();

        let requests = fake.requests();
        assert_eq!(requests.len(), 2);
        // Each request: exactly one System and it's first; the second
        // request also carries the previous round's conversation history.
        for request in &requests {
            let systems = request
                .messages
                .iter()
                .filter(|m| matches!(m, Message::System(_)))
                .count();
            assert_eq!(systems, 1);
            assert_eq!(request.messages[0], Message::system("You are an assistant"));
        }
        assert_eq!(requests[1].messages.len(), 4); // System + previous round's user/assistant + this round's user
        assert_eq!(requests[1].messages[3], Message::user("Any more?"));
    }

    /// Macro: arms without a system prompt — no System message is
    /// assembled when system_prompt is omitted.
    #[tokio::test]
    async fn macro_arms_without_system_prompt() {
        // Bare arm: no tools, no system prompt.
        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = crate::react_agent!(fake.clone());
        assert_eq!(agent.run("hi").await.unwrap(), "hi");
        assert_eq!(fake.requests()[0].messages.len(), 1); // user only, no System

        // Tool-list arm: no system prompt.
        let (t1, calls1) = FakeTool::new("t1", "one");
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "t1", "{}")],
            },
            FakeReply::Text("done".into()),
        ]);
        let mut agent = crate::react_agent!(fake.clone(), [t1]);
        assert_eq!(agent.run("x").await.unwrap(), "done");
        assert_eq!(calls1.load(Ordering::Relaxed), 1);
        assert_eq!(fake.requests()[1].messages.len(), 3); // user + assistant + tool, no System

        // Registry arm: no system prompt.
        let (t2, _calls2) = FakeTool::new("t2", "two");
        let mut registry = ToolRegistry::new();
        registry.register(t2);
        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = crate::react_agent!(fake.clone(), registry);
        assert_eq!(agent.run("hi").await.unwrap(), "hi");
        assert_eq!(fake.requests()[0].messages.len(), 1); // no System
    }

    /// Macro, three arms: no tools / heterogeneous tool list
    /// (auto-registered) / existing registry.
    #[tokio::test]
    async fn macro_three_arms() {
        // No-tool arm: empty registry, answers directly.
        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = crate::react_agent!(fake.clone(), "");
        assert_eq!(agent.run("hi").await.unwrap(), "hi");
        assert_eq!(fake.requests()[0].messages.len(), 1);

        // Heterogeneous tool-list arm: two tools of different types are
        // auto-registered and actually executed.
        let (t1, calls1) = FakeTool::new("t1", "one");
        let (t2, calls2) = FakeTool::new("t2", "two");
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "t1", "{}")],
            },
            FakeReply::Text("done".into()),
        ]);
        let mut agent = crate::react_agent!(fake.clone(), [t1, t2], "");
        assert_eq!(agent.run("x").await.unwrap(), "done");
        assert_eq!(calls1.load(Ordering::Relaxed), 1);
        assert_eq!(calls2.load(Ordering::Relaxed), 0);

        // Registry arm: the existing registry is passed through as-is.
        let (t3, calls3) = FakeTool::new("t3", "three");
        let mut registry = ToolRegistry::new();
        registry.register(t3);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "t3", "{}")],
            },
            FakeReply::Text("done".into()),
        ]);
        let mut agent = crate::react_agent!(fake, registry, "");
        assert_eq!(agent.run("x").await.unwrap(), "done");
        assert_eq!(calls3.load(Ordering::Relaxed), 1);
    }

    /// All six macro arms construct (isomorphic to the examples/
    /// react_agent.rs doc demo, ensuring the example doc code compiles).
    #[test]
    fn macro_all_arms_compile() {
        struct Echo;
        #[async_trait::async_trait]
        impl Tool for Echo {
            fn schema(&self) -> ToolSchema {
                ToolSchema::new("echo", "Echo", serde_json::json!({}))
            }
            async fn call(
                &self,
                _arguments: serde_json::Value,
                _context: ToolContext<'_>,
            ) -> Result<ToolResult, ToolError> {
                Ok(ToolOutput::text("echo").into())
            }
        }

        fn fake() -> SharedFake {
            SharedFake::new([FakeReply::Text("hi".into())])
        }

        let a1 = crate::react_agent!(fake()); // no tools, no system prompt
        let a2 = crate::react_agent!(fake(), "You are an assistant"); // no tools, with system prompt
        let a3 = crate::react_agent!(fake(), [Echo]); // tool list, no system prompt
        let a4 = crate::react_agent!(fake(), [Echo], "You are an assistant"); // tool list, with system prompt
        let mut registry = ToolRegistry::new();
        registry.register(Echo);
        let a5 = crate::react_agent!(fake(), registry.clone()); // existing registry, no system prompt
        let a6 = crate::react_agent!(fake(), registry, "You are an assistant"); // existing registry, with system prompt
        let _ = (a1, a2, a3, a4, a5, a6);
    }

    /// Empty system prompt: no System message is assembled.
    #[tokio::test]
    async fn empty_system_prompt_skips_system_message() {
        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = agent(fake.clone(), "");

        agent.run("hi").await.unwrap();

        let requests = fake.requests();
        assert_eq!(requests[0].messages.len(), 1);
        assert_eq!(requests[0].messages[0], Message::user("hi"));
    }

    /// Tool failures are not AgentError: the error text is fed back to the
    /// model and the loop continues.
    #[tokio::test]
    async fn tool_failure_returns_text_and_continues() {
        // Failing tool: always Err on execution.
        struct FailingTool;
        #[async_trait::async_trait]
        impl Tool for FailingTool {
            fn schema(&self) -> ToolSchema {
                ToolSchema::new("boom", "Tool that always fails", serde_json::json!({}))
            }
            async fn call(
                &self,
                _arguments: serde_json::Value,
                _context: ToolContext<'_>,
            ) -> Result<ToolResult, ToolError> {
                Err(ToolError::Execution("internal error".into()))
            }
        }

        let mut registry = ToolRegistry::new();
        registry.register(FailingTool);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "boom", "{}")],
            },
            FakeReply::Text("Got it".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());

        let answer = agent.run("Trigger failure").await.unwrap();
        assert_eq!(answer, "Got it");

        // The error was turned into text and fed back (registry
        // semantics); the loop didn't stop.
        let requests = fake.requests();
        assert!(requests[1].messages.iter().any(|m| matches!(
            m,
            Message::ToolResult { content, .. } if content.contains("internal error")
        )));
    }

    // ---- Streaming-path error semantics ----

    /// Streaming: exceeding the round limit terminates with
    /// Err(TooManyToolRounds); no Done / third request.
    #[tokio::test]
    async fn stream_too_many_tool_rounds_terminates() {
        let (calc, _calls) = FakeTool::new("calc", "42");
        let mut registry = ToolRegistry::new();
        registry.register(calc);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "calc", "{}")],
            },
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c2", "calc", "{}")],
            },
        ]);
        let mut agent = agent_with_registry(
            fake.clone(),
            registry,
            AgentConfig {
                max_tool_rounds: 2,
                ..Default::default()
            },
        );

        let mut stream = agent.run_stream("Keep computing").await.unwrap();
        let chunks: Vec<Result<MessageChunk, AgentError>> = stream.by_ref().collect().await;
        // Both rounds fully execute (tool rounds run before the limit
        // check: ToolCall+ToolResult ×2), and the top-of-round check of the
        // third round hits the limit: Err terminates, no Done, only two
        // conversation rounds sent.
        assert_eq!(chunks.len(), 5);
        assert!(matches!(
            chunks.last(),
            Some(Err(AgentError::TooManyToolRounds(2)))
        ));
        assert_eq!(fake.requests().len(), 2);
    }

    /// Streaming: entry record(user) failure → direct Err, no stream
    /// produced.
    #[tokio::test]
    async fn stream_entry_record_user_failure_returns_error() {
        struct FailingUserMemory;
        #[async_trait::async_trait]
        impl Memory for FailingUserMemory {
            async fn record(&mut self, _message: Message) -> Result<(), MemoryError> {
                Err(MemoryError::Storage("disk full".into()))
            }
            async fn context(&self) -> Result<Vec<Message>, MemoryError> {
                Ok(Vec::new())
            }
        }

        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = agent(fake.clone(), "").with_memory(FailingUserMemory);
        let err = match agent.run_stream("hi").await {
            Err(e) => e,
            Ok(_) => panic!("expected input recording failure to return Err directly"),
        };
        assert!(matches!(err, AgentError::Memory(MemoryError::Storage(_))));
    }

    /// Streaming: in-round context() failure → Err terminates the stream
    /// (no further chunks).
    #[tokio::test]
    async fn stream_context_failure_terminates_with_err() {
        struct FailingContextMemory;
        #[async_trait::async_trait]
        impl Memory for FailingContextMemory {
            async fn record(&mut self, _message: Message) -> Result<(), MemoryError> {
                Ok(())
            }
            async fn context(&self) -> Result<Vec<Message>, MemoryError> {
                Err(MemoryError::Storage("disk full".into()))
            }
        }

        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = agent(fake.clone(), "").with_memory(FailingContextMemory);
        let mut stream = agent.run_stream("hi").await.unwrap();
        let chunks: Vec<Result<MessageChunk, AgentError>> = stream.by_ref().collect().await;
        assert_eq!(chunks.len(), 1);
        assert!(matches!(
            chunks[0],
            Err(AgentError::Memory(MemoryError::Storage(_)))
        ));
    }

    /// Streaming: Assistant record failure → Err terminates the stream
    /// (Delta already produced, no Done).
    #[tokio::test]
    async fn stream_assistant_record_failure_terminates_with_err() {
        struct FailingAssistantMemory;
        #[async_trait::async_trait]
        impl Memory for FailingAssistantMemory {
            async fn record(&mut self, message: Message) -> Result<(), MemoryError> {
                if matches!(message, Message::Assistant { .. }) {
                    Err(MemoryError::Storage("disk full".into()))
                } else {
                    Ok(())
                }
            }
            async fn context(&self) -> Result<Vec<Message>, MemoryError> {
                Ok(Vec::new())
            }
        }

        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = agent(fake.clone(), "").with_memory(FailingAssistantMemory);
        let mut stream = agent.run_stream("hi").await.unwrap();
        let chunks: Vec<Result<MessageChunk, AgentError>> = stream.by_ref().collect().await;
        assert_eq!(chunks.len(), 2); // Delta + Err
        assert!(matches!(
            chunks.last(),
            Some(Err(AgentError::Memory(MemoryError::Storage(_))))
        ));
    }

    /// Streaming: tool failures turn into text fed back; the stream
    /// continues to Done (same semantics as run).
    #[tokio::test]
    async fn stream_tool_failure_returns_text_and_continues() {
        struct FailingTool;
        #[async_trait::async_trait]
        impl Tool for FailingTool {
            fn schema(&self) -> ToolSchema {
                ToolSchema::new("boom", "Tool that always fails", serde_json::json!({}))
            }
            async fn call(
                &self,
                _arguments: serde_json::Value,
                _context: ToolContext<'_>,
            ) -> Result<ToolResult, ToolError> {
                Err(ToolError::Execution("internal error".into()))
            }
        }

        let mut registry = ToolRegistry::new();
        registry.register(FailingTool);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "boom", "{}")],
            },
            FakeReply::Text("Got it".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());

        let mut stream = agent.run_stream("Trigger failure").await.unwrap();
        let chunks: Vec<MessageChunk> = stream.by_ref().map(|e| e.unwrap()).collect().await;
        assert!(matches!(
            &chunks[1],
            MessageChunk::ToolResult { content, .. } if content.contains("internal error")
        ));
        // The stream continues: the model answers directly in the second
        // round and Done follows.
        assert_done_summary(chunks.last().unwrap(), 2, 1, Usage::default(), true);
    }

