session files

1 files changed, 961 insertions, 0 deletions

diff --git a/libpanto/src/session.zig b/libpanto/src/session.zig
new file mode 100644
index 0000000..e77d121
--- /dev/null
+++ b/libpanto/src/session.zig
@@ -0,0 +1,961 @@
+//! On-disk session entry types and JSON serialization.
+//!
+//! These types are the wire format of pantograph's session log. They are
+//! intentionally separate from the in-memory `Conversation`/`Message`/
+//! `ContentBlock` model:
+//!
+//!   - The in-memory model holds only what providers need to serialize a
+//!     request (`role`, `content`).
+//!   - The on-disk model holds the full event-log story: provider/model
+//!     used per request, assistant stop reason and usage, timestamps, and
+//!     enough tree structure (`id`/`parent_id`) to allow future branching.
+//!
+//! Bridge functions at the bottom convert between the two. The bridge is
+//! lossy by design: assistant metadata (provider/model/stop_reason/usage)
+//! is recorded in entries but does NOT round-trip into the in-memory
+//! conversation, because providers don't need it for request serialization.
+//!
+//! Format version: 1 (see `CURRENT_VERSION`). Migrations live in
+//! `session_manager.zig`.
+
+const std = @import("std");
+const Allocator = std.mem.Allocator;
+const Writer = std.Io.Writer;
+
+const conversation = @import("conversation.zig");
+
+/// Bumped whenever the on-disk format changes in a way that older readers
+/// cannot tolerate. Older files are upgraded by `migrate()` on load and
+/// the file is rewritten once.
+pub const CURRENT_VERSION: u32 = 1;
+
+// =============================================================================
+// Header
+// =============================================================================
+
+/// First (and only) line of a session file. Metadata only — not part of
+/// the entry tree (no id/parent_id).
+pub const SessionHeader = struct {
+    version: u32,
+    id: []const u8, // UUIDv7 string, owned
+    timestamp: []const u8, // ISO 8601, owned
+    cwd: []const u8, // owned
+
+    pub fn deinit(self: SessionHeader, alloc: Allocator) void {
+        alloc.free(self.id);
+        alloc.free(self.timestamp);
+        alloc.free(self.cwd);
+    }
+};
+
+// =============================================================================
+// Entries
+// =============================================================================
+
+/// Fields shared by every non-header entry.
+pub const EntryBase = struct {
+    id: []const u8, // 8-char hex, owned
+    parent_id: ?[]const u8, // owned, null for first entry
+    timestamp: []const u8, // ISO 8601, owned
+
+    pub fn deinit(self: EntryBase, alloc: Allocator) void {
+        alloc.free(self.id);
+        if (self.parent_id) |p| alloc.free(p);
+        alloc.free(self.timestamp);
+    }
+};
+
+pub const SessionEntry = union(enum) {
+    message: MessageEntry,
+
+    pub fn base(self: SessionEntry) EntryBase {
+        return switch (self) {
+            .message => |m| m.base,
+        };
+    }
+
+    pub fn deinit(self: SessionEntry, alloc: Allocator) void {
+        switch (self) {
+            .message => |m| m.deinit(alloc),
+        }
+    }
+};
+
+pub const MessageEntry = struct {
+    base: EntryBase,
+    /// Recorded on user message entries (both human prompts and tool-result
+    /// messages). Both are submissions to a provider API; the stamp says
+    /// which one. Null on system and assistant entries.
+    provider: ?[]const u8 = null, // owned
+    model: ?[]const u8 = null, // owned
+    message: DiskMessage,
+
+    pub fn deinit(self: MessageEntry, alloc: Allocator) void {
+        self.base.deinit(alloc);
+        if (self.provider) |p| alloc.free(p);
+        if (self.model) |m| alloc.free(m);
+        self.message.deinit(alloc);
+    }
+};
+
+pub const DiskMessageRole = enum { system, user, assistant };
+
+pub const DiskMessage = struct {
+    role: DiskMessageRole,
+    content: []DiskContentBlock, // owned
+    // Assistant-only metadata. Null for system/user messages.
+    provider: ?[]const u8 = null, // owned
+    model: ?[]const u8 = null, // owned
+    stop_reason: ?[]const u8 = null, // owned
+    usage: ?Usage = null,
+
+    pub fn deinit(self: DiskMessage, alloc: Allocator) void {
+        for (self.content) |block| block.deinit(alloc);
+        alloc.free(self.content);
+        if (self.provider) |p| alloc.free(p);
+        if (self.model) |m| alloc.free(m);
+        if (self.stop_reason) |s| alloc.free(s);
+    }
+};
+
+/// Token usage reported by a provider for a single assistant turn.
+///
+/// All four input categories sum to the total prompt tokens that were
+/// billed for the turn:
+///   `input + cache_read + cache_write` = total prompt tokens.
