path: root/libpanto/src/agent.zig

//! The Agent owns the conversation-driving loop: provider streaming +
//! tool dispatch.
//!
//! On each turn, after the provider streams an assistant message, the
//! agent inspects it for ToolUse blocks. If any are present, the agent:
//!
//!   1. Groups them by their *owning registration* in the registry — a
//!      single `Tool` is its own group; every `ToolSource`-backed tool
//!      whose name maps to the same source forms one group.
//!   2. Spawns one concurrent task per group via `std.Io.Group`.
//!      A single-`Tool` group runs the tool's `invoke` once; a
//!      `ToolSource` group calls the source's `invoke_batch` with all
//!      of its calls at once. We use `Group.concurrent` (not `async`)
//!      because tool invocations may block on I/O and we need real
//!      concurrency, not just expressed asynchrony.
//!   3. Awaits the group. ToolResult blocks are assembled in the
//!      *original* call order (i.e. the order the LLM emitted them).
//!   4. Appends a user message containing the ToolResult blocks back
//!      into the conversation and loops.
//!
//! The "thread-safe" promise for single `Tool` registrations is
//! unchanged. For `ToolSource`-backed tools, the source's runtime
//! receives all of its calls on one thread per turn, so it can keep a
//! single-threaded interpreter (Lua, Python, ...) without further
//! synchronization.

const std = @import("std");
const Allocator = std.mem.Allocator;
const Io = std.Io;

const provider_mod = @import("provider.zig");
const config_mod = @import("config.zig");
const conversation = @import("conversation.zig");
const compaction_mod = @import("compaction.zig");
const tool_mod = @import("tool.zig");
const image_mod = @import("image.zig");
const tool_source_mod = @import("tool_source.zig");
const tool_registry_mod = @import("tool_registry.zig");

pub const Tool = tool_mod.Tool;
pub const ToolSource = tool_source_mod.ToolSource;
pub const ToolRegistry = tool_registry_mod.ToolRegistry;

const Entry = tool_registry_mod.Entry;

pub const Config = config_mod.Config;

/// Re-export for the `compact` usages parameter (provider-reported token
/// usage per message, used for retention sizing).
pub const conversation_Usage = @import("session.zig").Usage;

/// Deep-copy a message (role + all content blocks) into fresh owned
/// allocations. Used when rebuilding the conversation after compaction.
fn cloneMessage(alloc: Allocator, msg: conversation.Message) !conversation.Message {
    var content: std.ArrayList(conversation.ContentBlock) = .empty;
    errdefer {
        for (content.items) |*b| b.deinit(alloc);
        content.deinit(alloc);
    }
    try content.ensureTotalCapacity(alloc, msg.content.items.len);
    for (msg.content.items) |block| {
        content.appendAssumeCapacity(try cloneBlock(alloc, block));
    }
    return .{ .role = msg.role, .content = content, .usage = msg.usage };
}

fn cloneBlock(alloc: Allocator, block: conversation.ContentBlock) !conversation.ContentBlock {
    return switch (block) {
        .Text => |b| .{ .Text = try conversation.textualBlockFromSlice(alloc, b.items) },
        .Thinking => |b| blk: {
            const tb = try conversation.textualBlockFromSlice(alloc, b.text.items);
            errdefer {
                var mut = tb;
                mut.deinit(alloc);
            }
            const sig: ?[]const u8 = if (b.signature) |s| try alloc.dupe(u8, s) else null;
            break :blk .{ .Thinking = .{ .text = tb, .signature = sig } };
        },
        .ToolUse => |b| blk: {
            const id = try alloc.dupe(u8, b.id);
            errdefer alloc.free(id);
            const name = try alloc.dupe(u8, b.name);
            errdefer alloc.free(name);
            const input = try conversation.textualBlockFromSlice(alloc, b.input.items);
            break :blk .{ .ToolUse = .{ .id = id, .name = name, .input = input } };
        },
        .ToolResult => |b| blk: {
            const tuid = try alloc.dupe(u8, b.tool_use_id);
            errdefer alloc.free(tuid);
            var parts: std.ArrayList(conversation.ResultPartStored) = .empty;
            errdefer {
                for (parts.items) |*p| p.deinit(alloc);
                parts.deinit(alloc);
            }
            try parts.ensureTotalCapacity(alloc, b.parts.items.len);
            for (b.parts.items) |src| {
                switch (src) {
                    .text => |tb| {
                        const t = try conversation.textualBlockFromSlice(alloc, tb.items);
                        parts.appendAssumeCapacity(.{ .text = t });
                    },
                    .media => |m| {
                        const mt = try alloc.dupe(u8, m.media_type);
                        errdefer alloc.free(mt);
                        const data = try conversation.textualBlockFromSlice(alloc, m.data.items);
                        parts.appendAssumeCapacity(.{ .media = .{ .media_type = mt, .data = data } });
                    },
                }
            }
            break :blk .{ .ToolResult = .{ .tool_use_id = tuid, .parts = parts } };
        },
        .System => |b| .{ .System = .{
            .text = try conversation.textualBlockFromSlice(alloc, b.text.items),
            .mode = b.mode,
        } },
        .CompactionSummary => |b| .{ .CompactionSummary = .{
            .text = try conversation.textualBlockFromSlice(alloc, b.text.items),
        } },
    };
}

/// A minimal receiver that captures the assistant's streamed message for
/// compaction. We don't need incremental events — the assembled message is
/// read off the conversation after the turn — so all callbacks are no-ops.
const CompactionCapture = struct {
    allocator: Allocator,

    fn receiver(self: *CompactionCapture) provider_mod.Receiver {
        return .{ .ptr = self, .vtable = &vt };
    }
    fn deinit(self: *CompactionCapture) void {
        _ = self;
    }
    const vt: provider_mod.ReceiverVTable = .{
        .onMessageStart = onMessageStart,
        .onBlockStart = onBlockStart,
        .onToolDetails = onToolDetails,
        .onContentDelta = onContentDelta,
        .onBlockComplete = onBlockComplete,
        .onMessageComplete = onMessageComplete,
        .onError = onError,
    };
    fn onMessageStart(_: *anyopaque, _: conversation.MessageRole) anyerror!void {}
    fn onBlockStart(_: *anyopaque, _: provider_mod.ContentBlockType, _: usize) anyerror!void {}
    fn onToolDetails(_: *anyopaque, _: usize, _: []const u8, _: []const u8) anyerror!void {}
    fn onContentDelta(_: *anyopaque, _: usize, _: []const u8) anyerror!void {}
    fn onBlockComplete(_: *anyopaque, _: usize, _: conversation.ContentBlock) anyerror!void {}
    fn onMessageComplete(_: *anyopaque, _: conversation.Message, _: ?provider_mod.Usage) anyerror!void {}
    fn onError(_: *anyopaque, _: anyerror) void {}
};

fn isValidToolInput(input: []const u8) bool {
    if (input.len == 0) return true;
    if (input[0] != '{') return true; // legacy tests/tools may use opaque bytes
    var parsed = std.json.parseFromSlice(std.json.Value, std.heap.page_allocator, input, .{}) catch return false;
    defer parsed.deinit();
    return parsed.value == .object;
}

fn invalidInputResult(allocator: Allocator, input: []const u8) ![]tool_mod.ResultPart {
    const msg = try std.fmt.allocPrint(
        allocator,
        "Tool call was not executed: tool input was incomplete or invalid JSON. Partial input: {s}",
        .{input},
    );
    return tool_mod.ownedTextResult(allocator, msg);
}


pub const Agent = struct {
    allocator: Allocator,
    io: Io,
    /// The active configuration snapshot, consulted fresh at the top of
    /// every turn. Immutable while a turn is in flight; swap this pointer
    /// (`setConfig`) between turns to change provider/model/base_url and/or
    /// the visible tool set atomically. The pointee and its registry are
    /// owned by the embedder, not the agent.
    config: *const Config,
    /// Injectable streaming seam. Defaults to the real provider dispatch
    /// (`provider_mod.streamStep`); tests override it with a stub.
    stream_fn: provider_mod.StreamFn = provider_mod.streamStep,
    /// Compaction system prompt used for automatic compaction on context
    /// overflow. Borrowed; set by the embedder (resolved from its
    /// `COMPACTION.md` layers). When null, auto-compaction is disabled and
    /// a context-overflow error propagates unchanged.
    compaction_system_prompt: ?[]const u8 = null,
    /// Set by the embedder after `runStep` returns to learn whether an
    /// automatic compaction occurred this turn (so it can persist the
    /// rewritten conversation). Reset at the top of each `runStep`.
    auto_compacted: bool = false,

    pub fn init(allocator: Allocator, io: Io, config: *const Config) Agent {
        return .{
            .allocator = allocator,
            .io = io,
            .config = config,
        };
    }

    pub fn deinit(self: *Agent) void {
        // The agent owns neither the config snapshot nor the registry it
        // borrows; the embedder tears those down.
        _ = self;
    }

