graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs
Python
Keep empty output-limit replies out of continuation history
justrach34b5a6f6b140
diff --git a/scripts/test-headless-durability.py b/scripts/test-headless-durability.pyindex 96decd3f..2c60d7c6 100644--- a/scripts/test-headless-durability.py+++ b/scripts/test-headless-durability.py@@ -60,6 +60,13 @@ def run_case(binary, script, code=0, reason="completed", extra=(), resume=False,assert (work / "result.txt").read_text() == "saved\n"assert len(model.requests) == 3, model.requestsassert "increment" in json.dumps(model.requests[1]).lower() or "pieces" in json.dumps(model.requests[1]).lower(), model.requests[1]+ for request in model.requests[1:]:+ for message in request.get("messages", []):+ if message.get("role") == "assistant":+ assert message.get("content") or message.get("tool_calls"), message+ if script[0].get("stream_deltas"):+ assert "Think briefly" in json.dumps(model.requests[1]), model.requests[1]+ assert "unfinished thought" not in json.dumps(model.requests[1]), model.requests[1]if resume:assert (work / "result.txt").read_text() == "saved\n"resumed = subprocess.run(command + ["--resume", final["session_id"], "-p", "Continue from the saved checkpoint."],@@ -83,6 +90,9 @@ def main():print("PASS plain stdout preserved; startup resume handle on stderr", flush=True)run_case(binary, [{"text": "partial draft", "finish": "length"}, write, {"text": "saved"}])print("PASS length continues through file write; one terminal JSON result", flush=True)+ run_case(binary, [{"stream_deltas": [{"role": "assistant", "reasoning_content": "unfinished thought"}],+ "finish": "length"}, write, {"text": "saved"}])+ print("PASS reasoning-only length stop sends valid history and continues through file write", flush=True)run_case(binary, [write, {"text": "saved"}], code=3, reason="budget_exhausted", extra=("--max-model-calls", "1"), resume=True)print("PASS budget stop saves partial work and resumes by emitted session id", flush=True)run_case(binary, [{"text": "Partial work remains; continue with more budget."}], code=3, reason="budget_exhausted", extra=("--max-model-calls", "1"))diff --git a/src/agent_length.zig b/src/agent_length.zigindex 07d7e87d..a254f53c 100644--- a/src/agent_length.zig+++ b/src/agent_length.zig@@ -96,8 +96,9 @@ pub const State = struct {return error.OutputLengthExceeded;}var body: []const u8 = note;- // Never route a reasoning-only length stop through empty-completion- // rewind: preserve its history, but break the reasoning-budget loop.+ // Keep the length allowance separate from empty-completion retries.+ // Chat's non-replayable empty reply is excluded by stepOpenAI; still+ // use its reasoning to break the reasoning-budget loop.if (done) |text| if (std.mem.trim(u8, text, " \t\r\n").len == 0 and reasoningOnly(self.provider.kind, root)) {body = try std.fmt.allocPrint(self.arena, "{s} {s}", .{ note, @import("agent_empty_completion.zig").reasoning_budget_note });};diff --git a/src/agent_length_tests.zig b/src/agent_length_tests.zigindex ff89097d..d03f2b9b 100644--- a/src/agent_length_tests.zig+++ b/src/agent_length_tests.zig@@ -93,20 +93,45 @@ test "length continuation allowance stays bounded across tool progress and resettry std.testing.expectEqual(@as(u8, 1), next_turn.continuations);}-test "length continuation preserves reasoning-only output instead of empty-completion rewind" {+test "length continuation excludes empty chat replies from the next request without rewinding prior work" {var arena = std.heap.ArenaAllocator.init(std.testing.allocator);defer arena.deinit();- var buffer: [4096]u8 = undefined;+ var buffer: [16384]u8 = undefined;var writer = std.Io.Writer.fixed(&buffer);- var agent = fixture(arena.allocator(), &writer, .openai);- var state: length.State = .{};- const root = try parse(agent.arena, "{\"choices\":[{\"finish_reason\":\"length\",\"message\":{\"role\":\"assistant\",\"content\":null,\"reasoning_content\":\"unfinished thought\"}}]}");- try std.testing.expect((try state.step(&agent, root)) == null);- try std.testing.expectEqual(@as(usize, 2), agent.messages.items.len);- try std.testing.expectEqualStrings("unfinished thought", agent.messages.items[0].object.get("reasoning_content").?.string);- const serialized = try std.json.Stringify.valueAlloc(agent.arena, agent.messages.items[1], .{});- try std.testing.expect(std.mem.indexOf(u8, serialized, "Think briefly") != null);- try std.testing.expectEqual(@as(u8, 0), agent.empty_completion_retries);+ for ([_][]const u8{ "deepseek", "codegraff" }) |provider| {+ for ([_][]const u8{ "null", "\"\"", "\" \\n\"" }) |content| {+ var agent = fixture(arena.allocator(), &writer, .openai);+ agent.io = std.testing.io;+ agent.provider.id = provider;+ agent.provider.model = "deepseek-flash";+ var state: length.State = .{};+ try agent.messages.append(try messages.textMessage(agent.arena, "user", "Write the document"));+ const prior = try parse(agent.arena, "{\"role\":\"assistant\",\"content\":\"partial document\",\"reasoning_content\":\"prior useful reasoning\"}");+ try agent.messages.append(.