    /// Streaming: multiple tools in one round, chunks produced strictly in
    /// request order → result order, with correct summary counts.
    #[tokio::test]
    async fn stream_multiple_tools_same_round() {
        let (calc_a, _calls) = FakeTool::new("calc_a", "A");
        let (calc_b, _calls) = FakeTool::new("calc_b", "B");
        let mut registry = ToolRegistry::new();
        registry.register(calc_a).register(calc_b);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "calc_a", "{}"), call("c2", "calc_b", "{}")],
            },
            FakeReply::Text("Done".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());

        let mut stream = agent.run_stream("Compute").await.unwrap();
        let chunks: Vec<MessageChunk> = stream.by_ref().map(|e| e.unwrap()).collect().await;
        assert_eq!(chunks.len(), 6);
        assert_eq!(
            &chunks[..5],
            &[
                MessageChunk::ToolCall {
                    id: "c1".into(),
                    name: "calc_a".into(),
                    arguments: "{}".into()
                },
                MessageChunk::ToolCall {
                    id: "c2".into(),
                    name: "calc_b".into(),
                    arguments: "{}".into()
                },
                MessageChunk::ToolResult {
                    id: "c1".into(),
                    name: "calc_a".into(),
                    content: "A".into()
                },
                MessageChunk::ToolResult {
                    id: "c2".into(),
                    name: "calc_b".into(),
                    content: "B".into()
                },
                MessageChunk::Delta("Done".into()),
            ]
        );
        assert_done_summary(&chunks[5], 2, 2, Usage::default(), true);
    }

    /// Error semantics are equivalent across paths: with the same failure
    /// script (round limit), run and run_stream agree on error type and
    /// request history.
    #[tokio::test]
    async fn run_and_stream_error_semantics_equivalent() {
        let script = |fake: &SharedFake| {
            let (calc, _calls) = FakeTool::new("calc", "42");
            let mut registry = ToolRegistry::new();
            registry.register(calc);
            agent_with_registry(
                fake.clone(),
                registry,
                AgentConfig {
                    max_tool_rounds: 1,
                    ..Default::default()
                },
            )
        };

        // Non-streaming: round-limit Err.
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "calc", "{}")],
            },
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c2", "calc", "{}")],
            },
        ]);
        let mut agent = script(&fake);
        let run_err = agent.run("Compute").await.unwrap_err();

        // Streaming: same script, terminates with Err.
        let fake2 = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "calc", "{}")],
            },
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c2", "calc", "{}")],
            },
        ]);
        let mut agent = script(&fake2);
        let mut stream = agent.run_stream("Compute").await.unwrap();
        let chunks: Vec<Result<MessageChunk, AgentError>> = stream.by_ref().collect().await;
        let stream_err = chunks.into_iter().find_map(|e| e.err());

        // Same error type, same request count.
        assert_eq!(run_err, AgentError::TooManyToolRounds(1));
        assert_eq!(stream_err, Some(AgentError::TooManyToolRounds(1)));
        assert_eq!(fake.requests().len(), fake2.requests().len());
    }

    /// run_id is unique across instances: two back-to-back instances have
    /// different ids on their first run.
    #[tokio::test]
    async fn run_id_differs_across_instances() {
        let id_a = RunContext::generated().run_id;
        let id_b = RunContext::generated().run_id;
        assert_ne!(
            id_a, id_b,
            "back-to-back instances must not collide on run_id"
        );
        // Same format as the trace tests: run-{ts}-{n}.
        assert!(id_a.starts_with("run-") && id_b.starts_with("run-"));
    }

    /// Non-streaming: in-round context() failure → Err(Memory) passes
    /// through (symmetric with streaming).
    #[tokio::test]
    async fn run_context_failure_returns_memory_error() {
        struct FailingContextMemory;
        #[async_trait::async_trait]
        impl Memory for FailingContextMemory {
            async fn record(&mut self, _message: Message) -> Result<(), MemoryError> {
                Ok(())
            }
            async fn context(&self) -> Result<Vec<Message>, MemoryError> {
                Err(MemoryError::Storage("disk full".into()))
            }
        }

        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = agent(fake.clone(), "").with_memory(FailingContextMemory);
        let err = agent.run("hi").await.unwrap_err();
        assert!(matches!(err, AgentError::Memory(MemoryError::Storage(_))));
        // No conversation was started.
        assert!(fake.requests().is_empty());
    }

    /// Empty-stream semantics: end-of-round wrap-up with no answer
    /// recorded, and a normal Done.
    ///
    /// Note: reporting `Err` for truncated / empty streams is the Provider
    /// implementation's job; this test covers the Agent's behavior when the
    /// Provider delivers a legitimately empty stream.
    #[tokio::test]
    async fn stream_empty_provider_stream_yields_empty_answer() {
        struct EmptyStreamProvider;
        #[async_trait::async_trait]
        impl Provider for EmptyStreamProvider {
            async fn chat_with_context(
                &self,
                _r: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<ChatResponse, ProviderError> {
                unreachable!("this test uses streaming only")
            }
            async fn stream_chat_with_context(
                &self,
                _r: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError>
            {
                Ok(Box::pin(futures::stream::empty()))
            }
        }

        let mut agent = ReActAgent::new(EmptyStreamProvider, ToolRegistry::new(), "");
        let chunks: Vec<Result<MessageChunk, AgentError>> = {
            let mut stream = agent.run_stream("hi").await.unwrap();
            stream.by_ref().collect().await
        };
        // Empty answer: not recorded + normal Done (rounds counts 1).
        assert_eq!(chunks.len(), 1);
        let chunk = chunks[0].as_ref().unwrap();
        assert_done_summary_with_finish(chunk, 1, 0, Usage::default(), true, None);
        // The memory only holds the user input; no empty Assistant was
        // recorded.
        assert_eq!(
            agent.memory.context().await.unwrap(),
            vec![Message::user("hi")]
        );
    }

    /// Streaming: event order Delta → ToolCall → ToolResult → Done;
    /// reasoning never enters MessageChunk (only recorded with the
    /// message).
    #[tokio::test]
    async fn stream_event_order() {
        let (calc, _calls) = FakeTool::new("calc", "42");
        let mut registry = ToolRegistry::new();
        registry.register(calc);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "Thinking: ".into(),
                calls: vec![call("c1", "calc", "{}")],
            },
            FakeReply::TextWithReasoning {
                content: "The answer is 42".into(),
                reasoning: "Reasoning steps".into(),
            },
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());

        let mut stream = agent.run_stream("Compute").await.unwrap();
        let events: Vec<MessageChunk> = stream.by_ref().map(|e| e.unwrap()).collect().await;
        assert_eq!(events.len(), 5);
        assert_eq!(
            &events[..4],
            &[
                MessageChunk::Delta("Thinking: ".into()),
                MessageChunk::ToolCall {
                    id: "c1".into(),
                    name: "calc".into(),
                    arguments: "{}".into()
                },
                MessageChunk::ToolResult {
                    id: "c1".into(),
                    name: "calc".into(),
                    content: "42".into()
                },
                MessageChunk::Delta("The answer is 42".into()),
            ]
        );
        assert_done_summary(&events[4], 2, 1, Usage::default(), true);
    }

    /// Pure tool round: no text, no Delta dispatched; the tool call still
    /// executes and feeds back.
    #[tokio::test]
    async fn stream_pure_tool_round_no_delta() {
        let (calc, _calls) = FakeTool::new("calc", "42");
        let mut registry = ToolRegistry::new();
        registry.register(calc);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "calc", "{}")],
            },
            FakeReply::Text("42".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());

        let mut stream = agent.run_stream("Compute").await.unwrap();
        let events: Vec<MessageChunk> = stream.by_ref().map(|e| e.unwrap()).collect().await;
        assert_eq!(events.len(), 4);
        assert_eq!(
            &events[..3],
            &[
                MessageChunk::ToolCall {
                    id: "c1".into(),
                    name: "calc".into(),
                    arguments: "{}".into()
                },
                MessageChunk::ToolResult {
                    id: "c1".into(),
                    name: "calc".into(),
                    content: "42".into()
                },
                MessageChunk::Delta("42".into()),
            ]
        );
        assert_done_summary(&events[3], 2, 1, Usage::default(), true);
    }

    /// The Done summary carries real token usage (injected via
    /// FakeReply::WithUsage, summed across rounds), with rounds and tool
    /// counts tallied by execution.
    #[tokio::test]
    async fn stream_done_summary_accumulates_usage() {
        let (calc, _calls) = FakeTool::new("calc", "42");
        let mut registry = ToolRegistry::new();
        registry.register(calc);
        let fake = SharedFake::new([
            FakeReply::WithUsage {
                reply: Box::new(FakeReply::ToolCalls {
                    content: "".into(),
                    calls: vec![call("c1", "calc", "{}")],
                }),
                usage: Usage::new(10, 2),
            },
            FakeReply::text_with_usage("42", Usage::new(20, 5)),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());

        let mut stream = agent.run_stream("Compute").await.unwrap();
        let events: Vec<MessageChunk> = stream.by_ref().map(|e| e.unwrap()).collect().await;

        // Two conversation rounds (tool round + direct-answer round), one
        // tool execution; usage accumulates per round.
        assert_done_summary(events.last().unwrap(), 2, 1, Usage::new(30, 7), false);
    }

    /// Mixed usage presence: only reported parts are summed, and the
    /// summary marks the sum as a lower bound (usage_omitted).
    #[tokio::test]
    async fn summary_tracks_omitted_usage_rounds() {
        let (calc, _calls) = FakeTool::new("calc", "42");
        let mut registry = ToolRegistry::new();
        registry.register(calc);
        // Round 1 (tool round) reports nothing; round 2 (direct answer)
        // reports usage: the sum is the reported part only, and the summary
        // flags the omission.
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "calc", "{}")],
            },
            FakeReply::text_with_usage("42", Usage::new(20, 5)),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());

        let mut stream = agent.run_stream("Compute").await.unwrap();
        let events: Vec<MessageChunk> = stream.by_ref().map(|e| e.unwrap()).collect().await;

        assert_done_summary(events.last().unwrap(), 2, 1, Usage::new(20, 5), true);
    }

    /// run and run_stream with the same script share semantics: identical
    /// final text, identical request history (identical records).
    #[tokio::test]
    async fn run_and_stream_same_semantics() {
        let script = [
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "calc", "{}")],
            },
            FakeReply::Text("42".into()),
        ];
        let (calc, _calls) = FakeTool::new("calc", "42");

        // Non-streaming path.
        let mut registry = ToolRegistry::new();
        registry.register(calc.clone());
        let fake1 = SharedFake::new(script.clone());
        let mut agent1 = agent_with_registry(fake1.clone(), registry, AgentConfig::default());
        let answer1 = agent1.run("Compute").await.unwrap();

        // Streaming path.
        let mut registry = ToolRegistry::new();
        registry.register(calc);
        let fake2 = SharedFake::new(script);
        let mut agent2 = agent_with_registry(fake2.clone(), registry, AgentConfig::default());
        let events: Vec<MessageChunk> = agent2
            .run_stream("Compute")
            .await
            .unwrap()
            .map(|e| e.unwrap())
            .collect()
            .await;
        // Final answer = concatenation of the Deltas, verbatim.
        let answer2: String = events
            .iter()
            .filter_map(|e| match e {
                MessageChunk::Delta(d) => Some(d.as_str()),
                _ => None,
            })
            .collect();
        assert_eq!(answer1, answer2);
        assert_eq!(answer1, "42");

        // Both paths send the same message sequence to the model (same
        // recorded history).
        assert_eq!(fake1.requests(), fake2.requests());
    }