    /// Swap the active configuration snapshot. Takes effect at the start of
    /// the next turn. Safe to call between `runStep` invocations or from a
    /// tool handler that runs between provider steps; never mutates a
    /// snapshot a turn is currently reading.
    pub fn setConfig(self: *Agent, config: *const Config) void {
        self.config = config;
    }

    /// The registry exposed by the active snapshot.
    pub fn registry(self: *const Agent) *const ToolRegistry {
        return self.config.registry;
    }

    /// Drive the conversation forward until the model stops calling tools.
    pub fn runStep(
        self: *Agent,
        conv: *conversation.Conversation,
        receiver: *provider_mod.Receiver,
    ) !void {
        self.auto_compacted = false;
        while (true) {
            // Re-read the config snapshot at the top of each turn so a
            // mid-conversation swap takes effect here, never mid-stream.
            const cfg = self.config;
            self.stream_fn(self.allocator, self.io, cfg, conv, receiver) catch |err| {
                // Automatic compaction on context overflow: compact the
                // conversation once and retry the failed request a single
                // time. If the retry also overflows, surface the error.
                if (err != error.ContextOverflow) return err;
                if (self.auto_compacted) return err; // already retried once
                const sys = self.compaction_system_prompt orelse return err;
                const res = try self.compact(conv, sys, null);
                if (!res.compacted) return err; // nothing to shed; give up
                self.auto_compacted = true;
                // Retry the same request against the compacted context.
                try self.stream_fn(self.allocator, self.io, cfg, conv, receiver);
            };

            const last = conv.messages.items[conv.messages.items.len - 1];
            std.debug.assert(last.role == .assistant);

            // Defense-in-depth: a provider that silently committed an
            // empty assistant message means the turn made no observable
            // progress. Surface it instead of looping back to the prompt.
            if (last.content.items.len == 0) return error.EmptyAssistantResponse;

            if (!hasToolUseBlock(last)) return;

            try self.dispatchToolCalls(conv, last);
        }
    }

    fn hasToolUseBlock(msg: conversation.Message) bool {
        for (msg.content.items) |block| {
            if (block == .ToolUse) return true;
        }
        return false;
    }

    /// Outcome of a compaction attempt.
    pub const CompactionResult = struct {
        /// Whether the conversation was actually compacted. False means the
        /// active conversation already fit within the keep-verbatim budget
        /// (nothing to summarize) — the conversation is unchanged.
        compacted: bool,
        /// Number of whole turns kept verbatim after the summary.
        kept_turns: usize = 0,
        /// Number of conversation messages folded into the summary.
        summarized_messages: usize = 0,
    };

    /// Compact the conversation: summarize an older prefix into a single
    /// `.CompactionSummary` block and keep a recent suffix of whole turns
    /// verbatim. Mutates `conv` in place. The embedder is responsible for
    /// persisting the resulting new messages (the agent never touches the
    /// session log).
    ///
    /// The system prompt survives untouched: all `.system`-role messages
    /// are preserved in order, and no `replace` block is written. Only the
    /// conversation (user/assistant) prefix is summarized.
    ///
    /// Per-message provider usage is read directly off the conversation
    /// (`Message.usage`, set live by the provider and on replay from disk).
    /// `computeSplit` uses it to size the retention window; messages
    /// lacking usage fall back to word counting.
    ///
    /// `extra_instructions`, when non-null, is appended to the compaction
    /// system prompt for this run (the `/compact $ARGUMENTS` path).
    ///
    /// `system_prompt` is the compaction system prompt (resolved by the
    /// embedder from its `COMPACTION.md` layers, or a built-in default).
    pub fn compact(
        self: *Agent,
        conv: *conversation.Conversation,
        system_prompt: []const u8,
        extra_instructions: ?[]const u8,
    ) !CompactionResult {
        const messages = conv.messages.items;

        // Project per-message usage off the conversation for sizing.
        const usages = try self.allocator.alloc(?conversation_Usage, messages.len);
        defer self.allocator.free(usages);
        for (messages, 0..) |m, i| usages[i] = m.usage;

        const split = compaction_mod.computeSplit(messages, usages, self.config.compaction.keep_verbatim);

        // Determine the active conversation start (after any prior summary).
        const active_start: usize = if (conversation.latestCompactionIndex(messages)) |a| a + 1 else 0;

        // Nothing to summarize: the active conversation already fits, or the
        // prefix boundary is at/under the first active turn.
        if (split.prefix_end <= active_start) {
            return .{ .compacted = false };
        }
        // Count how many *conversation* (non-system) messages are in the
        // summarized prefix. If none, this is also a no-op.
        var summarized: usize = 0;
        for (messages[active_start..split.prefix_end]) |m| {
            if (m.role != .system) summarized += 1;
        }
        if (summarized == 0) return .{ .compacted = false };

        // Serialize the prefix transcript and carry forward the latest
        // existing summary (chained-compaction invariant).
        const transcript = try compaction_mod.serializeTranscript(
            self.allocator,
            messages[active_start..split.prefix_end],
        );
        defer self.allocator.free(transcript);

        const previous_summary = compaction_mod.latestSummaryText(messages);
        const body = try compaction_mod.buildRequestBody(self.allocator, transcript, previous_summary);
        defer self.allocator.free(body);

        const summary = try self.runCompactionRequest(system_prompt, body, extra_instructions);
        defer self.allocator.free(summary);

        try self.rewriteWithSummary(conv, split.prefix_end, summary);

        return .{
            .compacted = true,
            .kept_turns = split.kept_turns,
            .summarized_messages = summarized,
        };
    }

    /// Rewrite `conv.messages` to `[all system messages..., summary,
    /// kept-suffix...]`. The summarized conversation prefix (everything
    /// before `prefix_end` that isn't a system message) is dropped; system
    /// messages survive in order; a `.CompactionSummary` user message is
    /// inserted; the kept suffix (`messages[prefix_end..]`) is preserved.
    fn rewriteWithSummary(
        self: *Agent,
        conv: *conversation.Conversation,
        prefix_end: usize,
        summary: []const u8,
    ) !void {
        const alloc = self.allocator;
        const old = conv.messages.items;

        var rebuilt: std.ArrayList(conversation.Message) = .empty;
        errdefer {
            for (rebuilt.items) |*m| m.deinit(alloc);
            rebuilt.deinit(alloc);
        }

        // 1. All system messages from the summarized prefix survive, in
        //    order. (System messages in the kept suffix come along with it
        //    below, so only scan the prefix here.)
        for (old[0..prefix_end]) |*m| {
            if (m.role != .system) continue;
            try rebuilt.append(alloc, try cloneMessage(alloc, m.*));
        }

        // 2. The compaction summary, alone in a user message.
        {
            const tb = try conversation.textualBlockFromSlice(alloc, summary);
            var content: std.ArrayList(conversation.ContentBlock) = .empty;
            errdefer {
                for (content.items) |*b| b.deinit(alloc);
                content.deinit(alloc);
            }
            try content.append(alloc, .{ .CompactionSummary = .{ .text = tb } });
            try rebuilt.append(alloc, .{ .role = .user, .content = content });
        }

        // 3. The kept verbatim suffix.
        for (old[prefix_end..]) |*m| {
            try rebuilt.append(alloc, try cloneMessage(alloc, m.*));
        }

