Fix panel member streaming and persist full transcript with images

Three related panel-member improvements:

Tool-box streaming fix: correlate tool boxes by tool ID instead of a
single pointer (fixes orphaned spinners and empty boxes when a member
calls multiple tools in one message), forward CLI ToolEnd to capture the
tool input (command/path), normalize tool names, clean up orphaned
spinners on finalize, and add a blinking cursor during text streaming.

Image/screenshot support: CLI panel members now extract image refs from
the last user message (like the worker) and drop them as temp files
readable via the Read tool, so members can see screenshots and file
attachments instead of being blind on image-based questions.

Reload persistence: each member persists a structured JSON transcript
({name, model, segments:[text|tool]}) instead of plain text, so a reload
restores text AND tool boxes interleaved exactly like live. history.js
renders the transcript and keeps back-compat for the old [Panel: Name]
plaintext format.

- AnthropicService.Agents.cs: NonStreamingLoopCallbacks.TranscriptSink,
  loop collects text/tool segments, tool input added to combined event
- AnthropicService.Panel.cs: CLI member ID-correlation + ToolEnd input +
  image refs + transcript; RunPanelMember persists one JSON transcript
- stream-ui.js: tool boxes keyed by ID, name normalization, cursor,
  orphaned-spinner cleanup
- history.js: render JSON transcript with tool boxes, plaintext fallback

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
MelekiandClaude Opus 4.8 committed 2026-06-18 16:51:40 +02:00
1 parent aa0b8ee493
commit c32d1e004f
4 files changed
+211 -67

No files matched your search

+17 -1
View File
@@ -441,6 +441,14 @@ public partial class AnthropicService
/// Bei Fehler Exception werfen. Wenn null, wird Standard-Executor verwendet.
/// </summary>
public Func<string, JsonElement, CancellationToken, Task<string>> ExecuteTool { get; set; }
/// <summary>
/// Optionaler Transkript-Sammler (Panel-Member). Wenn gesetzt, sammelt der Loop
/// die interleavte Abfolge aus Text- und Tool-Segmenten, damit der Caller sie als
/// EINE strukturierte Message persistieren kann (Reload-Wiederherstellung der
/// Tool-Boxen). Segmente: {t:"text", v:...} bzw. {t:"tool", name, input, result, error}.
/// </summary>
public List<object> TranscriptSink { get; set; }
}
/// <summary>
@@ -570,6 +578,8 @@ public partial class AnthropicService
}
if (cb.SseResponse != null)
await SseHelper.TryWriteSse(cb.SseResponse, $"{cb.SsePrefix}_text", new { index = cb.MemberIndex, text, agentName = cb.AgentName }, cb.SseCt, cb.SseGate);
cb.TranscriptSink?.Add(new Dictionary<string, object> { ["t"] = "text", ["v"] = text });
}
else if (blockType == "tool_use")
{
@@ -628,7 +638,13 @@ public partial class AnthropicService
});
if (cb.SseResponse != null)
await SseHelper.TryWriteSse(cb.SseResponse, $"{cb.SsePrefix}_tool", new { index = cb.MemberIndex, name = toolName, result = TruncateToolResult(toolResult), isError, agentName = cb.AgentName }, cb.SseCt, cb.SseGate);
await SseHelper.TryWriteSse(cb.SseResponse, $"{cb.SsePrefix}_tool", new { index = cb.MemberIndex, name = toolName, input = inputEl.ToString(), result = TruncateToolResult(toolResult), isError, agentName = cb.AgentName }, cb.SseCt, cb.SseGate);
cb.TranscriptSink?.Add(new Dictionary<string, object>
{
["t"] = "tool", ["name"] = toolName, ["input"] = inputEl.ToString(),
["result"] = TruncateToolResult(toolResult), ["error"] = isError
});
}
}
+110 -47
View File
@@ -164,25 +164,30 @@ public partial class AnthropicService
// workerMessages oder parallele Member leckt.
var memberMessages = CloneForMember(workerMessages, previousLayerBlock);
string text;
List<object> segments = null; // interleavtes Transkript (Text + Tools) für die Reload-Persistenz
try
{
if (!string.IsNullOrEmpty(member.Provider) && _openAiService != null)
{
// OpenAI-kompatibler Member: tool-los (CompleteText), Proposer aus Kontext.
// Kein Tool-Transkript → unten als einzelnes Text-Segment persistiert.
var systemText = ExtractSystemPromptText(member.SystemPrompt);
text = await _openAiService.CompleteText(member.ModelId, systemText, memberMessages, apiToken);
}
else if (string.IsNullOrEmpty(apiToken) && _cliService != null && _cliService.IsAvailable)
{
// Kein API-Key → CLI-Member, read-only (Plan-Modus: nur Read/Glob/Grep).