+///
+/// `reasoning` is a **subset** of `output`, not additive — it's the
+/// portion of the output that the model spent on internal reasoning
+/// (OpenAI o-series, Anthropic extended thinking). Cost is computed
+/// from `output` only; `reasoning` is tracked for display.
+///
+/// All fields are `u64` and default to 0. Providers that don't report a
+/// given category leave it at 0. Future cache-tier breakdowns (e.g.
+/// Anthropic's 5m vs 1h tiers) can be added as new fields without
+/// breaking the format.
+pub const Usage = struct {
+    /// Fresh input tokens billed at the model's base input rate.
+    input: u64 = 0,
+    /// Output tokens billed at the model's base output rate.
+    output: u64 = 0,
+    /// Input tokens served from cache (typ. 0.1× base input rate on
+    /// Anthropic, 0.5× on OpenAI).
+    cache_read: u64 = 0,
+    /// Input tokens written to a new cache entry (Anthropic only at
+    /// time of writing; typ. 1.25× base input rate). OpenAI's cache
+    /// hits don't bill a write premium, so this stays 0 there.
+    cache_write: u64 = 0,
+    /// Subset of `output` spent on reasoning. For OpenAI o-series
+    /// models and Anthropic extended thinking. Display-only — cost is
+    /// already accounted for via `output`.
+    reasoning: u64 = 0,
+};
+
+// =============================================================================
+// Content blocks
+// =============================================================================
+
+pub const DiskContentBlock = union(enum) {
+    text: DiskTextBlock,
+    thinking: DiskThinkingBlock,
+    tool_use: DiskToolUseBlock,
+    tool_result: DiskToolResultBlock,
+
+    pub fn deinit(self: DiskContentBlock, alloc: Allocator) void {
+        switch (self) {
+            .text => |b| b.deinit(alloc),
+            .thinking => |b| b.deinit(alloc),
+            .tool_use => |b| b.deinit(alloc),
+            .tool_result => |b| b.deinit(alloc),
+        }
+    }
+};
+
+pub const DiskTextBlock = struct {
+    text: []const u8, // owned
+    pub fn deinit(self: DiskTextBlock, alloc: Allocator) void {
+        alloc.free(self.text);
+    }
+};
+
+pub const DiskThinkingBlock = struct {
+    thinking: []const u8, // owned
+    /// Anthropic's opaque integrity token. Other providers do not produce
+    /// one. Preserved here so resumed sessions can be sent back to
+    /// Anthropic with the original thinking block intact.
+    signature: ?[]const u8 = null, // owned
+    pub fn deinit(self: DiskThinkingBlock, alloc: Allocator) void {
+        alloc.free(self.thinking);
+        if (self.signature) |s| alloc.free(s);
+    }
+};
+
+pub const DiskToolUseBlock = struct {
+    id: []const u8, // owned
+    name: []const u8, // owned
+    input: []const u8, // raw JSON bytes, owned
+    pub fn deinit(self: DiskToolUseBlock, alloc: Allocator) void {
+        alloc.free(self.id);
+        alloc.free(self.name);
+        alloc.free(self.input);
+    }
+};
+
+pub const DiskToolResultBlock = struct {
+    tool_use_id: []const u8, // owned
+    content: []const u8, // owned
+    pub fn deinit(self: DiskToolResultBlock, alloc: Allocator) void {
+        alloc.free(self.tool_use_id);
+        alloc.free(self.content);
+    }
+};
+
+// =============================================================================
+// File entry (header or entry)
+// =============================================================================
+
+pub const FileEntry = union(enum) {
+    header: SessionHeader,
+    entry: SessionEntry,
+
+    pub fn deinit(self: FileEntry, alloc: Allocator) void {
+        switch (self) {
+            .header => |h| h.deinit(alloc),
+            .entry => |e| e.deinit(alloc),
+        }
+    }
+};
+
+// =============================================================================
+// Serialization
+// =============================================================================
+
+/// Serialize the header as a single JSON line. Caller owns returned bytes.
+/// The returned slice does NOT include a trailing newline.
+pub fn serializeHeader(allocator: Allocator, header: SessionHeader) ![]u8 {
+    var aw: Writer.Allocating = .init(allocator);
+    errdefer aw.deinit();
+
+    var s: std.json.Stringify = .{ .writer = &aw.writer };
+    try s.beginObject();
+    try s.objectField("type");
+    try s.write("session");
+    try s.objectField("version");
+    try s.write(header.version);
+    try s.objectField("id");
+    try s.write(header.id);
+    try s.objectField("timestamp");
+    try s.write(header.timestamp);
+    try s.objectField("cwd");
+    try s.write(header.cwd);
+    try s.endObject();
+
+    return try aw.toOwnedSlice();
+}
+
+/// Serialize an entry as a single JSON line. Caller owns returned bytes.