        // Swap in the rebuilt list and free the old one.
        for (conv.messages.items) |*m| m.deinit(alloc);
        conv.messages.deinit(alloc);
        conv.messages = rebuilt;
    }

    /// Run a single compaction provider call against a throwaway
    /// conversation. Returns the assistant's summary text (caller owns).
    ///
    /// Model selection: try `config.compaction.model` if set; on failure,
    /// fall back to the active chat model. Compaction runs with an empty
    /// tool registry and a single user message (the request body); no tools
    /// are exposed and no session logging occurs.
    fn runCompactionRequest(
        self: *Agent,
        system_prompt: []const u8,
        body: []const u8,
        extra_instructions: ?[]const u8,
    ) ![]u8 {
        const alloc = self.allocator;

        // Assemble the effective compaction system prompt (+ extra
        // instructions for a `/compact $ARGUMENTS` run).
        var sys_text: []const u8 = system_prompt;
        var sys_owned: ?[]u8 = null;
        defer if (sys_owned) |s| alloc.free(s);
        if (extra_instructions) |extra| {
            if (extra.len > 0) {
                const combined = try std.fmt.allocPrint(
                    alloc,
                    "{s}\n\n## Additional instructions for this compaction run\n\n{s}",
                    .{ system_prompt, extra },
                );
                sys_owned = combined;
                sys_text = combined;
            }
        }

        var empty_registry = ToolRegistry.init(alloc);
        defer empty_registry.deinit();

        // Try the configured compaction model first, then fall back to the
        // active chat model on any failure.
        if (self.config.compaction.model) |comp_provider| {
            const cfg: config_mod.Config = .{
                .provider = comp_provider,
                .registry = &empty_registry,
                .compaction = self.config.compaction,
            };
            if (self.runSingleCompactionTurn(&cfg, sys_text, body)) |summary| {
                return summary;
            } else |err| {
                std.log.warn("compaction model failed ({t}); falling back to active model", .{err});
            }
        }

        const cfg: config_mod.Config = .{
            .provider = self.config.provider,
            .registry = &empty_registry,
            .compaction = self.config.compaction,
        };
        return self.runSingleCompactionTurn(&cfg, sys_text, body);
    }

    /// One provider call for compaction. Builds a throwaway conversation
    /// (system prompt + one user message), streams a single turn through a
    /// capturing receiver, and returns the assembled assistant text.
    fn runSingleCompactionTurn(
        self: *Agent,
        cfg: *const config_mod.Config,
        system_prompt: []const u8,
        body: []const u8,
    ) ![]u8 {
        const alloc = self.allocator;
        var conv = conversation.Conversation.init(alloc);
        defer conv.deinit();
        try conv.addSystemMessage(system_prompt);
        try conv.addUserMessage(body);

        var capture = CompactionCapture{ .allocator = alloc };
        defer capture.deinit();
        var recv = capture.receiver();

        try self.stream_fn(alloc, self.io, cfg, &conv, &recv);

        // The provider appended an assistant message; gather its text.
        const last = conv.messages.items[conv.messages.items.len - 1];
        if (last.role != .assistant) return error.CompactionNoResponse;
        var out: std.ArrayList(u8) = .empty;
        errdefer out.deinit(alloc);
        for (last.content.items) |block| {
            if (block == .Text) try out.appendSlice(alloc, block.Text.items);
        }
        if (out.items.len == 0) return error.CompactionEmptySummary;
        return out.toOwnedSlice(alloc);
    }

    /// Dispatch every ToolUse block in `assistant_msg`. Groups by owning
    /// registration; one OS thread per group; results assembled in the
    /// original call order.
    fn dispatchToolCalls(
        self: *Agent,
        conv: *conversation.Conversation,
        assistant_msg: conversation.Message,
    ) !void {
        // Build the flat call list (in original order) and group calls
        // by owning registration.
        var calls: std.array_list.Managed(FlatCall) = .init(self.allocator);
        defer calls.deinit();

        for (assistant_msg.content.items) |block| {
            if (block != .ToolUse) continue;
            const tu = block.ToolUse;
            if (!isValidToolInput(tu.input.items)) {
                try calls.append(.{
                    .tool_use_id = tu.id,
                    .tool_name = tu.name,
                    .input = tu.input.items,
                    .entry = null,
                    .result = try invalidInputResult(self.allocator, tu.input.items),
                    .err = null,
                });
                continue;
            }
            const entry = self.config.registry.lookup(tu.name) orelse {
                // Unknown tool: abort the turn with a clear error.
                return error.UnknownTool;
            };
            try calls.append(.{
                .tool_use_id = tu.id,
                .tool_name = tu.name,
                .input = tu.input.items,
                .entry = entry.entry,
                .result = null,
                .err = null,
            });
        }
        std.debug.assert(calls.items.len > 0);

        // Partition into groups. A group's `kind` determines how it
        // runs; the `member_indices` are positions into `calls` (the
        // original call order) so we can write back results without
        // re-ordering.
        var groups: std.array_list.Managed(Group) = .init(self.allocator);
        defer {
            for (groups.items) |*g| g.deinit(self.allocator);
            groups.deinit();
        }
        try buildGroups(self.allocator, calls.items, &groups);

        // Spawn one concurrent task per group via `std.Io.Group`.
        // Single-tool groups run the tool's vtable; source groups run
        // the source's `invoke_batch`. We use `concurrent` rather than
        // `async` because tool work may block on I/O — under a
        // single-threaded `Io` `async` would deadlock; `concurrent`
        // forces real concurrency (or `error.ConcurrencyUnavailable`).
        var task_group: Io.Group = .init;
        // `cancel` is idempotent with `await`; if anything below this
        // point errors before we successfully `await`, this releases
        // the group's resources.
        defer task_group.cancel(self.io);
        errdefer {
            for (calls.items) |*c| {
                if (c.result) |r| tool_mod.freeResultParts(self.allocator, r);
            }
        }

        for (groups.items) |*g| {
            try task_group.concurrent(self.io, runGroup, .{ self, g, calls.items });
        }
        // `error.Canceled` here means cancellation propagated into this
        // dispatch from above; surface it like any other error.
        try task_group.await(self.io);

        // Assemble ToolResult blocks in original call order. If any
        // call errored, prefer to abort the turn — but only after the
        // standard errdefer above has freed remaining results.
        var content: std.ArrayList(conversation.ContentBlock) = .empty;
        errdefer {
            for (content.items) |*b| b.deinit(self.allocator);
            content.deinit(self.allocator);
        }
        try content.ensureTotalCapacity(self.allocator, calls.items.len);

        var first_err: ?anyerror = null;
        for (calls.items) |*c| {
            if (c.err) |e| {
                first_err = e;
                continue;
            }
            const result_parts = c.result orelse {
                // Internal error: every successful call should have left
                // parts behind. Treat as MissingToolResult.
                first_err = error.MissingToolResult;
                continue;
            };
            c.result = null; // ownership transferred below
            defer tool_mod.freeResultParts(self.allocator, result_parts);

            const id_copy = try self.allocator.dupe(u8, c.tool_use_id);
            errdefer self.allocator.free(id_copy);

            var stored: std.ArrayList(conversation.ResultPartStored) = .empty;
            errdefer {
                for (stored.items) |*p| p.deinit(self.allocator);
                stored.deinit(self.allocator);
            }
            try stored.ensureTotalCapacity(self.allocator, result_parts.len);
            for (result_parts) |part| {
                switch (part) {
                    .text => |t| {
                        var buf: conversation.TextualBlock = .empty;
                        errdefer buf.deinit(self.allocator);
                        try buf.appendSlice(self.allocator, t);
                        stored.appendAssumeCapacity(.{ .text = buf });
                    },
                    .media => |m| {
                        // libpanto owns the heavy lifting: detect the type
                        // (when the tool gave no hint), resize large
                        // rasters, then base64-encode for storage. Tools
                        // hand over raw bytes only.
                        const processed = image_mod.process(self.allocator, m.data, m.media_type) catch |e| {
                            // Unrecognized bytes: keep the turn alive by
                            // dropping the attachment and noting it as text.
                            if (e == error.UnknownMediaType) {
                                var note: conversation.TextualBlock = .empty;
                                errdefer note.deinit(self.allocator);
                                try note.appendSlice(self.allocator, "[unrecognized binary attachment dropped]");
                                stored.appendAssumeCapacity(.{ .text = note });
                                continue;
                            }
                            return e;
                        };
                        defer self.allocator.free(processed.data);