// Streamt LIVE (Token + Tool-Uses) über panel_member_*-Events.
text = await RunPanelMemberViaCli(response, member, memberMessages, index, sseGate, workCt, sseCt);
// Streamt LIVE (Token + Tool-Uses) über panel_member_*-Events und liefert das Transkript.
var (t, segs) = await RunPanelMemberViaCli(response, member, memberMessages, index, sseGate, workCt, sseCt);
text = t; segments = segs;
}
else
{
// Anthropic-API-Member: read-only (PlanMode) + NICHT persistieren (Caller persistiert einmal).
// Live-Streaming über den Loop: pro Tool-Runde panel_member_text / panel_member_tool.
// TranscriptSink sammelt Text-/Tool-Segmente für die Reload-Persistenz.
var memberUsage = new TokenUsage();
segments = new List<object>();
var (t, _) = await RunNonStreamingToolLoop(
member.ModelId, member.SystemPrompt,
new ToolPayloadOptions
@@ -197,28 +202,42 @@ public partial class AnthropicService
{
Persist = false,
SseResponse = response, SseCt = sseCt, SseGate = sseGate,
SsePrefix = "panel_member", MemberIndex = index, AgentName = name
SsePrefix = "panel_member", MemberIndex = index, AgentName = name,
TranscriptSink = segments
},
member.Effort);
text = t;
}
}
catch (OperationCanceledException) { throw; }
catch (Exception ex) { text = $"[Fehler bei {name}: {ex.Message}]"; }
catch (Exception ex) { text = $"[Fehler bei {name}: {ex.Message}]"; segments = null; }
text = (text ?? "").Trim();
// Proposal als EINE Message persistieren: Tag="panel" → aus Worker-History gefiltert,
// [Panel: Name]-Prefix → Label beim History-Rendering. Iteration=index+1 → stabile Reihenfolge.
// Nur die FINALE Runde persistieren (MoA-Zwischenrunden sind live-only, überleben kein F5).
if (persist && !string.IsNullOrEmpty(text))
// Member-Transkript als EINE strukturierte Message persistieren (Tag="panel" → aus
// Worker-History gefiltert, Iteration=index+1 → stabile Reihenfolge). Content = JSON
// {name, model, segments:[{t:"text"|"tool",...}]} → der Reload stellt Text UND Tool-Boxen
// wieder her (genau wie live). Fällt auf ein einzelnes Text-Segment zurück (OpenAI-Member,
// Fehler oder wenn nichts gestreamt wurde). Nur die FINALE Runde (MoA-Zwischenrunden sind
// live-only und überleben kein F5).
if (persist)
{
await _db.SaveMessage(new MessageRecord
if (segments == null || segments.Count == 0)
{
ChainId = chainId, Iteration = index + 1,
Role = "assistant", Type = "text",
Content = $"[Panel: {name}]\n{text}", Tag = "panel"
});
segments = new List<object>();
if (!string.IsNullOrEmpty(text))
segments.Add(new Dictionary<string, object> { ["t"] = "text", ["v"] = text });
}
if (segments.Count > 0)
{
var content = JsonSerializer.Serialize(new { name, model = member.ModelId, segments });
await _db.SaveMessage(new MessageRecord
{
ChainId = chainId, Iteration = index + 1,
Role = "assistant", Type = "text",
Content = content, Tag = "panel"
});
}
}
await SseHelper.TryWriteSse(response, "panel_member_done",
@@ -235,7 +254,7 @@ public partial class AnthropicService
/// laufen durch den geteilten <paramref name="sseGate"/> (parallele Member → sonst
/// vermischen sich die Frames). Persistiert nichts – der Caller speichert eine Proposal-Message.
/// </summary>
private async Task<string> RunPanelMemberViaCli(
private async Task<(string text, List<object> segments)> RunPanelMemberViaCli(
HttpResponse response,
SubAgentContext member,
List<object> memberMessages,
@@ -244,48 +263,92 @@ public partial class AnthropicService
CancellationToken workCt,
CancellationToken sseCt)
{
if (_cliService == null || !_cliService.IsAvailable) return "";
if (_cliService == null || !_cliService.IsAvailable) return ("", new List<object>());
var systemText = ExtractSystemPromptText(member.SystemPrompt);
var prompt = ExtractLastUserText(memberMessages) ?? "";
// Bild-/Datei-Anhänge der letzten User-Message als Temp-Dateien ablegen, damit der
// CLI-Member sie per Read-Tool sehen kann (Screenshots, PDFs etc.) – exakt wie der Worker.
