initial commit, moved libpanto over from the pantograph repo

1 files changed, 3925 insertions, 0 deletions

diff --git a/src/agent.zig b/src/agent.zig
new file mode 100644
index 0000000..09b0994
--- /dev/null
+++ b/src/agent.zig
@@ -0,0 +1,3925 @@
+//! 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 stream_mod = @import("stream.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");
+const session_store_mod = @import("session_store.zig");
+const null_store_mod = @import("null_store.zig");
+const turn_persist = @import("turn_persist.zig");
+
+pub const Tool = tool_mod.Tool;
+pub const ToolSource = tool_source_mod.ToolSource;
+pub const ToolRegistry = tool_registry_mod.ToolRegistry;
+
+const Event = stream_mod.Event;
+
+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;
+
+/// Append a single-text user message. `Conversation.addUserMessage` now
+/// takes a block slice (symmetric with `addAssistantMessage`); this wraps
+/// the common plain-text case used by the agent's interactive turn and the
+/// compaction prompt.
+fn addUserText(conv: *conversation.Conversation, text: []const u8) !void {
+    const tb = try conversation.textualBlockFromSlice(conv.allocator, text);
+    var block: conversation.ContentBlock = .{ .Text = tb };
+    errdefer block.deinit(conv.allocator);
+    try conv.addUserMessage(&.{block});
+}
+
+/// 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,
+        .metadata = if (msg.metadata) |m| try alloc.dupe(u8, m) else null,
+        // Preserve the producing identity so a kept-verbatim turn isn't
+        // re-stamped with the compaction model on persist.
+        .identity = if (msg.identity) |id| try conversation.dupeWireIdentity(alloc, id) else null,
+    };
+}
+
+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;
+            errdefer if (sig) |s| alloc.free(s);
+            const origin: ?conversation.SignatureOrigin = if (b.signature_origin) |o| try o.dupe(alloc) else null;
+            break :blk .{ .Thinking = .{ .text = tb, .signature = sig, .signature_origin = origin } };
+        },
+        .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, .is_error = b.is_error } };
+        },
+        .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),
+        } },
+    };
+}
+
+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.ResultParts {
+    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.ResultParts.fromTextOwned(allocator, msg);
+}
+
+/// What to do with an error returned by tool dispatch.
+const ToolErrorAction = enum {
+    /// Surface the failure to the model as an error `ToolResult`, then let
+    /// the agent loop continue.
+    tool_result,
+    /// Abort the whole turn and propagate to the embedder. Reserved for
+    /// failures that belong to the host, not the model/provider exchange.
+    hard_fail,
+};
+
+/// Decide how to handle a tool dispatch error. Only genuine host failures
+/// abort the turn; everything else becomes a model-visible tool result so
+/// the model can correct course (and so every `ToolUse` keeps its matching
+/// `ToolResult`, which providers require).
+fn classifyToolError(err: anyerror) ToolErrorAction {
+    return switch (err) {
+        error.Canceled, error.OutOfMemory => .hard_fail,
+        else => .tool_result,
+    };
+}
+
+/// Build an error `ResultPart` describing a failed tool call, in the
+/// model-readable form the plan specifies.
+fn toolErrorResult(
+    allocator: Allocator,
+    tool_name: []const u8,
+    err: anyerror,
+) !tool_mod.ResultParts {
+    const msg = try std.fmt.allocPrint(
+        allocator,
+        "Tool execution failed for `{s}`: {s}\n" ++
+            "You may fix the arguments, try a different tool, or explain the failure to the user.",
+        .{ tool_name, @errorName(err) },
+    );
+    return tool_mod.ResultParts.fromTextOwned(allocator, msg);
+}
+
+/// The user's submission that opens a turn: an ordered list of content
+/// blocks (text, tool results, images, …), the same shape used to rebuild a
+/// persisted user message. Ownership of `blocks` transfers to the agent's
+/// conversation on `run`; the caller must not deinit them afterwards. A plain
+/// chat turn is one `.Text` block; a turn that resumes after the embedder
+/// handled tool calls is one or more `.ToolResult` blocks.
+pub const UserMessage = struct {
+    blocks: []const conversation.ContentBlock,
+};
+
+/// 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,
+};
+
+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
+    /// atomically. The pointee is owned by the embedder, not the agent. The
+    /// tool set is no longer part of the snapshot — it lives on `_registry`.
+    _config: *const Config,
+    /// The tool set this agent exposes. Owned by the agent: created empty at
+    /// `init`, populated via `registerTool`/`registerToolSource`, torn down
+    /// in `deinit`. Read fresh by the agent loop each turn, so a
+    /// registration between turns is visible at the next turn boundary.
+    _registry: ToolRegistry,
+    /// The live conversation the agent drives. Owned by the agent (adopted
+    /// at `init`); torn down in `deinit`. Turn-driving methods operate on
+    /// this directly rather than taking a `*Conversation` parameter.
+    ///
+    /// This is the one intended-public field (every other field is
+    /// underscore-prefixed internal state): a borrowed handle on the live
+    /// conversation, for in-place context-management surgery. Valid for the
+    /// agent's lifetime; do not retain past it.
+    conversation: conversation.Conversation,
+    /// The session this agent appends to. Minted from the store at `init`
+    /// (fresh: `store.create()`) or supplied by the embedder on resume
+    /// (`resolve`/`latest`). `Session.append` proxies to the store and
+    /// updates the session's last-used wire identity. The embedder owns the
+    /// underlying store, which must outlive the agent.
+    _session: session_store_mod.Session,
+    /// Injectable streaming seam. Defaults to the real provider dispatch
+    /// (`provider_mod.openStream`); tests override it with a stub.
+    _open_stream_fn: provider_mod.OpenStreamFn = provider_mod.openStream,
+    /// 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,
+    /// PRNG state for backoff jitter. Seeded lazily on first retry. Only
+    /// touched from the single agent-loop thread (retries are serial), so
+    /// no synchronization is needed.
+    _retry_prng: ?std.Random.DefaultPrng = null,
+    /// High-water mark of messages durably handed to the store: the count of
+    /// conversation messages already persisted. `flushPersist` appends only
+    /// `[_persisted_through .. coherent_end)` and advances this, so the store
+    /// stays current at every `Stream.next()` boundary without re-appending.
+    /// Initialized to the loaded history length (already persisted) and
+    /// rewound by `rewriteWithSummary` so a compaction re-persists its summary
+    /// + restated suffix.
+    _persisted_through: usize = 0,
+    /// User-message blocks queued by an embedder seam (the CLI's
+    /// `panto.ext.agent:submit`) for the host front end to open the next
+    /// turn with, so the turn runs on the host's own driver (rendering,
+    /// interrupts) rather than inside the queuing callback. Blocks and the
+    /// list storage are owned by `conversation.allocator` — the same
+    /// ownership `run` adopts.
+    _pending_submission: std.ArrayList(conversation.ContentBlock) = .empty,
+
+    /// Construct an agent.
+    ///
+    /// `store` is the persistence backend (use `null_store.store()` to opt
+    /// out). `maybe_conversation` is adopted (ownership transferred) when
+    /// non-null — the resume path: open a store, ask it for the
+    /// conversation, hand it here. When null, a fresh empty conversation is
+    /// created. Either way the agent owns and tears down the conversation.
+    /// The inner is heap-pinned at `init`, so the returned `*Agent` is a
+    /// cheap, movable handle ("don't move the Agent" stops being a rule
+    /// anyone can violate). The caller owns it and must `deinit` it.
+    pub fn init(
+        allocator: Allocator,
+        io: Io,
+        config: *const Config,
+        session: session_store_mod.Session,
+        maybe_conversation: ?conversation.Conversation,
+    ) !*Agent {
+        const self = try allocator.create(Agent);
+        self.* = .{
+            ._allocator = allocator,
+            ._io = io,
+            ._config = config,
+            ._registry = ToolRegistry.init(allocator),
+            .conversation = maybe_conversation orelse conversation.Conversation.init(allocator),
+            ._session = session,
+        };
+        // Loaded history is already in the store; only messages produced from
+        // here on need persisting.
+        self._persisted_through = self.conversation.messages.items.len;
+        return self;
+    }
+
+    pub fn deinit(self: *Agent) void {
+        // The agent owns the conversation, the tool registry, and the
+        // session handle's `info` (minted by `store.create()` or resolved
+        // by the embedder and handed in). It borrows the config snapshot
+        // and the underlying store, which the embedder tears down. It is
+        // self-heap-pinned (`init` allocated it), so it frees itself last.
+        const allocator = self._allocator;
+        self._registry.deinit();
+        for (self._pending_submission.items) |*b| b.deinit(self.conversation.allocator);
+        self._pending_submission.deinit(self.conversation.allocator);
+        self.conversation.deinit();
+        self._session.info.deinit(allocator);
+        allocator.destroy(self);
+    }
+
+    /// The id of the session this agent appends to.
+    pub fn sessionId(self: *const Agent) []const u8 {
+        return self._session.info.id;
+    }
+
+    /// Add a single tool to this agent's tool set. Visible at the next turn.
+    pub fn registerTool(self: *Agent, tool: Tool) !void {
+        try self._registry.register(tool);
+    }
+
+    /// Add a tool source (a dynamic group of tools) to this agent's tool
+    /// set. Visible at the next turn.
+    pub fn registerToolSource(self: *Agent, src: ToolSource) !void {
+        try self._registry.registerSource(src);
+    }
+
+    /// The wire-format provider identity stamped on persisted entries,
+    /// derived from the active config snapshot. Ground truth: never a CLI
+    /// alias, never any `api_key` material.
+    fn wireIdentity(self: *const Agent) session_store_mod.WireIdentity {
+        return self._config.provider.wireIdentity();
+    }
+
+    /// 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;
+    }
+
+    /// Add a system message (append or replace mode) to the conversation
+    /// and persist it. The persisted entry records the mode so replay
+    /// reconstructs the same effective system prompt.
+    /// Append a system message (`.append` mode) and persist it. Adds to the
+    /// effective system prompt.
+    pub fn addSystemMessage(self: *Agent, text: []const u8) !void {
+        return self._persistSystemMessage(text, .append);
+    }
+
+    /// Replace the effective system prompt (`.replace` mode) and persist it.
+    /// Discards all prior system text; replay reconstructs the same prompt.
+    pub fn setSystemPrompt(self: *Agent, text: []const u8) !void {
+        return self._persistSystemMessage(text, .replace);
+    }
+
+    fn _persistSystemMessage(
+        self: *Agent,
+        text: []const u8,
+        mode: conversation.SystemMode,
+    ) !void {
+        switch (mode) {
+            .append => try self.conversation.addSystemMessage(text),
+            .replace => try self.conversation.replaceSystemMessage(text),
+        }
+        self.flushPersist();
+    }
+
+    /// Replace the text of the ToolResult block matching `tool_use_id` in
+    /// the conversation's LAST message. Media parts are preserved (the new
+    /// text becomes the single leading text part); `is_error` is untouched.
+    /// Returns false when the last message holds no such block.
+    ///
+    /// This is the application half of the embedder's "tool_result output
+    /// override" (a `tool_result` event handler assigning `ev.output`):
+    /// called between tool dispatch and the next provider request, the
+    /// override becomes canonical — wire, persistence, and compaction all
+    /// see the replaced text.