                        const mt = try self.allocator.dupe(u8, processed.media_type);
                        errdefer self.allocator.free(mt);

                        const enc = std.base64.standard.Encoder;
                        var buf: conversation.TextualBlock = .empty;
                        errdefer buf.deinit(self.allocator);
                        try buf.resize(self.allocator, enc.calcSize(processed.data.len));
                        _ = enc.encode(buf.items, processed.data);

                        stored.appendAssumeCapacity(.{ .media = .{ .media_type = mt, .data = buf } });
                    },
                }
            }

            content.appendAssumeCapacity(.{ .ToolResult = .{
                .tool_use_id = id_copy,
                .parts = stored,
            } });
        }

        if (first_err) |e| return e;

        try conv.messages.append(self.allocator, .{
            .role = .user,
            .content = content,
        });
    }
};

/// One ToolUse, as flattened into the agent's dispatch list. `result`
/// and `err` are filled in by the worker; exactly one is non-null on
/// successful task completion.
const FlatCall = struct {
    tool_use_id: []const u8, // borrowed from assistant_msg
    tool_name: []const u8, // borrowed from assistant_msg
    input: []const u8, // borrowed from assistant_msg
    entry: ?Entry,

    /// Owned result parts from `Tool.invoke` or `ToolSource.invoke_batch`.
    /// Allocated with the agent's allocator. Transferred into a
    /// ToolResultBlock on success.
    result: ?[]tool_mod.ResultPart,

    /// If non-null, the call failed and the turn must abort.
    err: ?anyerror,
};

/// One dispatch group. Either a single Tool invocation, or a batch of
/// calls headed to one ToolSource.
const Group = union(enum) {
    single: SingleGroup,
    source: SourceGroup,

    pub const SingleGroup = struct {
        tool: Tool,
        /// Index into the flat calls array.
        call_index: usize,
    };

    pub const SourceGroup = struct {
        source: *ToolSource,
        /// Indices into the flat calls array. Owned by the group.
        member_indices: []usize,
    };

    fn deinit(self: *Group, allocator: Allocator) void {
        switch (self.*) {
            .single => {},
            .source => |sg| allocator.free(sg.member_indices),
        }
    }
};

/// Partition the flat call list into groups. Order of groups is
/// arbitrary; order within a `source` group preserves the original
/// call order so that batch results can be written back positionally.
fn buildGroups(
    allocator: Allocator,
    calls: []const FlatCall,
    out: *std.array_list.Managed(Group),
) !void {
    // Map from source pointer to the index of its group in `out`.
    // Buffers per source, accumulated then frozen into slices.
    var pending: std.AutoHashMap(*ToolSource, std.array_list.Managed(usize)) =
        .init(allocator);
    defer {
        var it = pending.valueIterator();
        while (it.next()) |l| l.deinit();
        pending.deinit();
    }

    for (calls, 0..) |c, i| {
        const ent = c.entry orelse continue;
        switch (ent) {
            .single => |t| try out.append(.{ .single = .{ .tool = t, .call_index = i } }),
            .source => |sr| {
                const gop = try pending.getOrPut(sr.source);
                if (!gop.found_existing) {
                    gop.value_ptr.* = std.array_list.Managed(usize).init(allocator);
                }
                try gop.value_ptr.append(i);
            },
        }
    }

    // Freeze each pending list into a source-group entry. We move
    // ownership of the indices into `Group.source.member_indices`.
    var pit = pending.iterator();
    while (pit.next()) |entry| {
        const src = entry.key_ptr.*;
        const indices = try entry.value_ptr.toOwnedSlice();
        try out.append(.{ .source = .{ .source = src, .member_indices = indices } });
    }
}

/// Worker entry point. Runs one group to completion, populating
/// `calls[i].result` or `calls[i].err` for each member call.
///
/// Return type is `void`, which coerces to `Io.Cancelable!void` as
/// required by `Group.concurrent`. Tool errors are reported via
/// `FlatCall.err`, not by returning from this function.
fn runGroup(agent: *Agent, group: *Group, calls: []FlatCall) void {
    switch (group.*) {
        .single => |sg| {
            const i = sg.call_index;
            const c = &calls[i];
            const out = sg.tool.vtable.invoke(sg.tool.ctx, c.input, agent.allocator) catch |e| {
                c.err = e;
                return;
            };
            c.result = out;
        },
        .source => |sg| runSourceGroup(agent, sg, calls),
    }
}

fn runSourceGroup(agent: *Agent, sg: Group.SourceGroup, calls: []FlatCall) void {
    const n = sg.member_indices.len;

    const batch_calls = agent.allocator.alloc(tool_source_mod.Call, n) catch |e| {
        for (sg.member_indices) |i| calls[i].err = e;
        return;
    };
    defer agent.allocator.free(batch_calls);

    const batch_results = agent.allocator.alloc(tool_source_mod.CallResult, n) catch |e| {
        for (sg.member_indices) |i| calls[i].err = e;
        return;
    };
    defer agent.allocator.free(batch_results);

    for (sg.member_indices, 0..) |idx, j| {
        batch_calls[j] = .{
            .tool_name = calls[idx].tool_name,
            .input = calls[idx].input,
        };
        batch_results[j] = .{ .err = error.SourceDroppedCall };
    }

    sg.source.vtable.invoke_batch(
        sg.source.ctx,
        batch_calls,
        batch_results,
        agent.allocator,
    ) catch |e| {
        // Whole-batch failure: free any partial successes the source
        // already wrote, then mark every member as failed.
        for (batch_results) |r| switch (r) {
            .ok => |b| tool_mod.freeResultParts(agent.allocator, b),
            .err => {},
        };
        for (sg.member_indices) |i| calls[i].err = e;
        return;
    };

    // Per-call success/error.
    for (sg.member_indices, 0..) |i, j| {
        switch (batch_results[j]) {
            .ok => |b| calls[i].result = b,
            .err => |e| calls[i].err = e,
        }
    }
}

// -----------------------------------------------------------------------------
// Tests
// -----------------------------------------------------------------------------

const testing = std.testing;

/// Test helper: the items of a ToolResultBlock's first text part.
fn trText(tr: conversation.ToolResultBlock) []const u8 {
    for (tr.parts.items) |p| {
        if (p == .text) return p.text.items;
    }
    return "";
}

/// Test harness for the injectable `stream_fn` seam.
///
/// `provider_mod.StreamFn` carries no user context (it mirrors the real
/// free function exactly), so the stub parks its state in a module-level
/// pointer that `stubStreamStep` reads. The Zig test runner executes tests
/// serially in one process, so a single global slot is safe; each test
/// sets it via `install` before driving the agent.
var stub_active: ?*StubProvider = null;

const StubProvider = struct {
    allocator: Allocator,
    scripted: []const ScriptedTurn,
    next: usize = 0,
    /// Number of leading stream calls that should fail with
    /// `error.ContextOverflow` before any scripted turn is served. Used to
    /// drive the auto-compaction path. Decremented on each overflow.
    overflow_calls: usize = 0,

    const ScriptedTurn = struct {
        blocks: []const TestBlock,
    };

    const TestBlock = union(enum) {
        Text: []const u8,
        ToolUse: struct {
            id: []const u8,
            name: []const u8,
            input: []const u8,
        },
    };