+pub fn serializeEntry(allocator: Allocator, entry: SessionEntry) ![]u8 {
+    var aw: Writer.Allocating = .init(allocator);
+    errdefer aw.deinit();
+    var s: std.json.Stringify = .{ .writer = &aw.writer };
+    try writeEntry(&s, entry);
+    return try aw.toOwnedSlice();
+}
+
+fn writeEntry(s: *std.json.Stringify, entry: SessionEntry) !void {
+    switch (entry) {
+        .message => |m| try writeMessageEntry(s, m),
+    }
+}
+
+fn writeMessageEntry(s: *std.json.Stringify, m: MessageEntry) !void {
+    try s.beginObject();
+    try s.objectField("type");
+    try s.write("message");
+    try s.objectField("id");
+    try s.write(m.base.id);
+    try s.objectField("parentId");
+    if (m.base.parent_id) |p| try s.write(p) else try s.write(null);
+    try s.objectField("timestamp");
+    try s.write(m.base.timestamp);
+    // Top-level provider/model on user message entries.
+    if (m.provider) |p| {
+        try s.objectField("provider");
+        try s.write(p);
+    }
+    if (m.model) |mm| {
+        try s.objectField("model");
+        try s.write(mm);
+    }
+    try s.objectField("message");
+    try writeDiskMessage(s, m.message);
+    try s.endObject();
+}
+
+fn writeDiskMessage(s: *std.json.Stringify, msg: DiskMessage) !void {
+    try s.beginObject();
+    try s.objectField("role");
+    try s.write(@tagName(msg.role));
+    try s.objectField("content");
+    try s.beginArray();
+    for (msg.content) |block| {
+        try writeDiskBlock(s, block);
+    }
+    try s.endArray();
+    if (msg.provider) |p| {
+        try s.objectField("provider");
+        try s.write(p);
+    }
+    if (msg.model) |mm| {
+        try s.objectField("model");
+        try s.write(mm);
+    }
+    if (msg.stop_reason) |sr| {
+        try s.objectField("stopReason");
+        try s.write(sr);
+    }
+    if (msg.usage) |u| {
+        try s.objectField("usage");
+        try s.beginObject();
+        try s.objectField("input");
+        try s.write(u.input);
+        try s.objectField("output");
+        try s.write(u.output);
+        // Omit zero-valued auxiliary fields to keep older / unused
+        // sessions compact. Readers default missing fields to 0, so
+        // round-trip behavior is preserved.
+        if (u.cache_read != 0) {
+            try s.objectField("cacheRead");
+            try s.write(u.cache_read);
+        }
+        if (u.cache_write != 0) {
+            try s.objectField("cacheWrite");
+            try s.write(u.cache_write);
+        }
+        if (u.reasoning != 0) {
+            try s.objectField("reasoning");
+            try s.write(u.reasoning);
+        }
+        try s.endObject();
+    }
+    try s.endObject();
+}
+
+fn writeDiskBlock(s: *std.json.Stringify, block: DiskContentBlock) !void {
+    switch (block) {
+        .text => |b| {
+            try s.beginObject();
+            try s.objectField("type");
+            try s.write("text");
+            try s.objectField("text");
+            try s.write(b.text);
+            try s.endObject();
+        },
+        .thinking => |b| {
+            try s.beginObject();
+            try s.objectField("type");
+            try s.write("thinking");
+            try s.objectField("thinking");
+            try s.write(b.thinking);
+            if (b.signature) |sig| {
+                try s.objectField("signature");
+                try s.write(sig);
+            }
+            try s.endObject();
+        },
+        .tool_use => |b| {
+            try s.beginObject();
+            try s.objectField("type");
+            try s.write("toolUse");
+            try s.objectField("id");
+            try s.write(b.id);
+            try s.objectField("name");
+            try s.write(b.name);
+            try s.objectField("input");
+            try s.write(b.input);
+            try s.endObject();
+        },
+        .tool_result => |b| {
+            try s.beginObject();
+            try s.objectField("type");
+            try s.write("toolResult");
+            try s.objectField("toolUseId");
+            try s.write(b.tool_use_id);
+            try s.objectField("content");
+            try s.write(b.content);
+            try s.endObject();
+        },
+    }
+}
+
+// =============================================================================
+// Parsing
+// =============================================================================
+
+pub const ParseError = error{
+    InvalidJson,
+    MissingField,
+    UnknownType,
+    UnknownRole,
+    UnknownBlockType,
+} || Allocator.Error;
+
+/// Parse one JSON line into a `FileEntry`. Caller owns all bytes.