+    pub fn overrideToolResultOutput(self: *Agent, tool_use_id: []const u8, text: []const u8) !bool {
+        const msgs = self.conversation.messages.items;
+        if (msgs.len == 0) return false;
+        const last = &msgs[msgs.len - 1];
+        if (last.role != .user) return false;
+        for (last.content.items) |*block| {
+            if (block.* != .ToolResult) continue;
+            const tr = &block.ToolResult;
+            if (!std.mem.eql(u8, tr.tool_use_id, tool_use_id)) continue;
+
+            var new_parts: std.ArrayList(conversation.ResultPartStored) = .empty;
+            errdefer new_parts.deinit(self._allocator);
+            try new_parts.ensureTotalCapacity(self._allocator, tr.parts.items.len + 1);
+            var new_text: conversation.TextualBlock = .empty;
+            try new_text.appendSlice(self._allocator, text);
+            errdefer new_text.deinit(self._allocator);
+            new_parts.appendAssumeCapacity(.{ .text = new_text });
+            for (tr.parts.items) |*p| switch (p.*) {
+                .text => |*t| t.deinit(self._allocator),
+                .media => |m| new_parts.appendAssumeCapacity(.{ .media = m }),
+            };
+            tr.parts.deinit(self._allocator);
+            tr.parts = new_parts;
+            return true;
+        }
+        return false;
+    }
+
+    /// Replace the input JSON of the ToolUse block matching `tool_use_id`
+    /// in the conversation's LAST message (the committed assistant message,
+    /// between response completion and tool dispatch). Returns false when
+    /// no such block exists there.
+    ///
+    /// Application half of the "tool_call_complete input override": the
+    /// rewritten input is what the tool executes with AND what the model
+    /// sees of its own call in future requests.
+    pub fn overrideToolUseInput(self: *Agent, tool_use_id: []const u8, input_json: []const u8) !bool {
+        const msgs = self.conversation.messages.items;
+        if (msgs.len == 0) return false;
+        const last = &msgs[msgs.len - 1];
+        if (last.role != .assistant) return false;
+        for (last.content.items) |*block| {
+            if (block.* != .ToolUse) continue;
+            const tu = &block.ToolUse;
+            if (!std.mem.eql(u8, tu.id, tool_use_id)) continue;
+            tu.input.clearRetainingCapacity();
+            try tu.input.appendSlice(self._allocator, input_json);
+            return true;
+        }
+        return false;
+    }
+
+    /// Submit a user message and begin a turn, returning a resumable pull
+    /// `Stream`.
+    ///
+    /// The user message is appended to the conversation and durably persisted
+    /// *immediately* (before any provider call), so a crash before the model
+    /// replies leaves a recoverable dangling prompt in the store. No provider
+    /// I/O happens here: the request opens lazily on the first
+    /// `Stream.next()`.
+    ///
+    /// The returned `*Stream` is heap-allocated; the caller owns it and must
+    /// `deinit` it. Persistence of whatever the turn committed runs when the
+    /// stream reaches its terminal `turn_complete` or is `deinit`ed early
+    /// (so a partial turn is still durably logged), mirroring the previous
+    /// `runStep` exit-path guarantee.
+    ///
+    /// The agent re-reads its `config` snapshot at the top of each provider
+    /// response inside the stream, so a mid-conversation `setConfig` takes
+    /// effect at the next response boundary, never mid-stream.
+    /// Append blocks to the pending submission (see `_pending_submission`).
+    /// Adopts the block *contents*; the `blocks` slice itself stays the
+    /// caller's. Multiple calls accumulate into one pending user message.
+    pub fn queueSubmission(self: *Agent, blocks: []const conversation.ContentBlock) !void {
+        try self._pending_submission.appendSlice(self.conversation.allocator, blocks);
+    }
+
+    /// Take the pending submission, or null when none is queued. The caller
+    /// owns the returned slice: typically hand it to `run` (which adopts the
+    /// block contents) and free the slice itself with
+    /// `conversation.allocator`.
+    pub fn takeSubmission(self: *Agent) !?[]conversation.ContentBlock {
+        if (self._pending_submission.items.len == 0) return null;
+        return try self._pending_submission.toOwnedSlice(self.conversation.allocator);
+    }
+
+    pub fn run(self: *Agent, message: UserMessage) !*Stream {
+        self._auto_compacted = false;
+
+        // Append + persist the user prompt up front (the dangling-prompt
+        // recovery guarantee). `addMessage` adopts the blocks; persistence is
+        // brought current through the shared high-water flush.
+        try self.conversation.addMessage(.user, message.blocks, null);
+        self.flushPersist();
+
+        const s = try self._allocator.create(Stream);
+        s.* = Stream.init(self);
+        return s;
+    }
+
+    /// Bring the session store current with the conversation: append every
+    /// message in `[_persisted_through .. coherent_end)` as a single batch and
+    /// advance the high-water mark. `coherent_end` excludes a trailing
+    /// assistant message whose ToolUse blocks have no following results — a
+    /// turn interrupted between the assistant's tool call and dispatch. Leaving
+    /// it unpersisted keeps the log replayable and the live conversation in a
+    /// state `compact()` (or a tool-result `run`) can resume from. Called at
+    /// every `Stream.next()` boundary, so after any `next()` the store reflects
+    /// all committed, coherent messages.
+    pub fn flushPersist(self: *Agent) void {
+        const msgs = self.conversation.messages.items;
+        // An operation that rewrote the prefix (replace-system, cancel) can
+        // leave the mark past the new length; clamp before comparing.
+        if (self._persisted_through > msgs.len) self._persisted_through = msgs.len;
+        var end = msgs.len;
+        if (end > self._persisted_through and
+            msgs[end - 1].role == .assistant and
+            turn_persist.hasToolUseWithoutFollowingResults(&self.conversation, end - 1))
+        {
+            end -= 1;
+        }
+        if (end <= self._persisted_through) return;
+        turn_persist.persistRange(
+            self._allocator,
+            &self._session,
+            &self.conversation,
+            self._persisted_through,
+            end,
+            self.wireIdentity(),
+            &.{},
+        ) catch |e| {
+            std.log.err("session: failed to persist turn: {t}", .{e});
+            return;
+        };
+        self._persisted_through = end;
+    }
+
+    fn hasToolUseBlock(msg: conversation.Message) bool {
+        return toolUseCount(msg) > 0;
+    }
+
+    /// Open one provider response with the configured retry policy, pushing
+    /// any informational `provider_retry` events into `out`. Returns the
+    /// resumable `ProviderStream` once a request has been successfully
+    /// opened (headers received), or propagates a terminal error.
+    ///
+    /// Decision path for a failed open:
+    ///   - `ContextOverflow`: compact once, then retry the same request a
+    ///     single time against the compacted conversation (a one-shot path,
+    ///     independent of the transient-retry budget).
+    ///   - retryable provider error (rate limit, server, transport,
+    ///     malformed stream): sleep with exponential backoff + jitter
+    ///     (honoring `Retry-After` when present) and retry, up to
+    ///     `retry.max_attempts` total attempts.
+    ///   - anything else (auth, bad request, cancellation, local errors):
+    ///     propagate immediately.
+    ///
+    /// A failed open never mutates the conversation (providers commit the
+    /// assistant message only on success), so each retry runs against the
+    /// same snapshot. Mid-stream failures (surfaced from `produce`) re-enter
+    /// here via `Stream`, replaying the response from a fresh open — exactly
+    /// as the previous push loop replayed a failed `streamStep`.
+    fn openWithRetries(
+        self: *Agent,
+        cfg: *const Config,
+        out: *stream_mod.EventQueue,
+    ) !provider_mod.ProviderStream {
+        const policy = cfg.retry;
+        var attempt: usize = 1;
+        while (true) {
+            var diag: provider_mod.ProviderDiagnostic = .{};
+            const ps = self._open_stream_fn(self._allocator, self._io, cfg, &self._registry, &self.conversation, &diag) catch |err| {
+                if (err == error.ContextOverflow) {
+                    return self.handleContextOverflow(cfg, out, err);
+                }
+                if (!provider_mod.isRetryableProviderError(err)) return err;
+                // Out of attempts: hard-fail with the last error.
+                if (attempt >= policy.max_attempts) return err;
+
+                const delay_ms = self.backoffDelayMs(policy, attempt, diag.retry_after_ms);
+                try out.push(.{ .provider_retry = .{
+                    .attempt = attempt,
+                    .max_attempts = policy.max_attempts,
+                    .delay_ms = delay_ms,
+                    .err = err,
+                    .status_code = diag.status_code,
+                    .retry_after_ms = diag.retry_after_ms,
+                    .message = diag.message,
+                } });
+                if (delay_ms > 0) {
+                    const ms: i64 = @intCast(@min(delay_ms, std.math.maxInt(i64)));
+                    self._io.sleep(.fromMilliseconds(ms), .real) catch |e| return e;
+                }
+                attempt += 1;
+                continue;
+            };
+            return ps;
+        }
+    }
+
+    /// One-shot context-overflow recovery: compact once, retry once. Pushes
+    /// a `provider_retry` event with `compaction = true` and `delay_ms = 0`,
+    /// then re-opens the request against the compacted context. A second
+    /// overflow (or any other error) propagates.
+    fn handleContextOverflow(
+        self: *Agent,
+        cfg: *const Config,
+        out: *stream_mod.EventQueue,
+        err: anyerror,
+    ) !provider_mod.ProviderStream {
+        if (self._auto_compacted) return err; // already retried once this turn
+        const sys = self._config.compaction.compaction_prompt orelse return err;
+        const res = try self._compactInPlace(sys, null);
+        if (!res.compacted) return err; // nothing to shed; give up
+        self._auto_compacted = true;
+        try out.push(.{ .provider_retry = .{
+            .attempt = 1,
+            .max_attempts = 2,
+            .delay_ms = 0,
+            .err = err,
+            .compaction = true,
+        } });
+        // Retry the same request against the compacted context.
+        var diag: provider_mod.ProviderDiagnostic = .{};
+        return self._open_stream_fn(self._allocator, self._io, cfg, &self._registry, &self.conversation, &diag);
+    }
+
+    /// Compute the backoff delay (ms) for the just-failed `attempt`
+    /// (1-based). Prefers a provider `Retry-After` (capped by policy);
+    /// otherwise exponential `initial * multiplier^(attempt-1)`, capped,
+    /// with optional full jitter in `[0, delay)`.
+    fn backoffDelayMs(
+        self: *Agent,
+        policy: config_mod.RetryConfig,
+        attempt: usize,
+        retry_after_ms: ?u64,
+    ) u64 {
+        if (retry_after_ms) |ra| {
+            return @min(ra, policy.max_delay_ms);
+        }
+        const exp: f64 = @floatFromInt(attempt - 1);
+        const base: f64 = @as(f64, @floatFromInt(policy.initial_delay_ms)) *
+            std.math.pow(f64, policy.multiplier, exp);
+        const capped: f64 = @min(base, @as(f64, @floatFromInt(policy.max_delay_ms)));
+        var delay: u64 = @intFromFloat(capped);
+        if (policy.jitter and delay > 0) {
+            if (self._retry_prng == null) {
+                const ns = std.Io.Clock.now(.real, self._io).nanoseconds;
+                const seed: u64 = @truncate(@as(u128, @bitCast(@as(i128, ns))));
+                self._retry_prng = std.Random.DefaultPrng.init(seed);
+            }
+            const r = self._retry_prng.?.random();
+            delay = r.intRangeLessThan(u64, 0, delay + 1);
+        }
+        return delay;
+    }
+
+    /// Compact the conversation: summarize an older prefix into a single
+    /// `.CompactionSummary` block and keep a recent suffix of whole turns
+    /// verbatim. Mutates `self.conversation` in place.
+    ///
+    /// This is the pure transform — it does **not** persist. The explicit
+    /// `/compact` entry point is the public `compact` (which persists); the
+    /// automatic (context-overflow) path persists via `runStep`'s turn tail
+    /// (the rewritten window is logged as a fresh compaction window).
+    ///
+    /// 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).
+    fn _compactInPlace(
+        self: *Agent,
+        system_prompt: []const u8,
+        extra_instructions: ?[]const u8,
+    ) !CompactionResult {
+        const conv = &self.conversation;
+        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.text);
+
+        try self.rewriteWithSummary(conv, split.prefix_end, summary.text, summary.size);
+
+        return .{
+            .compacted = true,
+            .kept_turns = split.kept_turns,
+            .summarized_messages = summarized,
+        };
+    }
+
+    /// Compact and persist the result to the session store. This is the
+    /// explicit `/compact` entry point: it summarizes (via the private
+    /// `_compactInPlace` transform) and, if anything was compacted, appends
+    /// the new compaction window (summary + restated kept suffix) to the
+    /// store. Returns the `CompactionResult` for the embedder to report.