    /// Point the global seam at this stub and return the function to assign
    /// to `agent.stream_fn`. Call once per test, after constructing the
    /// stub on the stack.
    fn install(self: *StubProvider) provider_mod.StreamFn {
        stub_active = self;
        return stubStreamStep;
    }
};

fn stubStreamStep(
    allocator: Allocator,
    _: Io,
    _: *const config_mod.Config,
    conv: *conversation.Conversation,
    _: *provider_mod.Receiver,
) anyerror!void {
    const self = stub_active orelse return error.NoStubInstalled;
    _ = allocator;
    if (self.overflow_calls > 0) {
        self.overflow_calls -= 1;
        return error.ContextOverflow;
    }
    if (self.next >= self.scripted.len) return error.NoMoreScriptedTurns;
    const turn = self.scripted[self.next];
    self.next += 1;

    var blocks: std.ArrayList(conversation.ContentBlock) = .empty;
    errdefer {
        for (blocks.items) |*b| b.deinit(self.allocator);
        blocks.deinit(self.allocator);
    }
    for (turn.blocks) |tb| {
        switch (tb) {
            .Text => |s| {
                try blocks.append(self.allocator, .{
                    .Text = try conversation.textualBlockFromSlice(self.allocator, s),
                });
            },
            .ToolUse => |tu| {
                const id = try self.allocator.dupe(u8, tu.id);
                errdefer self.allocator.free(id);
                const name = try self.allocator.dupe(u8, tu.name);
                errdefer self.allocator.free(name);
                var input_buf: conversation.TextualBlock = .empty;
                errdefer input_buf.deinit(self.allocator);
                try input_buf.appendSlice(self.allocator, tu.input);
                try blocks.append(self.allocator, .{ .ToolUse = .{
                    .id = id,
                    .name = name,
                    .input = input_buf,
                } });
            },
        }
    }
    const moved = try blocks.toOwnedSlice(self.allocator);
    defer self.allocator.free(moved);
    try conv.addAssistantMessage(moved);
}

/// Build a stack registry + active `Config` snapshot wired together, for
/// tests that drive the agent. The caller owns both and must keep them
/// alive for the agent's lifetime.
const TestHarness = struct {
    registry: ToolRegistry,
    config: config_mod.Config,

    fn init(allocator: Allocator) TestHarness {
        return .{ .registry = ToolRegistry.init(allocator), .config = undefined };
    }

    /// Finalize the config snapshot to point at this harness's registry.
    /// Must be called after `init` and before constructing the agent, once
    /// the harness has a stable address.
    fn activate(self: *TestHarness) void {
        self.config = .{
            .provider = .{ .openai_chat = .{ .api_key = "k", .base_url = "u", .model = "m" } },
            .registry = &self.registry,
        };
    }

    fn deinit(self: *TestHarness) void {
        self.registry.deinit();
    }
};

const EchoTool = struct {
    prefix_owned: []u8,
    name_owned: []u8,

    fn create(allocator: Allocator, name: []const u8, prefix: []const u8) !Tool {
        const self = try allocator.create(EchoTool);
        errdefer allocator.destroy(self);
        self.name_owned = try allocator.dupe(u8, name);
        errdefer allocator.free(self.name_owned);
        self.prefix_owned = try allocator.dupe(u8, prefix);
        return .{
            .decl = .{
                .name = self.name_owned,
                .description = "echo",
                .schema_json = "{}",
            },
            .ctx = self,
            .vtable = &vt,
        };
    }

    const vt: Tool.VTable = .{ .invoke = invoke, .deinit = deinit };

    fn invoke(ctx: *anyopaque, input: []const u8, allocator: Allocator) anyerror![]tool_mod.ResultPart {
        const self: *EchoTool = @ptrCast(@alignCast(ctx));
        const msg = try std.fmt.allocPrint(allocator, "{s}{s}", .{ self.prefix_owned, input });
        return tool_mod.ownedTextResult(allocator, msg);
    }

    fn deinit(ctx: *anyopaque, allocator: Allocator) void {
        const self: *EchoTool = @ptrCast(@alignCast(ctx));
        allocator.free(self.name_owned);
        allocator.free(self.prefix_owned);
        allocator.destroy(self);
    }
};

const BarrierTool = struct {
    name_owned: []u8,
    barrier: *Barrier,

    const Barrier = struct {
        target: u32,
        arrived: std.atomic.Value(u32) = .init(0),
        thread_ids: [4]std.atomic.Value(u64) = .{
            .init(0), .init(0), .init(0), .init(0),
        },
    };

    fn create(allocator: Allocator, name: []const u8, barrier: *Barrier) !Tool {
        const self = try allocator.create(BarrierTool);
        errdefer allocator.destroy(self);
        self.name_owned = try allocator.dupe(u8, name);
        self.barrier = barrier;
        return .{
            .decl = .{
                .name = self.name_owned,
                .description = "barrier",
                .schema_json = "{}",
            },
            .ctx = self,
            .vtable = &vt,
        };
    }

    const vt: Tool.VTable = .{ .invoke = invoke, .deinit = deinit };

    fn invoke(ctx: *anyopaque, _: []const u8, allocator: Allocator) anyerror![]tool_mod.ResultPart {
        const self: *BarrierTool = @ptrCast(@alignCast(ctx));
        const arrived = self.barrier.arrived.fetchAdd(1, .acq_rel);
        if (arrived < self.barrier.thread_ids.len) {
            self.barrier.thread_ids[arrived].store(std.Thread.getCurrentId(), .release);
        }

        var i: usize = 0;
        while (self.barrier.arrived.load(.acquire) < self.barrier.target) : (i += 1) {
            if (i > 50_000) return error.BarrierTimeout;
            std.Thread.yield() catch {};
        }
        return tool_mod.textResult(allocator, "done");
    }

    fn deinit(ctx: *anyopaque, allocator: Allocator) void {
        const self: *BarrierTool = @ptrCast(@alignCast(ctx));
        allocator.free(self.name_owned);
        allocator.destroy(self);
    }
};

const FailingTool = struct {
    name_owned: []u8,

    fn create(allocator: Allocator, name: []const u8) !Tool {
        const self = try allocator.create(FailingTool);
        errdefer allocator.destroy(self);
        self.name_owned = try allocator.dupe(u8, name);
        return .{
            .decl = .{
                .name = self.name_owned,
                .description = "fails",
                .schema_json = "{}",
            },
            .ctx = self,
            .vtable = &vt,
        };
    }

    const vt: Tool.VTable = .{ .invoke = invoke, .deinit = deinit };

    fn invoke(_: *anyopaque, _: []const u8, _: Allocator) anyerror![]tool_mod.ResultPart {
        return error.ToolExploded;
    }

    fn deinit(ctx: *anyopaque, allocator: Allocator) void {
        const self: *FailingTool = @ptrCast(@alignCast(ctx));
        allocator.free(self.name_owned);
        allocator.destroy(self);
    }
};

const NoopReceiver = struct {
    fn make() provider_mod.Receiver {
        return .{ .ptr = @ptrCast(@constCast(&dummy)), .vtable = &vt };
    }
    var dummy: u8 = 0;
    const vt: provider_mod.ReceiverVTable = .{
        .onMessageStart = noop1,
        .onBlockStart = noop2,
        .onToolDetails = noopToolDetails,
        .onContentDelta = noop3,
        .onBlockComplete = noop4,
        .onMessageComplete = noop5,
        .onError = noop6,
    };
    fn noop1(_: *anyopaque, _: conversation.MessageRole) anyerror!void {}
    fn noop2(_: *anyopaque, _: provider_mod.ContentBlockType, _: usize) anyerror!void {}
    fn noopToolDetails(_: *anyopaque, _: usize, _: []const u8, _: []const u8) anyerror!void {}
    fn noop3(_: *anyopaque, _: usize, _: []const u8) anyerror!void {}
    fn noop4(_: *anyopaque, _: usize, _: conversation.ContentBlock) anyerror!void {}
    fn noop5(_: *anyopaque, _: conversation.Message, _: ?provider_mod.Usage) anyerror!void {}
    fn noop6(_: *anyopaque, _: anyerror) void {}
};