+pub fn parseLine(allocator: Allocator, line: []const u8) ParseError!FileEntry {
+    var parsed = std.json.parseFromSlice(std.json.Value, allocator, line, .{}) catch {
+        return error.InvalidJson;
+    };
+    defer parsed.deinit();
+    return parseValue(allocator, parsed.value);
+}
+
+fn parseValue(allocator: Allocator, v: std.json.Value) ParseError!FileEntry {
+    if (v != .object) return error.InvalidJson;
+    const type_v = v.object.get("type") orelse return error.MissingField;
+    if (type_v != .string) return error.MissingField;
+    const t = type_v.string;
+    if (std.mem.eql(u8, t, "session")) {
+        return .{ .header = try parseHeaderFromObject(allocator, v.object) };
+    } else if (std.mem.eql(u8, t, "message")) {
+        return .{ .entry = .{ .message = try parseMessageEntry(allocator, v.object) } };
+    } else {
+        return error.UnknownType;
+    }
+}
+
+fn parseHeaderFromObject(allocator: Allocator, obj: std.json.ObjectMap) ParseError!SessionHeader {
+    const version: u32 = blk: {
+        if (obj.get("version")) |vv| {
+            if (vv == .integer) break :blk @intCast(vv.integer);
+        }
+        break :blk 1;
+    };
+    const id = try dupeStringField(allocator, obj, "id");
+    errdefer allocator.free(id);
+    const timestamp = try dupeStringField(allocator, obj, "timestamp");
+    errdefer allocator.free(timestamp);
+    const cwd = try dupeStringField(allocator, obj, "cwd");
+    errdefer allocator.free(cwd);
+    return .{
+        .version = version,
+        .id = id,
+        .timestamp = timestamp,
+        .cwd = cwd,
+    };
+}
+
+fn parseMessageEntry(allocator: Allocator, obj: std.json.ObjectMap) ParseError!MessageEntry {
+    const id = try dupeStringField(allocator, obj, "id");
+    errdefer allocator.free(id);
+    const timestamp = try dupeStringField(allocator, obj, "timestamp");
+    errdefer allocator.free(timestamp);
+    const parent_id: ?[]const u8 = blk: {
+        const pv = obj.get("parentId") orelse break :blk null;
+        if (pv == .null) break :blk null;
+        if (pv != .string) return error.MissingField;
+        break :blk try allocator.dupe(u8, pv.string);
+    };
+    errdefer if (parent_id) |p| allocator.free(p);
+
+    const provider: ?[]const u8 = try dupeOptionalStringField(allocator, obj, "provider");
+    errdefer if (provider) |p| allocator.free(p);
+    const model: ?[]const u8 = try dupeOptionalStringField(allocator, obj, "model");
+    errdefer if (model) |m| allocator.free(m);
+
+    const msg_v = obj.get("message") orelse return error.MissingField;
+    if (msg_v != .object) return error.MissingField;
+    const msg = try parseDiskMessage(allocator, msg_v.object);
+
+    return .{
+        .base = .{ .id = id, .parent_id = parent_id, .timestamp = timestamp },
+        .provider = provider,
+        .model = model,
+        .message = msg,
+    };
+}
+
+fn parseDiskMessage(allocator: Allocator, obj: std.json.ObjectMap) ParseError!DiskMessage {
+    const role_v = obj.get("role") orelse return error.MissingField;
+    if (role_v != .string) return error.MissingField;
+    const role = std.meta.stringToEnum(DiskMessageRole, role_v.string) orelse return error.UnknownRole;
+
+    const content_v = obj.get("content") orelse return error.MissingField;
+    if (content_v != .array) return error.MissingField;
+    var content_list = try std.ArrayList(DiskContentBlock).initCapacity(allocator, content_v.array.items.len);
+    errdefer {
+        for (content_list.items) |b| b.deinit(allocator);
+        content_list.deinit(allocator);
+    }
+    for (content_v.array.items) |item| {
+        if (item != .object) return error.UnknownBlockType;
+        const block = try parseDiskBlock(allocator, item.object);
+        try content_list.append(allocator, block);
+    }
+    const content = try content_list.toOwnedSlice(allocator);
+    errdefer {
+        for (content) |b| b.deinit(allocator);
+        allocator.free(content);
+    }
+
+    const provider: ?[]const u8 = try dupeOptionalStringField(allocator, obj, "provider");
+    errdefer if (provider) |p| allocator.free(p);
+    const model: ?[]const u8 = try dupeOptionalStringField(allocator, obj, "model");
+    errdefer if (model) |m| allocator.free(m);
+    const stop_reason: ?[]const u8 = try dupeOptionalStringField(allocator, obj, "stopReason");
+    errdefer if (stop_reason) |s| allocator.free(s);
+
+    var usage: ?Usage = null;
+    if (obj.get("usage")) |uv| {
+        if (uv == .object) {
+            usage = .{
+                .input = readU64(uv.object, "input"),
+                .output = readU64(uv.object, "output"),