+    ///
+    /// `override_system_prompt`, when non-null, is the compaction system
+    /// prompt for this run; when null it falls back to
+    /// `config.compaction.compaction_prompt`. With neither set, this errors
+    /// (`error.NoCompactionPrompt`).
+    pub fn compact(
+        self: *Agent,
+        override_system_prompt: ?[]const u8,
+        extra_instructions: ?[]const u8,
+    ) !CompactionResult {
+        const system_prompt = override_system_prompt orelse
+            self._config.compaction.compaction_prompt orelse
+            return error.NoCompactionPrompt;
+        const res = try self._compactInPlace(system_prompt, extra_instructions);
+        // `rewriteWithSummary` rewound the high-water mark to the summary;
+        // flush appends the new compaction window (summary + restated suffix).
+        if (res.compacted) self.flushPersist();
+        return res;
+    }
+
+    /// 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,
+        summary_size: u64,
+    ) !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, with usage recomputed so the
+        //    restated window reads like a fresh conversation anchored at
+        //    the summary.
+        //
+        //    The post-compaction window is `[summary(user), kept_user,
+        //    kept_assistant, ...]`. `usage.input` is *cumulative* (the whole
+        //    prior prompt). We walk forward maintaining `running_input` —
+        //    the synthetic cumulative prompt size as of "just before the
+        //    next assistant output" — seeded with the summary's size:
+        //
+        //      - non-assistant message → it occupies context; add its size
+        //        (word-count heuristic, the same `messageTokenEstimate` the
+        //        splitter uses for usage-less messages).
+        //      - assistant message → its synthetic prompt is `running_input`.
+        //        The whole prompt total collapses into `input` (the rewrite
+        //        busts any provider prefix cache, so cache_read/cache_write
+        //        are zeroed). `output`/`reasoning` are copied verbatim (real
+        //        generation). Then `running_input += output` for the next
+        //        turn.
+        //
+        //    Assistants that had no usage stay null (we can't invent an
+        //    output we never measured; the splitter already tolerates null).
+        var running_input: u64 = summary_size;
+        for (old[prefix_end..]) |*m| {
+            var cloned = try cloneMessage(alloc, m.*);
+            if (cloned.role == .assistant) {
+                if (m.usage) |u| {
+                    cloned.usage = .{
+                        .input = running_input,
+                        .output = u.output,
+                        .cache_read = 0,
+                        .cache_write = 0,
+                        .reasoning = u.reasoning,
+                    };
+                    running_input += u.output;
+                }
+            } else {
+                running_input += compaction_mod.messageTokenEstimate(m.*, null);
+            }
+            try rebuilt.append(alloc, cloned);
+        }
+
+        // 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;
+
+        // The rewrite invalidated the old message indices. Rewind the
+        // persistence high-water mark to the inserted summary so the next
+        // `flushPersist` re-appends the new compaction window (summary +
+        // restated kept suffix) as fresh entries.
+        self._persisted_through =
+            conversation.latestCompactionIndex(conv.messages.items) orelse 0;
+    }
+
+    /// 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,
+    ) !CompactionSummary {
+        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,
+                .compaction = self._config.compaction,
+            };
+            if (self.runSingleCompactionTurn(&cfg, &empty_registry, 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,
+            .compaction = self._config.compaction,
+        };
+        return self.runSingleCompactionTurn(&cfg, &empty_registry, sys_text, body);
+    }
+
+    /// A generated compaction summary plus the token size to attribute to
+    /// it when it becomes the new conversation anchor. `size` is the
+    /// provider-reported `output` token count for the summary turn (the
+    /// real generated length), falling back to the word-count heuristic
+    /// when the provider emitted no usage. `text` is caller-owned.
+    const CompactionSummary = struct {
+        text: []u8,
+        size: u64,
+    };
+
+    /// 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 plus
+    /// the summary's token size (for usage anchoring on the rewrite).
+    fn runSingleCompactionTurn(
+        self: *Agent,
+        cfg: *const config_mod.Config,
+        registry: *const ToolRegistry,
+        system_prompt: []const u8,
+        body: []const u8,
+    ) !CompactionSummary {
+        const alloc = self._allocator;
+        var conv = conversation.Conversation.init(alloc);
+        defer conv.deinit();
+        try conv.addSystemMessage(system_prompt);
+        try addUserText(&conv, body);
+
+        // Drive one provider response to completion, ignoring every event.
+        // Compaction doesn't need incremental output — the assembled message
+        // is read off the conversation below — so we just pump the pull
+        // stream until it commits the assistant message.
+        var queue = stream_mod.EventQueue.init(alloc);
+        defer queue.deinit();
+        var ps = try self._open_stream_fn(alloc, self._io, cfg, registry, &conv, null);
+        defer ps.deinit();
+        while (true) {
+            const status = try ps.produce(&queue);
+            // Drain (and discard) any events to bound the queue/arena.
+            while (queue.pop()) |_| {}
+            if (status == .response_complete) break;
+        }
+
+        // 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;
+
+        // Summary size: prefer the provider's reported output token count
+        // for this turn; otherwise word-count the assembled summary text
+        // (the same fallback the splitter uses for usage-less messages).
+        const size: u64 = if (last.usage) |u|
+            u.output
+        else
+            compaction_mod.messageTokenEstimate(last, null);
+
+        return .{ .text = try out.toOwnedSlice(alloc), .size = size };
+    }
+
+    /// 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,
+        assistant_msg: conversation.Message,
+    ) !void {
+        const conv = &self.conversation;
+        // 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,
+                    .is_error = true,
+                });
+                continue;
+            }
+            const entry = self._registry.lookup(tu.name) orelse {
+                // Unknown tool: don't abort. Synthesize an error result so
+                // the model can correct, and so this ToolUse still gets its
+                // matching ToolResult (providers reject a follow-up request
+                // otherwise).
+                try calls.append(.{
+                    .tool_use_id = tu.id,
+                    .tool_name = tu.name,
+                    .input = tu.input.items,
+                    .entry = null,
+                    .result = try toolErrorResult(self._allocator, tu.name, error.UnknownTool),
+                    .err = null,
+                    .is_error = true,
+                });
+                continue;
+            };
+            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| r.deinit(self._allocator);
+            }
+        }
+
+        // Try real concurrency first. If the `Io` implementation can't
+        // provide it (`error.ConcurrencyUnavailable`), fall back to running
+        // every group sequentially on this thread — tool batches are small
+        // (rarely more than a handful of calls) so the serial path is a fine
+        // safety net rather than a hard failure.
+        var ran_concurrently = true;
+        for (groups.items) |*g| {
+            task_group.concurrent(self._io, runGroup, .{ self, g, calls.items }) catch |e| {
+                if (e == error.ConcurrencyUnavailable) {
+                    ran_concurrently = false;
+                    break;
+                }
+                return e;
+            };
+        }
+        if (ran_concurrently) {
+            // `error.Canceled` here means cancellation propagated into this
+            // dispatch from above; surface it like any other error.
+            try task_group.await(self._io);
+        } else {
+            // Cancel any tasks that were spawned before the failure, then
+            // run all groups serially. Only entry-bearing calls are touched
+            // by `runGroup`; the pre-seeded error results (unknown tool,
+            // invalid input) have `entry == null` and must be left intact.
+            task_group.cancel(self._io);
+            for (calls.items) |*c| {
+                if (c.entry == null) continue;
+                if (c.result) |r| r.deinit(self._allocator);
+                c.result = null;
+                c.err = null;
+            }
+            for (groups.items) |*g| runGroup(self, g, calls.items);
+        }
+
+        // Pre-pass: resolve worker-reported errors. A hard host failure
+        // (cancellation, OOM) aborts the whole turn. Every other failure is
+        // converted into a model-visible error `ToolResult` so the model can
+        // recover and so each `ToolUse` keeps its matching `ToolResult`
+        // (providers reject the next request otherwise).
+        for (calls.items) |*c| {
+            const e = c.err orelse continue;
+            if (classifyToolError(e) == .hard_fail) return e;
+            // Replace any partial result with a synthesized error result.
+            if (c.result) |r| {
+                r.deinit(self._allocator);
+                c.result = null;
+            }
+            c.result = try toolErrorResult(self._allocator, c.tool_name, e);
+            c.err = null;
+            c.is_error = true;
+        }
+
+        // Assemble ToolResult blocks in original call order.
+        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);
+
+        for (calls.items) |*c| {
+            const result_parts = c.result orelse {
+                // Internal invariant: every call should now have a result
+                // (success, synthesized error, or pre-seeded error).
+                return error.MissingToolResult;
+            };
+            c.result = null; // ownership transferred below
+            defer result_parts.deinit(self._allocator);
+
+            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.items.len);
+            for (result_parts.items) |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| {
+                            // Media processing failure: keep the turn alive by
+                            // dropping the attachment and noting it as text,
+                            // rather than aborting. `UnknownMediaType` gets a
+                            // friendly note; other failures name the error.
+                            var note: conversation.TextualBlock = .empty;
+                            errdefer note.deinit(self._allocator);
+                            if (e == error.UnknownMediaType) {
+                                try note.appendSlice(self._allocator, "[unrecognized binary attachment dropped]");
+                            } else {
+                                const txt = try std.fmt.allocPrint(
+                                    self._allocator,
+                                    "[media attachment dropped: {s}]",
+                                    .{@errorName(e)},
+                                );
+                                defer self._allocator.free(txt);
+                                try note.appendSlice(self._allocator, txt);
+                            }
+                            stored.appendAssumeCapacity(.{ .text = note });
+                            continue;
+                        };
+                        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,
+                .is_error = c.is_error,
+            } });
+        }
+
+        try conv.messages.append(self._allocator, .{
+            .role = .user,
+            .content = content,
+        });
+    }
+};
+
+/// A resumable pull handle over one agent turn.
+///
+/// `next()` pulls one `Event` at a time, driving the agent loop
+/// incrementally: open a provider response, stream its events, dispatch any
+/// tool calls between responses, and repeat until the model stops calling
+/// tools. The whole loop's state lives here (not on a stack frame), so the
+/// turn can suspend and resume between events.
+///
+/// Contract (see `stream.zig`):
+///   - an `Event` value is streaming progress, including `turn_complete`;
+///   - `null` means exhausted (already past `turn_complete`), never before;
+///   - an error is a genuine failure (network/parse/provider).
+///
+/// Event payloads borrow from the stream's decode state or the conversation
+/// and are valid only until the next `next()` call.
+pub const Stream = struct {
+    // Internal state (underscore-prefixed by convention: not part of the
+    // public surface, even though Zig fields are always reachable). The only
+    // intended-public field is `state`, the transparent turn state machine.
+    _agent: *Agent,
+    _queue: stream_mod.EventQueue,
+    state: State,
+    /// The active provider response, when in `.streaming`.
+    _response: ?provider_mod.ProviderStream = null,
+    /// First message index of this turn (the boundary `cancel` rolls back to).
+    _start: usize,
+    /// A terminal error to surface once any already-queued events (e.g.
+    /// `provider_retry` notices pushed before the failing attempt) have been
+    /// drained. `next()` yields the queue first, then this error.
+    _pending_error: ?anyerror = null,
+    /// Count of mid-stream retries attempted for the CURRENT response. Unlike
+    /// `openWithRetries` (which loops internally with a local counter), a
+    /// mid-stream failure unwinds back to `.turn_start` and re-opens, so the
+    /// attempt count must live on the stream to survive across re-opens.
+    /// Reset to 0 each time a response completes. A persistently-broken
+    /// stream would otherwise re-open forever without ever exhausting
+    /// `retry.max_attempts`.