/// A configurable ToolSource for testing the grouped-dispatch path.
/// Stores every batch it receives so tests can assert "calls X and Y
/// arrived in the same batch on the same thread".
const TestSource = struct {
    name_owned: []u8,
    decls: []tool_source_mod.ToolDecl,
    decl_strings: std.array_list.Managed([]u8),
    /// Sequence of (thread_id, [tool_name; n]) per batch received.
    /// Only mutated inside `invoke_batch`. Because libpanto guarantees
    /// at most one outstanding `invoke_batch` per source at any time
    /// (one batch per turn per source), no synchronization is needed.
    batches: std.array_list.Managed(Batch),
    allocator: Allocator,

    const Batch = struct {
        thread_id: u64,
        names: std.array_list.Managed([]u8),
    };

    fn create(
        allocator: Allocator,
        source_name: []const u8,
        tool_names: []const []const u8,
    ) !ToolSource {
        const self = try allocator.create(TestSource);
        errdefer allocator.destroy(self);

        var strings = std.array_list.Managed([]u8).init(allocator);
        errdefer {
            for (strings.items) |s| allocator.free(s);
            strings.deinit();
        }

        const name_owned = try allocator.dupe(u8, source_name);
        try strings.append(name_owned);

        const decls = try allocator.alloc(tool_source_mod.ToolDecl, tool_names.len);
        errdefer allocator.free(decls);
        for (tool_names, 0..) |tn, i| {
            const n = try allocator.dupe(u8, tn);
            try strings.append(n);
            const d = try allocator.dupe(u8, "test src tool");
            try strings.append(d);
            const s = try allocator.dupe(u8, "{}");
            try strings.append(s);
            decls[i] = .{ .name = n, .description = d, .schema_json = s };
        }

        self.* = .{
            .name_owned = name_owned,
            .decls = decls,
            .decl_strings = strings,
            .batches = std.array_list.Managed(Batch).init(allocator),
            .allocator = allocator,
        };

        return ToolSource{
            .name = self.name_owned,
            .tools = self.decls,
            .ctx = self,
            .vtable = &vt,
        };
    }

    const vt: ToolSource.VTable = .{
        .invoke_batch = invokeBatch,
        .deinit = deinitSrc,
    };

    fn invokeBatch(
        ctx: *anyopaque,
        calls: []const tool_source_mod.Call,
        results: []tool_source_mod.CallResult,
        allocator: Allocator,
    ) anyerror!void {
        const self: *TestSource = @ptrCast(@alignCast(ctx));
        var batch: Batch = .{
            .thread_id = std.Thread.getCurrentId(),
            .names = std.array_list.Managed([]u8).init(self.allocator),
        };
        for (calls) |c| {
            const copy = try self.allocator.dupe(u8, c.tool_name);
            try batch.names.append(copy);
        }
        try self.batches.append(batch);

        for (calls, 0..) |c, i| {
            const msg = std.fmt.allocPrint(
                allocator,
                "{s}->{s}",
                .{ c.tool_name, c.input },
            ) catch |e| {
                results[i] = .{ .err = e };
                continue;
            };
            results[i] = .{
                .ok = tool_mod.ownedTextResult(allocator, msg) catch |e| {
                    results[i] = .{ .err = e };
                    continue;
                },
            };
        }
    }

    fn deinitSrc(ctx: *anyopaque, _: Allocator) void {
        const self: *TestSource = @ptrCast(@alignCast(ctx));
        for (self.decl_strings.items) |s| self.allocator.free(s);
        self.decl_strings.deinit();
        for (self.batches.items) |*b| {
            for (b.names.items) |n| self.allocator.free(n);
            b.names.deinit();
        }
        self.batches.deinit();
        self.allocator.free(self.decls);
        self.allocator.destroy(self);
    }
};

/// A source that always fails the whole batch by returning an error
/// from invoke_batch (rather than recording per-call errors). Used to
/// verify libpanto's whole-batch-failure path.
const FailingSource = struct {
    name_owned: []u8,
    decls: []tool_source_mod.ToolDecl,
    decl_strings: std.array_list.Managed([]u8),
    allocator: Allocator,

    fn create(allocator: Allocator, source_name: []const u8, tool_names: []const []const u8) !ToolSource {
        const self = try allocator.create(FailingSource);
        errdefer allocator.destroy(self);

        var strings = std.array_list.Managed([]u8).init(allocator);
        errdefer {
            for (strings.items) |s| allocator.free(s);
            strings.deinit();
        }

        const name_owned = try allocator.dupe(u8, source_name);
        try strings.append(name_owned);

        const decls = try allocator.alloc(tool_source_mod.ToolDecl, tool_names.len);
        errdefer allocator.free(decls);
        for (tool_names, 0..) |tn, i| {
            const n = try allocator.dupe(u8, tn);
            try strings.append(n);
            const d = try allocator.dupe(u8, "fails");
            try strings.append(d);
            const s = try allocator.dupe(u8, "{}");
            try strings.append(s);
            decls[i] = .{ .name = n, .description = d, .schema_json = s };
        }

        self.* = .{
            .name_owned = name_owned,
            .decls = decls,
            .decl_strings = strings,
            .allocator = allocator,
        };
        return ToolSource{ .name = self.name_owned, .tools = self.decls, .ctx = self, .vtable = &vt };
    }

    const vt: ToolSource.VTable = .{ .invoke_batch = invokeBatch, .deinit = deinitSrc };

    fn invokeBatch(
        _: *anyopaque,
        _: []const tool_source_mod.Call,
        _: []tool_source_mod.CallResult,
        _: Allocator,
    ) anyerror!void {
        return error.SourceExploded;
    }

    fn deinitSrc(ctx: *anyopaque, _: Allocator) void {
        const self: *FailingSource = @ptrCast(@alignCast(ctx));
        for (self.decl_strings.items) |s| self.allocator.free(s);
        self.decl_strings.deinit();
        self.allocator.free(self.decls);
        self.allocator.destroy(self);
    }
};

test "registry register and lookup" {
    var h = TestHarness.init(testing.allocator);
    defer h.deinit();
    try h.registry.register(try EchoTool.create(testing.allocator, "echo", "ECHO:"));
    try testing.expectEqual(@as(usize, 1), h.registry.count());
    try testing.expect(h.registry.lookup("echo") != null);
}

test "duplicate register returns error" {
    var h = TestHarness.init(testing.allocator);
    defer h.deinit();
    try h.registry.register(try EchoTool.create(testing.allocator, "echo", "A:"));

    var dup = try EchoTool.create(testing.allocator, "echo", "B:");
    try testing.expectError(error.DuplicateTool, h.registry.register(dup));
    dup.vtable.deinit(dup.ctx, testing.allocator);
}

test "runStep dispatches a tool call and loops to a final text turn" {
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{
            .{ .ToolUse = .{ .id = "tc_1", .name = "echo", .input = "hello" } },
        } },
        .{ .blocks = &.{
            .{ .Text = "ok" },
        } },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    try h.registry.register(try EchoTool.create(allocator, "echo", "ECHO:"));
    h.activate();
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("call a tool");

    var recv = NoopReceiver.make();
    try agent.runStep(&conv, &recv);

    try testing.expectEqual(@as(usize, 4), conv.messages.items.len);

    try testing.expectEqual(conversation.MessageRole.assistant, conv.messages.items[1].role);
    try testing.expectEqual(@as(usize, 1), conv.messages.items[1].content.items.len);
    try testing.expectEqualStrings("tc_1", conv.messages.items[1].content.items[0].ToolUse.id);

    try testing.expectEqual(conversation.MessageRole.user, conv.messages.items[2].role);
    try testing.expectEqual(@as(usize, 1), conv.messages.items[2].content.items.len);
    const tr = conv.messages.items[2].content.items[0].ToolResult;
    try testing.expectEqualStrings("tc_1", tr.tool_use_id);
    try testing.expectEqualStrings("ECHO:hello", trText(tr));

    try testing.expectEqual(conversation.MessageRole.assistant, conv.messages.items[3].role);
    try testing.expectEqualStrings("ok", conv.messages.items[3].content.items[0].Text.items);
}

test "runStep dispatches multiple tool calls in parallel" {
    const allocator = testing.allocator;

    var barrier: BarrierTool.Barrier = .{ .target = 3 };