+                .cache_read = readU64(uv.object, "cacheRead"),
+                .cache_write = readU64(uv.object, "cacheWrite"),
+                .reasoning = readU64(uv.object, "reasoning"),
+            };
+        }
+    }
+
+    return .{
+        .role = role,
+        .content = content,
+        .provider = provider,
+        .model = model,
+        .stop_reason = stop_reason,
+        .usage = usage,
+    };
+}
+
+fn parseDiskBlock(allocator: Allocator, obj: std.json.ObjectMap) ParseError!DiskContentBlock {
+    const type_v = obj.get("type") orelse return error.MissingField;
+    if (type_v != .string) return error.MissingField;
+    const t = type_v.string;
+    if (std.mem.eql(u8, t, "text")) {
+        const text = try dupeStringField(allocator, obj, "text");
+        return .{ .text = .{ .text = text } };
+    } else if (std.mem.eql(u8, t, "thinking")) {
+        const text = try dupeStringField(allocator, obj, "thinking");
+        errdefer allocator.free(text);
+        const sig = try dupeOptionalStringField(allocator, obj, "signature");
+        return .{ .thinking = .{ .thinking = text, .signature = sig } };
+    } else if (std.mem.eql(u8, t, "toolUse")) {
+        const id = try dupeStringField(allocator, obj, "id");
+        errdefer allocator.free(id);
+        const name = try dupeStringField(allocator, obj, "name");
+        errdefer allocator.free(name);
+        const input = try dupeStringField(allocator, obj, "input");
+        return .{ .tool_use = .{ .id = id, .name = name, .input = input } };
+    } else if (std.mem.eql(u8, t, "toolResult")) {
+        const tuid = try dupeStringField(allocator, obj, "toolUseId");
+        errdefer allocator.free(tuid);
+        const content = try dupeStringField(allocator, obj, "content");
+        return .{ .tool_result = .{ .tool_use_id = tuid, .content = content } };
+    } else {
+        return error.UnknownBlockType;
+    }
+}
+
+fn readU64(obj: std.json.ObjectMap, name: []const u8) u64 {
+    const v = obj.get(name) orelse return 0;
+    if (v != .integer) return 0;
+    if (v.integer < 0) return 0;
+    return @intCast(v.integer);
+}
+
+fn dupeStringField(allocator: Allocator, obj: std.json.ObjectMap, name: []const u8) ParseError![]const u8 {
+    const v = obj.get(name) orelse return error.MissingField;
+    if (v != .string) return error.MissingField;
+    return try allocator.dupe(u8, v.string);
+}
+
+fn dupeOptionalStringField(allocator: Allocator, obj: std.json.ObjectMap, name: []const u8) ParseError!?[]const u8 {
+    const v = obj.get(name) orelse return null;
+    if (v == .null) return null;
+    if (v != .string) return error.MissingField;
+    return try allocator.dupe(u8, v.string);
+}
+
+// =============================================================================
+// Bridge between in-memory and on-disk content blocks
+// =============================================================================
+
+/// Convert an in-memory `ContentBlock` to a `DiskContentBlock`. All strings
+/// are duplicated; the source block remains untouched and the resulting
+/// disk block is independently owned.
+pub fn contentBlockToDisk(
+    allocator: Allocator,
+    block: conversation.ContentBlock,
+) !DiskContentBlock {
+    switch (block) {
+        .Text => |tb| {
+            const text = try allocator.dupe(u8, tb.items);
+            return .{ .text = .{ .text = text } };
+        },
+        .Thinking => |tb| {
+            const text = try allocator.dupe(u8, tb.text.items);
+            errdefer allocator.free(text);
+            const sig: ?[]const u8 = if (tb.signature) |s| try allocator.dupe(u8, s) else null;
+            return .{ .thinking = .{ .thinking = text, .signature = sig } };
+        },
+        .ToolUse => |tu| {
+            const id = try allocator.dupe(u8, tu.id);
+            errdefer allocator.free(id);
+            const name = try allocator.dupe(u8, tu.name);
+            errdefer allocator.free(name);
+            const input = try allocator.dupe(u8, tu.input.items);
+            return .{ .tool_use = .{ .id = id, .name = name, .input = input } };
+        },
+        .ToolResult => |tr| {
+            const tuid = try allocator.dupe(u8, tr.tool_use_id);
+            errdefer allocator.free(tuid);
+            const content = try allocator.dupe(u8, tr.content.items);
+            return .{ .tool_result = .{ .tool_use_id = tuid, .content = content } };
+        },
+    }
+}
+
+/// Convert a `DiskContentBlock` to an in-memory `ContentBlock`. Allocates
+/// fresh owned buffers for every string field. The returned block is
+/// independently owned.
