Better file upload

docs/ARCHITECTURE.md

@@ -116,6 +116,7 @@ codex exec --json -m {model} -C {cwd} --skip-git-repo-check \
 codex exec resume {sessionId} --json -m {model} --skip-git-repo-check "{prompt}"
 ```
 (`exec resume` does not use `-C` or `-s`.)
+Image attachments are passed via `-i /path/to/image`; non-image attachments are appended to the prompt as readable file paths.
 
 **Ollama Provider:** Stateless HTTP requests to Ollama API:
 - POST to `/api/chat` with full message history
@@ -241,6 +242,8 @@ Conversations default to sandboxed mode for safety. Sandbox configuration:
 | `GET` | `/api/files/content` | Get structured file content (standalone cwd) |
 | `GET` | `/api/files/download` | Download file |
 | `POST` | `/api/files/upload` | Upload file |
+| `POST` | `/api/conversations/:id/upload` | Upload local file as conversation attachment |
+| `POST` | `/api/conversations/:id/attachments/from-files` | Copy existing cwd file(s) into conversation attachments |
 | `GET` | `/api/conversations/:id/files` | List files in cwd |
 | `GET` | `/api/conversations/:id/files/content` | Get file content |
 | `GET` | `/api/conversations/:id/files/search` | Git grep search |
@@ -371,7 +374,7 @@ public/js/
 Five mutually exclusive views with CSS transform transitions:
 
 1. **List View** — Conversation browser grouped by cwd, search (keyword + semantic), archive toggle
-2. **Chat View** — Messages, input bar, file panel with preview
+2. **Chat View** — Messages, input bar, file panel with preview, and direct file-to-chat attachment from file tree/viewer actions
 3. **Stats View** — Analytics dashboard with cost tracking, activity charts
 4. **Branches View** — Fork tree visualization with parent/child navigation
 5. **Memory View** — Memory management (global + project-scoped)

docs/REFERENCE.md

@@ -126,7 +126,7 @@ public/css/
   role: 'user',
   text: string,
   timestamp: number,
-  attachments: [{ filename, url }]  // optional
+  attachments: [{ path, filename, url }]  // optional
 }
 ```
 
@@ -139,8 +139,13 @@ public/css/
   cost: number,           // USD
   duration: number,       // ms
   sessionId: string,
-  inputTokens: number,
-  outputTokens: number
+  inputTokens: number,    // Provider-reported input token count (raw when available)
+  outputTokens: number,
+  netInputTokens?: number,    // rawInputTokens - cachedInputTokens
+  typedInputTokens?: number,  // rough estimate of user-typed prompt tokens
+  rawInputTokens?: number,    // full provider-reported prompt/context tokens
+  cachedInputTokens?: number, // provider cache hit tokens (if available)
+  reasoningTokens?: number    // provider-reported reasoning/thinking tokens
 }
 ```
 
@@ -235,7 +240,7 @@ public/css/
 | `loadConversations()` | conversations.js | Fetch + render list |
 | `openConversation(id)` | conversations.js | Load + display conversation |
 | `forkConversation(idx)` | conversations.js | Fork from message |
-| `sendMessage(text)` | ui.js | Send with attachments |
+| `sendMessage(text)` | ui.js | Send with local uploads or pre-attached server files |
 | `renderMarkdown(text)` | markdown.js | Markdown → HTML |
 
 ---

lib/providers/claude.js

@@ -117,6 +117,10 @@ function calculateMessageCost(inputTokens, outputTokens, modelId) {
   return inputCost + outputCost;
 }
 
+function estimateTypedInputTokens(text) {
+  return text ? Math.max(1, Math.ceil(String(text).length / 4)) : 0;
+}
+
 // Active Claude processes per conversation
 const activeProcesses = new Map();
 const PROCESS_TIMEOUT = 5 * 60 * 1000; // 5 minutes
@@ -190,6 +194,7 @@ function buildInlineHistoryContext(messages = [], latestUserText = '', options =
   let totalChars = 0;
   for (let i = priorMessages.length - 1; i >= 0; i--) {
     const msg = priorMessages[i];
+    if (msg?.summarized) continue;
     let text = String(msg?.text || '').trim();
     for (const replacement of pathReplacements) {
       text = text.split(replacement.from).join(replacement.to);
@@ -376,8 +381,20 @@ function handleResultEvent(
     conv.claudeForkSessionId = null;
   }
 
-  const inputTokens = event.total_input_tokens ?? event.input_tokens ?? event.usage?.input_tokens ?? 0;
+  const rawInputTokens = event.total_input_tokens ?? event.input_tokens ?? event.usage?.input_tokens ?? 0;
+  const cachedInputTokens =
+    event.cache_read_input_tokens ??