// Ohne das wären die Member bei bildbasierten Fragen blind → read-only Tools nutzlos.
var tempImagePaths = new List<string>();
var imageRefs = ExtractLastUserImageRefs(memberMessages, tempImagePaths);
if (!string.IsNullOrEmpty(imageRefs))
prompt = imageRefs + "\n" + prompt;
var historyContext = BuildHistoryContext(memberMessages);
if (!string.IsNullOrEmpty(historyContext))
prompt = $"<conversation_history>\n{historyContext}\n</conversation_history>\n\n<current_request>\n{prompt}\n</current_request>";
var fullText = new StringBuilder();
string currentToolName = null; // ToolResult trägt keinen Namen → vom ToolStart merken
await _cliService.ChatStreamingAsync(
prompt,
async evt =>
{
switch (evt.Type)
// Transkript-Segmente (Text + Tools interleaved) für die Reload-Persistenz.
var segments = new List<object>();
var segText = new StringBuilder();
void FlushText()
{
if (segText.Length == 0) return;
segments.Add(new Dictionary<string, object> { ["t"] = "text", ["v"] = segText.ToString() });
segText.Clear();
}
// Tool-Uses per ID korrelieren: Ein Member kann MEHRERE Tools in einer Assistant-Message
// aufrufen (ToolStart#1,#2,… ToolResult#1,#2,…). Ein einzelner "currentTool"-Zeiger würde
// dann Boxen vertauschen/verwaisen lassen. Außerdem ist der Tool-Input erst bei ToolEnd
// (content_block_stop) vollständig – ToolStart/ToolResult tragen ihn nicht.
var toolNames = new Dictionary<string, string>(); // toolId -> CLI-Name
var toolInputs = new Dictionary<string, string>(); // toolId -> Input-JSON
try
{
await _cliService.ChatStreamingAsync(
prompt,
async evt =>
{
case CliEventType.TextDelta:
fullText.Append(evt.Text);
await SseHelper.TryWriteSse(response, "panel_member_text",
new { index, text = evt.Text }, sseCt, sseGate);
break;
case CliEventType.ToolStart:
currentToolName = evt.ToolName;
await SseHelper.TryWriteSse(response, "panel_member_tool_use",
new { index, name = evt.ToolName }, sseCt, sseGate);
break;
case CliEventType.ToolResult:
await SseHelper.TryWriteSse(response, "panel_member_tool_end",
new { index, name = currentToolName, result = evt.ToolResultContent, isError = evt.ToolResultIsError }, sseCt, sseGate);
currentToolName = null;
break;
}
},
modelId: member.ModelId,
systemPrompt: systemText,
workspace: member.Workspace,
hasWorkspace: !string.IsNullOrEmpty(member.Workspace),
planMode: true,
effort: member.Effort,
cancellationToken: workCt
);
return fullText.ToString();
switch (evt.Type)
{
case CliEventType.TextDelta:
fullText.Append(evt.Text);
segText.Append(evt.Text);
await SseHelper.TryWriteSse(response, "panel_member_text",
new { index, text = evt.Text }, sseCt, sseGate);
break;
case CliEventType.ToolStart:
FlushText(); // Text-Segment vor der Tool-Box schließen (Reihenfolge wie live)
if (!string.IsNullOrEmpty(evt.ToolId)) toolNames[evt.ToolId] = evt.ToolName;
await SseHelper.TryWriteSse(response, "panel_member_tool_use",
new { index, id = evt.ToolId, name = evt.ToolName }, sseCt, sseGate);
break;
case CliEventType.ToolEnd:
// Input streamt erst nach ToolStart ein → hier vollständig. Für ToolResult merken.