+    _stream_retries: usize = 0,
+    /// Owned copy of the provider's diagnostic message for the in-flight
+    /// mid-stream retry notice. Borrowed by the queued `provider_retry`
+    /// event, so it must outlive that event: freed on the next mid-stream
+    /// failure and on `deinit`.
+    _retry_message: ?[]u8 = null,
+
+    pub const State = enum {
+        /// Open the next provider response (with retries).
+        turn_start,
+        /// Pump the active provider response into events.
+        streaming,
+        /// A provider response completed; decide tools-vs-done.
+        after_response,
+        /// Tool dispatch has been announced to callers; run the blocking
+        /// dispatch on the next pull so `tool_dispatch_start` is observable
+        /// before long-running tools complete.
+        dispatching_tools,
+        /// The turn reached its terminal `turn_complete`.
+        done,
+        /// A failure already propagated; `next()` is poisoned.
+        failed,
+    };
+
+    fn init(agent: *Agent) Stream {
+        return .{
+            ._agent = agent,
+            ._queue = stream_mod.EventQueue.init(agent._allocator),
+            .state = .turn_start,
+            ._start = agent.conversation.messages.items.len,
+        };
+    }
+
+    pub fn deinit(self: *Stream) void {
+        // Persist whatever the turn committed, on every exit path — including
+        // dropping the stream mid-turn after some messages were committed.
+        // `next()` already flushes at every boundary, so this is usually a
+        // no-op; it catches a stream dropped without a final `next()`.
+        self._agent.flushPersist();
+        if (self._response) |ps| ps.deinit();
+        if (self._retry_message) |m| self._agent._allocator.free(m);
+        self._queue.deinit();
+        self._agent._allocator.destroy(self);
+    }
+
+    /// Abort this turn from the embedder side (for example, Escape in a TUI).
+    /// Provider/tool work that is currently inside `next()` cannot be unwound
+    /// until that call returns, but once control reaches the embedder this
+    /// drops every assistant/tool-result message produced by this stream before
+    /// `deinit()` has a chance to persist them. The already-persisted user
+    /// prompt (written by `Agent.run`) is intentionally kept, leaving a valid
+    /// conversation prefix with no dangling ToolUse blocks.
+    pub fn cancel(self: *Stream) void {
+        if (self._response) |ps| {
+            ps.deinit();
+            self._response = null;
+        }
+        const conv = &self._agent.conversation;
+        while (conv.messages.items.len > self._start) {
+            var msg = conv.messages.pop().?;
+            msg.deinit(conv.allocator);
+        }
+        // Rolled the conversation back; clamp the persistence mark to match.
+        self._agent.flushPersist();
+        self.state = .done;
+    }
+
+    /// Reset a `.failed` stream back to `.turn_start` so the caller can resume
+    /// `next()` after changing the agent config (e.g. an embedder that catches
+    /// a terminal `ProviderBadRequest`, rewrites the provider config, and
+    /// wants to retry the SAME turn without re-appending the user message).
+    ///
+    /// Safe only at the OPEN boundary: a failed open never mutates the
+    /// conversation, so the already-committed user prompt at `_start` stays
+    /// intact and the re-open re-reads the (now-swapped) config snapshot.
+    /// Returns `error.StreamNotFailed` if the stream is not in `.failed`.
+    pub fn reopen(self: *Stream) !void {
+        if (self.state != .failed) return error.StreamNotFailed;
+        if (self._response) |ps| {
+            ps.deinit();
+            self._response = null;
+        }
+        self._pending_error = null;
+        self.state = .turn_start;
+    }
+
+    /// Pull the next event, or null past the terminal. See the contract
+    /// above.
+    pub fn next(self: *Stream) !?Event {
+        // Bring persistence current on every return path (events, terminal,
+        // errors). This is the interruptibility guarantee: after any `next()`
+        // the store reflects all committed, coherent messages, so the caller
+        // can `break` and immediately `compact()` or resume with a new turn.
+        defer self._agent.flushPersist();
+        // Always drain queued events first; they borrow decode/conversation
+        // state valid until this call returns.
+        if (self._queue.pop()) |ev| return ev;
+        // Queue drained: a deferred terminal error surfaces now.
+        if (self._pending_error) |err| {
+            self._pending_error = null;
+            self.state = .failed;
+            return err;
+        }
+
+        while (true) {
+            switch (self.state) {
+                .done => return null,
+                .failed => return error.StreamPoisoned,
+                .turn_start => {
+                    // Re-read the config snapshot at each response boundary
+                    // so a mid-conversation swap takes effect here, never
+                    // mid-stream.
+                    const cfg = self._agent._config;
+                    const ps = self._agent.openWithRetries(cfg, &self._queue) catch |err| {
+                        // Surface the failure after any queued retry notices
+                        // (pushed before each backoff) are drained.
+                        if (self._queue.pop()) |ev| {
+                            self._pending_error = err;
+                            return ev;
+                        }
+                        self.state = .failed;
+                        return err;
+                    };
+                    self._response = ps;
+                    self.state = .streaming;
+                    if (self._queue.pop()) |ev| return ev; // retry notices
+                },
+                .streaming => {
+                    const ps = self._response.?;
+                    const status = ps.produce(&self._queue) catch |err| {
+                        // Mid-stream failure. The conversation was not
+                        // mutated (commit happens only at response
+                        // completion), so retry by re-opening from scratch,
+                        // exactly as the prior push loop replayed a failed
+                        // streamStep. Non-retryable errors propagate.
+                        //
+                        // Capture the provider's diagnostic message (e.g. an
+                        // Anthropic `overloaded_error`) BEFORE deinit frees
+                        // it, dup'ing it so it outlives the response. Owned
+                        // here and freed after the retry notice is queued.
+                        if (self._retry_message) |old| {
+                            self._agent._allocator.free(old);
+                            self._retry_message = null;
+                        }
+                        if (ps.lastError()) |m| {
+                            self._retry_message = self._agent._allocator.dupe(u8, m) catch null;
+                        }
+                        ps.deinit();
+                        self._response = null;
+                        if (!provider_mod.isRetryableProviderError(err)) {
+                            self.state = .failed;
+                            return err;
+                        }
+                        // Out of attempts: hard-fail with the last error
+                        // instead of re-opening forever. `_stream_retries`
+                        // counts mid-stream retries already made for this
+                        // response; the initial attempt is the open itself,
+                        // so the Nth retry is attempt N+1.
+                        const cfg = self._agent._config;
+                        self._stream_retries += 1;
+                        if (self._stream_retries >= cfg.retry.max_attempts) {
+                            self.state = .failed;
+                            return err;
+                        }
+                        self.state = .turn_start;
+                        // Emit a retry notice so consumers see the stall.
+                        const delay_ms = self._agent.backoffDelayMs(cfg.retry, self._stream_retries, null);
+                        self._queue.push(.{ .provider_retry = .{
+                            .attempt = self._stream_retries,
+                            .max_attempts = cfg.retry.max_attempts,
+                            .delay_ms = delay_ms,
+                            .err = err,
+                            .message = self._retry_message,
+                        } }) catch |e| {
+                            self.state = .failed;
+                            return e;
+                        };
+                        if (delay_ms > 0) {
+                            const ms: i64 = @intCast(@min(delay_ms, std.math.maxInt(i64)));
+                            self._agent._io.sleep(.fromMilliseconds(ms), .real) catch |e| {
+                                self.state = .failed;
+                                return e;
+                            };
+                        }
+                        if (self._queue.pop()) |ev| return ev;
+                        continue;
+                    };
+                    if (self._queue.pop()) |ev| return ev;
+                    if (status == .response_complete) {
+                        ps.deinit();
+                        self._response = null;
+                        // A response completed: reset the mid-stream retry
+                        // budget so a later turn starts fresh.
+                        self._stream_retries = 0;
+                        self.state = .after_response;
+                    }
+                    // else `.more`: loop and pump again.
+                },
+                .after_response => {
+                    const conv = &self._agent.conversation;
+                    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.
+                    if (last.content.items.len == 0) {
+                        self.state = .failed;
+                        return error.EmptyAssistantResponse;
+                    }
+
+                    if (!Agent.hasToolUseBlock(last)) {
+                        self.state = .done;
+                        return .turn_complete;
+                    }
+
+                    // Announce tool dispatch and return immediately. The
+                    // dispatch itself can block for a long time (especially
+                    // with parallel tools); yielding this boundary event first
+                    // lets UIs/renderers show all tool calls as running instead
+                    // of appearing frozen until the slowest tool completes.
+                    const count = toolUseCount(last);
+                    self._queue.push(.{ .tool_dispatch_start = .{ .count = count } }) catch |e| {
+                        self.state = .failed;
+                        return e;
+                    };
+                    self.state = .dispatching_tools;
+                    if (self._queue.pop()) |ev| return ev;
+                },
+                .dispatching_tools => {
+                    const conv = &self._agent.conversation;
+                    const last = conv.messages.items[conv.messages.items.len - 1];
+                    std.debug.assert(last.role == .assistant);
+                    std.debug.assert(Agent.hasToolUseBlock(last));
+
+                    self._agent.dispatchToolCalls(last) catch |err| {
+                        self.state = .failed;
+                        return err;
+                    };
+                    const result_msg = conv.messages.items[conv.messages.items.len - 1];
+                    self._queue.push(.{ .tool_dispatch_complete = .{ .message = result_msg } }) catch |e| {
+                        self.state = .failed;
+                        return e;
+                    };
+                    self.state = .turn_start;
+                    if (self._queue.pop()) |ev| return ev;
+                },
+            }
+        }
+    }
+};
+
+fn toolUseCount(msg: conversation.Message) usize {
+    var n: usize = 0;
+    for (msg.content.items) |block| {
+        if (block == .ToolUse) n += 1;
+    }
+    return n;
+}
+
+/// 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.ResultParts,
+
+    /// If non-null, the worker reported a failure for this call. After
+    /// dispatch it is classified: host failures abort the turn, everything
+    /// else is converted into an error `ToolResult`.
+    err: ?anyerror,
+
+    /// True when `result` already holds a synthesized error result (unknown
+    /// tool, invalid input). Worker-reported `err`s are folded into this
+    /// during assembly.
+    is_error: bool = false,
+};
+
+/// 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| b.deinit(agent._allocator),
+            .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: submit `text` and drive the whole turn to completion via the
+/// pull `Stream`, discarding every event (the agent tests assert on
+/// conversation/store state, not on the event stream). Mirrors the old
+/// `submitUserMessage` + `runStep`: it returns the same terminal error a
+/// turn would raise.
+fn drainTurn(agent: *Agent, text: []const u8) !void {
+    var s = try runUserText(agent, text);
+    defer s.deinit();
+    while (try s.next()) |_| {}
+}
+
+/// Test helper: open a turn from a single user `.Text` block, mirroring the
+/// old text-only `run(.{ .text = ... })`. The block is adopted by the agent's
+/// conversation.
+fn runUserText(agent: *Agent, text: []const u8) !*Stream {
+    var blocks = [_]conversation.ContentBlock{
+        .{ .Text = try conversation.textualBlockFromSlice(agent._allocator, text) },
+    };
+    return agent.run(.{ .blocks = &blocks });
+}
+
+/// 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 `_open_stream_fn` seam.
+///
+/// `provider_mod.OpenStreamFn` carries no user context (it mirrors the real
+/// free function exactly), so the stub parks its state in a module-level
+/// pointer that `stubOpenStream` 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 opens 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,
+    /// A queue of provider errors to return, in order, before any scripted
+    /// turn is served. Each entry is consumed on one open. Used to drive the
+    /// transient-retry path.
+    scripted_errors: []const ScriptedError = &.{},
+    error_idx: usize = 0,
+    /// A queue of MID-STREAM errors. Unlike `scripted_errors` (which fail at
+    /// open time), each entry here lets the open succeed but makes the
+    /// returned response fail on its first `produce`, exercising the
+    /// `.streaming` retry path. Consumed in order, one per open.