    /// In-stream error: after an Err event the stream terminates; no Done.
    ///
    /// Note: this path (an error mid-event-stream after establishment)
    /// can't be built with FakeReply::Error — its semantics are "this round
    /// failed; the method returns Err directly"; here a custom Provider
    /// produces a stream that errors midway.
    #[tokio::test]
    async fn stream_error_terminates_without_done() {
        struct FailInStream;
        #[async_trait::async_trait]
        impl Provider for FailInStream {
            async fn chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<ChatResponse, ProviderError> {
                unreachable!("this test uses streaming path only")
            }
            async fn stream_chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError>
            {
                Ok(Box::pin(futures::stream::iter(vec![
                    Ok(StreamEvent::Delta("hi".into())),
                    Err(ProviderError::Protocol {
                        message: "boom".into(),
                    }),
                ])))
            }
        }

        let mut agent = ReActAgent::new(FailInStream, ToolRegistry::new(), "");

        let mut stream = agent.run_stream("two").await.unwrap();
        assert_eq!(
            stream.next().await.unwrap().unwrap(),
            MessageChunk::Delta("hi".into())
        );
        assert!(matches!(
            stream.next().await.unwrap(),
            Err(AgentError::Provider(ProviderError::Protocol { message: m })) if m == "boom"
        ));
        assert!(stream.next().await.is_none()); // terminated, no Done
    }

    /// Script exhausted (one extra round): Err(ProviderError::Protocol("script
    /// exhausted")).
    #[tokio::test]
    async fn script_exhausted_fails_explicitly() {
        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = agent(fake, "");
        agent.run("one").await.unwrap();

        let err = agent.run("two").await.unwrap_err();
        assert!(
            matches!(err, AgentError::Provider(ProviderError::Protocol { message: m }) if m.contains("exhausted"))
        );
    }

    /// Shared state flows: the loop injects the agent-held state into every
    /// tool call, and tools read/write the same instance (the application
    /// side can also read/write across runs).
    #[tokio::test]
    async fn shared_state_flows_to_tools() {
        struct CounterTool;
        #[async_trait::async_trait]
        impl Tool for CounterTool {
            fn schema(&self) -> ToolSchema {
                ToolSchema::new("counter", "Count", serde_json::json!({}))
            }
            async fn call(
                &self,
                _arguments: serde_json::Value,
                context: ToolContext<'_>,
            ) -> Result<ToolResult, ToolError> {
                let state = context.state;
                state.with_mut::<usize>(|n| *n += 1);
                Ok(ToolOutput::text(format!("count={}", state.get::<usize>().unwrap_or(0))).into())
            }
        }

        let mut registry = ToolRegistry::new();
        registry.register(CounterTool);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "counter", "{}"), call("c2", "counter", "{}")],
            },
            FakeReply::Text("done".into()),
        ]);
        let state = SharedState::new();
        state.insert(0usize);
        let mut agent = ReActAgent::new(fake, registry, "").with_state(state.clone());

        agent.run("Count").await.unwrap();

        // Two calls in the same round share one instance: the count
        // accumulates, and the application side reads the final cross-run
        // value.
        assert_eq!(state.get::<usize>(), Some(2));
    }

    /// Memory errors pass through (custom Memory injected via
    /// with_memory).
    #[tokio::test]
    async fn memory_error_passthrough() {
        struct FailingMemory;
        #[async_trait::async_trait]
        impl Memory for FailingMemory {
            async fn record(&mut self, _message: Message) -> Result<(), MemoryError> {
                Err(MemoryError::Storage("disk full".into()))
            }
            async fn context(&self) -> Result<Vec<Message>, MemoryError> {
                Ok(Vec::new())
            }
        }

        let mut agent = ReActAgent::new(
            FakeProvider::new([FakeReply::Text("hi".into())]),
            ToolRegistry::new(),
            "",
        )
        .with_memory(FailingMemory);
        let err = agent.run("hi").await.unwrap_err();
        assert!(matches!(err, AgentError::Memory(MemoryError::Storage(_))));
    }

    /// Streaming: tool-result record failure → an Err item terminates the
    /// stream; no further chunks.
    /// Errors are produced as Err items and terminate the stream — the
    /// break must exit the whole round loop, otherwise the stream would
    /// keep producing the next round after the Err (the memory would lack
    /// the ToolResult, leaving the next round's message sequence
    /// incomplete).
    #[tokio::test]
    async fn stream_tool_result_record_failure_terminates_stream() {
        struct FailingToolResultMemory;
        #[async_trait::async_trait]
        impl Memory for FailingToolResultMemory {
            async fn record(&mut self, message: Message) -> Result<(), MemoryError> {
                if matches!(message, Message::ToolResult { .. }) {
                    Err(MemoryError::Storage("disk full".into()))
                } else {
                    Ok(())
                }
            }
            async fn context(&self) -> Result<Vec<Message>, MemoryError> {
                Ok(Vec::new())
            }
        }

        let (calc, _calls) = FakeTool::new("calc", "42");
        let mut registry = ToolRegistry::new();
        registry.register(calc);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "calc", "{}")],
            },
            FakeReply::Text("42".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default())
            .with_memory(FailingToolResultMemory);

        let mut stream = agent.run_stream("Compute").await.unwrap();
        let chunks: Vec<Result<MessageChunk, AgentError>> = stream.by_ref().collect().await;
        // ToolCall → ToolResult → Err (terminates): nothing after the Err.
        assert_eq!(chunks.len(), 3);
        assert!(matches!(
            chunks[2],
            Err(AgentError::Memory(MemoryError::Storage(_)))
        ));
    }

    /// AgentConfig.options passes through: every round's ChatRequest
    /// carries the configured model parameters.
    #[tokio::test]
    async fn config_options_forwarded_to_chat_request() {
        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = agent(fake.clone(), "").with_config(AgentConfig {
            max_tool_rounds: 10,
            options: ModelOptions {
                temperature: Some(0.2),
                max_tokens: Some(128),
                extra: Default::default(),
                structured: None,
            },
            ..Default::default()
        });
        agent.run("hi").await.unwrap();

        let req = &fake.requests()[0];
        assert_eq!(req.options.temperature, Some(0.2));
        assert_eq!(req.options.max_tokens, Some(128));
    }

    #[tokio::test]
    async fn request_options_replace_config_options() {
        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = agent(fake.clone(), "").with_config(AgentConfig {
            options: ModelOptions {
                temperature: Some(0.2),
                max_tokens: Some(128),
                ..Default::default()
            },
            ..Default::default()
        });
        agent
            .run_request(RunRequest::text("hi").with_options(ModelOptions {
                max_tokens: Some(64),
                ..Default::default()
            }))
            .await
            .unwrap();

        let req = &fake.requests()[0];
        assert_eq!(req.options.temperature, None);
        assert_eq!(req.options.max_tokens, Some(64));
    }

    // ---- Cancellation through RunContext ----

    /// Pre-round cancellation: with an already-cancelled token,
    /// run_request_with_context returns Cancelled immediately, having started no
    /// conversation; the user message is already recorded (kept, consistent
    /// with the main path).
    #[tokio::test]
    async fn cancelled_before_run_returns_cancelled() {
        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = agent(fake.clone(), "");
        let token = CancellationToken::new();
        token.cancel();

        let err = agent
            .run_request_with_context(RunRequest::text("hi"), cancellation_context(&token))
            .await
            .unwrap_err();
        assert!(matches!(err, AgentError::Cancelled));
        assert_eq!(fake.requests().len(), 0); // no conversation started
        assert_eq!(agent.memory.context().await.unwrap().len(), 1); // user only
    }

    /// Pre-cancelled token: both paths report the same rounds in RunEnded
    /// (incremented at the start of the round, both report 1).
    #[tokio::test]
    async fn pre_cancelled_token_rounds_consistent_across_paths() {
        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let token = CancellationToken::new();
        token.cancel();

        // Non-streaming path.
        let (mut run_agent, mut rx) = attach_channel(agent(fake.clone(), ""));
        let err = run_agent
            .run_request_with_context(RunRequest::text("hi"), cancellation_context(&token))
            .await
            .unwrap_err();
        assert!(matches!(err, AgentError::Cancelled));
        drop(run_agent);
        let events = drain(&mut rx).await;
        let rounds_run = match react_event(&**events.last().unwrap()) {
            ReActEvent::RunEnded { summary, error, .. } => {
                assert_eq!(error, &Some(AgentError::Cancelled));
                summary.rounds
            }
            _ => panic!("expected RunEnded"),
        };

        // Streaming path.
        let (mut stream_agent, mut rx) = attach_channel(agent(fake, ""));
        let mut stream = stream_agent
            .run_stream_request_with_context(RunRequest::text("hi"), cancellation_context(&token))
            .await
            .unwrap();
        let chunks: Vec<MessageChunk> = stream.by_ref().map(|e| e.unwrap()).collect().await;
        assert!(chunks.contains(&MessageChunk::Cancelled));
        drop(stream);
        drop(stream_agent);
        let events = drain(&mut rx).await;
        let rounds_stream = match react_event(&**events.last().unwrap()) {
            ReActEvent::RunEnded { summary, error, .. } => {
                assert_eq!(error, &Some(AgentError::Cancelled));
                summary.rounds
            }
            _ => panic!("expected RunEnded"),
        };

        assert_eq!(rounds_run, rounds_stream);
        assert_eq!(rounds_run, 1);
    }

    /// Immediate cancellation mid-conversation: cancelling while chat is
    /// pending makes run return Cancelled immediately (dropping the
    /// in-flight request); nothing is recorded for this round, leaving no
    /// orphan Assistant in the memory.
    #[tokio::test]
    async fn cancel_during_chat_drops_inflight() {
        struct PendingProvider;
        #[async_trait::async_trait]
        impl Provider for PendingProvider {
            async fn chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<ChatResponse, ProviderError> {
                std::future::pending().await // never completes: simulates a slow LLM
            }
            async fn stream_chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError>
            {
                unreachable!("this test uses non-streaming path only")
            }
        }

        let mut agent = ReActAgent::new(PendingProvider, ToolRegistry::new(), "");
        let token = CancellationToken::new();
        let result = tokio::select! {
            r = agent.run_request_with_context(RunRequest::text("hi"), cancellation_context(&token)) => r.map(|output| output.answer),
            _ = async {
                tokio::time::sleep(Duration::from_millis(20)).await;
                token.cancel();
                std::future::pending::<()>().await; // stay pending so the run branch wins
            } => unreachable!("cancellation branch only sends a signal"),
        };
        assert!(matches!(result, Err(AgentError::Cancelled)));
        assert_eq!(agent.memory.context().await.unwrap().len(), 1); // user only
    }

    /// Non-cancellation error propagation: when the Provider returns a
    /// Timeout, both run / run_stream return AgentError::Provider(Timeout),
    /// and the memory holds no orphan messages (user only, no half-recorded
    /// residue).
    #[tokio::test]
    async fn provider_error_propagates_from_both_paths() {
        struct FailProvider(ProviderError);
        #[async_trait::async_trait]
        impl Provider for FailProvider {
            async fn chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<ChatResponse, ProviderError> {
                Err(self.0.clone())
            }
            async fn stream_chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError>
            {
                Err(self.0.clone())
            }
        }

        let err = ProviderError::Timeout(TimeoutStage::Request);
        // Non-streaming: the error is returned directly.
        let mut agent = ReActAgent::new(FailProvider(err.clone()), ToolRegistry::new(), "");
        let got = agent.run("hi").await.unwrap_err();
        assert!(matches!(
            got,
            AgentError::Provider(ProviderError::Timeout(_))
        ));
        assert_eq!(agent.memory.context().await.unwrap().len(), 1); // user only

        // Streaming: the error terminates the stream as an in-stream Err
        // item; nothing follows (no Done).
        let mut agent = ReActAgent::new(FailProvider(err.clone()), ToolRegistry::new(), "");
        let mut stream = agent.run_stream("hi").await.unwrap();
        let item = stream.next().await.unwrap().unwrap_err();
        assert!(matches!(
            item,
            AgentError::Provider(ProviderError::Timeout(_))
        ));
        assert!(stream.next().await.is_none());
        drop(stream); // release the borrow of agent before inspecting the memory
        assert_eq!(agent.memory.context().await.unwrap().len(), 1);
    }

