Compact long sessions into fresh context windows

README.md

@@ -93,6 +93,8 @@ Run `/feedback` to open the feedback form at `fx.sh/feedback`. It does not creat
 
 Run `/trace` to create a private Markdown diagnostic with logs, session context, runtime state, permissions, and recent activity. On macOS, fx copies the `.md` file to the clipboard; on other platforms, it saves the file and prints its path. Review and redact the trace before sharing it.
 
+fx automatically summarizes a long session into a fresh context window when the active model request reaches 80% of its usable input capacity, then continues the same turn. Run `/compact` to create the same durable handoff immediately and wait for your next prompt.
+
 Use `fx ask` for a single request:
 
 ```bash

src/acp/prompt.zig

@@ -338,6 +338,10 @@ const AcpContext = struct {
             .max_tool_result_bytes = session.max_tool_result_bytes,
             .api_key = session.api_key,
             .agent_stream_provider = server.streamProviderFor(self.state, session.provider),
+            .compaction_route = self.state.cfg.provider_set.compactionRoute(
+                session.provider,
+                session.credential_source,
+            ),
             .credential_source = session.credential_source,
             .account_id = session.account_id,
             .provider = session.provider,
@@ -773,7 +777,7 @@ pub fn handlePrompt(
     if (explicit_skills.diagnostic_notice) |notice| try pushContextNotice(@ptrCast(&ctx), notice);
 
     session.session_rt.setConversationLanguageFromUserMessage(owned_prompt);
-    const context_history = try session.session_rt.snapshotContextHistory(alloc);
+    const context_history = try session.session_rt.snapshotHistory(alloc);
     defer types.freeHistoryTurnSlice(alloc, context_history);
     var context_snapshot = try state.context_snapshot.dupe(alloc);
     defer context_snapshot.deinit(alloc);
@@ -801,6 +805,8 @@ pub fn handlePrompt(
         .gateway_team = state.gateway_team,
         .permission_mode = captured_permission_mode,
         .history = context_history,
+        .context_history_start = session.session_rt.contextHistoryStart(),
+        .unversioned_history_count = session.session_rt.unversionedHistoryEnd(),
         .root_user_intent_context = root_user_intent_context,
         .grants = session.session_grants,
         .context_snapshot = context_snapshot,
@@ -1311,6 +1317,10 @@ fn agentRuntimeDeps(ctx: *AcpContext) agent_runtime.AgentRuntimeDeps {
     return .{
         .ctx = @ptrCast(ctx),
         .agent_stream_provider = server.streamProviderFor(ctx.state, ctx.state.active_session.?.provider),
+        .compaction_route = ctx.state.cfg.provider_set.compactionRoute(
+            ctx.state.active_session.?.provider,
+            ctx.state.active_session.?.credential_source,
+        ),
         .flush_assistant_stream_per_content_chunk = host_target.is_wasm,
         .tool_registry = ctx.toolRegistry(),
         .context_registry = ctx.state.cfg.context_registry,

src/builtins/commands.zig

@@ -444,7 +444,7 @@ pub const slash_specs = [_]SlashSpec{
     .{ .kind = .copy, .command = "/copy", .help_entry = "/copy", .completion_description = "copy the last assistant response", .presentation_category = .session },
     .{ .kind = .feedback, .command = "/feedback", .help_entry = "/feedback", .completion_description = "open the fx feedback form", .presentation_category = .product, .show_in_welcome = true },
     .{ .kind = .trace, .command = "/trace", .help_entry = "/trace", .completion_description = "copy a private diagnostic trace", .presentation_category = .product },
-    .{ .kind = .compact, .command = "/compact", .help_entry = "/compact", .completion_description = "compact older conversation turns", .presentation_category = .session },
+    .{ .kind = .compact, .command = "/compact", .help_entry = "/compact", .completion_description = "summarize context into a fresh window", .presentation_category = .session },
     .{ .kind = .settings, .command = "/settings", .help_entry = "/settings [startup-scrollback [on|off]]", .completion_description = "browse and update settings", .presentation_category = .appearance, .has_args = true, .accepts_payload = true },
     .{ .kind = .alias, .command = "/alias", .aliases = &.{}, .help_entry = "/alias [name] [command]", .completion_description = "show alias availability", .presentation_category = .extensions, .has_args = true, .accepts_payload = true },
     .{ .kind = .credits, .command = "/credits", .aliases = &.{"/balance"}, .help_entry = "/credits (/balance)", .completion_description = "show gateway credits balance", .presentation_category = .account, .requires_prompt_credential = true },

src/builtins/gateway.zig

@@ -148,10 +148,12 @@ pub const generation_usage_provider = gateway_generation_usage.provider;
 
 pub const agent_stream_provider = agent_stream_provider_contract.Provider{
     .stream_fn = streamAgentCompletion,
+    .build_request_fn = buildAgentRequestForProvider,
 };
 
 pub const provider_bundle = provider_set.Bundle{
     .capabilities = .{ .fx_search = true, .vision_fallback = true },
+    .compaction_model = "openai/gpt-5.6-luna",
     .presentation = provider_catalog.find(.gateway),
     .auth_strategy = .vercel,
     .fallback_model_capabilities_fn = vercel_model_policy.capabilitiesForModel,
@@ -250,6 +252,14 @@ pub fn buildAgentRequest(
     unreachable;
 }
 
+fn buildAgentRequestForProvider(
+    _: ?*anyopaque,
+    alloc: Allocator,