if (!string.IsNullOrEmpty(evt.ToolId)) toolInputs[evt.ToolId] = evt.ToolInput;
break;
case CliEventType.ToolResult:
var rid = evt.ToolId ?? "";
toolNames.TryGetValue(rid, out var rName);
toolInputs.TryGetValue(rid, out var rInput);
segments.Add(new Dictionary<string, object>
{
["t"] = "tool", ["name"] = rName ?? "", ["input"] = rInput ?? "",
["result"] = TruncateToolResult(evt.ToolResultContent), ["error"] = evt.ToolResultIsError
});
await SseHelper.TryWriteSse(response, "panel_member_tool_end",
new { index, id = rid, name = rName, input = rInput, result = evt.ToolResultContent, isError = evt.ToolResultIsError }, sseCt, sseGate);
break;
}
},
modelId: member.ModelId,
systemPrompt: systemText,
workspace: member.Workspace,
hasWorkspace: !string.IsNullOrEmpty(member.Workspace),
planMode: true,
effort: member.Effort,
cancellationToken: workCt
);
}
finally
{
CleanupTempImages(tempImagePaths);
}
FlushText(); // restliches Text-Segment am Ende sichern
return (fullText.ToString(), segments);
}
/// <summary>Extrahiert den System-Prompt-Text aus den List&lt;object&gt;-Blöcken eines SubAgentContext.</summary>
+45 -2
View File
@@ -481,13 +481,56 @@ function _renderReviewGroup(tagGroup, archived) {
}
/**
* Renders a single panel member's proposal (tag="panel", content "[Panel: Name]\n...")
* as its own panel bubble – matches the live panel_member bubble.
* Renders a single panel member (tag="panel") as its own panel bubble – matches the live
* panel_member bubble. Two persisted formats:
* - NEW: JSON transcript {name, model, segments:[{t:"text",v} | {t:"tool",name,input,result,error}]}
* → restores text AND tool boxes interleaved (just like live).
* - OLD (back-compat): plaintext "[Panel: Name]\n..." → text only.
*/
function _renderPanelMemberBubble(msg, archived) {
var bubble = document.createElement('div');
bubble.className = 'msg review-bubble panel-bubble' + (archived ? ' archived' : '');
var content = msg.content || '';
// ── Neues Format: JSON-Transkript (Text + Tool-Boxen) ──
var data = null;
if (content.charAt(0) === '{') {
try { data = JSON.parse(content); } catch (e) { data = null; }
}
if (data && Array.isArray(data.segments)) {
var model = (typeof prettyModelName === 'function') ? prettyModelName(data.model) : data.model;
bubble.appendChild(_createSubAgentHeader('panel', data.name, model ? { model: model } : undefined));
var rendered = false;
for (var i = 0; i < data.segments.length; i++) {
var seg = data.segments[i];
if (!seg) continue;
if (seg.t === 'text') {
if (!seg.v) continue;
var tdiv = document.createElement('div');
tdiv.className = 'review-content markdown';
tdiv.innerHTML = renderMarkdown(seg.v);
parseFluentEmojis(tdiv);
bubble.appendChild(tdiv);
rendered = true;
} else if (seg.t === 'tool') {
if (typeof isToolVisible === 'function' && !isToolVisible(seg.name)) continue;
var box = document.createElement('div');
box.className = 'tool-box review-tool-box';
renderToolBoxContent(box, seg.name, seg.input || null, seg.result, seg.error, null);
bubble.appendChild(box);
rendered = true;
}
}
if (!rendered) {
var empty = document.createElement('div');
empty.className = 'review-content markdown';
empty.innerHTML = renderMarkdown('_(keine Antwort)_');
bubble.appendChild(empty);
}
return bubble;
}
// ── Alt-Format (Back-Compat): "[Panel: Name]\n..." Plaintext ──
var name = null;
var nameMatch = content.match(/^\[Panel: ([^\]]+)\]/);
if (nameMatch) name = nameMatch[1];
+39 -17
View File
@@ -247,7 +247,7 @@ function addPanelMemberBubble(data) {
bubble._panelSpinner = spinner; // wird beim ersten Text/Tool entfernt
bubble._segText = null; // aktuell offenes Text-Segment (oder null)
bubble._segTextStr = ''; // dessen akkumulierter Text
bubble._tool = null; // aktuell offene Tool-Box (CLI: Spinner→Füllen)
bubble._toolBoxes = {}; // toolId -> offene Tool-Box (CLI: per ID korreliert → robust bei parallelen Tools)
bubble._fullText = ''; // gesamter gestreamter Text (für Finalize-Fallback)
_panelActiveLayer[data.index] = layer; // ab jetzt zeigen Events für diesen index auf diese Runde
_panelBubbles[_panelKey(data.index, layer)] = bubble;