    #[tokio::test]
    async fn deadline_exceeded_during_chat_is_distinct_from_provider_timeout() {
        struct PendingProvider;
        #[async_trait::async_trait]
        impl Provider for PendingProvider {
            async fn chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<ChatResponse, ProviderError> {
                std::future::pending().await
            }

            async fn stream_chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError>
            {
                std::future::pending().await
            }
        }

        let mut agent = ReActAgent::new(PendingProvider, ToolRegistry::new(), "");
        let err = agent
            .run_request_with_context(
                RunRequest::text("hi"),
                RunContext::new("deadline").with_timeout(Duration::from_millis(10)),
            )
            .await
            .unwrap_err();
        assert_eq!(err, AgentError::DeadlineExceeded);
        assert_eq!(agent.memory.context().await.unwrap().len(), 1);
    }

    #[tokio::test]
    async fn streaming_deadline_exceeded_terminates_with_error_item() {
        struct PendingStreamProvider;
        #[async_trait::async_trait]
        impl Provider for PendingStreamProvider {
            async fn chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<ChatResponse, ProviderError> {
                unreachable!("this test uses streaming only")
            }

            async fn stream_chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError>
            {
                Ok(Box::pin(futures::stream::pending()))
            }
        }

        let mut agent = ReActAgent::new(PendingStreamProvider, ToolRegistry::new(), "");
        let mut stream = agent
            .run_stream_request_with_context(
                RunRequest::text("hi"),
                RunContext::new("stream-deadline").with_timeout(Duration::from_millis(10)),
            )
            .await
            .unwrap();
        assert_eq!(
            stream.next().await.unwrap().unwrap_err(),
            AgentError::DeadlineExceeded
        );
    }

    /// Structured output: an answer conforming to the schema is returned
    /// directly (validation passes in a single round).
    #[cfg(feature = "structured")]
    #[tokio::test]
    async fn structured_output_valid_answer_passes() {
        let schema = serde_json::json!({
            "type": "object",
            "properties": { "city": { "type": "string" } },
            "required": ["city"],
        });
        let fake = SharedFake::new([FakeReply::Text(r#"{"city":"Beijing"}"#.into())]);
        let mut agent = agent(fake.clone(), "").with_structured_output(schema);
        let answer = agent.run("Beijing weather").await.unwrap();
        assert_eq!(answer, r#"{"city":"Beijing"}"#);
        assert_eq!(fake.requests().len(), 1); // passed in one round, no retry
    }

    /// Structured output: invalid JSON is first fed back to the model for
    /// retry; the corrected second round passes.
    #[cfg(feature = "structured")]
    #[tokio::test]
    async fn structured_output_retries_after_invalid_answer() {
        let schema = serde_json::json!({
            "type": "object",
            "properties": { "city": { "type": "string" } },
            "required": ["city"],
        });
        let fake = SharedFake::new([
            FakeReply::Text("Not JSON".into()),
            FakeReply::Text(r#"{"city":"Beijing"}"#.into()),
        ]);
        let mut agent = agent(fake.clone(), "").with_structured_output(schema);
        let answer = agent.run("Beijing weather").await.unwrap();
        assert_eq!(answer, r#"{"city":"Beijing"}"#);
        assert_eq!(fake.requests().len(), 2); // round one fails, round two passes after feedback
        // The validation-failure feedback was recorded (a User message);
        // the memory contains the retry round.
        let context = agent.memory.context().await.unwrap();
        assert!(context.iter().any(|m| matches!(
            m,
            Message::User(blocks) if blocks.iter().any(|b| matches!(b, ContentBlock::Text(t) if t.contains("JSON schema validation")))
        )));
    }

    /// Structured output: validation retries have an independent budget
    /// (`max_structured_retries`); exhausting it without success fails the
    /// run — the tool-round limit is not consumed, and no "tool round limit
    /// exceeded" is reported.
    #[cfg(feature = "structured")]
    #[tokio::test]
    async fn structured_output_exhausts_retry_budget() {
        let schema = serde_json::json!({ "type": "object" });
        let fake = SharedFake::new([
            FakeReply::Text("bad1".into()),
            FakeReply::Text("bad2".into()),
            FakeReply::Text("bad3".into()),
            FakeReply::Text("bad4".into()),
        ]);
        // Even a tool-round limit of 1 doesn't affect structured retries
        // (independent budget).
        let mut agent = agent(fake.clone(), "").with_config(AgentConfig {
            max_tool_rounds: 1,
            max_structured_retries: 3,
            ..Default::default()
        });
        agent = agent.with_structured_output(schema);
        let err = agent.run("hi").await.unwrap_err();
        assert!(matches!(err, AgentError::StructuredRetriesExhausted(3)));
        assert_eq!(fake.requests().len(), 4); // 4 attempts: the 4th failure hits the limit
    }

    /// Typed run: the schema is auto-generated from the type and injected
    /// into the request; after validation passes, the answer deserializes
    /// into the target type.
    #[cfg(feature = "structured")]
    #[tokio::test]
    async fn typed_output_parses_valid_answer() {
        #[derive(Debug, Deserialize, JsonSchema)]
        struct Weather {
            city: String,
        }
        let fake = SharedFake::new([FakeReply::Text(r#"{"city":"Beijing"}"#.into())]);
        let mut agent = agent(fake.clone(), "");
        let weather: Weather = agent.run_typed("Beijing weather").await.unwrap();
        assert_eq!(weather.city, "Beijing");
        assert_eq!(fake.requests().len(), 1); // passed in one round, no retry
        // This run's generated schema was injected into the request (the
        // config is untouched — no pollution of later text runs).
        assert!(fake.requests()[0].options.structured.is_some());
        assert!(agent.config.options.structured.is_none());
    }

    #[cfg(feature = "structured")]
    #[tokio::test]
    async fn typed_run_request_returns_value_and_output() {
        #[derive(Debug, Deserialize, JsonSchema, PartialEq)]
        struct Weather {
            city: String,
        }
        let fake = SharedFake::new([FakeReply::Text(r#"{"city":"Beijing"}"#.into())]);
        let mut agent = agent(fake.clone(), "");
        let typed = agent
            .run_typed_request_with_context::<Weather>(
                RunRequest::text("Beijing weather"),
                RunContext::new("typed-1"),
            )
            .await
            .unwrap();

        assert_eq!(
            typed.value,
            Weather {
                city: "Beijing".into()
            }
        );
        assert_eq!(typed.output.run_id, "typed-1");
        assert_eq!(typed.output.answer, r#"{"city":"Beijing"}"#);
        assert_eq!(
            typed.output.final_message,
            Message::assistant(r#"{"city":"Beijing"}"#)
        );
    }

    #[cfg(feature = "structured")]
    #[tokio::test]
    async fn typed_schema_overrides_request_and_config_schema() {
        #[derive(Debug, Deserialize, JsonSchema)]
        struct Weather {
            city: String,
        }
        let config_schema = serde_json::json!({
            "type": "object",
            "properties": { "config_only": { "type": "string" } },
            "required": ["config_only"],
        });
        let request_schema = serde_json::json!({
            "type": "object",
            "properties": { "request_only": { "type": "string" } },
            "required": ["request_only"],
        });
        let fake = SharedFake::new([FakeReply::Text(r#"{"city":"Beijing"}"#.into())]);
        let mut agent = agent(fake.clone(), "").with_structured_output(config_schema);
        let options = ModelOptions {
            structured: Some(request_schema),
            ..Default::default()
        };

        let weather: Weather = agent
            .run_typed_request(RunRequest::text("Beijing weather").with_options(options))
            .await
            .unwrap()
            .value;

        assert_eq!(weather.city, "Beijing");
        let requests = fake.requests();
        let structured = requests[0]
            .options
            .structured
            .as_ref()
            .expect("typed schema should be sent");
        let props = &structured["properties"];
        assert!(props.get("city").is_some());
        assert!(props.get("request_only").is_none());
        assert!(props.get("config_only").is_none());
    }

    /// Typed run: an invalid answer is fed back for retry first; the
    /// corrected second round deserializes successfully.
    #[cfg(feature = "structured")]
    #[tokio::test]
    async fn typed_output_retries_then_parses() {
        #[derive(Debug, Deserialize, JsonSchema)]
        struct Weather {
            city: String,
        }
        let fake = SharedFake::new([
            FakeReply::Text("Not JSON".into()),
            FakeReply::Text(r#"{"city":"Beijing"}"#.into()),
        ]);
        let mut agent = agent(fake.clone(), "");
        let weather: Weather = agent.run_typed("Beijing weather").await.unwrap();
        assert_eq!(weather.city, "Beijing");
        assert_eq!(fake.requests().len(), 2);
    }

    /// Typed run: schema disagrees with the serde type (the custom schema
    /// says string while the type expects number) → validation passes but
    /// deserialization fails → StructuredParse (the schemars-generated
    /// path agrees with serde by default; conflicts only come from
    /// user-custom schemas).
    #[cfg(feature = "structured")]
    #[tokio::test]
    async fn typed_output_parse_failure_on_schema_mismatch() {
        fn string_schema(_: &mut schemars::SchemaGenerator) -> schemars::Schema {
            serde_json::from_value(serde_json::json!({ "type": "string" })).unwrap()
        }
        #[derive(Debug, Deserialize, JsonSchema)]
        #[allow(dead_code)] // error-path test: the field serves deserialization validation, never read
        struct Weather {
            #[schemars(schema_with = "string_schema")]
            temperature: i32,
        }
        let fake = SharedFake::new([FakeReply::Text(r#"{"temperature":"30"}"#.into())]);
        let mut agent = agent(fake.clone(), "");
        let err: AgentError = agent
            .run_typed::<Weather>("Beijing weather")
            .await
            .unwrap_err();
        assert!(matches!(err, AgentError::StructuredParse(_)));
    }

    /// TypedAgent interface: code with the generic bound `A: TypedAgent`
    /// can call run_typed on any implementation (independent of the
    /// concrete type; Box<dyn Agent> is unaffected).
    #[cfg(feature = "structured")]
    #[tokio::test]
    async fn typed_agent_trait_generic_call() {
        #[derive(Debug, Deserialize, JsonSchema)]
        struct Weather {
            city: String,
        }
        async fn typed_run<A: TypedAgent + Send>(
            agent: &mut A,
            input: &str,
        ) -> Result<Weather, AgentError> {
            agent.run_typed(input).await
        }

        let fake = SharedFake::new([FakeReply::Text(r#"{"city":"Beijing"}"#.into())]);
        let mut agent = agent(fake.clone(), "");
        let weather = typed_run(&mut agent, "Beijing weather").await.unwrap();
        assert_eq!(weather.city, "Beijing");
    }

    /// Typed and text runs can be mixed: run_typed yields typed results,
    /// Agent::run yields text (no structured constraint; free text is
    /// unaffected by this run's schema).
    #[cfg(feature = "structured")]
    #[tokio::test]
    async fn typed_output_agent_trait_run_returns_text() {
        #[derive(Debug, Deserialize, JsonSchema)]
        struct Weather {
            city: String,
        }
        let fake = SharedFake::new([
            FakeReply::Text(r#"{"city":"Beijing"}"#.into()),
            FakeReply::Text("Hello".into()),
        ]);
        let mut agent = agent(fake.clone(), "");
        let weather: Weather = agent.run_typed("Beijing weather").await.unwrap();
        assert_eq!(weather.city, "Beijing");
        // Later text run: unaffected by run_typed's schema (no structured
        // constraint).
        let text = Agent::run(&mut agent, "Say hi").await.unwrap();
        assert_eq!(text, "Hello");
    }