+pub fn diskContentBlockToInternal(
+    allocator: Allocator,
+    block: DiskContentBlock,
+) !conversation.ContentBlock {
+    switch (block) {
+        .text => |b| {
+            const tb = try conversation.textualBlockFromSlice(allocator, b.text);
+            return .{ .Text = tb };
+        },
+        .thinking => |b| {
+            const tb = try conversation.textualBlockFromSlice(allocator, b.thinking);
+            errdefer {
+                var mut = tb;
+                mut.deinit(allocator);
+            }
+            const sig: ?[]const u8 = if (b.signature) |s| try allocator.dupe(u8, s) else null;
+            return .{ .Thinking = .{ .text = tb, .signature = sig } };
+        },
+        .tool_use => |b| {
+            const id = try allocator.dupe(u8, b.id);
+            errdefer allocator.free(id);
+            const name = try allocator.dupe(u8, b.name);
+            errdefer allocator.free(name);
+            const input = try conversation.textualBlockFromSlice(allocator, b.input);
+            return .{ .ToolUse = .{ .id = id, .name = name, .input = input } };
+        },
+        .tool_result => |b| {
+            const tuid = try allocator.dupe(u8, b.tool_use_id);
+            errdefer allocator.free(tuid);
+            const content = try conversation.textualBlockFromSlice(allocator, b.content);
+            return .{ .ToolResult = .{ .tool_use_id = tuid, .content = content } };
+        },
+    }
+}
+
+// =============================================================================
+// Tests
+// =============================================================================
+
+const testing = std.testing;
+
+fn dupe(allocator: Allocator, s: []const u8) ![]const u8 {
+    return try allocator.dupe(u8, s);
+}
+
+test "serialize/parse header round-trip" {
+    const a = testing.allocator;
+    const header: SessionHeader = .{
+        .version = 1,
+        .id = try dupe(a, "019dc5ba-53f6-71a5-ab8f-b1f8709c2572"),
+        .timestamp = try dupe(a, "2026-04-25T17:40:15.990Z"),
+        .cwd = try dupe(a, "/Users/travis/Code/pantograph"),
+    };
+    defer header.deinit(a);
+
+    const line = try serializeHeader(a, header);
+    defer a.free(line);
+
+    var fe = try parseLine(a, line);
+    defer fe.deinit(a);
+    try testing.expect(fe == .header);
+    try testing.expectEqual(@as(u32, 1), fe.header.version);
+    try testing.expectEqualStrings(header.id, fe.header.id);
+    try testing.expectEqualStrings(header.cwd, fe.header.cwd);
+}
+
+test "serialize/parse user message entry round-trip (with provider/model stamp)" {
+    const a = testing.allocator;
+
+    var content = try a.alloc(DiskContentBlock, 1);
+    content[0] = .{ .text = .{ .text = try dupe(a, "hello world") } };
+
+    const entry: SessionEntry = .{ .message = .{
+        .base = .{
+            .id = try dupe(a, "a1b2c3d4"),
+            .parent_id = try dupe(a, "00000000"),
+            .timestamp = try dupe(a, "2026-04-25T17:40:16.000Z"),
+        },
+        .provider = try dupe(a, "openai"),
+        .model = try dupe(a, "gpt-4o"),
+        .message = .{
+            .role = .user,
+            .content = content,
+        },
+    } };
+    defer entry.deinit(a);
+
+    const line = try serializeEntry(a, entry);
+    defer a.free(line);
+
+    var fe = try parseLine(a, line);
+    defer fe.deinit(a);
+    try testing.expect(fe == .entry);
+    const got = fe.entry.message;
+    try testing.expectEqualStrings("a1b2c3d4", got.base.id);
+    try testing.expectEqualStrings("00000000", got.base.parent_id.?);
+    try testing.expectEqualStrings("openai", got.provider.?);
+    try testing.expectEqualStrings("gpt-4o", got.model.?);
+    try testing.expectEqual(DiskMessageRole.user, got.message.role);
+    try testing.expectEqual(@as(usize, 1), got.message.content.len);
+    try testing.expectEqualStrings("hello world", got.message.content[0].text.text);
+}
+
+test "serialize/parse assistant message entry with metadata" {
+    const a = testing.allocator;
+
+    var content = try a.alloc(DiskContentBlock, 3);
+    content[0] = .{ .thinking = .{
+        .thinking = try dupe(a, "let me think"),
+        .signature = try dupe(a, "sig-xyz"),
+    } };
+    content[1] = .{ .text = .{ .text = try dupe(a, "I'll check.") } };
+    content[2] = .{ .tool_use = .{
+        .id = try dupe(a, "tool_abc"),