{ .object = prior });+ const raw = try std.fmt.allocPrint(agent.arena, "{{\"choices\":[{{\"finish_reason\":\"length\",\"message\":{{\"role\":\"assistant\",\"content\":{s},\"reasoning_content\":\"unfinished thought\",\"tool_calls\":[]}}}}]}}", .{content});+ const root = try parse(agent.arena, raw);+ // Even the last allowed reply must not poison saved/resumed history.+ for (0..length.max_continuations) |_| try std.testing.expect((try state.step(&agent, root)) == null);+ try std.testing.expectError(error.OutputLengthExceeded, state.step(&agent, root));+ try std.testing.expectEqual(@as(usize, 2 + length.max_continuations), agent.messages.items.len);+ try std.testing.expectEqual(length.max_continuations, state.continuations);+ try std.testing.expectEqual(@as(u8, 0), agent.empty_completion_retries);+ const body = try agent.buildBody(null, false, false, false);+ defer std.testing.allocator.free(body);+ try std.testing.expect(std.mem.indexOf(u8, body, "unfinished thought") == null);+ try std.testing.expect(std.mem.indexOf(u8, body, "prior useful reasoning") != null);+ try std.testing.expect(std.mem.indexOf(u8, body, "Think briefly") != null);+ const request = try parse(agent.arena, body);+ var assistants: usize = 0;+ for (request.get("messages").?.array.items) |message| {+ if (!std.mem.eql(u8, message.object.get("role").?.string, "assistant")) continue;+ assistants += 1;+ try std.testing.expectEqualStrings("partial document", message.object.get("content").?.string);+ }+ try std.testing.expectEqual(@as(usize, 1), assistants);+ try std.testing.expectEqualStrings("saved the document", (try state.step(&agent, try parse(agent.arena, chat_done))).?);+ }+ }}test "length continuation respects run budgets and both engine cancellation flags without granting another call" {@@ -160,7 +185,7 @@ test "length continuation never dispatches truncated tool calls and leaves replavar buffer: [4096]u8 = undefined;var writer = std.Io.Writer.fixed(&buffer);const cases = [_]struct { kind: Kind, raw: []const u8 }{- .{ .kind = .openai, .raw = "{\"choices\":[{\"finish_reason\":\"length\",\"message\":{\"role\":\"assistant\",\"content\":null,\"tool_calls\":[{\"id\":\"c\",\"type\":\"function\",\"function\":{\"name\":\"read_file\",\"arguments\":\"\"}}]}}]}" },+ .{ .kind = .openai, .raw = "{\"choices\":[{\"finish_reason\":\"length\",\"message\":{\"role\":\"assistant\",\"content\":null,\"reasoning_content\":\"keep tool reasoning\",\"tool_calls\":[{\"id\":\"c\",\"type\":\"function\",\"function\":{\"name\":\"read_file\",\"arguments\":\"\"}}]}}]}" },.{ .kind = .responses, .raw = "{\"incomplete_details\":{\"reason\":\"max_output_tokens\"},\"output\":[{\"type\":\"function_call\",\"call_id\":\"c\",\"name\":\"read_file\",\"arguments\":\"\"}]}" },.{ .kind = .anthropic, .raw = "{\"stop_reason\":\"max_tokens\",\"content\":[{\"type\":\"tool_use\",\"id\":\"c\",\"name\":\"read_file\"}]}" },};@@ -174,6 +199,13 @@ test "length continuation never dispatches truncated tool calls and leaves replatry std.testing.expect(std.mem.indexOf(u8, result, "truncated: the response ended") != null);const history = try std.json.Stringify.valueAlloc(agent.arena, agent.messages.items[0], .{});try std.testing.expect(std.mem.indexOf(u8, history, "graff_invalid_input") == null);+ if (case.kind == .openai) {+ const body = try agent.buildBody(null, false, false, false);+ defer std.testing.allocator.free(body);+ try std.testing.expect(std.mem.indexOf(u8, body, "keep tool reasoning") != null);+ try std.testing.expect(std.mem.indexOf(u8, body, "\"tool_calls\"") != null);+ try std.testing.expect(std.mem.indexOf(u8, body, "\"tool_call_id\":\"c\"") != null);+ }try std.testing.expectEqual(@as(u8, 1), state.continuations);}}diff --git a/src/agent_steps.zig b/src/agent_steps.zigindex e527bf93..b3d27b5c 100644--- a/src/agent_steps.zig+++ b/src/agent_steps.zig@@ -291,6 +291,20 @@ pub fn stepOpenAI(self: *Agent, root: std.json.ObjectMap) !?[]const u8 {try self.sayApiError("api error: choice message was not an object", .{});return error.ApiError;};+ // A length stop may spend its entire budget reasoning. Chat rejects+ // replaying an assistant with neither content nor tool calls (#1646).+ // Keep real partial answers and tool-call reasoning verbatim; only this+ // empty reply stays out of history, including on budget/cancel exits.+ const has_content = if (msg_obj.get("content")) |c| switch (c) {+ .string => std.mem.trim(u8, c.string, " \t\r\n").len > 0,+ .array => c.array.items.len > 0,+ else => false,+ } else false;+ const has_calls = if (msg_obj.get("tool_calls")) |t| t == .array and t.array.items.len > 0 else false;+ if (length.limited(.openai, root) and !has_content and !has_calls) {+ self.pairContextMeterWithCurrentLocal();+ return "";+ }try self.messages.append(message); // echo verbatim (content may be null)var final_text: []const u8 = "";