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{
            .{ .ToolUse = .{ .id = "a", .name = "barrierA", .input = "" } },
            .{ .ToolUse = .{ .id = "b", .name = "barrierB", .input = "" } },
            .{ .ToolUse = .{ .id = "c", .name = "barrierC", .input = "" } },
        } },
        .{ .blocks = &.{
            .{ .Text = "done" },
        } },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    try h.registry.register(try BarrierTool.create(allocator, "barrierA", &barrier));
    try h.registry.register(try BarrierTool.create(allocator, "barrierB", &barrier));
    try h.registry.register(try BarrierTool.create(allocator, "barrierC", &barrier));
    h.activate();
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("go");

    var recv = NoopReceiver.make();
    try agent.runStep(&conv, &recv);

    const tr_msg = conv.messages.items[2];
    try testing.expectEqual(@as(usize, 3), tr_msg.content.items.len);
    try testing.expectEqualStrings("a", tr_msg.content.items[0].ToolResult.tool_use_id);
    try testing.expectEqualStrings("b", tr_msg.content.items[1].ToolResult.tool_use_id);
    try testing.expectEqualStrings("c", tr_msg.content.items[2].ToolResult.tool_use_id);

    const t0 = barrier.thread_ids[0].load(.acquire);
    const t1 = barrier.thread_ids[1].load(.acquire);
    const t2 = barrier.thread_ids[2].load(.acquire);
    try testing.expect(t0 != 0 and t1 != 0 and t2 != 0);
    try testing.expect(t0 != t1 and t1 != t2 and t0 != t2);
}

test "runStep propagates tool errors and aborts the turn" {
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{
            .{ .ToolUse = .{ .id = "x", .name = "boom", .input = "" } },
        } },
        .{ .blocks = &.{.{ .Text = "should-not-see" }} },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    try h.registry.register(try FailingTool.create(allocator, "boom"));
    h.activate();
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("break it");

    var recv = NoopReceiver.make();
    try testing.expectError(error.ToolExploded, agent.runStep(&conv, &recv));

    try testing.expectEqual(@as(usize, 2), conv.messages.items.len);
}

test "runStep errors UnknownTool when the model calls something unregistered" {
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{
            .{ .ToolUse = .{ .id = "z", .name = "ghost", .input = "" } },
        } },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    h.activate();
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("call a ghost");

    var recv = NoopReceiver.make();
    try testing.expectError(error.UnknownTool, agent.runStep(&conv, &recv));
}

test "runStep with no tool calls returns after one provider step" {
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{.{ .Text = "hi" }} },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    h.activate();
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("hello");

    var recv = NoopReceiver.make();
    try agent.runStep(&conv, &recv);

    try testing.expectEqual(@as(usize, 2), conv.messages.items.len);
    try testing.expectEqualStrings("hi", conv.messages.items[1].content.items[0].Text.items);
}

test "runStep surfaces EmptyAssistantResponse when provider commits an empty message" {
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{} },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    h.activate();
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("hi");

    var recv = NoopReceiver.make();
    try testing.expectError(error.EmptyAssistantResponse, agent.runStep(&conv, &recv));
}

// ------------ ToolSource tests ------------

test "runStep delivers all source-backed calls in one batch on one thread" {
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{
            .{ .ToolUse = .{ .id = "a", .name = "lua_x", .input = "1" } },
            .{ .ToolUse = .{ .id = "b", .name = "lua_y", .input = "2" } },
            .{ .ToolUse = .{ .id = "c", .name = "lua_x", .input = "3" } },
        } },
        .{ .blocks = &.{.{ .Text = "done" }} },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    try h.registry.registerSource(try TestSource.create(allocator, "panto-lua", &.{ "lua_x", "lua_y" }));
    h.activate();
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("go");

    var recv = NoopReceiver.make();
    try agent.runStep(&conv, &recv);

    // Locate the source and inspect its observed batches.
    const view = h.registry.lookup("lua_x") orelse return error.NotFound;
    const src_ptr = view.entry.source.source;
    const test_src: *TestSource = @ptrCast(@alignCast(src_ptr.ctx));

    try testing.expectEqual(@as(usize, 1), test_src.batches.items.len);
    const b = test_src.batches.items[0];
    try testing.expectEqual(@as(usize, 3), b.names.items.len);
    try testing.expectEqualStrings("lua_x", b.names.items[0]);
    try testing.expectEqualStrings("lua_y", b.names.items[1]);
    try testing.expectEqualStrings("lua_x", b.names.items[2]);

    // ToolResults arrived in the original call order.
    const tr_msg = conv.messages.items[2];
    try testing.expectEqual(@as(usize, 3), tr_msg.content.items.len);
    try testing.expectEqualStrings("a", tr_msg.content.items[0].ToolResult.tool_use_id);
    try testing.expectEqualStrings("lua_x->1", trText(tr_msg.content.items[0].ToolResult));
    try testing.expectEqualStrings("b", tr_msg.content.items[1].ToolResult.tool_use_id);
    try testing.expectEqualStrings("lua_y->2", trText(tr_msg.content.items[1].ToolResult));
    try testing.expectEqualStrings("c", tr_msg.content.items[2].ToolResult.tool_use_id);
    try testing.expectEqualStrings("lua_x->3", trText(tr_msg.content.items[2].ToolResult));
}

test "runStep: distinct sources run on distinct threads in parallel" {
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{
            .{ .ToolUse = .{ .id = "a", .name = "src_a_t", .input = "" } },
            .{ .ToolUse = .{ .id = "b", .name = "src_b_t", .input = "" } },
        } },
        .{ .blocks = &.{.{ .Text = "done" }} },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    try h.registry.registerSource(try TestSource.create(allocator, "src_a", &.{"src_a_t"}));
    try h.registry.registerSource(try TestSource.create(allocator, "src_b", &.{"src_b_t"}));
    h.activate();
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("go");

    var recv = NoopReceiver.make();
    try agent.runStep(&conv, &recv);

    const view_a = h.registry.lookup("src_a_t") orelse return error.NotFound;
    const view_b = h.registry.lookup("src_b_t") orelse return error.NotFound;
    const sa: *TestSource = @ptrCast(@alignCast(view_a.entry.source.source.ctx));
    const sb: *TestSource = @ptrCast(@alignCast(view_b.entry.source.source.ctx));

    try testing.expectEqual(@as(usize, 1), sa.batches.items.len);
    try testing.expectEqual(@as(usize, 1), sb.batches.items.len);
    // The two sources ran on distinct OS threads.
    try testing.expect(sa.batches.items[0].thread_id != sb.batches.items[0].thread_id);
}

test "runStep: source whole-batch error aborts the turn" {
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{
            .{ .ToolUse = .{ .id = "a", .name = "fa", .input = "" } },
            .{ .ToolUse = .{ .id = "b", .name = "fb", .input = "" } },
        } },
        .{ .blocks = &.{.{ .Text = "never" }} },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    try h.registry.registerSource(try FailingSource.create(allocator, "fs", &.{ "fa", "fb" }));
    h.activate();
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("kaboom");

    var recv = NoopReceiver.make();
    try testing.expectError(error.SourceExploded, agent.runStep(&conv, &recv));

    // Conversation stops at user + assistant(tool_use). No ToolResult appended.
    try testing.expectEqual(@as(usize, 2), conv.messages.items.len);
}

test "runStep: mixed single Tools and source-backed tools coexist in one turn" {
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{
            .{ .ToolUse = .{ .id = "a", .name = "single", .input = "X" } },
            .{ .ToolUse = .{ .id = "b", .name = "src_t1", .input = "Y" } },
            .{ .ToolUse = .{ .id = "c", .name = "src_t2", .input = "Z" } },
        } },
        .{ .blocks = &.{.{ .Text = "done" }} },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    try h.registry.register(try EchoTool.create(allocator, "single", "S:"));
    try h.registry.registerSource(try TestSource.create(allocator, "src", &.{ "src_t1", "src_t2" }));
    h.activate();
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("go");

    var recv = NoopReceiver.make();
    try agent.runStep(&conv, &recv);

    const tr_msg = conv.messages.items[2];
    try testing.expectEqual(@as(usize, 3), tr_msg.content.items.len);
    try testing.expectEqualStrings("S:X", trText(tr_msg.content.items[0].ToolResult));
    try testing.expectEqualStrings("src_t1->Y", trText(tr_msg.content.items[1].ToolResult));
    try testing.expectEqualStrings("src_t2->Z", trText(tr_msg.content.items[2].ToolResult));
}

test "setConfig swaps the visible tool set between turns" {
    // The core RCU promise: the agent reads `*const Config` fresh each
    // turn, so swapping the pointer mid-conversation changes the tool set
    // the next turn sees. Config A exposes only `echo`; config B only
    // `late`. After `setConfig(&cfg_b)`, a turn that calls `late` resolves
    // — proving both the swap and per-turn re-consultation.
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{.{ .ToolUse = .{ .id = "2", .name = "late", .input = "B" } }} },
        .{ .blocks = &.{.{ .Text = "done" }} },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();