+        .name = try dupe(a, "bash"),
+        .input = try dupe(a, "{\"command\":\"ls\"}"),
+    } };
+
+    const entry: SessionEntry = .{ .message = .{
+        .base = .{
+            .id = try dupe(a, "b2c3d4e5"),
+            .parent_id = try dupe(a, "a1b2c3d4"),
+            .timestamp = try dupe(a, "2026-04-25T17:40:17.000Z"),
+        },
+        .message = .{
+            .role = .assistant,
+            .content = content,
+            .provider = try dupe(a, "anthropic"),
+            .model = try dupe(a, "claude-sonnet-4-20250514"),
+            .stop_reason = try dupe(a, "toolUse"),
+            .usage = .{ .input = 1500, .output = 85 },
+        },
+    } };
+    defer entry.deinit(a);
+
+    const line = try serializeEntry(a, entry);
+    defer a.free(line);
+
+    var fe = try parseLine(a, line);
+    defer fe.deinit(a);
+    const got = fe.entry.message;
+    try testing.expectEqual(DiskMessageRole.assistant, got.message.role);
+    try testing.expectEqual(@as(usize, 3), got.message.content.len);
+    try testing.expectEqualStrings("let me think", got.message.content[0].thinking.thinking);
+    try testing.expectEqualStrings("sig-xyz", got.message.content[0].thinking.signature.?);
+    try testing.expectEqualStrings("bash", got.message.content[2].tool_use.name);
+    try testing.expectEqualStrings("{\"command\":\"ls\"}", got.message.content[2].tool_use.input);
+    try testing.expectEqualStrings("anthropic", got.message.provider.?);
+    try testing.expectEqualStrings("toolUse", got.message.stop_reason.?);
+    try testing.expect(got.message.usage != null);
+    try testing.expectEqual(@as(u64, 1500), got.message.usage.?.input);
+    try testing.expectEqual(@as(u64, 85), got.message.usage.?.output);
+}
+
+test "serialize/parse tool result message entry" {
+    const a = testing.allocator;
+
+    var content = try a.alloc(DiskContentBlock, 1);
+    content[0] = .{ .tool_result = .{
+        .tool_use_id = try dupe(a, "tool_abc"),
+        .content = try dupe(a, "file1.txt\nfile2.txt"),
+    } };
+
+    const entry: SessionEntry = .{ .message = .{
+        .base = .{
+            .id = try dupe(a, "c3d4e5f6"),
+            .parent_id = try dupe(a, "b2c3d4e5"),
+            .timestamp = try dupe(a, "2026-04-25T17:40:18.000Z"),
+        },
+        .provider = try dupe(a, "anthropic"),
+        .model = try dupe(a, "claude-sonnet-4-20250514"),
+        .message = .{
+            .role = .user,
+            .content = content,
+        },
+    } };
+    defer entry.deinit(a);
+
+    const line = try serializeEntry(a, entry);
+    defer a.free(line);
+
+    var fe = try parseLine(a, line);
+    defer fe.deinit(a);
+    const got = fe.entry.message;
+    try testing.expectEqual(DiskMessageRole.user, got.message.role);
+    try testing.expectEqualStrings("tool_abc", got.message.content[0].tool_result.tool_use_id);
+    try testing.expectEqualStrings("file1.txt\nfile2.txt", got.message.content[0].tool_result.content);
+    try testing.expectEqualStrings("anthropic", got.provider.?);
+}
+
+test "parse: null parentId is handled" {
+    const a = testing.allocator;
+    const line =
+        \\{"type":"message","id":"abcdefab","parentId":null,"timestamp":"2026-04-25T17:40:00Z","message":{"role":"system","content":[{"type":"text","text":"hi"}]}}
+    ;
+    var fe = try parseLine(a, line);
+    defer fe.deinit(a);
+    try testing.expect(fe.entry.message.base.parent_id == null);
+}
+
+test "parse: malformed JSON is reported" {
+    const a = testing.allocator;
+    try testing.expectError(error.InvalidJson, parseLine(a, "not json"));
+    try testing.expectError(error.InvalidJson, parseLine(a, "{\"type\":\"message\""));
+}
+
+test "parse: unknown entry type is reported" {
+    const a = testing.allocator;
+    const line =
+        \\{"type":"future_entry","id":"abcdefab","parentId":null,"timestamp":"2026-04-25T17:40:00Z"}
+    ;
+    try testing.expectError(error.UnknownType, parseLine(a, line));
+}
+
+test "contentBlockToDisk: Text round-trips via in-memory" {
+    const a = testing.allocator;
+
+    var tb = try conversation.textualBlockFromSlice(a, "hello");
+    defer tb.deinit(a);
+    const block: conversation.ContentBlock = .{ .Text = tb };
+
+    const disk = try contentBlockToDisk(a, block);
+    defer disk.deinit(a);
+    try testing.expectEqualStrings("hello", disk.text.text);
