René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"BUG：sometime new message and old message all in same id","articleBody":"\nAn occasional bug in the agentapi caused old message history and new user messages to be mixed together in the same SSE message body.\n\ndata: {\"id\":68,\"role\":\"agent\",\"message\":\"old + new\"}\n\nNote!\n\nThe \"old + new\" above\n\n\"old\" refers to a very, very large collection of historical messages, not a streaming update of the screen.\n\n\n\n![Image](https://github.com/user-attachments/assets/cacb5a86-3018-48ad-8cd8-c9c7fc99dbc9)\n\nBug 1: A \"ghost update\" at the start of sendMessage\n\n  Before appending the user message, sendMessage calls updateLastAgentMessageLocked once:\n\n  // pty_conversation.go:386-391\n  c.lock.Lock()\n  screenBeforeMessage := c.cfg.AgentIO.ReadScreen()\n  now := c.cfg.Clock.Now()\n  c.updateLastAgentMessageLocked(screenBeforeMessage, now)  // ← user msg NOT yet appended\n  c.writingMessage = true\n  c.lock.Unlock()\n\n  At this point c.messages still ends with the agent message (id=68), so:\n\n  // line 324\n  shouldCreateNewMessage := len(c.messages) == 0 ||\n      c.messages[len(c.messages)-1].Role == ConversationRoleUser\n  // = false  ← overwrites existing slot instead of creating a new one\n\n  Any content change here writes into messages[68] — no new ID is created.\n\n  ---\n  Bug 2: screenDiff set-based matching bleeds historical content in\n\n  // diff.go:32-38\n  firstNonMatchingLine := len(newLines)\n  for i, line := range newLines[dynamicHeaderEnd+1:] {\n      if !oldLinesMap[line] {      // ← set lookup by line content\n          firstNonMatchingLine = i\n          break\n      }\n  }\n  newSectionLines := newLines[firstNonMatchingLine:]  // ← takes EVERYTHING from that line onward\n\n  oldLinesMap is a content-based set of old screen lines (unordered). The algorithm finds the first line not in that set, then takes all subsequent lines — including any historical content that reappears later on screen.\n\n  When the terminal accumulates output without clearing, the new screen looks like:\n\n  [historical response lines]   ← some line here is NOT in oldLinesMap\n  [new user input echo]\n  [new agent response lines]\n\n  So screenDiff returns \"historical_content\\nnew_content\" as a single blob.\n\n  ---\n  How the two bugs combine\n\n  ┌─────────────────────────────────────────────────────────────┬──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┬─────────────────────────────────────────────┐\n  │                            Step                             │                                                        State                                                         │                   Effect                    │\n  ├─────────────────────────────────────────────────────────────┼──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┼─────────────────────────────────────────────┤\n  │ sendMessage starts, calls updateLastAgentMessageLocked      │ writingMessage=false, last message is agent (id=68), shouldCreateNewMessage=false                                    │ Any content difference overwrites id=68     │\n  ├─────────────────────────────────────────────────────────────┼──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┼─────────────────────────────────────────────┤\n  │ screenDiff(screenBeforeLastUserMessage, currentScreen) runs │ New screen contains old response lines + new response lines; set-match finds a boundary in the middle of old content │ Returns \"historical + new content\"          │\n  ├─────────────────────────────────────────────────────────────┼──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┼─────────────────────────────────────────────┤\n  │ Write path                                                  │ shouldCreateNewMessage=false → takes else branch: messages[last] = conversationMessage                               │ id=68 now contains both old and new content │\n  └─────────────────────────────────────────────────────────────┴──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┴─────────────────────────────────────────────┘\n\n  Only after this does the user message get appended (id=69), and the snapshot loop eventually creates a proper new agent message (id=70). But id=68 has already been corrupted and emitted via SSE.\n\n  ---\n  Fix directions\n\n  Fix 1 — Remove the redundant updateLastAgentMessageLocked call at the start of sendMessage. The screen is stable at this point (that's why stableSignal fired), so this call produces no useful work:\n\n  // In sendMessage, remove this line from the first lock block:\n  - c.updateLastAgentMessageLocked(screenBeforeMessage, now)\n\n  Fix 2 — Replace the set-based screenDiff with a positional diff. Instead of \"find the first line whose content isn't in a set\", compare line-by-line from the end of the old screen forward, so historical content that reappears mid-screen doesn't shift\n  the boundary:\n\n  // Current (buggy): content-set lookup\n  if !oldLinesMap[line] { firstNonMatchingLine = i; break }\n\n  // Better: compare the tail of newLines against the tail of oldLines positionally,\n  // or use a suffix/LCS approach so repeated content doesn't cause false boundaries.","author":{"url":"https://github.com/mrchanglinux-spec","@type":"Person","name":"mrchanglinux-spec"},"datePublished":"2026-03-20T16:10:25.000Z","interactionStatistic":{"@type":"InteractionCounter","interactionType":"https://schema.org/CommentAction","userInteractionCount":3},"url":"https://github.com/205/agentapi/issues/205"}