+    event.usage?.cache_read_input_tokens ??
+    event.usage?.input_tokens_details?.cached_tokens ??
+    0;
+  const netInputTokens = Math.max(0, rawInputTokens - cachedInputTokens);
+  const inputTokens = rawInputTokens;
+  const typedInputTokens = conv._typedInputTokens ?? null;
   const outputTokens = event.total_output_tokens ?? event.output_tokens ?? event.usage?.output_tokens ?? 0;
+  const reasoningTokens =
+    event.usage?.output_tokens_details?.reasoning_tokens ??
+    event.usage?.completion_tokens_details?.reasoning_tokens ??
+    0;
   if (!resultText.trim() && inputTokens === 0 && outputTokens === 0) {
     if (canRetryWithFreshSession) {
       // Resumed sessions can occasionally become invalid and return empty 0/0.
@@ -417,7 +434,12 @@ function handleResultEvent(
     duration: event.duration_ms,
     sessionId: event.session_id,
     inputTokens,
+    netInputTokens,
+    typedInputTokens,
     outputTokens,
+    reasoningTokens,
+    rawInputTokens,
+    cachedInputTokens,
   });
   conv.status = 'idle';
   conv.thinkingStartTime = null;
@@ -437,8 +459,14 @@ function handleResultEvent(
     duration: event.duration_ms,
     sessionId: event.session_id,
     inputTokens,
+    netInputTokens,
+    typedInputTokens,
     outputTokens,
+    reasoningTokens,
+    rawInputTokens,
+    cachedInputTokens,
   });
+  delete conv._typedInputTokens;
   broadcastStatus(conversationId, 'idle');
 }
 
@@ -542,6 +570,7 @@ class ClaudeProvider extends LLMProvider {
     const promptText = inlineHistory
       ? `${inlineHistory}\n\n[New user message]\n${text}`
       : text;
+    conv._typedInputTokens = estimateTypedInputTokens(text);
 
     const args = [
       '-p', promptText,
@@ -741,6 +770,7 @@ Please continue the conversation naturally, using the above context as backgroun
       ) {
         retryScheduled = true;
         delete conv._retryAfterEmptyResultMode;
+        delete conv._typedInputTokens;
         void this.chat(ws, conversationId, conv, text, attachments, uploadDir, callbacks, memories, {
           ...runtime,
           retried: true,
@@ -776,6 +806,7 @@ Please continue the conversation naturally, using the above context as backgroun
         isSlashOnlyPrompt,
       });
       delete conv._retryAfterEmptyResultMode;
+      delete conv._typedInputTokens;
     });
 
     proc.on('error', (err) => {
@@ -784,6 +815,7 @@ Please continue the conversation naturally, using the above context as backgroun
 
       if (!retryScheduled && conv.status === 'thinking' && canRetryWithCompactHistory && isContextOverflowError(err?.message || err?.code)) {
         retryScheduled = true;
+        delete conv._typedInputTokens;
         void this.chat(ws, conversationId, conv, text, attachments, uploadDir, callbacks, memories, {
           ...runtime,
           retried: true,
@@ -801,6 +833,7 @@ Please continue the conversation naturally, using the above context as backgroun
       });
       broadcastStatus(conversationId, 'idle');
       delete conv._retryAfterEmptyResultMode;
+      delete conv._typedInputTokens;
     });
   }
 

lib/providers/codex.js

@@ -54,6 +54,10 @@ function calculateMessageCost(inputTokens, outputTokens, modelId) {
   return inputCost + outputCost;
 }
 
+function estimateTypedInputTokens(text) {
+  return text ? Math.max(1, Math.ceil(String(text).length / 4)) : 0;
+}
+
 // Active Codex processes per conversation
 const activeProcesses = new Map();
 const PROCESS_TIMEOUT = 5 * 60 * 1000; // 5 minutes
@@ -112,6 +116,30 @@ function formatMemoriesForPrompt(memories) {
   return `\n\n[User memories]\n${sections.join('\n\n')}\n[/User memories]`;
 }
 
+function partitionAttachments(attachments = []) {
+  const imageAttachments = [];
+  const fileAttachments = [];
+
+  for (const attachment of attachments || []) {
+    const filePath = attachment?.path;
+    if (!filePath || typeof filePath !== 'string') continue;
+
+    if (/\.(png|jpg|jpeg|gif|webp)$/i.test(filePath)) {
+      imageAttachments.push(attachment);
+    } else {
+      fileAttachments.push(attachment);
+    }