+}
+
+test "diskContentBlockToInternal: ToolUse preserves id/name/input" {
+    const a = testing.allocator;
+
+    const disk: DiskContentBlock = .{ .tool_use = .{
+        .id = try a.dupe(u8, "tu_1"),
+        .name = try a.dupe(u8, "bash"),
+        .input = try a.dupe(u8, "{\"command\":\"ls\"}"),
+    } };
+    defer disk.deinit(a);
+
+    var inmem = try diskContentBlockToInternal(a, disk);
+    defer inmem.deinit(a);
+    try testing.expectEqualStrings("tu_1", inmem.ToolUse.id);
+    try testing.expectEqualStrings("bash", inmem.ToolUse.name);
+    try testing.expectEqualStrings("{\"command\":\"ls\"}", inmem.ToolUse.input.items);
+}
+
+test "Usage: all five fields round-trip; zero-valued fields omitted from JSON" {
+    const a = testing.allocator;
+
+    var content = try a.alloc(DiskContentBlock, 1);
+    content[0] = .{ .text = .{ .text = try dupe(a, "hi") } };
+
+    const entry: SessionEntry = .{ .message = .{
+        .base = .{
+            .id = try dupe(a, "deadbeef"),
+            .parent_id = null,
+            .timestamp = try dupe(a, "2026-04-25T17:40:17.000Z"),
+        },
+        .message = .{
+            .role = .assistant,
+            .content = content,
+            .provider = try dupe(a, "anthropic"),
+            .model = try dupe(a, "claude-sonnet-4-20250514"),
+            .stop_reason = try dupe(a, "stop"),
+            .usage = .{
+                .input = 100,
+                .output = 50,
+                .cache_read = 800,
+                .cache_write = 200,
+                .reasoning = 30,
+            },
+        },
+    } };
+    defer entry.deinit(a);
+
+    const line = try serializeEntry(a, entry);
+    defer a.free(line);
+
+    // Every non-zero field should appear in the serialized JSON.
+    try testing.expect(std.mem.indexOf(u8, line, "\"input\":100") != null);
+    try testing.expect(std.mem.indexOf(u8, line, "\"output\":50") != null);
+    try testing.expect(std.mem.indexOf(u8, line, "\"cacheRead\":800") != null);
+    try testing.expect(std.mem.indexOf(u8, line, "\"cacheWrite\":200") != null);
+    try testing.expect(std.mem.indexOf(u8, line, "\"reasoning\":30") != null);
+
+    var fe = try parseLine(a, line);
+    defer fe.deinit(a);
+    const u = fe.entry.message.message.usage.?;
+    try testing.expectEqual(@as(u64, 100), u.input);
+    try testing.expectEqual(@as(u64, 50), u.output);
+    try testing.expectEqual(@as(u64, 800), u.cache_read);
+    try testing.expectEqual(@as(u64, 200), u.cache_write);
+    try testing.expectEqual(@as(u64, 30), u.reasoning);
+}
+
+test "Usage: zero-valued auxiliary fields are omitted but parse back as 0" {
+    const a = testing.allocator;
+
+    var content = try a.alloc(DiskContentBlock, 1);
+    content[0] = .{ .text = .{ .text = try dupe(a, "hi") } };
+
+    const entry: SessionEntry = .{ .message = .{
+        .base = .{
+            .id = try dupe(a, "deadbeef"),
+            .parent_id = null,
+            .timestamp = try dupe(a, "2026-04-25T17:40:17.000Z"),
+        },
+        .message = .{
+            .role = .assistant,
+            .content = content,
+            .usage = .{ .input = 100, .output = 50 },
+        },
+    } };
+    defer entry.deinit(a);
+
+    const line = try serializeEntry(a, entry);
+    defer a.free(line);
+
+    try testing.expect(std.mem.indexOf(u8, line, "cacheRead") == null);
+    try testing.expect(std.mem.indexOf(u8, line, "cacheWrite") == null);
+    try testing.expect(std.mem.indexOf(u8, line, "reasoning") == null);
+
+    var fe = try parseLine(a, line);
+    defer fe.deinit(a);
+    const u = fe.entry.message.message.usage.?;
+    try testing.expectEqual(@as(u64, 0), u.cache_read);
+    try testing.expectEqual(@as(u64, 0), u.cache_write);
+    try testing.expectEqual(@as(u64, 0), u.reasoning);
+}
+
+test "diskContentBlockToInternal: Thinking preserves signature" {
+    const a = testing.allocator;
+
+    const disk: DiskContentBlock = .{ .thinking = .{
+        .thinking = try a.dupe(u8, "reasoning..."),
+        .signature = try a.dupe(u8, "sig123"),
+    } };
+    defer disk.deinit(a);
+
+    var inmem = try diskContentBlockToInternal(a, disk);
+    defer inmem.deinit(a);
+    try testing.expectEqualStrings("reasoning...", inmem.Thinking.text.items);
+    try testing.expectEqualStrings("sig123", inmem.Thinking.signature.?);
+}