+    stream_errors: []const StreamFailure = &.{},
+    stream_error_idx: usize = 0,
+    /// Count of opens observed (failed + succeeded). Lets tests assert the
+    /// exact number of attempts.
+    calls_made: usize = 0,
+
+    const ScriptedError = struct {
+        err: anyerror,
+        status_code: ?u16 = null,
+        retry_after_ms: ?u64 = null,
+    };
+
+    const StreamFailure = struct {
+        err: anyerror,
+        /// Optional provider diagnostic surfaced via `ProviderStream.lastError`.
+        message: ?[]const u8 = null,
+    };
+
+    const ScriptedTurn = struct {
+        blocks: []const TestBlock,
+        /// Optional provider usage to stamp on the committed assistant
+        /// message (mirrors a real provider's terminal usage).
+        usage: ?conversation.Usage = null,
+    };
+
+    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._open_stream_fn`. Call once per test, after constructing the
+    /// stub on the stack.
+    fn install(self: *StubProvider) provider_mod.OpenStreamFn {
+        stub_active = self;
+        return stubOpenStream;
+    }
+};
+
+/// A canned resumable response: on the first `produce` it commits the
+/// scripted assistant message to the conversation and pushes the terminal
+/// `message_complete`, then reports `.response_complete`. It does not emit
+/// per-block events (the agent tests assert on conversation state, not the
+/// event stream), which is sufficient for driving the agent loop.
+const StubResponse = struct {
+    allocator: Allocator,
+    conv: *conversation.Conversation,
+    turn: StubProvider.ScriptedTurn,
+    done: bool = false,
+    /// If set, the first `produce` returns this error instead of committing
+    /// (simulates a mid-stream provider failure after a successful open).
+    produce_error: ?anyerror = null,
+    /// Optional provider diagnostic surfaced via `lastError` after the
+    /// mid-stream failure (borrowed; static test strings).
+    error_message: ?[]const u8 = null,
+
+    fn create(
+        allocator: Allocator,
+        conv: *conversation.Conversation,
+        turn: StubProvider.ScriptedTurn,
+    ) !provider_mod.ProviderStream {
+        const self = try allocator.create(StubResponse);
+        self.* = .{ .allocator = allocator, .conv = conv, .turn = turn };
+        return .{ .ptr = self, .vtable = &vtable };
+    }
+
+    fn createFailing(
+        allocator: Allocator,
+        conv: *conversation.Conversation,
+        err: anyerror,
+        message: ?[]const u8,
+    ) !provider_mod.ProviderStream {
+        const self = try allocator.create(StubResponse);
+        self.* = .{
+            .allocator = allocator,
+            .conv = conv,
+            .turn = .{ .blocks = &.{} },
+            .produce_error = err,
+            .error_message = message,
+        };
+        return .{ .ptr = self, .vtable = &vtable };
+    }
+
+    const vtable: provider_mod.ProviderStream.VTable = .{
+        .produce = produceVT,
+        .deinit = deinitVT,
+        .last_error = lastErrorVT,
+    };
+
+    fn lastErrorVT(ptr: *anyopaque) ?[]const u8 {
+        const self: *StubResponse = @ptrCast(@alignCast(ptr));
+        return self.error_message;
+    }
+
+    fn produceVT(ptr: *anyopaque, out: *stream_mod.EventQueue) anyerror!provider_mod.ProviderStream.ProduceStatus {
+        const self: *StubResponse = @ptrCast(@alignCast(ptr));
+        if (self.produce_error) |err| return err;
+        if (self.done) return .response_complete;
+        self.done = true;
+
+        var blocks: std.ArrayList(conversation.ContentBlock) = .empty;
+        errdefer {
+            for (blocks.items) |*b| b.deinit(self.allocator);
+            blocks.deinit(self.allocator);
+        }
+        for (self.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 self.conv.addAssistantMessage(moved, self.turn.usage);
+
+        const msg = self.conv.messages.items[self.conv.messages.items.len - 1];
+        try out.push(.{ .message_complete = .{ .message = msg, .usage = self.turn.usage } });
+        return .response_complete;
+    }
+
+    fn deinitVT(ptr: *anyopaque) void {
+        const self: *StubResponse = @ptrCast(@alignCast(ptr));
+        self.allocator.destroy(self);
+    }
+};
+
+fn stubOpenStream(
+    allocator: Allocator,
+    _: Io,
+    _: *const config_mod.Config,
+    _: *const ToolRegistry,
+    conv: *conversation.Conversation,
+    diag: ?*provider_mod.ProviderDiagnostic,
+) anyerror!provider_mod.ProviderStream {
+    const self = stub_active orelse return error.NoStubInstalled;
+    self.calls_made += 1;
+    if (self.error_idx < self.scripted_errors.len) {
+        const e = self.scripted_errors[self.error_idx];
+        self.error_idx += 1;
+        if (diag) |d| {
+            d.status_code = e.status_code;
+            d.retry_after_ms = e.retry_after_ms;
+        }
+        return e.err;
+    }
+    if (self.stream_error_idx < self.stream_errors.len) {
+        const f = self.stream_errors[self.stream_error_idx];
+        self.stream_error_idx += 1;
+        return StubResponse.createFailing(allocator, conv, f.err, f.message);
+    }
+    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;
+    return StubResponse.create(allocator, conv, turn);
+}
+
+/// Build a stack registry + active `Config` snapshot for tests that drive
+/// the agent. Post-R1 the registry no longer lives on `Config` — it lives
+/// on the `Agent`. The harness still owns a registry so tests can pre-stage
+/// tools and copy them onto the agent (`seed`) after `init`. 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. 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" } },
+        };
+    }
+
+    /// Move the tools pre-staged on the harness registry onto a freshly
+    /// `init`ed agent's own registry. Post-R1 the agent owns its tool set,
+    /// so tests stage tools on the harness then transplant them here. The
+    /// agent's empty registry is freed and replaced; the harness registry
+    /// is left empty (its `deinit` becomes a no-op free).
+    fn seedInto(self: *TestHarness, agent: *Agent) void {
+        agent._registry.deinit();
+        agent._registry = self.registry;
+        self.registry = ToolRegistry.init(self.registry.allocator);
+    }
+
+    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.ResultParts {
+        const self: *EchoTool = @ptrCast(@alignCast(ctx));
+        const msg = try std.fmt.allocPrint(allocator, "{s}{s}", .{ self.prefix_owned, input });
+        return tool_mod.ResultParts.fromTextOwned(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.ResultParts {
+        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.ResultParts.fromText(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.ResultParts {
+        return error.ToolExploded;
+    }
+
+    fn deinit(ctx: *anyopaque, allocator: Allocator) void {
+        const self: *FailingTool = @ptrCast(@alignCast(ctx));
+        allocator.free(self.name_owned);
+        allocator.destroy(self);
+    }
+};
+
+/// A tool that returns a hard host failure (`error.Canceled`), which must
+/// abort the whole turn rather than degrade into a tool result.
+const HardFailTool = struct {
+    name_owned: []u8,
+
+    fn create(allocator: Allocator, name: []const u8) !Tool {
+        const self = try allocator.create(HardFailTool);
+        errdefer allocator.destroy(self);
+        self.name_owned = try allocator.dupe(u8, name);
+        return .{
+            .decl = .{ .name = self.name_owned, .description = "hard fail", .schema_json = "{}" },
+            .ctx = self,
+            .vtable = &vt,
+        };
+    }
+
+    const vt: Tool.VTable = .{ .invoke = invoke, .deinit = deinit };
+
+    fn invoke(_: *anyopaque, _: []const u8, _: Allocator) anyerror!tool_mod.ResultParts {
+        return error.Canceled;
+    }
+
+    fn deinit(ctx: *anyopaque, allocator: Allocator) void {
+        const self: *HardFailTool = @ptrCast(@alignCast(ctx));
+        allocator.free(self.name_owned);
+        allocator.destroy(self);
+    }
+};
+
+/// An in-memory `SessionStore` test double: records every appended
+/// `StoredMessage` (role + provider/model stamp) so tests can assert the
+/// agent persisted the right turn without touching disk. Honors the store
+/// ownership contract by freeing each consumed message after recording its
+/// salient fields.
+const CapturingStore = struct {
+    allocator: Allocator,
+    roles: std.ArrayList(conversation.MessageRole) = .empty,
+    base_urls: std.ArrayList([]const u8) = .empty,
+
+    fn init(allocator: Allocator) CapturingStore {
+        return .{ .allocator = allocator };
+    }
+
+    fn deinit(self: *CapturingStore) void {
+        for (self.base_urls.items) |s| self.allocator.free(s);
+        self.base_urls.deinit(self.allocator);
+        self.roles.deinit(self.allocator);
+    }
+
+    fn createVT(ctx: *anyopaque) session_store_mod.Session {
+        const self: *CapturingStore = @ptrCast(@alignCast(ctx));
+        const a = self.allocator;
+        const info: session_store_mod.SessionInfo = .{
+            .id = a.dupe(u8, "cap") catch "cap",
+            .created = a.dupe(u8, "") catch "",
+            .modified = a.dupe(u8, "") catch "",
+            .message_count = 0,
+            .last_user_message = a.dupe(u8, "") catch "",
+            .api_style = .openai_chat,
+            .base_url = a.dupe(u8, "") catch "",
+            .model = a.dupe(u8, "") catch "",
+            .reasoning = .default,
+        };
+        return .{ .info = info, .store = self.store() };
+    }
+
+    fn listVT(ctx: *anyopaque) anyerror![]session_store_mod.SessionInfo {
+        const self: *CapturingStore = @ptrCast(@alignCast(ctx));
+        return self.allocator.alloc(session_store_mod.SessionInfo, 0);
+    }
+    fn freeSessionInfosVT(ctx: *anyopaque, infos: []session_store_mod.SessionInfo) void {
+        const self: *CapturingStore = @ptrCast(@alignCast(ctx));
+        for (infos) |i| i.deinit(self.allocator);
+        self.allocator.free(infos);
+    }
+    fn resolveVT(_: *anyopaque, _: []const u8) anyerror!?session_store_mod.Session {
+        return null;
+    }
+    fn latestVT(_: *anyopaque) anyerror!?session_store_mod.Session {
+        return null;
+    }
+    fn loadVT(_: *anyopaque, _: []const u8) anyerror!?conversation.Conversation {
+        return null;
+    }
+
+    fn appendMessagesVT(
+        ctx: *anyopaque,
+        _: []const u8,
+        messages: []session_store_mod.PersistentMessage,
+    ) anyerror!void {
+        const self: *CapturingStore = @ptrCast(@alignCast(ctx));
+        for (messages) |m| {
+            try self.roles.append(self.allocator, m.message.role);
+            try self.base_urls.append(self.allocator, try self.allocator.dupe(u8, m.identity.base_url));
+        }
+    }
+
+    const vtable: session_store_mod.SessionStore.VTable = .{
+        .create = createVT,
+        .list = listVT,
+        .freeSessionInfos = freeSessionInfosVT,
+        .resolve = resolveVT,
+        .latest = latestVT,
+        .load = loadVT,
+        .appendMessages = appendMessagesVT,
+    };
+
+    fn store(self: *CapturingStore) session_store_mod.SessionStore {
+        return .{ .ptr = self, .vtable = &vtable };
+    }
+};
+
+test "agent persists user, assistant, and tool-result messages of a 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 cap = CapturingStore.init(allocator);
+    defer cap.deinit();
+    const agent = try Agent.init(allocator, io, &h.config, cap.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    try drainTurn(agent, "call a tool");
+
+    // Persisted, in order: user prompt, assistant(ToolUse), user(ToolResult),
+    // assistant(text).