    /// Streaming: structured validation retries have an independent budget;
    /// exceeding it terminates with an in-stream error (same semantics as
    /// run).
    #[cfg(feature = "structured")]
    #[tokio::test]
    async fn structured_output_stream_exhausts_retry_budget() {
        let schema = serde_json::json!({ "type": "object" });
        let fake = SharedFake::new([
            FakeReply::Text("bad1".into()),
            FakeReply::Text("bad2".into()),
        ]);
        let mut agent = agent(fake.clone(), "").with_config(AgentConfig {
            max_structured_retries: 1,
            ..Default::default()
        });
        agent = agent.with_structured_output(schema);
        let mut stream = agent.run_stream("hi").await.unwrap();
        let mut saw_err = false;
        while let Some(item) = stream.next().await {
            if let Err(e) = item {
                assert!(matches!(e, AgentError::StructuredRetriesExhausted(1)));
                saw_err = true;
                break;
            }
        }
        assert!(saw_err);
    }

    /// Tool-round atomicity: cancellation during a tool round doesn't stop
    /// the tool from finishing and recording its result (no interruption,
    /// no half-recorded residue); cancellation takes effect before the next
    /// round's conversation.
    #[tokio::test]
    async fn tool_round_atomic_under_cancel() {
        // Slow tool: 100ms execution, counts calls.
        struct SlowTool {
            calls: Arc<AtomicUsize>,
        }
        #[async_trait::async_trait]
        impl Tool for SlowTool {
            fn schema(&self) -> ToolSchema {
                ToolSchema::new("slow", "Slow tool", serde_json::json!({}))
            }
            async fn call(
                &self,
                _arguments: serde_json::Value,
                _context: ToolContext<'_>,
            ) -> Result<ToolResult, ToolError> {
                self.calls.fetch_add(1, Ordering::Relaxed);
                tokio::time::sleep(Duration::from_millis(100)).await;
                Ok(ToolOutput::text("42").into())
            }
        }

        let calls = Arc::new(AtomicUsize::new(0));
        let mut registry = ToolRegistry::new();
        registry.register(SlowTool {
            calls: calls.clone(),
        });
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "slow", "{}")],
            },
            FakeReply::Text("42".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());
        let token = CancellationToken::new();

        // Cancel at 50ms: right in the middle of the tool round (tool takes
        // 100ms).
        let result = tokio::select! {
            r = agent.run_request_with_context(RunRequest::text("Compute"), cancellation_context(&token)) => r.map(|output| output.answer),
            _ = async {
                tokio::time::sleep(Duration::from_millis(50)).await;
                token.cancel();
                std::future::pending::<()>().await; // stay pending so the run branch wins
            } => unreachable!("cancellation branch only sends a signal"),
        };
        assert!(matches!(result, Err(AgentError::Cancelled)));
        assert_eq!(calls.load(Ordering::Relaxed), 1); // tool finished (atomic)
        assert_eq!(fake.requests().len(), 1); // only one conversation round sent
        // Memory is complete: user + assistant (with tool requests) +
        // tool_result paired.
        let ctx = agent.memory.context().await.unwrap();
        assert_eq!(ctx.len(), 3);
        assert!(matches!(
            &ctx[1],
            Message::Assistant { tool_calls, .. } if tool_calls.len() == 1
        ));
        assert!(matches!(&ctx[2], Message::ToolResult { content, .. } if content == "42"));
    }

    /// Custom tool-round executor injection (non-streaming): the round
    /// policy is the application's to own — here, calls run in reversed
    /// order, and a call the policy refuses is answered with a synthetic
    /// denial outcome (the tool never runs). Memory records results in
    /// the order the policy yields them.
    #[tokio::test]
    async fn custom_tool_round_executor() {
        // Deny the "alpha" call on the first round, run everything else;
        // yield outcomes in reversed order.
        #[derive(Default)]
        struct DenyAlphaToolRoundExecutor {
            denied: bool,
        }
        #[async_trait::async_trait]
        impl ToolRoundExecutor for DenyAlphaToolRoundExecutor {
            async fn execute_round<'a>(
                &'a mut self,
                ctx: ToolRoundCtx<'a>,
                calls: Vec<ToolCall>,
            ) -> BoxStream<'a, ToolCallOutcome> {
                let mut outcomes = Vec::with_capacity(calls.len());
                for call in calls.into_iter().rev() {
                    if call.name == "alpha" && !self.denied {
                        self.denied = true;
                        // Policy refuses: a synthetic outcome feeds the
                        // denial text back through the normal ToolResult
                        // channel, without running the tool.
                        outcomes.push(ToolCallOutcome {
                            call,
                            content: "denied by policy".into(),
                            memory_policy: ToolMemoryPolicy::Normal,
                            effect: None,
                        });
                    } else {
                        outcomes.push(ctx.run(call).await);
                    }
                }
                Box::pin(futures::stream::iter(outcomes))
            }
        }

        let (alpha_tool, alpha_calls) = FakeTool::new("alpha", "A");
        let (beta_tool, beta_calls) = FakeTool::new("beta", "B");
        let mut registry = ToolRegistry::new();
        registry.register(alpha_tool);
        registry.register(beta_tool);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "alpha", "{}"), call("c2", "beta", "{}")],
            },
            FakeReply::Text("done".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default())
            .with_tool_round_executor(DenyAlphaToolRoundExecutor::default());

        let answer = agent.run("Compute").await.unwrap();
        assert_eq!(answer, "done");
        // alpha never ran (denied); beta ran once.
        assert_eq!(alpha_calls.load(Ordering::Relaxed), 0);
        assert_eq!(beta_calls.load(Ordering::Relaxed), 1);
        // The second request carries the round's results in the order the
        // policy yielded them: beta (executed) first, then alpha (denied).
        let requests = fake.requests();
        assert_eq!(requests.len(), 2);
        assert_eq!(requests[1].messages[2], Message::tool_result("c2", "B"),);
        assert_eq!(
            requests[1].messages[3],
            Message::tool_result("c1", "denied by policy"),
        );
    }

    /// Custom tool-round executor injection (streaming): each outcome is
    /// dispatched as a ToolResult chunk as the policy yields it, in the
    /// policy's order.
    #[tokio::test]
    async fn custom_tool_round_executor_streaming() {
        #[derive(Default)]
        struct ReversedToolRoundExecutor;
        #[async_trait::async_trait]
        impl ToolRoundExecutor for ReversedToolRoundExecutor {
            async fn execute_round<'a>(
                &'a mut self,
                ctx: ToolRoundCtx<'a>,
                calls: Vec<ToolCall>,
            ) -> BoxStream<'a, ToolCallOutcome> {
                let mut outcomes = Vec::with_capacity(calls.len());
                for call in calls.into_iter().rev() {
                    outcomes.push(ctx.run(call).await);
                }
                Box::pin(futures::stream::iter(outcomes))
            }
        }

        let (alpha_tool, _) = FakeTool::new("alpha", "A");
        let (beta_tool, _) = FakeTool::new("beta", "B");
        let mut registry = ToolRegistry::new();
        registry.register(alpha_tool);
        registry.register(beta_tool);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "alpha", "{}"), call("c2", "beta", "{}")],
            },
            FakeReply::Text("done".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default())
            .with_tool_round_executor(ReversedToolRoundExecutor);

        let mut stream = agent.run_stream("Compute").await.unwrap();
        let mut results = Vec::new();
        let mut done = false;
        while let Some(item) = stream.next().await {
            match item.unwrap() {
                MessageChunk::ToolResult { id, content, .. } => {
                    results.push((id, content));
                }
                MessageChunk::Done(_) => done = true,
                _ => {}
            }
        }
        assert!(done);
        // Dispatched in policy order: beta (reversed) before alpha.
        assert_eq!(
            results,
            vec![
                ("c2".to_string(), "B".to_string()),
                ("c1".to_string(), "A".to_string())
            ]
        );
    }

    #[tokio::test]
    async fn kernel_batches_effect_requests_from_same_round() {
        let mut registry = ToolRegistry::new();
        registry
            .register(EffectTool::new("read_a", "effect-a", "read A"))
            .register(EffectTool::new("read_b", "effect-b", "read B"));
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "read_a", "{}"), call("c2", "read_b", "{}")],
            },
            FakeReply::Text("done".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());
        let context = RunContext::new("kernel-batch-effects");

        let action = agent
            .start(RunRequest::text("read both"), &context)
            .await
            .unwrap();
        let AgentAction::RequestModel { request } = action else {
            panic!("expected initial model request");
        };
        let action = agent
            .observe(
                Observation::Model(ModelObservation::new(
                    request.id,
                    fake.chat(request.chat).await.unwrap(),
                )),
                &context,
            )
            .await
            .unwrap();
        let AgentAction::RequestEffects { requests } = action else {
            panic!("expected batch effect request");
        };
        assert_eq!(
            requests
                .iter()
                .map(|request| request.id.as_str())
                .collect::<Vec<_>>(),
            vec!["effect-a", "effect-b"]
        );
        assert_eq!(
            requests
                .iter()
                .map(|request| request.source.tool_call_id.as_deref())
                .collect::<Vec<_>>(),
            vec![Some("c1"), Some("c2")]
        );

        // Harnesses may execute in parallel; observations are matched by id
        // and recorded back in the model's tool-call order.
        let action = agent
            .observe(
                Observation::Effects(vec![
                    EffectObservation::succeeded("effect-b", "observed B"),
                    EffectObservation::succeeded("effect-a", "observed A"),
                ]),
                &context,
            )
            .await
            .unwrap();
        let AgentAction::RequestModel { request } = action else {
            panic!("expected next model request");
        };
        assert!(matches!(
            &request.chat.messages[2],
            Message::ToolResult { id, content } if id == "c1" && content == "observed A"
        ));
        assert!(matches!(
            &request.chat.messages[3],
            Message::ToolResult { id, content } if id == "c2" && content == "observed B"
        ));

        let action = agent
            .observe(
                Observation::Model(ModelObservation::new(
                    request.id,
                    fake.chat(request.chat).await.unwrap(),
                )),
                &context,
            )
            .await
            .unwrap();
        let AgentAction::Respond { output } = action else {
            panic!("expected final response");
        };
        assert_eq!(output.answer, "done");
    }

    #[tokio::test]
    async fn kernel_records_mixed_outputs_and_effects_in_tool_call_order() {
        let mut registry = ToolRegistry::new();
        registry
            .register(FakeTool::new("before", "plain before").0)
            .register(EffectTool::new("read", "effect-read", "read"))
            .register(FakeTool::new("after", "plain after").0);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![
                    call("c1", "before", "{}"),
                    call("c2", "read", "{}"),
                    call("c3", "after", "{}"),
                ],
            },
            FakeReply::Text("done".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());
        let context = RunContext::new("kernel-mixed-effects");

        let AgentAction::RequestModel { request } = agent
            .start(RunRequest::text("read with context"), &context)
            .await
            .unwrap()
        else {
            panic!("expected initial model request");
        };
        let AgentAction::RequestEffect { request } = agent
            .observe(
                Observation::Model(ModelObservation::new(
                    request.id,
                    fake.chat(request.chat).await.unwrap(),
                )),
                &context,
            )
            .await
            .unwrap()
        else {
            panic!("expected single effect request");
        };
        assert_eq!(request.id, "effect-read");

        let action = agent
            .observe(
                Observation::Effect(EffectObservation::succeeded("effect-read", "observed read")),
                &context,
            )
            .await
            .unwrap();
        let AgentAction::RequestModel { request } = action else {
            panic!("expected next model request");
        };
        assert!(matches!(
            &request.chat.messages[2],
            Message::ToolResult { id, content } if id == "c1" && content == "plain before"
        ));
        assert!(matches!(
            &request.chat.messages[3],
            Message::ToolResult { id, content } if id == "c2" && content == "observed read"
        ));
        assert!(matches!(
            &request.chat.messages[4],
            Message::ToolResult { id, content } if id == "c3" && content == "plain after"
        ));
    }