+  }
+
+  return { imageAttachments, fileAttachments };
+}
+
+function buildFileAttachmentPrompt(fileAttachments = []) {
+  if (!Array.isArray(fileAttachments) || fileAttachments.length === 0) return '';
+  const paths = fileAttachments.map((a) => a.path).join('\n');
+  return `\n\n[Attached file${fileAttachments.length > 1 ? 's' : ''} — read for context:]\n${paths}`;
+}
+
 function buildInlineHistoryContext(messages = [], latestUserText = '', options = {}) {
   if (!Array.isArray(messages) || messages.length === 0) return '';
   const maxChars = Number.isFinite(Number(options.maxChars)) && Number(options.maxChars) > 0
@@ -128,6 +156,7 @@ function buildInlineHistoryContext(messages = [], latestUserText = '', options =
   let totalChars = 0;
   for (let i = priorMessages.length - 1; i >= 0; i--) {
     const msg = priorMessages[i];
+    if (msg?.summarized) continue;
     const text = String(msg?.text || '').trim();
     if (!text) continue;
 
@@ -496,8 +525,14 @@ function processCodexEvent(
         0;
       // For UX, show net new input tokens instead of full cached prompt tokens.
       const inputTokens = Math.max(0, rawInputTokens - cachedInputTokens);
+      const netInputTokens = inputTokens;
       const displayInputTokens = conv._codexDisplayInputTokens ?? inputTokens;
+      const typedInputTokens = conv._typedInputTokens ?? displayInputTokens;
       const outputTokens = usage.output_tokens || 0;
+      const reasoningTokens =
+        usage.output_tokens_details?.reasoning_tokens ??
+        usage.completion_tokens_details?.reasoning_tokens ??
+        0;
       if (!assistantText.trim() && inputTokens === 0 && outputTokens === 0) {
         if (canRetryWithFreshSession) {
           // Resumed sessions can occasionally become invalid and return empty 0/0.
@@ -522,6 +557,7 @@ function processCodexEvent(
           error: 'Model returned an empty response. Please retry.',
         });
         delete conv._codexDisplayInputTokens;
+        delete conv._typedInputTokens;
         broadcastStatus(conversationId, 'idle');
         break;
       }
@@ -538,8 +574,11 @@ function processCodexEvent(
         duration: event.duration_ms,
         sessionId: conv.codexSessionId,
         inputTokens,
+        netInputTokens,
         displayInputTokens,
+        typedInputTokens,
         outputTokens,
+        reasoningTokens,
         rawInputTokens,
         cachedInputTokens,
       });
@@ -555,12 +594,16 @@ function processCodexEvent(
         duration: event.duration_ms,
         sessionId: conv.codexSessionId,
         inputTokens,
+        netInputTokens,
         displayInputTokens,
+        typedInputTokens,
         outputTokens,
+        reasoningTokens,
         rawInputTokens,
         cachedInputTokens,
       });
       delete conv._codexDisplayInputTokens;
+      delete conv._typedInputTokens;
       broadcastStatus(conversationId, 'idle');
       break;
     }
@@ -623,14 +666,15 @@ class CodexProvider extends LLMProvider {
       args.push('-s', allowWrites ? 'workspace-write' : 'read-only');
     }
 
+    const { imageAttachments, fileAttachments } = partitionAttachments(attachments);
+
     // Image attachments
-    const images = (attachments || []).filter(a => /\.(png|jpg|jpeg|gif|webp)$/i.test(a.path));
-    for (const img of images) {
+    for (const img of imageAttachments) {
       args.push('-i', img.path);
     }
 
     // Grant access to uploads directory
-    if (!isResume && attachments && attachments.length > 0) {
+    if (!isResume && (imageAttachments.length > 0 || fileAttachments.length > 0)) {
       args.push('--add-dir', path.join(uploadDir, conversationId));
     }
 
@@ -651,8 +695,11 @@ class CodexProvider extends LLMProvider {
     const promptBase = inlineHistory
       ? `${inlineHistory}\n\n[New user message]\n${text}`
       : text;
-    const prompt = promptBase + formatMemoriesForPrompt(enabledMemories);
-    conv._codexDisplayInputTokens = text ? Math.max(1, Math.ceil(text.length / 4)) : 0;
+    const prompt = promptBase
+      + buildFileAttachmentPrompt(fileAttachments)