+    try testing.expectEqual(@as(usize, 4), cap.roles.items.len);
+    try testing.expectEqual(conversation.MessageRole.user, cap.roles.items[0]);
+    try testing.expectEqual(conversation.MessageRole.assistant, cap.roles.items[1]);
+    try testing.expectEqual(conversation.MessageRole.user, cap.roles.items[2]);
+    try testing.expectEqual(conversation.MessageRole.assistant, cap.roles.items[3]);
+
+    // The wire identity (base_url from the active config) rode through on
+    // every entry.
+    for (cap.base_urls.items) |b| {
+        try testing.expectEqualStrings("u", b);
+    }
+}
+
+test "interruption: persistence stays current and excludes a dangling tool call" {
+    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 cap = CapturingStore.init(allocator);
+    defer cap.deinit();
+    const agent = try Agent.init(allocator, io, &h.config, cap.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    var s = try runUserText(agent, "call a tool");
+
+    // `run` persisted the user prompt up front, and nothing else yet.
+    try testing.expectEqual(@as(usize, 1), cap.roles.items.len);
+    try testing.expectEqual(conversation.MessageRole.user, cap.roles.items[0]);
+
+    // Pull until the assistant's tool-call message completes, then break — the
+    // embedder's interruption point (e.g. to handle the tool calls itself).
+    while (try s.next()) |ev| {
+        if (ev == .message_complete) break;
+    }
+
+    // The tool-call message is live in the conversation...
+    const msgs = agent.conversation.messages.items;
+    try testing.expectEqual(conversation.MessageRole.assistant, msgs[msgs.len - 1].role);
+    try testing.expect(Agent.hasToolUseBlock(msgs[msgs.len - 1]));
+    // ...but it is a dangling tool call (no results), so the store still holds
+    // only the coherent prefix: the interruption guarantee.
+    try testing.expectEqual(@as(usize, 1), cap.roles.items.len);
+
+    // Dropping the stream mid-turn must not persist the dangling call either.
+    s.deinit();
+    try testing.expectEqual(@as(usize, 1), cap.roles.items.len);
+
+    // The conversation is coherent enough to resume from: feeding the tool
+    // result as a fresh user turn resolves the dangling call, and now both the
+    // assistant tool-call and the user result become persistable.
+    var parts: std.ArrayList(conversation.ResultPartStored) = .empty;
+    try parts.append(allocator, .{ .text = try conversation.textualBlockFromSlice(allocator, "42") });
+    var s2 = try agent.run(.{ .blocks = &.{
+        .{ .ToolResult = .{
+            .tool_use_id = try allocator.dupe(u8, "tc_1"),
+            .parts = parts,
+            .is_error = false,
+        } },
+    } });
+    defer s2.deinit();
+    // user prompt, assistant(ToolUse), user(ToolResult) are all persisted once
+    // the dangling call is resolved by the new user turn.
+    try testing.expectEqual(@as(usize, 3), cap.roles.items.len);
+    try testing.expectEqual(conversation.MessageRole.user, cap.roles.items[0]);
+    try testing.expectEqual(conversation.MessageRole.assistant, cap.roles.items[1]);
+    try testing.expectEqual(conversation.MessageRole.user, cap.roles.items[2]);
+}
+
+test "agent runs a turn against NullStore without persisting or erroring" {
+    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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    try drainTurn(agent, "hello");
+
+    // Nothing crashed; the conversation has the user + assistant messages.
+    try testing.expectEqual(@as(usize, 2), agent.conversation.messages.items.len);
+}
+
+test "override mutators rewrite the conversation tail's tool blocks" {
+    const allocator = testing.allocator;
+    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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+
+    const conv = &agent.conversation;
+    // Committed assistant message carrying one ToolUse.
+    var a_content: std.ArrayList(conversation.ContentBlock) = .empty;
+    try a_content.append(allocator, .{ .ToolUse = .{
+        .id = try allocator.dupe(u8, "tc_1"),
+        .name = try allocator.dupe(u8, "read"),
+        .input = try conversation.textualBlockFromSlice(allocator, "{\"path\":\"orig\"}"),
+    } });
+    try conv.messages.append(allocator, .{ .role = .assistant, .content = a_content });
+
+    // Input override rewrites the tail's matching ToolUse; unknown ids miss.
+    try testing.expect(try agent.overrideToolUseInput("tc_1", "{\"path\":\"new\"}"));
+    try testing.expect(!try agent.overrideToolUseInput("tc_missing", "{}"));
+    const tu = conv.messages.items[0].content.items[0].ToolUse;
+    try testing.expectEqualStrings("{\"path\":\"new\"}", tu.input.items);
+
+    // Tool-result user message: a text part followed by a media part.
+    var parts: std.ArrayList(conversation.ResultPartStored) = .empty;
+    try parts.append(allocator, .{ .text = try conversation.textualBlockFromSlice(allocator, "original output") });
+    try parts.append(allocator, .{ .media = .{
+        .media_type = try allocator.dupe(u8, "image/png"),
+        .data = try conversation.textualBlockFromSlice(allocator, "aW1n"),
+    } });
+    var u_content: std.ArrayList(conversation.ContentBlock) = .empty;
+    try u_content.append(allocator, .{ .ToolResult = .{
+        .tool_use_id = try allocator.dupe(u8, "tc_1"),
+        .parts = parts,
+    } });
+    try conv.messages.append(allocator, .{ .role = .user, .content = u_content });
+
+    // The ToolUse override only targets the tail — now a user message.
+    try testing.expect(!try agent.overrideToolUseInput("tc_1", "{}"));
+
+    // Output override replaces the text, keeps the media part.
+    try testing.expect(try agent.overrideToolResultOutput("tc_1", "replaced output"));
+    try testing.expect(!try agent.overrideToolResultOutput("tc_other", "x"));
+    const tr = conv.messages.items[1].content.items[0].ToolResult;
+    try testing.expectEqual(@as(usize, 2), tr.parts.items.len);
+    try testing.expectEqualStrings("replaced output", tr.parts.items[0].text.items);
+    try testing.expectEqualStrings("image/png", tr.parts.items[1].media.media_type);
+}
+
+/// 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.ResultParts.fromTextOwned(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);
+    }
+};
+
+/// A source that succeeds the first member call and fails the rest with a
+/// per-call error (returning void from `invoke_batch`). Exercises the
+/// per-call error path distinct from a whole-batch failure.
+const PartialSource = 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(PartialSource);
+        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, "partial");
+            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,
+        calls: []const tool_source_mod.Call,
+        results: []tool_source_mod.CallResult,
+        allocator: Allocator,
+    ) anyerror!void {
+        for (calls, 0..) |_, j| {
+            if (j == 0) {
+                results[j] = .{ .ok = try tool_mod.ResultParts.fromText(allocator, "ok") };
+            } else {
+                results[j] = .{ .err = error.PerCallBoom };
+            }
+        }
+    }
+
+    fn deinitSrc(ctx: *anyopaque, _: Allocator) void {
+        const self: *PartialSource = @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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    try drainTurn(agent, "call a tool");
+
+    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 "Stream emits tool_dispatch_start before running tools" {
+    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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    var s = try runUserText(agent, "call a tool");
+    defer s.deinit();
+
+    const first = (try s.next()).?;
+    try testing.expect(first == .message_complete);
+    try testing.expectEqual(@as(usize, 2), agent.conversation.messages.items.len);
+
+    const start = (try s.next()).?;
+    try testing.expect(start == .tool_dispatch_start);
+    try testing.expectEqual(@as(usize, 1), start.tool_dispatch_start.count);
+    // The ToolResult user message must not exist yet; otherwise callers do
+    // not get a chance to render the tool as running before dispatch blocks.
+    try testing.expectEqual(@as(usize, 2), agent.conversation.messages.items.len);
+
+    const complete = (try s.next()).?;
+    try testing.expect(complete == .tool_dispatch_complete);
+    try testing.expectEqual(@as(usize, 3), agent.conversation.messages.items.len);
+}
+
+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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    try drainTurn(agent, "go");
+
+    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: native tool handler error becomes an error result and the model gets another turn" {
+    const allocator = testing.allocator;
+
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{
+            .{ .ToolUse = .{ .id = "x", .name = "boom", .input = "" } },
+        } },
+        .{ .blocks = &.{.{ .Text = "i will recover" }} },
+    };
+    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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    try drainTurn(agent, "break it");
+
+    // user, assistant(tool_use), user(tool_result), assistant(text)
+    try testing.expectEqual(@as(usize, 4), conv.messages.items.len);
+    const tr = conv.messages.items[2].content.items[0].ToolResult;
+    try testing.expectEqualStrings("x", tr.tool_use_id);
+    try testing.expect(tr.is_error);
+    try testing.expect(std.mem.indexOf(u8, trText(tr), "ToolExploded") != null);
+}
+
+test "runStep: unknown tool becomes an error tool result and the loop continues" {
+    const allocator = testing.allocator;
+
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{
+            .{ .ToolUse = .{ .id = "z", .name = "ghost", .input = "" } },
+        } },
+        .{ .blocks = &.{.{ .Text = "ok, that tool does not exist" }} },
+    };
+    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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    try drainTurn(agent, "call a ghost");
+
+    // messages: user, assistant(tool_use), user(tool_result), assistant(text)
+    try testing.expectEqual(@as(usize, 4), conv.messages.items.len);
+    const tr_msg = conv.messages.items[2];
+    try testing.expectEqual(conversation.MessageRole.user, tr_msg.role);
+    const tr = tr_msg.content.items[0].ToolResult;
+    try testing.expectEqualStrings("z", tr.tool_use_id);
+    try testing.expect(tr.is_error);
+    try testing.expect(std.mem.indexOf(u8, trText(tr), "UnknownTool") != null);
+}
+
+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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    try drainTurn(agent, "hello");
+
+    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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    try testing.expectError(error.EmptyAssistantResponse, drainTurn(agent, "hi"));
+}
+
+// ------------ 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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    try drainTurn(agent, "go");
+
+    // Locate the source and inspect its observed batches.
+    const view = agent._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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    try drainTurn(agent, "go");
+
+    const view_a = agent._registry.lookup("src_a_t") orelse return error.NotFound;
+    const view_b = agent._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 becomes per-call error results and continues" {
+    const allocator = testing.allocator;
+
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{
+            .{ .ToolUse = .{ .id = "a", .name = "fa", .input = "" } },
+            .{ .ToolUse = .{ .id = "b", .name = "fb", .input = "" } },
+        } },
+        .{ .blocks = &.{.{ .Text = "recovered" }} },
+    };
+    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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    try drainTurn(agent, "kaboom");
+
+    // user, assistant(tool_use x2), user(tool_result x2), assistant(text)
+    try testing.expectEqual(@as(usize, 4), conv.messages.items.len);
+    const tr_msg = conv.messages.items[2];
+    try testing.expectEqual(@as(usize, 2), tr_msg.content.items.len);
+    // Every member of the failed batch produced an error result, in order.
+    const tr_a = tr_msg.content.items[0].ToolResult;
+    const tr_b = tr_msg.content.items[1].ToolResult;
+    try testing.expectEqualStrings("a", tr_a.tool_use_id);
+    try testing.expectEqualStrings("b", tr_b.tool_use_id);
+    try testing.expect(tr_a.is_error);
+    try testing.expect(tr_b.is_error);
+    try testing.expect(std.mem.indexOf(u8, trText(tr_a), "SourceExploded") != null);
+}
+
+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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    try drainTurn(agent, "go");
+
+    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 provider between turns; agent tool set persists" {
+    // Post-R1 the tool set lives on the `Agent`, not on `Config`. Swapping
+    // the config pointer (`setConfig`) changes provider/model at the next
+    // turn boundary but leaves the agent's registered tools intact: a turn
+    // after the swap still resolves a tool registered before it.
