Optimize YoloCoder with Jev context selection and bounded Muse editing

.gitignore

@@ -40,3 +40,5 @@ go.work.sum
 # on disk and the compiler still reads it. `find . -name '* [0-9]*'` finds
 # them when a build fails for no apparent reason.
 * [0-9]*
+
+/dist/

README.md

@@ -168,6 +168,10 @@ Windows PowerShell:
 irm https://raw.githubusercontent.com/mindsdb/yolocoder/main/install.ps1 | iex
 ```
 
+## Muse + Jev build
+
+See [the Muse + Jev guide](docs/jev-muse.md) for build instructions and model configuration.
+
 ## Develop
 
 Requires Go 1.24 or newer.

docs/jev-muse.md

@@ -0,0 +1,25 @@
+# Muse + Jev build
+
+Use Jev to select initial context and route small UI edits to a bounded Muse conversation. Other coding tasks use the configured primary model, with explicit completion checks and bounded recovery from supported provider errors.
+
+## Build and configure
+
+Requires Go 1.24+, Node.js 22 for web projects, and a MindsHub account serving `muse-spark-1-3` and `jev-1.13.0`.
+
+```sh
+./scripts/build-jev-muse.sh
+YOLOCODER_NO_AUTOUPDATE=1 ./dist/yolocoder-jev-muse config
+YOLOCODER_NO_AUTOUPDATE=1 ./dist/yolocoder-jev-muse model
+```
+
+Configure MindsHub and select **muse-spark-1-3** as the primary coding model. The build selects **jev-1.13.0** for routing and **muse-spark-1-3** for the small-edit writer. The model picker changes the primary model; the small-edit model is set at build time.
+
+From an empty app directory or an existing YoloCoder web project, launch the absolute path to the built binary with `--web`. The normal CLI also works. Keep `YOLOCODER_NO_AUTOUPDATE=1` when running this custom build so an update does not replace its compiled model configuration.
+
+## Behaviour
+
+Jev selects bounded initial context; uncertainty or an unavailable router lets the coding model obtain its own context. Small UI edits can use exact literal substitutions, checked for a unique match before the batch is written. The project's detected check still runs after edits.
+
+Normal tasks finish only after an explicit completion signal, successful tool results, a complete response and the project's detected check. Rejected edits receive path and remaining-budget feedback. A bounded diagnostic check can help with an already-required continuation, and narrowly recognized provider errors have a finite recovery budget.
+
+The Jev router takes precedence over `/preselect` in this build, avoiding two selectors per request. Plain builds still support `/preselect`; their normal completion and recovery improvements apply without the new router. Architecture notes, pattern-based reads, search pre-reading and rejected-patch logging remain available.

internal/agent/chat_test.go

@@ -191,6 +191,13 @@ func TestRunnerOverChatCompletions(t *testing.T) {
 		if err := json.NewDecoder(request.Body).Decode(&body); err != nil {
 			t.Fatal(err)
 		}
+		// Normal request timing follows the intact conversation in either
+		// dialect; inspect the original tool result immediately before it.
+		notes := timingNotes(t, body.Messages)
+		if len(notes) != 1 || body.Messages[len(body.Messages)-1].Content != notes[0] {
+			t.Fatal("missing or accumulated normal timing metadata")
+		}
+		body.Messages = body.Messages[:len(body.Messages)-1]
 		requests++
 		writer.Header().Set("Content-Type", "application/json")
 		reply := func(message chatMessage) {
@@ -214,16 +221,21 @@ func TestRunnerOverChatCompletions(t *testing.T) {
 				t.Fatalf("tool result did not carry the file: %q", content)
 			}
 			patch := "@index.html\n-<title>Old</title>\n+<title>New</title>\n"
-			arguments, _ := json.Marshal(map[string]string{"patch": patch})
+			arguments, _ := json.Marshal(map[string]any{"patch": patch, "response_comment_for_user": "", "task_complete": false})
 			reply(chatMessage{Role: "assistant", ToolCalls: []chatToolCall{{
 				ID: "call_2", Type: "function",
 				Function: chatCallFunction{Name: "apply_diff", Arguments: string(arguments)},
 			}}})
 		case 3:
-			// And the edit's own result comes back the same way.
+			// Preserve the edit's tool result before the diagnostic user message.
 			last := body.Messages[len(body.Messages)-1]
-			if last.Role != "tool" || last.ToolCallID != "call_2" {
-				t.Fatalf("last message = %+v, want the edit's result", last)
+			content, _ := last.Content.(string)
+			if last.Role != "user" || !strings.Contains(content, "Diagnostic checkpoint skipped") {
+				t.Fatalf("last message = %+v, want the honest skipped diagnostic", last)
+			}
+			tool := body.Messages[len(body.Messages)-2]
+			if tool.Role != "tool" || tool.ToolCallID != "call_2" {
+				t.Fatalf("preceding message = %+v, want the edit's result", tool)
 			}
 			reply(chatMessage{Role: "assistant", Content: "Retitle"})
 		default:

internal/agent/client.go

@@ -159,6 +159,12 @@ func (client *Client) Dialect() string {
 }
 
 func (client *Client) create(ctx context.Context, request responseRequest) (responseEnvelope, error) {
+	return client.createWithRetry(ctx, request, nil)
+}
+
+// retry is a turn-owned, one-use permission, shared by compatibility fallbacks.