    // Config A: only `echo`.
    var reg_a = ToolRegistry.init(allocator);
    defer reg_a.deinit();
    try reg_a.register(try EchoTool.create(allocator, "echo", "A:"));
    const cfg_a: config_mod.Config = .{
        .provider = .{ .openai_chat = .{ .api_key = "k", .base_url = "u", .model = "m" } },
        .registry = &reg_a,
    };

    // Config B: only `late`.
    var reg_b = ToolRegistry.init(allocator);
    defer reg_b.deinit();
    try reg_b.register(try EchoTool.create(allocator, "late", "B:"));
    const cfg_b: config_mod.Config = .{
        .provider = .{ .openai_chat = .{ .api_key = "k", .base_url = "u", .model = "m" } },
        .registry = &reg_b,
    };

    var agent = Agent.init(allocator, io, &cfg_a);
    agent.stream_fn = stub.install();

    // Under A: `echo` visible, `late` not.
    try testing.expect(agent.config.registry.lookup("echo") != null);
    try testing.expect(agent.config.registry.lookup("late") == null);

    // Swap. Under B: the visibility inverts.
    agent.setConfig(&cfg_b);
    try testing.expect(agent.config.registry.lookup("echo") == null);
    try testing.expect(agent.config.registry.lookup("late") != null);

    // A real turn under B resolves `late` (which would have been
    // UnknownTool under A), then loops to the final text turn.
    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("go");
    var recv = NoopReceiver.make();
    try agent.runStep(&conv, &recv);

    const tr = conv.messages.items[2].content.items[0].ToolResult;
    try testing.expectEqualStrings("2", tr.tool_use_id);
    try testing.expectEqualStrings("B:B", trText(tr));
}

test "compact: summarizes prefix, keeps suffix, system survives" {
    const allocator = testing.allocator;

    // The stub returns a single text turn — used as the summary text.
    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{.{ .Text = "SUMMARY OF EARLIER" }} },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    h.activate();
    // keep_verbatim sized so only the last (short) turn fits: q2+a2 are
    // 3 words each => ceil(3*1.3)=4 tokens each => 8 total <= 10, while
    // adding the longer first turn exceeds it.
    h.config.compaction = .{ .keep_verbatim = 10 };
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addSystemMessage("you are helpful");
    try conv.addUserMessage("first question here with several words");
    try conv.addAssistantMessage(&.{
        .{ .Text = try conversation.textualBlockFromSlice(allocator, "first answer with several words") },
    });
    try conv.addUserMessage("second recent question");
    try conv.addAssistantMessage(&.{
        .{ .Text = try conversation.textualBlockFromSlice(allocator, "second recent answer") },
    });

    const res = try agent.compact(&conv, "Summarize the conversation.", null);
    try testing.expect(res.compacted);

    // Expected rebuilt: [system, compaction summary(user), user q2, asst a2]
    try testing.expectEqual(@as(usize, 4), conv.messages.items.len);
    try testing.expectEqual(conversation.MessageRole.system, conv.messages.items[0].role);
    try testing.expectEqualStrings(
        "you are helpful",
        conv.messages.items[0].content.items[0].System.text.items,
    );
    try testing.expectEqual(conversation.MessageRole.user, conv.messages.items[1].role);
    try testing.expectEqualStrings(
        "SUMMARY OF EARLIER",
        conv.messages.items[1].content.items[0].CompactionSummary.text.items,
    );
    try testing.expectEqualStrings(
        "second recent question",
        conv.messages.items[2].content.items[0].Text.items,
    );
    try testing.expectEqualStrings(
        "second recent answer",
        conv.messages.items[3].content.items[0].Text.items,
    );
}

test "compact: no-op when conversation already fits the budget" {
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{.{ .Text = "should not be used" }} },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    h.activate();
    h.config.compaction = .{ .keep_verbatim = 1_000_000 };
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addSystemMessage("sys");
    try conv.addUserMessage("hi");
    try conv.addAssistantMessage(&.{
        .{ .Text = try conversation.textualBlockFromSlice(allocator, "hello") },
    });

    const res = try agent.compact(&conv, "Summarize.", null);
    try testing.expect(!res.compacted);
    try testing.expectEqual(@as(usize, 3), conv.messages.items.len);
    // Stub was never consumed.
    try testing.expectEqual(@as(usize, 0), stub.next);
}

test "compact: extra instructions are appended to the system prompt" {
    const allocator = testing.allocator;

    // Capture the system prompt the stub sees by scripting a turn and
    // inspecting the throwaway conversation isn't directly possible via the
    // current stub; instead we just assert compaction succeeds with extra
    // instructions present (smoke test of the append path).
    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{.{ .Text = "S" }} },
    };
    var stub = StubProvider{ .allocator = allocator, .scripted = &scripted };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    h.activate();
    h.config.compaction = .{ .keep_verbatim = 1 };
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("question one two three");
    try conv.addAssistantMessage(&.{
        .{ .Text = try conversation.textualBlockFromSlice(allocator, "answer one two three") },
    });
    try conv.addUserMessage("question two");
    try conv.addAssistantMessage(&.{
        .{ .Text = try conversation.textualBlockFromSlice(allocator, "answer two") },
    });

    const res = try agent.compact(&conv, "Base prompt.", "keep bug #3 details");
    try testing.expect(res.compacted);
}

test "runStep: auto-compacts on context overflow and retries once" {
    const allocator = testing.allocator;

    // First stream call overflows; then the compaction request returns a
    // summary; then the retried main request returns a final text turn.
    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{.{ .Text = "COMPACTED SUMMARY" }} }, // compaction call
        .{ .blocks = &.{.{ .Text = "final answer" }} }, // retried main call
    };
    var stub = StubProvider{
        .allocator = allocator,
        .scripted = &scripted,
        .overflow_calls = 1,
    };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    h.activate();
    h.config.compaction = .{ .keep_verbatim = 10 };
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();
    agent.compaction_system_prompt = "Summarize the conversation.";

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addSystemMessage("you are helpful");
    try conv.addUserMessage("first question with several words here");
    try conv.addAssistantMessage(&.{
        .{ .Text = try conversation.textualBlockFromSlice(allocator, "first answer with several words") },
    });
    try conv.addUserMessage("second recent question");

    var recv = NoopReceiver.make();
    try agent.runStep(&conv, &recv);

    try testing.expect(agent.auto_compacted);
    // After compaction + retry: [system, summary, user q2, assistant final].
    const msgs = conv.messages.items;
    try testing.expectEqual(conversation.MessageRole.system, msgs[0].role);
    try testing.expectEqualStrings(
        "COMPACTED SUMMARY",
        msgs[1].content.items[0].CompactionSummary.text.items,
    );
    try testing.expectEqualStrings("second recent question", msgs[2].content.items[0].Text.items);
    try testing.expectEqualStrings("final answer", msgs[msgs.len - 1].content.items[0].Text.items);
}

test "runStep: context overflow without compaction prompt propagates" {
    const allocator = testing.allocator;

    const scripted = [_]StubProvider.ScriptedTurn{
        .{ .blocks = &.{.{ .Text = "unused" }} },
    };
    var stub = StubProvider{
        .allocator = allocator,
        .scripted = &scripted,
        .overflow_calls = 1,
    };
    var threaded: std.Io.Threaded = .init(allocator, .{});
    defer threaded.deinit();
    const io = threaded.io();
    var h = TestHarness.init(allocator);
    defer h.deinit();
    h.activate();
    var agent = Agent.init(allocator, io, &h.config);
    agent.stream_fn = stub.install();
    // No compaction_system_prompt set -> overflow propagates.

    var conv = conversation.Conversation.init(allocator);
    defer conv.deinit();
    try conv.addUserMessage("hi");

    var recv = NoopReceiver.make();
    try testing.expectError(error.ContextOverflow, agent.runStep(&conv, &recv));
}