    #[tokio::test]
    async fn kernel_rejects_partial_effect_batch_observation() {
        let mut registry = ToolRegistry::new();
        registry
            .register(EffectTool::new("read_a", "effect-a", "read A"))
            .register(EffectTool::new("read_b", "effect-b", "read B"));
        let fake = SharedFake::new([FakeReply::ToolCalls {
            content: "".into(),
            calls: vec![call("c1", "read_a", "{}"), call("c2", "read_b", "{}")],
        }]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());
        let context = RunContext::new("kernel-batch-effects-partial");

        let AgentAction::RequestModel { request } = agent
            .start(RunRequest::text("read both"), &context)
            .await
            .unwrap()
        else {
            panic!("expected initial model request");
        };
        let AgentAction::RequestEffects { .. } = agent
            .observe(
                Observation::Model(ModelObservation::new(
                    request.id,
                    fake.chat(request.chat).await.unwrap(),
                )),
                &context,
            )
            .await
            .unwrap()
        else {
            panic!("expected batch effect request");
        };
        let err = agent
            .observe(
                Observation::Effects(vec![EffectObservation::succeeded("effect-a", "observed A")]),
                &context,
            )
            .await
            .unwrap_err();
        assert!(
            matches!(err, AgentError::InvalidStep(message) if message.contains("count mismatch"))
        );
    }

    #[tokio::test]
    async fn kernel_rejects_duplicate_effect_batch_observation_and_can_retry() {
        let mut registry = ToolRegistry::new();
        registry
            .register(EffectTool::new("read_a", "effect-a", "read A"))
            .register(EffectTool::new("read_b", "effect-b", "read B"));
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "read_a", "{}"), call("c2", "read_b", "{}")],
            },
            FakeReply::Text("done".into()),
        ]);
        let mut agent = agent_with_registry(fake.clone(), registry, AgentConfig::default());
        let context = RunContext::new("kernel-batch-effects-duplicate");

        let AgentAction::RequestModel { request } = agent
            .start(RunRequest::text("read both"), &context)
            .await
            .unwrap()
        else {
            panic!("expected initial model request");
        };
        let AgentAction::RequestEffects { .. } = agent
            .observe(
                Observation::Model(ModelObservation::new(
                    request.id,
                    fake.chat(request.chat).await.unwrap(),
                )),
                &context,
            )
            .await
            .unwrap()
        else {
            panic!("expected batch effect request");
        };
        let err = agent
            .observe(
                Observation::Effects(vec![
                    EffectObservation::succeeded("effect-a", "observed A"),
                    EffectObservation::succeeded("effect-a", "observed A again"),
                ]),
                &context,
            )
            .await
            .unwrap_err();
        assert!(matches!(err, AgentError::InvalidStep(message) if message.contains("duplicate")));

        let action = agent
            .observe(
                Observation::Effects(vec![
                    EffectObservation::succeeded("effect-b", "observed B"),
                    EffectObservation::succeeded("effect-a", "observed A"),
                ]),
                &context,
            )
            .await
            .unwrap();
        let AgentAction::RequestModel { request } = action else {
            panic!("expected next model request");
        };
        assert!(matches!(
            &request.chat.messages[2],
            Message::ToolResult { id, content } if id == "c1" && content == "observed A"
        ));
        assert!(matches!(
            &request.chat.messages[3],
            Message::ToolResult { id, content } if id == "c2" && content == "observed B"
        ));
    }

    /// Mid-stream cancellation: already-dispatched Deltas are kept, a
    /// Cancelled terminal event wraps up, no Done; nothing is recorded for
    /// this round (recording happens at the end of the round).
    #[tokio::test]
    async fn stream_cancel_mid_generation() {
        // Slow-stream Provider: 100ms between Deltas (wide gaps, so the
        // cancellation landing point is stable).
        struct SlowStreamProvider;
        #[async_trait::async_trait]
        impl Provider for SlowStreamProvider {
            async fn chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<ChatResponse, ProviderError> {
                unreachable!("this test uses streaming path only")
            }
            async fn stream_chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError>
            {
                Ok(Box::pin(async_stream::stream! {
                    yield Ok(StreamEvent::Delta("d0".into()));
                    tokio::time::sleep(Duration::from_millis(100)).await;
                    yield Ok(StreamEvent::Delta("d1".into()));
                    tokio::time::sleep(Duration::from_millis(100)).await;
                    yield Ok(StreamEvent::Done {
                        reason: FinishReason::Stop,
                        usage: None,
                    });
                }))
            }
        }

        let mut agent = ReActAgent::new(SlowStreamProvider, ToolRegistry::new(), "");
        let token = CancellationToken::new();
        let mut stream = agent
            .run_stream_request_with_context(RunRequest::text("hi"), cancellation_context(&token))
            .await
            .unwrap();

        // Cancel at 50ms: d0 already received (0ms), d1 still pending
        // (100ms) — cancellation lands in the gap.
        let mut first = Vec::new();
        tokio::select! {
            _ = async {
                while let Some(ev) = stream.next().await {
                    let ev = ev.unwrap();
                    if ev == MessageChunk::Cancelled {
                        break;
                    }
                    first.push(ev);
                }
            } => {}
            _ = async {
                tokio::time::sleep(Duration::from_millis(50)).await;
                token.cancel();
            } => {}
        }
        assert_eq!(first, vec![MessageChunk::Delta("d0".into())]);

        // Keep consuming: after cancellation the stream terminates with
        // Cancelled, no Done.
        let rest: Vec<MessageChunk> = stream.by_ref().map(|e| e.unwrap()).collect().await;
        assert_eq!(rest, vec![MessageChunk::Cancelled]);
        drop(stream); // release the borrow of agent
        // Nothing recorded for this round: the memory only holds the user.
        assert_eq!(agent.memory.context().await.unwrap().len(), 1);
    }

    /// No cancellation residue: after a cancellation, a fresh token runs
    /// normally (each run has an independent cancellation source — the
    /// basis of multi-round conversations).
    #[tokio::test]
    async fn cancelled_then_fresh_token_run_works() {
        let fake = SharedFake::new([FakeReply::Text("hi".into())]);
        let mut agent = agent(fake, "");

        let t1 = CancellationToken::new();
        t1.cancel();
        assert!(matches!(
            agent
                .run_request_with_context(RunRequest::text("one"), cancellation_context(&t1))
                .await,
            Err(AgentError::Cancelled)
        ));

        let t2 = CancellationToken::new();
        assert_eq!(
            agent
                .run_request_with_context(RunRequest::text("two"), cancellation_context(&t2))
                .await
                .unwrap()
                .answer,
            "hi"
        );
    }

    // ---- Observation channel: the loop pushes process events once an
    // EventChannel is attached ----

    use crate::event_channel::{BroadcastEventChannel, EventReceiver};
    use crate::tool::RegistryError;

    /// Attach a broadcast channel and return a receiver; after the run,
    /// dropping the agent (the channel's last holder) closes the channel so
    /// the receiver can drain the final events.
    fn attach_channel(agent: ReActAgent) -> (ReActAgent, Box<dyn EventReceiver>) {
        let channel = BroadcastEventChannel::new(64);
        let rx = channel.subscribe();
        (agent.with_event_channel(channel), rx)
    }

    /// Drain: all events until the channel closes (None).
    async fn drain(rx: &mut Box<dyn EventReceiver>) -> Vec<Arc<dyn AgentEvent>> {
        let mut out = Vec::new();
        while let Some(ev) = rx.recv().await {
            out.push(ev);
        }
        out
    }

    fn names(events: &[Arc<dyn AgentEvent>]) -> Vec<&'static str> {
        events.iter().map(|e| e.name()).collect()
    }

    /// Downcast an event to ReActEvent (enum event set: one downcast, then
    /// an exhaustive match).
    fn react_event(ev: &dyn AgentEvent) -> &ReActEvent {
        ev.as_any()
            .downcast_ref::<ReActEvent>()
            .expect("test event should be ReActEvent")
    }

    /// Non-streaming run event sequence: RunStarted → ToolStarted →
    /// ToolCompleted(Ok) → RunEnded (summary, error=None); bystanders can
    /// observe a non-streaming run.
    #[tokio::test]
    async fn events_published_on_run() {
        let (calc, _calls) = FakeTool::new("calc", "42");
        let mut registry = ToolRegistry::new();
        registry.register(calc);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "calc", r#"{"a":1}"#)],
            },
            FakeReply::Text("The answer is 42".into()),
        ]);
        let (mut agent, mut rx) =
            attach_channel(agent_with_registry(fake, registry, AgentConfig::default()));
        let answer = agent.run("Compute").await.unwrap();
        assert_eq!(answer, "The answer is 42");
        drop(agent);
        let events = drain(&mut rx).await;

        assert_eq!(
            names(&events),
            ["run.started", "tool.started", "tool.completed", "run.ended"]
        );
        // One downcast + exhaustive match (the consumption style of an
        // enum event set).
        match react_event(&*events[1]) {
            // ToolStarted: carries id / arguments, paired with
            // ToolCompleted.
            ReActEvent::ToolStarted {
                id,
                name,
                arguments,
            } => {
                assert_eq!(id, "c1");
                assert_eq!(name, "calc");
                assert_eq!(arguments, r#"{"a":1}"#);
            }
            _ => panic!("expected ToolStarted"),
        }
        match react_event(&*events[2]) {
            // ToolCompleted: Ok carries the result text.
            ReActEvent::ToolCompleted { result, .. } => {
                assert_eq!(result, &Ok(ToolOutput::text("42").into()));
            }
            _ => panic!("expected ToolCompleted"),
        }
        match react_event(&*events[3]) {
            // RunEnded: normal end with error=None; the non-streaming
            // summary is real (2 rounds + 1 tool).
            ReActEvent::RunEnded { summary, error } => {
                assert_eq!(error, &None);
                assert_eq!(summary.rounds, 2);
                assert_eq!(summary.tool_calls, 1);
            }
            _ => panic!("expected RunEnded"),
        }
    }

    #[tokio::test]
    async fn run_started_event_preserves_block_input() {
        let blocks = vec![ContentBlock::Text("look".into())];
        let fake = SharedFake::new([FakeReply::Text("done".into())]);
        let (mut agent, mut rx) = attach_channel(agent(fake, ""));
        agent
            .run_request(RunRequest::blocks(blocks.clone()))
            .await
            .unwrap();
        drop(agent);
        let events = drain(&mut rx).await;

        match react_event(&*events[0]) {
            ReActEvent::RunStarted { input, .. } => {
                assert_eq!(input, &crate::UserInput::Blocks(blocks));
            }
            _ => panic!("expected RunStarted"),
        }
    }

    /// Tool failure: ToolCompleted's Err carries the registry
    /// classification (NotFound), whose Display is the error text fed back
    /// to the model.
    #[tokio::test]
    async fn events_carry_tool_failure() {
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "".into(),
                calls: vec![call("c1", "nope", "{}")],
            },
            FakeReply::Text("Got it".into()),
        ]);
        let (mut agent, mut rx) = attach_channel(agent(fake, ""));
        agent.run("Trigger").await.unwrap();
        drop(agent);

        let events = drain(&mut rx).await;
        let completed = events
            .iter()
            .find_map(|e| match react_event(&**e) {
                ReActEvent::ToolCompleted { result, .. } => Some(result),
                _ => None,
            })
            .expect("expected ToolCompleted event");
        assert!(matches!(completed, Err(RegistryError::NotFound(n)) if n == "nope"));
        assert_eq!(
            completed.as_ref().unwrap_err().to_string(),
            "tool not found: nope"
        );
    }

    /// Streaming path: Delta / Reasoning events push along with the stream;
    /// the RunEnded summary matches the pulling path's semantics.
    #[tokio::test]
    async fn stream_events_include_delta_and_reasoning() {
        let (calc, _calls) = FakeTool::new("calc", "42");
        let mut registry = ToolRegistry::new();
        registry.register(calc);
        let fake = SharedFake::new([
            FakeReply::ToolCalls {
                content: "Thinking: ".into(),
                calls: vec![call("c1", "calc", "{}")],
            },
            FakeReply::TextWithReasoning {
                content: "The answer is 42".into(),
                reasoning: "Reasoning steps".into(),
            },
        ]);
        let (mut agent, mut rx) =
            attach_channel(agent_with_registry(fake, registry, AgentConfig::default()));