@@ -259,6 +259,11 @@ function _panelClearSpinner(bubble) {
if (bubble._panelSpinner) { bubble._panelSpinner.remove(); bubble._panelSpinner = null; }
}
function _panelRemoveCursor(bubble) {
var c = bubble.querySelector('.cursor');
if (c) c.remove();
}
function appendPanelMemberText(data) {
var bubble = _panelGetBubble(data.index);
if (!bubble || !data.text) return;
@@ -273,6 +278,12 @@ function appendPanelMemberText(data) {
bubble._segTextStr += data.text;
bubble._fullText += data.text;
bubble._segText.innerHTML = renderMarkdown(bubble._segTextStr);
// Blinkender Cursor ans Ende des Segments (wie beim Worker-Stream) → "schreibt gerade"
var cursor = document.createElement('span');
cursor.className = 'cursor';
var lastBlock = bubble._segText.lastElementChild;
if (lastBlock) lastBlock.appendChild(cursor);
else bubble._segText.appendChild(cursor);
scrollDown();
}
@@ -280,15 +291,17 @@ function addPanelMemberToolUse(data) {
var bubble = _panelGetBubble(data.index);
if (!bubble) return;
_panelClearSpinner(bubble);
bubble._segText = null; // Text-Segment schließen → Folge-Text kommt UNTER die Tool-Box
if (!isToolVisible(data.name)) { bubble._tool = null; return; }
_panelRemoveCursor(bubble); // Cursor aus dem laufenden Text-Segment entfernen
bubble._segText = null; // Text-Segment schließen → Folge-Text kommt UNTER die Tool-Box
var nm = normalizeToolName(data.name); // CLI-Name (Bash) → gott-Konvention (execute) – konsistent zur fertigen Box
if (!isToolVisible(nm)) return;
var toolDiv = document.createElement('div');
toolDiv.className = 'tool-box review-tool-box streamed';
toolDiv.dataset.toolName = data.name;
var displayName = TOOL_DISPLAY_NAMES[data.name] || data.name;
toolDiv.dataset.toolName = nm;
var displayName = TOOL_DISPLAY_NAMES[nm] || nm;
toolDiv.innerHTML = '<span class="tool-name">' + esc(displayName) + '</span> <span class="tool-spinner"><i class="fa fa-hourglass-half"></i></span>';
bubble.appendChild(toolDiv);
bubble._tool = toolDiv;
if (data.id) bubble._toolBoxes[data.id] = toolDiv; // per ID korreliert (mehrere Tools parallel)
scrollDown();
}
@@ -297,18 +310,17 @@ function finishPanelMemberTool(data) {
if (!bubble) return;
_panelClearSpinner(bubble);
bubble._segText = null; // nach Tool-Result neuen Text-Block beginnen
if (bubble._tool) {
// CLI: offene Box (mit Spinner) füllen
renderToolBoxContent(bubble._tool, data.name, null, data.result, data.isError, null);
bubble._tool = null;
} else {
// Anthropic-Loop (kombiniertes panel_member_tool-Event): fertige Box direkt anlegen
if (!isToolVisible(data.name)) return;
var toolDiv = document.createElement('div');
toolDiv.className = 'tool-box review-tool-box streamed';
bubble.appendChild(toolDiv);
renderToolBoxContent(toolDiv, data.name, null, data.result, data.isError, null);
// CLI: offene Box per ID finden; Anthropic-Loop (kombiniert, ohne ID) → neue Box anlegen
var box = (data.id && bubble._toolBoxes && bubble._toolBoxes[data.id]) ? bubble._toolBoxes[data.id] : null;
if (!box) {
if (!isToolVisible(normalizeToolName(data.name))) return;
box = document.createElement('div');
box.className = 'tool-box review-tool-box streamed';
bubble.appendChild(box);
}
// data.name = roher Name, data.input = roher Input-JSON → renderToolBoxContent normalisiert beides
renderToolBoxContent(box, data.name, data.input || null, data.result, data.isError, null);
if (data.id && bubble._toolBoxes) delete bubble._toolBoxes[data.id];
scrollDown();
}
@@ -316,6 +328,16 @@ function finalizePanelMember(data) {
var bubble = _panelGetBubble(data.index);
if (!bubble) return;
_panelClearSpinner(bubble);
_panelRemoveCursor(bubble);
// Verwaiste Tool-Boxen (Tool ohne Result, z.B. in Plan-Mode geblockt) → Spinner entfernen,
// damit keine Box mit endlosem "⏳" stehen bleibt.
if (bubble._toolBoxes) {
for (var tid in bubble._toolBoxes) {
var sp = bubble._toolBoxes[tid].querySelector('.tool-spinner');
if (sp) sp.remove();
}
bubble._toolBoxes = {};
}
// Wenn LIVE nichts gestreamt wurde (z.B. OpenAI-Member = one-shot): finalen Text rendern.
if (!bubble._fullText || !bubble._fullText.trim()) {
var text = (data.text && data.text.trim()) ? data.text : '_(keine Antwort)_';