+    request: agent_stream_provider_contract.RequestData,
+) anyerror![]u8 {
+    return buildAgentRequest(alloc, request);
+}
+
 fn resolveGatewayProviderOptions(
     model: []const u8,
     effort: shared_types.ReasoningEffort,
@@ -522,35 +532,48 @@ fn streamAgentCompletion(
             error.SubscriptionCredentialCannotAuthorizeGateway,
         );
     }
-    const payload = try buildAgentRequest(alloc, request.data());
-    defer alloc.free(payload);
+    const payload = request.prepared_request_body orelse
+        try buildAgentRequest(alloc, request.data());
+    defer if (request.prepared_request_body == null) alloc.free(payload);
     var events = request.events;
-    const result = gateway_client.streamGatewayCompletion(
-        alloc,
-        .{
-            .api_key = request.credential.secret,
-            .team = request.credential.tenant,
-            .session_id = request.session_id,
-            .model = request.model,
-            .retry_count = request.retry_count,
-            .chat_url = agentChatUrl(),
-            .payload = payload,
-            .trace_ctx = request.trace_ctx,
-            .content_capture_limit = request.content_capture_limit,
-            .delivery = request.delivery,
-            .admission = request.admission,
-            .on_reasoning_chunk = EventBridge.reasoning,
-            .on_tool_input_chunk = EventBridge.toolInput,
-            .provider_attempt_owner = switch (request.provider_attempt_owner) {
-                .transport => .transport,
-                .agent => .agent,
-            },
+    const stream_request = gateway_client.StreamRequest{
+        .api_key = request.credential.secret,
+        .team = request.credential.tenant,
+        .session_id = request.session_id,
+        .model = request.model,
+        .retry_count = request.retry_count,
+        .chat_url = agentChatUrl(),
+        .payload = payload,
+        .trace_ctx = request.trace_ctx,
+        .content_capture_limit = request.content_capture_limit,
+        .delivery = request.delivery,
+        .admission = request.admission,
+        .on_reasoning_chunk = EventBridge.reasoning,
+        .on_tool_input_chunk = EventBridge.toolInput,
+        .provider_attempt_owner = switch (request.provider_attempt_owner) {
+            .transport => .transport,
+            .agent => .agent,
         },
-        &events,
-        EventBridge.content,
-        EventBridge.toolStart,
-        request.cancel_flag,
-    ) catch |err| {
+    };
+    const result = (if (request.deadline) |deadline|
+        gateway_client.streamGatewayCompletionBounded(
+            alloc,
+            stream_request,
+            &events,
+            EventBridge.content,
+            EventBridge.toolStart,
+            deadline,
+            request.cancel_flag,
+        )
+    else
+        gateway_client.streamGatewayCompletion(
+            alloc,
+            stream_request,
+            &events,
+            EventBridge.content,
+            EventBridge.toolStart,
+            request.cancel_flag,
+        )) catch |err| {
         request.attempt_evidence.network_failure = gateway_client.networkFailureEvidence(
             err,
             request.delivery.load(),

src/builtins/providers.zig

@@ -11,6 +11,7 @@ const provider_catalog = @import("../core/auth/provider_catalog.zig");
 pub const native = provider_set.Set{
     .gateway = gateway.provider_bundle,
     .codex = .{
+        .compaction_model = "gpt-5.6-luna",
         .presentation = provider_catalog.find(.codex),
         .auth_strategy = .chatgpt,
         .agent_stream = openai_codex.agent_stream_provider,
@@ -19,6 +20,7 @@ pub const native = provider_set.Set{
         .permission_reviewer = openai_codex_permission_reviewer.provider,
     },
     .grok = .{
+        .compaction_model = "grok-4.5",
         .presentation = provider_catalog.find(.grok),
         .auth_strategy = .grok,
         .agent_stream = xai_grok.agent_stream_provider,

src/core/agent/agent_runtime.zig

@@ -39,12 +39,14 @@ pub const dispatchAttentionRequiredCheckpoint = runtime_lifecycle.dispatchAttent
 pub const TurnFinalizationGuard = runtime_finalization.TurnFinalizationGuard;
 pub const Config = runtime_config.Config;
 pub const processAgentPrompt = runtime_orchestrator.processAgentPrompt;
+pub const compactContextTransaction = runtime_orchestrator.compactContextTransaction;
 pub const persistedStatusForCurrentFxLocalResult = runtime_execution_memory.persistedStatusForCurrentFxLocalResult;
 pub const classifyProviderExecutedResultStatus = runtime_execution_memory.classifyProviderExecutedResultStatus;
 pub const normalizeAssistantTextForDisplay = runtime_assistant_stream.normalizeAssistantTextForDisplay;
 
 test {
     _ = @import("stream_provider.zig");
+    _ = @import("runtime/context_compaction.zig");
     _ = @import("runtime/tests/gateway_flow.zig");
     _ = @import("runtime/tests/tool_flow.zig");
     _ = @import("runtime/tests/interruption_flow.zig");

src/core/agent/runtime/config.zig

@@ -12,9 +12,6 @@ const stream_provider = @import("../stream_provider.zig");
 
 const ReasoningEffort = types.ReasoningEffort;
 
-pub const default_history_context_budget_tokens: usize = 24_000;
-pub const history_context_budget_window_divisor: usize = 4;
-
 /// Who owns the prompt driving this run. A root turn's prompt is real user
 /// input; a subagent turn's prompt is assistant-authored delegation and can
 /// never establish user authorization for the automatic reviewer.