+// Coding sessions supply it; routing and other callers stay unchanged.
+func (client *Client) createWithRetry(ctx context.Context, request responseRequest, retry func(int) bool) (responseEnvelope, error) {
 	request.Model = client.model
 	body := any(request)
 	if client.chat {
@@ -176,23 +182,10 @@ func (client *Client) create(ctx context.Context, request responseRequest) (resp
 	// Authorization header), so it is safe to trace.
 	label := requestLabel(request)
 	debug.Log("REQUEST "+label, string(payload))
-	httpRequest, err := http.NewRequestWithContext(ctx, http.MethodPost, client.endpoint, bytes.NewReader(payload))
-	if err != nil {
-		return responseEnvelope{}, err
-	}
-	httpRequest.Header.Set("Authorization", "Bearer "+client.apiKey)
-	httpRequest.Header.Set("Content-Type", "application/json")
-	httpRequest.Header.Set("Accept", "application/json")
-	response, err := client.http.Do(httpRequest)
-	if err != nil {
-		return responseEnvelope{}, fmt.Errorf("call LLM: %w", err)
-	}
-	defer response.Body.Close()
-	replyBody, err := io.ReadAll(io.LimitReader(response.Body, 8<<20))
+	response, replyBody, err := client.post(ctx, payload, label, retry)
 	if err != nil {
 		return responseEnvelope{}, err
 	}
-	debug.Log(fmt.Sprintf("RESPONSE %s (%s)", label, response.Status), string(replyBody))
 
 	// Check the status before parsing. Decoding first turned a 404 with
 	// an empty body into "decode LLM response: unexpected end of JSON
@@ -204,7 +197,7 @@ func (client *Client) create(ctx context.Context, request responseRequest) (resp
 		debug.Log("DIALECT", "the Responses API returned 404; switching to /v1/chat/completions")
 		client.chat = true
 		client.endpoint = chatEndpoint(client.baseURL)
-		return client.create(ctx, request)
+		return client.createWithRetry(ctx, request, retry)
 	}
 
 	// Some providers constrain generation to a tool call whenever tools
@@ -216,7 +209,7 @@ func (client *Client) create(ctx context.Context, request responseRequest) (resp
 		debug.Log("TOOLS", "the provider demanded a tool call the model would not make; asking again without tools")
 		request.Tools = nil
 		request.ToolChoice = ""
-		return client.create(ctx, request)
+		return client.createWithRetry(ctx, request, retry)
 	}
 
 	// Some providers won't take tool definitions and a JSON schema in the
@@ -225,7 +218,7 @@ func (client *Client) create(ctx context.Context, request responseRequest) (resp
 	if client.chat && !client.dropSchema && rejectsSchemaWithTools(replyBody) {
 		debug.Log("SCHEMA", "the provider rejects response_format alongside tools; describing the shape in the instructions instead")
 		client.dropSchema = true
-		return client.create(ctx, request)
+		return client.createWithRetry(ctx, request, retry)
 	}
 
 	envelope, parseErr := client.decode(replyBody)
@@ -238,6 +231,54 @@ func (client *Client) create(ctx context.Context, request responseRequest) (resp
 	return envelope, nil
 }
 
+// post reuses the exact encoded request and context after an eligible error.
+// Transport/read errors and successful responses never grant a replay: delivery
+// is ambiguous there, or the response may already contain usable tool output.
+func (client *Client) post(ctx context.Context, payload []byte, label string, retry func(int) bool) (*http.Response, []byte, error) {
+	const responseLimit = 8 << 20
+	endpoint := client.endpoint
+	for {
+		httpRequest, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewReader(payload))
+		if err != nil {
+			return nil, nil, err
+		}
+		httpRequest.Header.Set("Authorization", "Bearer "+client.apiKey)
+		httpRequest.Header.Set("Content-Type", "application/json")
+		httpRequest.Header.Set("Accept", "application/json")
+		response, err := client.http.Do(httpRequest)
+		if err != nil {
+			return nil, nil, fmt.Errorf("call LLM: %w", err)
+		}
+		replyBody, err := io.ReadAll(io.LimitReader(response.Body, responseLimit))
+		response.Body.Close()
+		if err != nil {
+			return nil, nil, err
+		}
+		debug.Log(fmt.Sprintf("RESPONSE %s (%s)", label, response.Status), string(replyBody))
+		// Reaching the read cap may hide an incomplete body, so do not replay it.
