agent: Zed's max_tokens is the window a prompt and its reply share, so a prompt keeps the room magpie's catalog gives it (#850) — Zed's openai_compatible max_tokens is the whole window: it keeps max_output_tokens of it for the reply and lets the prompt fill only the rest before it compacts (its thread's input_token_capacity = max_tokens - max_output_tokens). magpie wrote its catalog's Context there, which is what a prompt may hold, so Zed took the reply out of it: glm-4.6 (204800 / 131072) left a prompt 73728, deepseek-chat (128000 / 128000) none at all, gpt-5 (input 272000 / 128000) 144000 of its 400000 window. max_tokens is now Context + max_output_tokens: 335872, 256000, 400000 (OpenAI's own), and a model with no output limit still gets its Context alone. In a sandbox HOME with the real models.dev catalog, setting Zed to magpie wrote glm-4.6 335872/131072, gpt-5 400000/128000, deepseek-chat 256000/128000, claude-sonnet-4-5 264000/64000, kimi-k2 229376/98304. TestZedWindowHoldsPromptAndReply (gpt-5 400000 with a 272000 prompt; a model without an output limit) and TestZedRestoresProviderAndModelLimits fail on the old code. go vet, builds (darwin, GOOS=linux nogui, GOOS=windows) and go test -tags nogui -count=1 ./... under a temp HOME pass (TestProvidersAnswerWhileListsComeIn, TestUserPath and TestClaudeSignedOut timed out under load and pass on their own).

internal/agent/zed.go

@@ -125,6 +125,12 @@ func zedProviderJSON() map[string]any {
 		if context == 0 {
 			context = 128000 // Zed requires a context window for every custom model.
 		}
+		// Zed's max_tokens is the window a prompt and its reply share: it
+		// keeps max_output_tokens of it for the reply and lets the prompt
+		// fill the rest before it compacts. Context is what a prompt may
+		// hold, so the window is it and the reply together (#850); written
+		// as Context alone, a model whose reply may be as long as its
+		// prompt (glm-4.6) left Zed no room for a prompt at all.
 		entry := map[string]any{
 			"name": m.ID, "display_name": m.Name, "max_tokens": context,
 			"capabilities": map[string]any{
@@ -134,7 +140,9 @@ func zedProviderJSON() map[string]any {
 			},
 		}
 		if output := maxTokens(m); output > 0 {
-			entry["max_output_tokens"] = min(output, context)
+			output = min(output, context)
+			entry["max_output_tokens"] = output
+			entry["max_tokens"] = context + output
 		}
 		models = append(models, entry)
 	}

internal/agent/zed_test.go

@@ -6,6 +6,7 @@ import (
 	"os"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
 	"testing"
 
@@ -139,6 +140,54 @@ func TestZedFreshAndOwn(t *testing.T) {
 	}
 }
 
+// Zed's max_tokens is the window a prompt and its reply share, and it lets
+// a prompt fill max_tokens - max_output_tokens before it compacts (its
+// thread's input_token_capacity). That room is the prompt magpie's catalog
+// says the model takes (#850): GPT-5's input 272000 with its 128000 reply,
+// a 400000 window as OpenAI gives it, and a model with no output limit
+// gets its context alone.
+func TestZedWindowHoldsPromptAndReply(t *testing.T) {
+	home := syncHome(t)
+	credential := zedCredential
+	t.Cleanup(func() { zedCredential = credential; catalog.Reset() })
+	zedCredential = func(string) error { return nil }
+	writeFile(t, catalog.CachePath(), `{"openai":{"models":{
+		"gpt-5":{"id":"gpt-5","name":"GPT-5","limit":{"context":400000,"input":272000,"output":128000}},
+		"plain":{"id":"plain","name":"Plain","limit":{"context":64000}}}}}`)
+	catalog.Reset()
+	if err := provider.Save(provider.Provider{ID: "relay", Name: "Relay", Key: "k", Chat: "http://127.0.0.1:1/v1", Models: []string{"gpt-5", "plain"}}); err != nil {
+		t.Fatal(err)
+	}
+	a := zedAt(filepath.Join(home, "zed"))
+	if err := a.Field("model").Set("magpie/relay/gpt-5"); err != nil {
+		t.Fatal(err)
+	}
+	var p struct {
+		Models []struct {
+			Name   string `json:"name"`
+			Max    int    `json:"max_tokens"`
+			Output *int   `json:"max_output_tokens"`
+		} `json:"available_models"`
+	}
+	raw, _ := edit.GetJSON(a.Path, zedProvider)
+	if err := json.Unmarshal([]byte(raw), &p); err != nil {
+		t.Fatal(err)
+	}
+	got := map[string]string{}
+	for _, m := range p.Models {
+		s := strconv.Itoa(m.Max)
+		if m.Output != nil {
+			s += " " + strconv.Itoa(*m.Output) + " prompt " + strconv.Itoa(m.Max-*m.Output)
+		}
+		got[m.Name] = s
+	}
+	for name, want := range map[string]string{"relay/gpt-5": "400000 128000 prompt 272000", "relay/plain": "64000"} {
+		if got[name] != want {
+			t.Errorf("%s: max_tokens, max_output_tokens = %q, want %q (all %v)", name, got[name], want, got)
+		}
+	}
+}
+
 func TestZedRestoresProviderAndModelLimits(t *testing.T) {
 	home := syncHome(t)
 	writeFile(t, catalog.CachePath(), `{"zai":{"models":{"glm-4.6":{"id":"glm-4.6","name":"GLM","modalities":{"input":["text","image"]},"limit":{"context":204800,"output":300000}}}}}`)
@@ -152,7 +201,9 @@ func TestZedRestoresProviderAndModelLimits(t *testing.T) {
 	if err := a.Field("model").Set("magpie/relay/glm-4.6"); err != nil {
 		t.Fatal(err)
 	}
-	for key, want := range map[string]string{"max_tokens": "204800", "max_output_tokens": "204800", "capabilities.images": "true"} {
+	// Zed keeps max_output_tokens of max_tokens for the reply: the prompt's
+	// 204800 stays (#850), not 204800 - 204800 = 0
+	for key, want := range map[string]string{"max_tokens": "409600", "max_output_tokens": "204800", "capabilities.images": "true"} {
 		if got, _ := edit.GetJSON(a.Path, zedProvider+".available_models.0."+key); got != want {
 			t.Errorf("%s = %q, want %q", key, got, want)
 		}