        // Pulling path as usual: text = concatenated Deltas.
        let answer: String = agent
            .run_stream("Compute")
            .await
            .unwrap()
            .map(|e| e.unwrap())
            .filter_map(|e| async move {
                match e {
                    MessageChunk::Delta(d) => Some(d),
                    _ => None,
                }
            })
            .collect()
            .await;
        assert_eq!(answer, "Thinking: The answer is 42");
        drop(agent);

        let events = drain(&mut rx).await;
        // Order matches the StreamEvent docs: content Deltas first,
        // Reasoning at the end of the round (OpenAiProvider and
        // FakeProvider share the order).
        assert_eq!(
            names(&events),
            [
                "run.started",
                "delta",
                "tool.started",
                "tool.completed",
                "delta",
                "reasoning",
                "run.ended",
            ]
        );
        let reasoning = events
            .iter()
            .find_map(|e| match react_event(&**e) {
                ReActEvent::Reasoning { text } => Some(text.as_str()),
                _ => None,
            })
            .expect("expected Reasoning event");
        assert_eq!(reasoning, "Reasoning steps");
        match react_event(&**events.last().unwrap()) {
            ReActEvent::RunEnded { summary, error } => {
                assert_eq!(error, &None);
                assert_eq!(summary.rounds, 2);
                assert_eq!(summary.tool_calls, 1);
            }
            _ => panic!("expected RunEnded"),
        }
    }

    /// Cancellation: RunEnded carries error=Some(Cancelled), so bystanders
    /// know the run's outcome.
    #[tokio::test]
    async fn cancelled_run_publishes_run_ended_with_error() {
        // Slow-stream Provider: 100ms between Deltas (stable cancellation landing point).
        struct SlowStreamProvider;
        #[async_trait::async_trait]
        impl Provider for SlowStreamProvider {
            async fn chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<ChatResponse, ProviderError> {
                unreachable!("this test uses streaming path only")
            }
            async fn stream_chat_with_context(
                &self,
                _request: ChatRequest,
                _context: &ProviderRequestContext,
            ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError>
            {
                Ok(Box::pin(async_stream::stream! {
                    yield Ok(StreamEvent::Delta("d0".into()));
                    tokio::time::sleep(Duration::from_millis(100)).await;
                    yield Ok(StreamEvent::Delta("d1".into()));
                    tokio::time::sleep(Duration::from_millis(100)).await;
                    yield Ok(StreamEvent::Done {
                        reason: FinishReason::Stop,
                        usage: None,
                    });
                }))
            }
        }

        let (mut agent, mut rx) =
            attach_channel(ReActAgent::new(SlowStreamProvider, ToolRegistry::new(), ""));
        let token = CancellationToken::new();
        let mut stream = agent
            .run_stream_request_with_context(RunRequest::text("hi"), cancellation_context(&token))
            .await
            .unwrap();

        // Cancel at 50ms: the select cancellation branch wins and the
        // consumption branch is dropped; after cancellation, keep consuming
        // the stream and wrap up with Cancelled (isomorphic to
        // stream_cancel_mid_generation).
        tokio::select! {
            _ = async {
                while let Some(ev) = stream.next().await {
                    let ev = ev.unwrap();
                    if ev == MessageChunk::Cancelled {
                        break;
                    }
                }
            } => {}
            _ = async {
                tokio::time::sleep(Duration::from_millis(50)).await;
                token.cancel();
            } => {}
        }
        let rest: Vec<MessageChunk> = stream.by_ref().map(|e| e.unwrap()).collect().await;
        assert!(rest.contains(&MessageChunk::Cancelled));
        drop(stream);
        drop(agent);

        let events = drain(&mut rx).await;
        assert!(matches!(names(&events).as_slice(), [.., "run.ended"]));
        match react_event(&**events.last().unwrap()) {
            ReActEvent::RunEnded { error, .. } => {
                assert_eq!(error, &Some(AgentError::Cancelled));
            }
            _ => panic!("expected RunEnded"),
        }
    }

    #[cfg(feature = "tracing")]
    mod tracing_tests {
        use super::*;

        // ---- Observability spans: structure and fields (asserted with a
        // collecting subscriber) ----

        use std::collections::HashMap;
        use std::sync::atomic::AtomicU64;
        use tracing::field::{Field, Visit};
        use tracing::subscriber::Subscriber;
        use tracing::{Event, Id, Level, Metadata};

        /// Creation info of one span (name / level / field values at
        /// creation).
        #[derive(Debug, Clone, PartialEq, Eq)]
        struct SpanInfo {
            name: &'static str,
            level: Level,
            fields: Vec<(String, String)>,
        }

        /// The operation sequence received by the subscriber (enter / exit /
        /// close / field record).
        #[derive(Debug, Clone, PartialEq, Eq)]
        enum Op {
            Enter(String),
            Exit(String),
            Close(String),
            Record(String, String),
        }

        /// Collecting subscriber: gathers span creation, enter/exit, and field
        /// records into Vecs for the tests to assert hierarchy and fields.
        /// The tracing tests below force a current-thread runtime because the
        /// collecting subscriber is installed with a thread-local dispatch
        /// guard. A multi-thread runtime can resume the test future on a worker
        /// that does not have that guard.
        #[derive(Debug, Default)]
        struct CollectSubscriber {
            spans: std::sync::Mutex<Vec<SpanInfo>>,
            ops: std::sync::Mutex<Vec<Op>>,
            names: std::sync::Mutex<HashMap<Id, String>>,
            next_id: AtomicU64,
        }

        /// Collect field values as debug text (all of Visit's default methods
        /// land in record_debug).
        struct FieldCollector<'a>(&'a mut Vec<(String, String)>);

        impl Visit for FieldCollector<'_> {
            fn record_debug(&mut self, field: &Field, value: &dyn fmt::Debug) {
                self.0
                    .push((field.name().to_string(), format!("{value:?}")));
            }
        }

        impl Subscriber for CollectSubscriber {
            fn enabled(&self, _metadata: &Metadata<'_>) -> bool {
                true
            }

            fn new_span(&self, span: &tracing::span::Attributes<'_>) -> Id {
                // fetch_add returns the old value (0 on the first call), and
                // tracing Ids must be non-zero.
                let id = Id::from_u64(self.next_id.fetch_add(1, Ordering::Relaxed) + 1);
                let mut fields = Vec::new();
                span.record(&mut FieldCollector(&mut fields));
                self.spans.lock().unwrap().push(SpanInfo {
                    name: span.metadata().name(),
                    level: *span.metadata().level(),
                    fields,
                });
                self.names
                    .lock()
                    .unwrap()
                    .insert(id.clone(), span.metadata().name().to_string());
                id
            }

            fn record(&self, id: &Id, values: &tracing::span::Record<'_>) {
                let name = self.names.lock().unwrap().get(id).cloned();
                let Some(name) = name else { return };
                let mut fields = Vec::new();
                values.record(&mut FieldCollector(&mut fields));
                let mut ops = self.ops.lock().unwrap();
                for (field, value) in fields {
                    ops.push(Op::Record(name.clone(), format!("{field}={value}")));
                }
            }

            fn enter(&self, id: &Id) {
                let name = self.names.lock().unwrap().get(id).cloned();
                if let Some(name) = name {
                    self.ops.lock().unwrap().push(Op::Enter(name));
                }
            }

            fn exit(&self, id: &Id) {
                let name = self.names.lock().unwrap().get(id).cloned();
                if let Some(name) = name {
                    self.ops.lock().unwrap().push(Op::Exit(name));
                }
            }

            fn try_close(&self, id: Id) -> bool {
                let name = self.names.lock().unwrap().get(&id).cloned();
                if let Some(name) = name {
                    self.ops.lock().unwrap().push(Op::Close(name));
                }
                true
            }

            fn clone_span(&self, id: &Id) -> Id {
                id.clone()
            }

            fn event(&self, _event: &Event<'_>) {}
            fn record_follows_from(&self, _span: &Id, _follows: &Id) {}
        }

        /// Swap the current thread's default dispatch to the collecting
        /// subscriber (the guard is held across awaits).
        fn collect_guard(sub: &Arc<CollectSubscriber>) -> tracing::dispatcher::DefaultGuard {
            tracing::dispatcher::set_default(&tracing::Dispatch::new(sub.clone()))
        }

        fn enter_names(ops: &[Op]) -> Vec<String> {
            ops.iter()
                .filter_map(|op| match op {
                    Op::Enter(name) => Some(name.clone()),
                    _ => None,
                })
                .collect()
        }

        fn records_of(ops: &[Op], span: &str) -> Vec<String> {
            ops.iter()
                .filter_map(|op| match op {
                    Op::Record(s, kv) if s == span => Some(kv.clone()),
                    _ => None,
                })
                .collect()
        }

        /// Assert hierarchy invariants (shared by both paths): (1) whenever a
        /// span is entered, `agent.run` must be on the enter/exit stack (no
        /// orphan spans — llm/tool must render under run);
        /// (2) no span with the same name on the stack (the same span can't be
        /// entered twice); (3) enter/exit strictly pair up.
        fn assert_nesting_invariants(ops: &[Op]) {
            let mut stack: Vec<String> = Vec::new();
            for op in ops {
                match op {
                    Op::Enter(name) => {
                        if name != "agent.run" {
                            assert!(
                                stack.contains(&"agent.run".to_string()),
                                "span {name} requires agent.run on the stack when entering (stack: {stack:?})"
                            );
                        }
                        assert!(
                            !stack.contains(name),
                            "same span entered twice (double instrumenting): {name} (stack: {stack:?})"
                        );
                        stack.push(name.clone());
                    }
                    Op::Exit(name) => {
                        assert_eq!(
                            stack.pop().as_deref(),
                            Some(name.as_str()),
                            "exit must pair with enter: {name}"
                        );
                    }
                    _ => {}
                }
            }
        }

        /// Non-streaming path: span tree agent.run → llm_request / tool (two
        /// levels, grouped by the round attribute); hierarchy invariants hold
        /// (no orphans, no double enters); llm_request records usage at wrap-up
        /// (dual channel); run.id is consistent across the tree.
        #[tokio::test(flavor = "current_thread")]
        async fn trace_span_tree_non_stream() {
            let sub = Arc::new(CollectSubscriber::default());
            let _guard = collect_guard(&sub);

            let (calc, _calls) = FakeTool::new("calc", "42");
            let mut registry = ToolRegistry::new();
            registry.register(calc);
            let fake = SharedFake::new([
                FakeReply::WithUsage {
                    reply: Box::new(FakeReply::ToolCalls {
                        content: "".into(),
                        calls: vec![call("c1", "calc", "{}")],
                    }),
                    usage: Usage::new(10, 2),
                },
                FakeReply::text_with_usage("42", Usage::new(20, 5)),
            ]);
            let mut agent = agent_with_registry(fake, registry, AgentConfig::default());
            assert_eq!(agent.run("Compute").await.unwrap(), "42");

            let ops = sub.ops.lock().unwrap().clone();
            // Hierarchy invariants (no orphan spans, no double enters, paired
            // exits).
            assert_nesting_invariants(&ops);
            // Structure: two rounds, one llm_request each (round one also has a
            // tool). Note every instrumented future enters and exits twice (one
            // poll + one more enter inside Instrumented's drop; tracing
            // guarantees inner's Drop also runs in the span context), so the
            // assertions use "first-appearance order" and lower-bound counts
            // rather than exact enter/exit sequences.
            let enters = enter_names(&ops);
            let mut first_seen = Vec::new();
            for name in enters.iter() {
                if !first_seen.contains(name) {
                    first_seen.push(name.clone());
                }
            }
            assert_eq!(first_seen, ["agent.run", "llm_request", "tool"]);
            assert!(enters.iter().filter(|n| *n == "llm_request").count() >= 2);
            assert!(enters.iter().filter(|n| *n == "tool").count() >= 1);

