Run tasks on Codex beside Claude; drop turn caps and mid-turn snapshots

internal/codex/session_internal_test.go

@@ -18,6 +18,8 @@ import (
 	"sync"
 	"testing"
 	"time"
+
+	"g.fc.run/misc/byre/internal/mcp"
 )
 
 const (
@@ -40,6 +42,9 @@ var scenarios = []scenario{
 	{"Resume", scenarioResume},
 	{"ModelSwitch", scenarioModelSwitch},
 	{"Compact", scenarioCompact},
+	{"Tools", scenarioTools},
+	{"Fork", scenarioFork},
+	{"Steer", scenarioSteer},
 }
 
 func TestLive(t *testing.T) {
@@ -267,6 +272,87 @@ func scenarioCompact(h *harness) {
 	}
 }
 
+func lookupServer(calls *[]string) *mcp.Server {
+	return &mcp.Server{Name: "byre", Version: "0.0.0", Tools: []mcp.Tool{{
+		Name:        "lookup",
+		Description: "Look up the current codeword for a topic.",
+		InputSchema: json.RawMessage(`{"type":"object","properties":{"topic":{"type":"string"}},"required":["topic"]}`),
+		Handler: func(ctx context.Context, args json.RawMessage) (string, error) {
+			*calls = append(*calls, string(args))
+			return "The codeword is TEAL-9.", nil
+		},
+	}}}
+}
+
+func scenarioTools(h *harness) {
+	calls := []string{}
+	srv := lookupServer(&calls)
+
+	first := h.start("first", Options{MCPServers: []*mcp.Server{srv}, Reminder: "This is a byre test session."})
+	h.send(first, "Call byre's lookup tool with topic birds, then reply with just the codeword it gives.")
+	h.waitTurn(first, nil)
+	if len(calls) != 1 || !strings.Contains(h.lastText(first), "TEAL-9") {
+		h.t.Fatalf("tool calls %v; assistant said %q", calls, h.lastText(first))
+	}
+	id := first.ID()
+	h.close(first)
+
+	resumed := h.start("resumed", Options{Resume: id, MCPServers: []*mcp.Server{srv}})
+	h.send(resumed, "Call byre's lookup tool again with topic fish, then reply with just the codeword it gives.")
+	h.waitTurn(resumed, nil)
+	if len(calls) != 2 || !strings.Contains(h.lastText(resumed), "TEAL-9") {
+		h.t.Fatalf("resumed thread lost the tool: calls %v; assistant said %q", calls, h.lastText(resumed))
+	}
+}
+
+func scenarioFork(h *harness) {
+	src := h.start("source", Options{})
+	h.send(src, "Remember the codeword AZURE-17. Reply with the single word ok.")
+	keep := h.waitTurn(src, nil)
+	h.send(src, "The codeword is now BEIGE-3; forget the old one. Reply with the single word ok.")
+	h.waitTurn(src, nil)
+	id := src.ID()
+	h.close(src)
+
+	fork := h.start("fork", Options{ForkFrom: id, ForkAt: keep.ID})
+	if fork.ID() == "" || fork.ID() == id {
+		h.t.Fatalf("fork thread id %q, source %q", fork.ID(), id)
+	}
+	h.send(fork, "What is the codeword? Reply with just the codeword.")
+	h.waitTurn(fork, nil)
+	text := h.lastText(fork)
+	if !strings.Contains(text, "AZURE-17") || strings.Contains(text, "BEIGE-3") {
+		h.t.Fatalf("fork didn't stop after the first turn; assistant said %q", text)
+	}
+}
+
+func scenarioSteer(h *harness) {
+	s := h.start("", Options{})
+	h.send(s, "Run the shell command `sleep 5 && echo finished`, then reply with what it printed.")
+
+	steered := false
+	h.waitTurn(s, func(ev Event) {
+		if steered || ev.Method != MethodItemStarted {
+			return
+		}
+		if decode[*ItemStartedNotification](h.t, ev).Item.Type != ThreadItemTypeCommandExecution {
+			return
+		}
+		steered = true
+		err := s.Steer(h.ctx, "Also end your reply with the word PELICAN.")
+		if err != nil {
+			h.t.Errorf("steer: %v", err)
+		}
+	})
+
+	if !steered {
+		h.t.Fatalf("never saw a command to steer during")
+	}
+	if !strings.Contains(h.lastText(s), "PELICAN") {
+		h.t.Fatalf("steer didn't reach the turn; assistant said %q", h.lastText(s))
+	}
+}
+
 type harness struct {
 	t    *testing.T
 	name string