+    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();
+
+    const cfg_a: config_mod.Config = .{
+        .provider = .{ .openai_chat = .{ .api_key = "k", .base_url = "a", .model = "m" } },
+    };
+    const cfg_b: config_mod.Config = .{
+        .provider = .{ .openai_chat = .{ .api_key = "k", .base_url = "b", .model = "m" } },
+    };
+
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &cfg_a, ns.store().create(), null);
+    defer agent.deinit();
+    try agent.registerTool(try EchoTool.create(allocator, "late", "B:"));
+    agent._open_stream_fn = stub.install();
+
+    // The tool is visible regardless of which config is active.
+    try testing.expect(agent._registry.lookup("late") != null);
+    agent.setConfig(&cfg_b);
+    try testing.expect(agent._registry.lookup("late") != null);
+
+    // A real turn after the swap still resolves `late`, then loops to the
+    // final text turn.
+    const conv = &agent.conversation;
+    try drainTurn(agent, "go");
+
+    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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+    try conv.addSystemMessage("you are helpful");
+    try addUserText(conv, "first question here with several words");
+    try conv.addAssistantMessage(&.{
+        .{ .Text = try conversation.textualBlockFromSlice(allocator, "first answer with several words") },
+    }, null);
+    try addUserText(conv, "second recent question");
+    try conv.addAssistantMessage(&.{
+        .{ .Text = try conversation.textualBlockFromSlice(allocator, "second recent answer") },
+    }, null);
+
+    const res = try agent._compactInPlace("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 "interruption: resuming next() after a break dispatches the pending tools" {
+    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 cap = CapturingStore.init(allocator);
+    defer cap.deinit();
+    const agent = try Agent.init(allocator, io, &h.config, cap.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    var s = try runUserText(agent, "call a tool");
+    defer s.deinit();
+
+    // Break right after the assistant's tool-call message completes — the
+    // stream is paused, NOT closed. The tool has not been dispatched.
+    while (try s.next()) |ev| {
+        if (ev == .message_complete) break;
+    }
+    try testing.expectEqual(@as(usize, 1), cap.roles.items.len); // only the prompt
+
+    // Resuming `next()` on the same stream picks up where it left off and runs
+    // the agent's registered-tool machinery: no new user turn required.
+    var saw_dispatch = false;
+    while (try s.next()) |ev| {
+        if (ev == .tool_dispatch_complete) saw_dispatch = true;
+    }
+    try testing.expect(saw_dispatch);
+
+    // The full turn is now persisted: prompt, assistant(ToolUse),
+    // user(ToolResult), assistant(text).
+    try testing.expectEqual(@as(usize, 4), cap.roles.items.len);
+    try testing.expectEqual(conversation.MessageRole.user, cap.roles.items[2]);
+}
+
+test "compact: tolerates a trailing dangling tool call (interrupted turn)" {
+    const allocator = testing.allocator;
+
+    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();
+    h.config.compaction = .{ .keep_verbatim = 10 };
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+    try conv.addSystemMessage("you are helpful");
+    try addUserText(conv, "first question here with several words");
+    try conv.addAssistantMessage(&.{
+        .{ .Text = try conversation.textualBlockFromSlice(allocator, "first answer with several words") },
+    }, null);
+    // A fresh turn whose assistant message ends in a tool call with no results
+    // yet — exactly the state a break right after `message_complete` leaves.
+    try addUserText(conv, "second recent question");
+    try conv.addAssistantMessage(&.{
+        .{ .ToolUse = .{
+            .id = try allocator.dupe(u8, "tc_1"),
+            .name = try allocator.dupe(u8, "echo"),
+            .input = try conversation.textualBlockFromSlice(allocator, "hello"),
+        } },
+    }, null);
+
+    // Compaction must not choke on the dangling tail: it summarizes the older
+    // turn and keeps the interrupted turn (incl. the dangling call) verbatim.
+    const res = try agent._compactInPlace("Summarize the conversation.", null);
+    try testing.expect(res.compacted);
+
+    // The dangling tool call survives as the tail, intact and resumable.
+    const last = conv.messages.items[conv.messages.items.len - 1];
+    try testing.expectEqual(conversation.MessageRole.assistant, last.role);
+    try testing.expect(Agent.hasToolUseBlock(last));
+    try testing.expectEqualStrings("tc_1", last.content.items[0].ToolUse.id);
+}
+
+test "compact: restated suffix usage reconstructs a fresh cumulative chain" {
+    const allocator = testing.allocator;
+
+    // The compaction turn reports a known output token count (the summary
+    // size used as the new anchor).
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{
+            .blocks = &.{.{ .Text = "SUMMARY" }},
+            .usage = .{ .input = 9999, .output = 100 }, // input is ignored; only output anchors
+        },
+    };
+    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();
+    // Budget keeps the last two turns verbatim (deltas 8 + 6 = 14) but
+    // summarizes the prefix.
+    h.config.compaction = .{ .keep_verbatim = 20 };
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+    // Prefix turn (will be summarized). Cumulative footprint = 500+40+10+50
+    // = 600. Its real usage (with cache buckets) is irrelevant
+    // post-compaction.
+    try addUserText(conv, "first question here with several words");
+    try conv.addAssistantMessage(
+        &.{.{ .Text = try conversation.textualBlockFromSlice(allocator, "first answer with words") }},
+        .{ .input = 500, .output = 50, .cache_read = 40, .cache_write = 10 },
+    );
+    // Kept turn 1: cumulative 608 (delta 8 over prefix). Reasoning is a
+    // subset of output.
+    try addUserText(conv, "kept question"); // 2 words => ceil(2*1.3)=3 tokens
+    try conv.addAssistantMessage(
+        &.{.{ .Text = try conversation.textualBlockFromSlice(allocator, "kept answer") }},
+        .{ .input = 600, .output = 8, .reasoning = 5 },
+    );
+    // Kept turn 2: cumulative 614 (delta 6). Cache buckets present to prove
+    // they collapse into `input` on restatement.
+    try addUserText(conv, "two more words here"); // 4 words => ceil(4*1.3)=6
+    try conv.addAssistantMessage(
+        &.{.{ .Text = try conversation.textualBlockFromSlice(allocator, "final answer") }},
+        .{ .input = 600, .output = 4, .cache_read = 8, .cache_write = 2 },
+    );
+
+    const res = try agent._compactInPlace("Summarize.", null);
+    try testing.expect(res.compacted);
+
+    // Rebuilt: [summary(user), u1, a1, u2, a2].
+    try testing.expectEqual(@as(usize, 5), conv.messages.items.len);
+    const asst1 = conv.messages.items[2];
+    const asst2 = conv.messages.items[4];
+    try testing.expect(conv.messages.items[1].usage == null); // kept user
+
+    // First restated assistant: input = summary_size(100) +
+    // user_size("kept question" => 3) = 103. Cache zeroed; output/reasoning
+    // verbatim.
+    const us1 = asst1.usage.?;
+    try testing.expectEqual(@as(u64, 103), us1.input);
+    try testing.expectEqual(@as(u64, 0), us1.cache_read);
+    try testing.expectEqual(@as(u64, 0), us1.cache_write);
+    try testing.expectEqual(@as(u64, 8), us1.output);
+    try testing.expectEqual(@as(u64, 5), us1.reasoning);
+
+    // Second restated assistant: input = prev.input(103) + prev.output(8) +
+    // user_size("two more words here" => 6) = 117. Original cache buckets
+    // (8+2) collapse away; only the synthetic cumulative total survives in
+    // `input`.
+    const us2 = asst2.usage.?;
+    try testing.expectEqual(@as(u64, 117), us2.input);
+    try testing.expectEqual(@as(u64, 0), us2.cache_read);
+    try testing.expectEqual(@as(u64, 0), us2.cache_write);
+    try testing.expectEqual(@as(u64, 4), us2.output);
+}
+
+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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+    try conv.addSystemMessage("sys");
+    try addUserText(conv, "hi");
+    try conv.addAssistantMessage(&.{
+        .{ .Text = try conversation.textualBlockFromSlice(allocator, "hello") },
+    }, null);
+
+    const res = try agent._compactInPlace("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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+    try addUserText(conv, "question one two three");
+    try conv.addAssistantMessage(&.{
+        .{ .Text = try conversation.textualBlockFromSlice(allocator, "answer one two three") },
+    }, null);
+    try addUserText(conv, "question two");
+    try conv.addAssistantMessage(&.{
+        .{ .Text = try conversation.textualBlockFromSlice(allocator, "answer two") },
+    }, null);
+
+    const res = try agent._compactInPlace("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, .compaction_prompt = "Summarize the conversation." };
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+    try conv.addSystemMessage("you are helpful");
+    try addUserText(conv, "first question with several words here");
+    try conv.addAssistantMessage(&.{
+        .{ .Text = try conversation.textualBlockFromSlice(allocator, "first answer with several words") },
+    }, null);
+
+    try drainTurn(agent, "second recent question");
+
+    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 ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+    // No compaction_system_prompt set -> overflow propagates.
+
+    try testing.expectError(error.ContextOverflow, drainTurn(agent, "hi"));
+}
+
+// -----------------------------------------------------------------------------
+// Phase 6: provider retry + tool-error holistic tests
+// -----------------------------------------------------------------------------
+
+/// Records the `provider_retry` events a turn emits, so tests can assert
+/// retry scheduling without a live provider.
+const RetryRecorder = struct {
+    infos: std.ArrayList(provider_mod.ProviderRetryInfo) = .empty,
+    allocator: Allocator,
+
+    /// Append a copy of `info`, dup'ing its borrowed `message` so assertions
+    /// stay valid after the source `Stream` is deinit'd.
+    fn append(self: *RetryRecorder, info: provider_mod.ProviderRetryInfo) !void {
+        var owned = info;
+        if (info.message) |m| owned.message = try self.allocator.dupe(u8, m);
+        try self.infos.append(self.allocator, owned);
+    }
+
+    fn deinit(self: *RetryRecorder) void {
+        for (self.infos.items) |info| {
+            if (info.message) |m| self.allocator.free(m);
+        }
+        self.infos.deinit(self.allocator);
+    }
+};
+
+/// Drive a whole turn via the pull `Stream`, recording every
+/// `provider_retry` event into `rr` and discarding the rest. Returns the
+/// same terminal error the turn would raise.
+fn drainTurnRecording(agent: *Agent, text: []const u8, rr: *RetryRecorder) !void {
+    var s = try runUserText(agent, text);
+    defer s.deinit();
+    while (try s.next()) |ev| {
+        if (ev == .provider_retry) try rr.append(ev.provider_retry);
+    }
+}
+
+/// Build an agent + harness with near-zero backoff so retry tests don't
+/// actually sleep. Caller owns the harness and must keep it alive.
+fn fastRetryHarness(h: *TestHarness) void {
+    h.activate();
+    // Make sleeps negligible and deterministic (no jitter).
+    h.config.retry = .{
+        .max_attempts = 4,
+        .initial_delay_ms = 0,
+        .max_delay_ms = 0,
+        .multiplier = 2.0,
+        .jitter = false,
+    };
+}
+
+test "runStep: provider 429 retries then succeeds without duplicate messages" {
+    const allocator = testing.allocator;
+
+    const errs = [_]StubProvider.ScriptedError{
+        .{ .err = error.ProviderRateLimited, .status_code = 429 },
+        .{ .err = error.ProviderRateLimited, .status_code = 429 },
+    };
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{.{ .Text = "finally" }} },
+    };
+    var stub = StubProvider{
+        .allocator = allocator,
+        .scripted = &scripted,
+        .scripted_errors = &errs,
+    };
+    var threaded: std.Io.Threaded = .init(allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    var h = TestHarness.init(allocator);
+    defer h.deinit();
+    fastRetryHarness(&h);
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    var rr = RetryRecorder{ .allocator = allocator };
+    defer rr.deinit();
+    try drainTurnRecording(agent, "hi", &rr);
+
+    // Two failures + one success.
+    try testing.expectEqual(@as(usize, 3), stub.calls_made);
+    // No duplicate assistant messages: user + single assistant.
+    try testing.expectEqual(@as(usize, 2), conv.messages.items.len);
+    try testing.expectEqual(conversation.MessageRole.assistant, conv.messages.items[1].role);
+    // Two retry notifications, delivered before each delayed retry.