            // usage via both channels: llm_request records on the span at
            // wrap-up (the usage injected per round).
            assert_eq!(
                records_of(&ops, "llm_request"),
                [
                    "usage.prompt_tokens=10",
                    "usage.completion_tokens=2",
                    "usage.prompt_tokens=20",
                    "usage.completion_tokens=5",
                ]
            );

            // Levels: agent.run is INFO (skeleton visible by default), detail
            // spans are DEBUG.
            let spans = sub.spans.lock().unwrap().clone();
            let run_span = spans.iter().find(|s| s.name == "agent.run").unwrap();
            assert_eq!(run_span.level, Level::INFO);
            assert_eq!(
                spans
                    .iter()
                    .find(|s| s.name == "llm_request")
                    .unwrap()
                    .level,
                Level::DEBUG
            );
            assert_eq!(
                spans.iter().find(|s| s.name == "tool").unwrap().level,
                Level::DEBUG
            );

            // round attribute: llm_request carries the round number (two
            // rounds = 1, 2 — the grouping key).
            let llm_rounds: Vec<u64> = spans
                .iter()
                .filter(|s| s.name == "llm_request")
                .map(|s| {
                    s.fields
                        .iter()
                        .find(|(f, _)| f == "round")
                        .map(|(_, v)| v.parse().unwrap())
                        .unwrap()
                })
                .collect();
            assert_eq!(llm_rounds, vec![1, 2]);

            // run.id: every span carries the same run id (the correlation key
            // with the event stream's RunStarted).
            let run_ids: Vec<String> = spans
                .iter()
                .map(|s| {
                    s.fields
                        .iter()
                        .find(|(f, _)| f == "run.id")
                        .map(|(_, v)| v.clone())
                        .unwrap_or_else(|| panic!("span {} must carry a run.id field", s.name))
                })
                .collect();
            assert!(run_ids.iter().all(|id| id == &run_ids[0]));
            assert!(run_ids[0].starts_with("run-"));
        }

        /// Streaming path: the same span structure; the run span enters on
        /// every poll (covering the whole consumption period); usage is
        /// recorded at wrap-up when the Done event arrives.
        #[tokio::test(flavor = "current_thread")]
        async fn trace_span_tree_stream() {
            let sub = Arc::new(CollectSubscriber::default());
            let _guard = collect_guard(&sub);

            let (calc, _calls) = FakeTool::new("calc", "42");
            let mut registry = ToolRegistry::new();
            registry.register(calc);
            let fake = SharedFake::new([
                FakeReply::WithUsage {
                    reply: Box::new(FakeReply::ToolCalls {
                        content: "".into(),
                        calls: vec![call("c1", "calc", "{}")],
                    }),
                    usage: Usage::new(10, 2),
                },
                FakeReply::text_with_usage("42", Usage::new(20, 5)),
            ]);
            let mut agent = agent_with_registry(fake, registry, AgentConfig::default());
            agent
                .run_stream("Compute")
                .await
                .unwrap()
                .for_each(|_| async {})
                .await;

            let ops = sub.ops.lock().unwrap().clone();
            // Hierarchy invariants: any change that breaks the span hierarchy
            // (e.g. losing an instrument on the tool path) is caught by this
            // assertion.
            assert_nesting_invariants(&ops);
            // First-appearance enter order = the hierarchy; per-await
            // instrumenting inside the generator makes the same span enter and
            // exit many times (across polls + drops), so only first-appearance
            // order and multiple enters of run are asserted.
            let enters = enter_names(&ops);
            let mut first_seen = Vec::new();
            for name in enters.iter() {
                if !first_seen.contains(name) {
                    first_seen.push(name.clone());
                }
            }
            assert_eq!(first_seen, ["agent.run", "llm_request", "tool"]);
            // The run span enters on every poll: it enters many times during
            // stream consumption, far beyond the 2 of poll+drop.
            assert!(enters.iter().filter(|n| *n == "agent.run").count() > 2);

            // usage recorded at wrap-up (the per-round injected usage).
            assert_eq!(
                records_of(&ops, "llm_request"),
                [
                    "usage.prompt_tokens=10",
                    "usage.completion_tokens=2",
                    "usage.prompt_tokens=20",
                    "usage.completion_tokens=5",
                ]
            );
        }

        /// Two runs: run.id differs (per-instance increment); the event stream's
        /// RunStarted carries the same id — the correlation key between trace
        /// and the event stream.
        #[tokio::test(flavor = "current_thread")]
        async fn trace_run_id_differs_between_runs() {
            let sub = Arc::new(CollectSubscriber::default());
            let _guard = collect_guard(&sub);

            let fake =
                SharedFake::new([FakeReply::Text("hi".into()), FakeReply::Text("bye".into())]);
            let (mut agent, mut rx) = attach_channel(agent(fake, ""));
            agent.run("one").await.unwrap();
            agent.run("two").await.unwrap();
            drop(agent);

            let spans = sub.spans.lock().unwrap().clone();
            let run_ids: Vec<String> = spans
                .iter()
                .filter(|s| s.name == "agent.run")
                .map(|s| {
                    s.fields
                        .iter()
                        .find(|(f, _)| f == "run.id")
                        .map(|(_, v)| v.clone())
                        .unwrap()
                })
                .collect();
            assert_eq!(run_ids.len(), 2);
            assert_ne!(run_ids[0], run_ids[1]);

            // Event-stream side: RunStarted carries the same run id as the span
            // (one per run).
            let events = drain(&mut rx).await;
            let event_ids: Vec<String> = events
                .iter()
                .filter_map(|e| match react_event(&**e) {
                    ReActEvent::RunStarted { run_id, .. } => Some(run_id.clone()),
                    _ => None,
                })
                .collect();
            assert_eq!(event_ids, run_ids);
        }

        /// Run failure: the agent.run span records an error at wrap-up
        /// (observability spots the failed run at a glance).
        #[tokio::test(flavor = "current_thread")]
        async fn trace_run_error_recorded() {
            let sub = Arc::new(CollectSubscriber::default());
            let _guard = collect_guard(&sub);

            let (calc, _calls) = FakeTool::new("calc", "42");
            let mut registry = ToolRegistry::new();
            registry.register(calc);
            let fake = SharedFake::new([
                FakeReply::ToolCalls {
                    content: "".into(),
                    calls: vec![call("c1", "calc", "{}")],
                },
                FakeReply::ToolCalls {
                    content: "".into(),
                    calls: vec![call("c2", "calc", "{}")],
                },
            ]);
            let mut agent = agent_with_registry(
                fake,
                registry,
                AgentConfig {
                    max_tool_rounds: 2,
                    ..Default::default()
                },
            );
            assert!(matches!(
                agent.run("Keep computing").await,
                Err(AgentError::TooManyToolRounds(2))
            ));

            let ops = sub.ops.lock().unwrap().clone();
            // Values are Debug-formatted, so strings carry quotes; the error
            // text is asserted with starts_with (extensible, not locked to the
            // full text).
            let records = records_of(&ops, "agent.run");
            assert_eq!(records.len(), 1);
            assert!(
                records[0].starts_with("error=\"model requested tools for more than 2 rounds"),
                "unexpected records: {records:?}"
            );
        }

        /// Tool failure: the tool span records an error (the failure is fed
        /// back as text; the run ends normally, agent.run has no error).
        #[tokio::test(flavor = "current_thread")]
        async fn trace_tool_error_recorded() {
            struct FailingTool;
            #[async_trait::async_trait]
            impl Tool for FailingTool {
                fn schema(&self) -> ToolSchema {
                    ToolSchema::new("boom", "Tool that always fails", serde_json::json!({}))
                }
                async fn call(
                    &self,
                    _arguments: serde_json::Value,
                    _context: ToolContext<'_>,
                ) -> Result<ToolResult, ToolError> {
                    Err(ToolError::Execution("internal error".into()))
                }
            }

            let sub = Arc::new(CollectSubscriber::default());
            let _guard = collect_guard(&sub);

            let mut registry = ToolRegistry::new();
            registry.register(FailingTool);
            let fake = SharedFake::new([
                FakeReply::ToolCalls {
                    content: "".into(),
                    calls: vec![call("c1", "boom", "{}")],
                },
                FakeReply::Text("Got it".into()),
            ]);
            let mut agent = agent_with_registry(fake, registry, AgentConfig::default());
            agent.run("Trigger failure").await.unwrap();

            let ops = sub.ops.lock().unwrap().clone();
            // The tool span records the failure (the registry classification's
            // Display); the run ends normally, with no error.
            let tool_errors = records_of(&ops, "tool");
            assert_eq!(tool_errors.len(), 1);
            assert!(tool_errors[0].contains("internal error"));
            assert!(records_of(&ops, "agent.run").is_empty());
        }

        /// Provider failure (non-streaming): the llm_request span records the
        /// error — observability can pinpoint which call of which round failed;
        /// the run span records too (the whole run failed).
        #[tokio::test(flavor = "current_thread")]
        async fn trace_llm_error_recorded() {
            let sub = Arc::new(CollectSubscriber::default());
            let _guard = collect_guard(&sub);

            // Script exhausted: the second run's conversation fails outright.
            let fake = SharedFake::new([FakeReply::Text("hi".into())]);
            let mut agent = agent(fake, "");
            agent.run("one").await.unwrap();
            let err = agent.run("two").await.unwrap_err();
            assert!(matches!(err, AgentError::Provider(_)));

            let ops = sub.ops.lock().unwrap().clone();
            // The failed call: the llm_request span has an error record (the
            // first run's successful call only has usage records; the two spans'
            // records land in the same collector).
            let llm_records = records_of(&ops, "llm_request");
            assert!(
                llm_records.iter().any(|r| r.starts_with("error=")),
                "the failed round's llm span should have an error: {llm_records:?}"
            );
            // The run span also records the failed outcome.
            assert!(
                records_of(&ops, "agent.run")
                    .iter()
                    .any(|r| r.starts_with("error="))
            );
        }

        /// Provider failure mid-stream (streaming): the consumption loop's Err
        /// event records the error on the llm_request span, and the run span
        /// records in sync; hierarchy invariants still hold.
        #[tokio::test(flavor = "current_thread")]
        async fn trace_stream_llm_error_recorded() {
            struct FailInStream;
            #[async_trait::async_trait]
            impl Provider for FailInStream {
                async fn chat_with_context(
                    &self,
                    _request: ChatRequest,
                    _context: &ProviderRequestContext,
                ) -> Result<ChatResponse, ProviderError> {
                    unreachable!("this test uses streaming path only")
                }
                async fn stream_chat_with_context(
                    &self,
                    _request: ChatRequest,
                    _context: &ProviderRequestContext,
                ) -> Result<BoxStream<'static, Result<StreamEvent, ProviderError>>, ProviderError>
                {
                    Ok(Box::pin(futures::stream::iter(vec![
                        Ok(StreamEvent::Delta("hi".into())),
                        Err(ProviderError::Protocol {
                            message: "boom".into(),
                        }),
                    ])))
                }
            }

            let sub = Arc::new(CollectSubscriber::default());
            let _guard = collect_guard(&sub);

            let mut agent = ReActAgent::new(FailInStream, ToolRegistry::new(), "");
            let mut stream = agent.run_stream("two").await.unwrap();
            while stream.next().await.is_some() {}

            let ops = sub.ops.lock().unwrap().clone();
            assert_nesting_invariants(&ops);
            assert!(
                records_of(&ops, "llm_request")
                    .iter()
                    .any(|r| r.contains("boom"))
            );
            assert!(
                records_of(&ops, "agent.run")
                    .iter()
                    .any(|r| r.contains("boom"))
            );
        }
    }
}