+      + formatMemoriesForPrompt(enabledMemories);
+    conv._typedInputTokens = estimateTypedInputTokens(text);
+    conv._codexDisplayInputTokens = conv._typedInputTokens;
 
     // Prompt must be last
     args.push(prompt);
@@ -758,6 +805,7 @@ class CodexProvider extends LLMProvider {
         delete conv._codexOpenToolIds;
         delete conv._codexOpenTraceCount;
         delete conv._codexDisplayInputTokens;
+        delete conv._typedInputTokens;
         void this.chat(ws, conversationId, conv, text, attachments, uploadDir, callbacks, memories, {
           ...runtime,
           retried: true,
@@ -797,6 +845,7 @@ class CodexProvider extends LLMProvider {
       delete conv._codexOpenToolIds;
       delete conv._codexOpenTraceCount;
       delete conv._codexDisplayInputTokens;
+      delete conv._typedInputTokens;
     });
 
     proc.on('error', (err) => {
@@ -808,6 +857,7 @@ class CodexProvider extends LLMProvider {
         delete conv._codexOpenToolIds;
         delete conv._codexOpenTraceCount;
         delete conv._codexDisplayInputTokens;
+        delete conv._typedInputTokens;
         void this.chat(ws, conversationId, conv, text, attachments, uploadDir, callbacks, memories, {
           ...runtime,
           retried: true,
@@ -828,6 +878,7 @@ class CodexProvider extends LLMProvider {
       delete conv._codexOpenToolIds;
       delete conv._codexOpenTraceCount;
       delete conv._codexDisplayInputTokens;
+      delete conv._typedInputTokens;
     });
   }
 
@@ -951,3 +1002,8 @@ module.exports.MODELS = MODELS;
 module.exports.activeProcesses = activeProcesses;
 module.exports.processCodexEvent = processCodexEvent;
 module.exports.handleNoOutputClose = handleNoOutputClose;
+module.exports.partitionAttachments = partitionAttachments;
+module.exports.buildFileAttachmentPrompt = buildFileAttachmentPrompt;
+module.exports._private = {
+  buildInlineHistoryContext,
+};

lib/providers/ollama.js

@@ -11,6 +11,10 @@ const OLLAMA_HOST = process.env.OLLAMA_HOST || 'http://localhost:11434';
 // Active requests per conversation (for cancellation)
 const activeRequests = new Map();
 
+function estimateTypedInputTokens(text) {
+  return text ? Math.max(1, Math.ceil(String(text).length / 4)) : 0;
+}
+
 // Default models if Ollama is not available
 const DEFAULT_MODELS = [
   { id: 'llama3.2', name: 'Llama 3.2', context: 128000 },
@@ -122,6 +126,7 @@ ${summaryMsg.text}
    */
   async chat(ws, conversationId, conv, text, attachments, uploadDir, callbacks, memories = []) {
     const { onSave, broadcastStatus } = callbacks;
+    conv._typedInputTokens = estimateTypedInputTokens(text);
 
     // Note: Ollama doesn't support file attachments natively
     // We could read file contents and append to the message, but skip for now
@@ -198,7 +203,11 @@ ${summaryMsg.text}
                 cost: 0, // Ollama is free/local
                 duration,
                 inputTokens: json.prompt_eval_count || 0,
+                netInputTokens: json.prompt_eval_count || 0,
+                typedInputTokens: conv._typedInputTokens ?? null,
                 outputTokens: json.eval_count || 0,
+                rawInputTokens: json.prompt_eval_count || 0,
+                cachedInputTokens: 0,
               });
               conv.status = 'idle';
               conv.thinkingStartTime = null;
@@ -211,7 +220,11 @@ ${summaryMsg.text}
                 cost: 0,
                 duration,
                 inputTokens: json.prompt_eval_count || 0,
+                netInputTokens: json.prompt_eval_count || 0,
+                typedInputTokens: conv._typedInputTokens ?? null,
                 outputTokens: json.eval_count || 0,
+                rawInputTokens: json.prompt_eval_count || 0,
+                cachedInputTokens: 0,
               });
               broadcastStatus(conversationId, 'idle');
             }
@@ -249,6 +262,7 @@ ${summaryMsg.text}
       broadcastStatus(conversationId, 'idle');
     } finally {
       activeRequests.delete(conversationId);
+      delete conv._typedInputTokens;
     }
   }