+    try testing.expectEqual(@as(usize, 2), rr.infos.items.len);
+    try testing.expectEqual(@as(?u16, 429), rr.infos.items[0].status_code);
+    try testing.expectEqual(@as(usize, 1), rr.infos.items[0].attempt);
+    try testing.expectEqual(@as(usize, 2), rr.infos.items[1].attempt);
+}
+
+test "runStep: provider 500 retries with backoff notification" {
+    const allocator = testing.allocator;
+
+    const errs = [_]StubProvider.ScriptedError{
+        .{ .err = error.ProviderServerError, .status_code = 500 },
+    };
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{.{ .Text = "ok" }} },
+    };
+    var stub = StubProvider{
+        .allocator = allocator,
+        .scripted = &scripted,
+        .scripted_errors = &errs,
+    };
+    var threaded: std.Io.Threaded = .init(allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    var h = TestHarness.init(allocator);
+    defer h.deinit();
+    fastRetryHarness(&h);
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    var rr = RetryRecorder{ .allocator = allocator };
+    defer rr.deinit();
+    try drainTurnRecording(agent, "hi", &rr);
+
+    try testing.expectEqual(@as(usize, 2), stub.calls_made);
+    try testing.expectEqual(@as(usize, 1), rr.infos.items.len);
+    try testing.expectEqual(error.ProviderServerError, rr.infos.items[0].err);
+    try testing.expectEqual(@as(usize, 4), rr.infos.items[0].max_attempts);
+    try testing.expect(!rr.infos.items[0].compaction);
+}
+
+test "runStep: provider auth failure does not retry" {
+    const allocator = testing.allocator;
+
+    const errs = [_]StubProvider.ScriptedError{
+        .{ .err = error.ProviderAuthFailed, .status_code = 401 },
+    };
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{.{ .Text = "unreachable" }} },
+    };
+    var stub = StubProvider{
+        .allocator = allocator,
+        .scripted = &scripted,
+        .scripted_errors = &errs,
+    };
+    var threaded: std.Io.Threaded = .init(allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    var h = TestHarness.init(allocator);
+    defer h.deinit();
+    fastRetryHarness(&h);
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    var rr = RetryRecorder{ .allocator = allocator };
+    defer rr.deinit();
+    try testing.expectError(error.ProviderAuthFailed, drainTurnRecording(agent, "hi", &rr));
+
+    // Exactly one attempt, no retry notification.
+    try testing.expectEqual(@as(usize, 1), stub.calls_made);
+    try testing.expectEqual(@as(usize, 0), rr.infos.items.len);
+}
+
+test "runStep: retries exhaust and hard-fail after max_attempts" {
+    const allocator = testing.allocator;
+
+    const errs = [_]StubProvider.ScriptedError{
+        .{ .err = error.ProviderUnavailable, .status_code = 503 },
+        .{ .err = error.ProviderUnavailable, .status_code = 503 },
+        .{ .err = error.ProviderUnavailable, .status_code = 503 },
+        .{ .err = error.ProviderUnavailable, .status_code = 503 },
+    };
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{.{ .Text = "unreachable" }} },
+    };
+    var stub = StubProvider{
+        .allocator = allocator,
+        .scripted = &scripted,
+        .scripted_errors = &errs,
+    };
+    var threaded: std.Io.Threaded = .init(allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    var h = TestHarness.init(allocator);
+    defer h.deinit();
+    fastRetryHarness(&h);
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    var rr = RetryRecorder{ .allocator = allocator };
+    defer rr.deinit();
+    try testing.expectError(error.ProviderUnavailable, drainTurnRecording(agent, "hi", &rr));
+
+    // 4 attempts total (max_attempts), 3 retry notifications.
+    try testing.expectEqual(@as(usize, 4), stub.calls_made);
+    try testing.expectEqual(@as(usize, 3), rr.infos.items.len);
+}
+
+test "runStep: mid-stream failure retries then succeeds" {
+    const allocator = testing.allocator;
+
+    // The open succeeds every time; the FIRST two responses fail during
+    // `produce` (mid-stream), the third is a normal scripted turn.
+    const stream_errs = [_]StubProvider.StreamFailure{
+        .{ .err = error.ProviderOverloaded, .message = "overloaded_error: Overloaded" },
+        .{ .err = error.ProviderOverloaded, .message = "overloaded_error: Overloaded" },
+    };
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{.{ .Text = "finally" }} },
+    };
+    var stub = StubProvider{
+        .allocator = allocator,
+        .scripted = &scripted,
+        .stream_errors = &stream_errs,
+    };
+    var threaded: std.Io.Threaded = .init(allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    var h = TestHarness.init(allocator);
+    defer h.deinit();
+    fastRetryHarness(&h);
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    var rr = RetryRecorder{ .allocator = allocator };
+    defer rr.deinit();
+    try drainTurnRecording(agent, "hi", &rr);
+
+    // Three opens: two that fail mid-stream + one that succeeds.
+    try testing.expectEqual(@as(usize, 3), stub.calls_made);
+    // No duplicate assistant messages: the mid-stream failures committed
+    // nothing, so we end with user + a single assistant.
+    try testing.expectEqual(@as(usize, 2), conv.messages.items.len);
+    try testing.expectEqual(conversation.MessageRole.assistant, conv.messages.items[1].role);
+    // Two retry notifications, with a monotonically increasing attempt count
+    // (the bug was that this stayed pinned at 1 forever).
+    try testing.expectEqual(@as(usize, 2), rr.infos.items.len);
+    try testing.expectEqual(@as(usize, 1), rr.infos.items[0].attempt);
+    try testing.expectEqual(@as(usize, 2), rr.infos.items[1].attempt);
+    // The provider's diagnostic is surfaced on the retry notice (was the
+    // whole point: the user should see "overloaded", not a bare error name).
+    try testing.expectEqualStrings("overloaded_error: Overloaded", rr.infos.items[0].message.?);
+}
+
+test "runStep: mid-stream failures exhaust and hard-fail after max_attempts" {
+    const allocator = testing.allocator;
+
+    // Every open succeeds but every response fails mid-stream. Without the
+    // attempt counter this would re-open forever; it must hard-fail after
+    // `max_attempts` opens.
+    const stream_errs = [_]StubProvider.StreamFailure{
+        .{ .err = error.ProviderStreamMalformed },
+        .{ .err = error.ProviderStreamMalformed },
+        .{ .err = error.ProviderStreamMalformed },
+        .{ .err = error.ProviderStreamMalformed },
+    };
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{.{ .Text = "unreachable" }} },
+    };
+    var stub = StubProvider{
+        .allocator = allocator,
+        .scripted = &scripted,
+        .stream_errors = &stream_errs,
+    };
+    var threaded: std.Io.Threaded = .init(allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    var h = TestHarness.init(allocator);
+    defer h.deinit();
+    fastRetryHarness(&h);
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    var rr = RetryRecorder{ .allocator = allocator };
+    defer rr.deinit();
+    try testing.expectError(error.ProviderStreamMalformed, drainTurnRecording(agent, "hi", &rr));
+
+    // 4 opens total (max_attempts), 3 retry notifications — then hard-fail
+    // instead of looping endlessly.
+    try testing.expectEqual(@as(usize, 4), stub.calls_made);
+    try testing.expectEqual(@as(usize, 3), rr.infos.items.len);
+}
+
+test "runStep: Retry-After is honored and reported" {
+    const allocator = testing.allocator;
+
+    const errs = [_]StubProvider.ScriptedError{
+        .{ .err = error.ProviderRateLimited, .status_code = 429, .retry_after_ms = 7000 },
+    };
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{.{ .Text = "ok" }} },
+    };
+    var stub = StubProvider{
+        .allocator = allocator,
+        .scripted = &scripted,
+        .scripted_errors = &errs,
+    };
+    var threaded: std.Io.Threaded = .init(allocator, .{});
+    defer threaded.deinit();
+    const io = threaded.io();
+    var h = TestHarness.init(allocator);
+    defer h.deinit();
+    h.activate();
+    // Cap below the Retry-After to verify the policy cap applies.
+    h.config.retry = .{ .initial_delay_ms = 0, .max_delay_ms = 1, .jitter = false };
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    var rr = RetryRecorder{ .allocator = allocator };
+    defer rr.deinit();
+    try drainTurnRecording(agent, "hi", &rr);
+
+    try testing.expectEqual(@as(usize, 1), rr.infos.items.len);
+    // Reported Retry-After is the raw provider value...
+    try testing.expectEqual(@as(?u64, 7000), rr.infos.items[0].retry_after_ms);
+    // ...but the actual delay is capped by policy.max_delay_ms.
+    try testing.expectEqual(@as(u64, 1), rr.infos.items[0].delay_ms);
+}
+
+test "runStep: cancellation from a tool still hard-fails" {
+    const allocator = testing.allocator;
+
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{
+            .{ .ToolUse = .{ .id = "x", .name = "hard", .input = "" } },
+        } },
+        .{ .blocks = &.{.{ .Text = "unreachable" }} },
+    };
+    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 HardFailTool.create(allocator, "hard"));
+    h.activate();
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    try testing.expectError(error.Canceled, drainTurn(agent, "go"));
+    // Turn aborts: no tool result appended (user + assistant only).
+    try testing.expectEqual(@as(usize, 2), conv.messages.items.len);
+}
+
+test "runStep: source per-call error produces a per-call error result and continues" {
+    const allocator = testing.allocator;
+
+    const scripted = [_]StubProvider.ScriptedTurn{
+        .{ .blocks = &.{
+            .{ .ToolUse = .{ .id = "a", .name = "pa", .input = "" } },
+            .{ .ToolUse = .{ .id = "b", .name = "pb", .input = "" } },
+        } },
+        .{ .blocks = &.{.{ .Text = "moving on" }} },
+    };
+    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 PartialSource.create(allocator, "ps", &.{ "pa", "pb" }));
+    h.activate();
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+
+    try drainTurn(agent, "go");
+
+    try testing.expectEqual(@as(usize, 4), conv.messages.items.len);
+    const tr_msg = conv.messages.items[2];
+    const tr_a = tr_msg.content.items[0].ToolResult; // first call succeeded
+    const tr_b = tr_msg.content.items[1].ToolResult; // second failed
+    try testing.expect(!tr_a.is_error);
+    try testing.expectEqualStrings("ok", trText(tr_a));
+    try testing.expect(tr_b.is_error);
+    try testing.expect(std.mem.indexOf(u8, trText(tr_b), "PerCallBoom") != null);
+}
+
+test "runStep: context-overflow compaction fires a compaction retry notification" {
+    const allocator = testing.allocator;
+
+    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, .compaction_prompt = "Summarize the conversation." };
+    var ns = null_store_mod.NullStore.init(allocator);
+    const agent = try Agent.init(allocator, io, &h.config, ns.store().create(), null);
+    defer agent.deinit();
+    h.seedInto(agent);
+    agent._open_stream_fn = stub.install();
+
+    const conv = &agent.conversation;
+    try conv.addSystemMessage("you are helpful");
+    try addUserText(conv, "first question with several words here");
+    try conv.addAssistantMessage(&.{
+        .{ .Text = try conversation.textualBlockFromSlice(allocator, "first answer with several words") },
+    }, null);
+
+    var rr = RetryRecorder{ .allocator = allocator };
+    defer rr.deinit();
+    try drainTurnRecording(agent, "second recent question", &rr);
+
+    try testing.expect(agent._auto_compacted);
+    // Exactly one notification, flagged as a compaction retry with no delay.
+    try testing.expectEqual(@as(usize, 1), rr.infos.items.len);
+    try testing.expect(rr.infos.items[0].compaction);
+    try testing.expectEqual(@as(u64, 0), rr.infos.items[0].delay_ms);
+    try testing.expectEqual(error.ContextOverflow, rr.infos.items[0].err);
+}