+		eligible := response.StatusCode == http.StatusBadGateway || response.StatusCode == 524 ||
+			(response.StatusCode == http.StatusServiceUnavailable && client.model == "muse-spark-1-3" && musePolicyUnavailable(replyBody))
+		if eligible && len(replyBody) < responseLimit && ctx.Err() == nil && retry != nil && retry(response.StatusCode) {
+			if err := ctx.Err(); err != nil {
+				return nil, nil, fmt.Errorf("call LLM: %w", err)
+			}
+			continue
+		}
+		return response, replyBody, nil
+	}
+}
+
+func musePolicyUnavailable(body []byte) bool {
+	const message = "The model policy is temporarily unavailable. Please retry shortly."
+	if strings.TrimSpace(string(body)) == message {
+		return true
+	}
+	var reply struct {
+		Error *apiError `json:"error"`
+	}
+	return json.Unmarshal(body, &reply) == nil && reply.Error != nil && reply.Error.Message == message
+}
+
 // decode reads a reply in whichever dialect this client speaks, returning
 // it in the Responses shape the rest of the agent works with.
 func (client *Client) decode(body []byte) (responseEnvelope, error) {
@@ -316,9 +357,8 @@ func (response responseEnvelope) calls() []responseItem {
 }
 
 // cutOff reports whether the model ran out of output room before it
-// finished, rather than deciding anything. The reply then carries no
-// usable text — an empty message, or none at all — and the turn should
-// ask again, shorter, instead of failing over it.
+// finished, rather than deciding anything. An incomplete envelope may
+// still contain text or tool calls; callers decide how to continue.
 func (response responseEnvelope) cutOff() bool {
 	if response.IncompleteDetails == nil {
 		return false

internal/agent/complete_test.go

@@ -25,7 +25,7 @@ func applyWith(t *testing.T, session *changeSession, arguments map[string]any) s
 	if err != nil {
 		t.Fatal(err)
 	}
-	output, _ := session.applyDiff(responseItem{Name: "apply_diff", Arguments: string(encoded)})
+	output, _, _ := session.applyDiff(responseItem{Name: "apply_diff", Arguments: string(encoded)})
 	return output
 }
 
@@ -34,34 +34,32 @@ func TestAFinishedEditEndsTheTurn(t *testing.T) {
 	session := sessionOn(t, "a.css", ".a {\n  color: red;\n}\n")
 	applyWith(t, session, map[string]any{
 		"patch":                     "@a.css\n-  color: red;\n+  color: blue;\n",
-		"turn_is_complete":          true,
+		"task_complete":             true,
 		"response_comment_for_user": "Made it blue.",
 	})
 	if session.closing != "Made it blue." {
 		t.Fatalf("closing = %q, so the turn would spend a round saying done", session.closing)
 	}
 }
 
-// A model that fills the reply but leaves the flag false is still
-// finished — that was the only signal there was, and honouring it costs
-// nothing.
-func TestAReplyWithoutTheFlagStillEndsTheTurn(t *testing.T) {
+// A closing note alone must not end an unfinished normal task.
+func TestAReplyWithoutTheFlagDoesNotCompleteTheTask(t *testing.T) {
 	session := sessionOn(t, "a.css", ".a {\n  color: red;\n}\n")
 	applyWith(t, session, map[string]any{
 		"patch":                     "@a.css\n-  color: red;\n+  color: blue;\n",
-		"turn_is_complete":          false,
+		"task_complete":             false,
 		"response_comment_for_user": "Made it blue.",
 	})
-	if session.closing != "Made it blue." {
-		t.Fatalf("closing = %q", session.closing)
+	if session.complete {
+		t.Fatal("closing note completed an unfinished task")
 	}
 }
 
 func TestAnUnfinishedEditDoesNotEndTheTurn(t *testing.T) {
 	session := sessionOn(t, "a.css", ".a {\n  color: red;\n}\n")
 	applyWith(t, session, map[string]any{
 		"patch":                     "@a.css\n-  color: red;\n+  color: blue;\n",
-		"turn_is_complete":          false,
+		"task_complete":             false,
 		"response_comment_for_user": "",
 	})
 	if session.closing != "" {
@@ -74,7 +72,7 @@ func TestAFailedEditCannotEndTheTurn(t *testing.T) {
 	session := sessionOn(t, "a.css", ".a {\n  color: red;\n}\n")
 	applyWith(t, session, map[string]any{
 		"patch":                     "@a.css\n-  color: nowhere;\n+  color: blue;\n",
-		"turn_is_complete":          true,
+		"task_complete":             true,
 		"response_comment_for_user": "Made it blue.",
 	})
 	if session.closing != "" {