GitHub Issue

Title: WebSocket delta corruption with inline directive tags

Labels: bug, gateway, websocket

Bug Description

WebSocket delta and final messages are corrupted when using inline directive tags (e.g., [[reply_to_current]]) at the beginning of LLM output. Characters are truncated due to leading whitespace mismatch between Pi embedded and Gateway processing.

Observed Behavior

Example corruption:

• Expected: www.6.xumum.xyz

• Received: www.6.xum.xyz (missing um)

• Expected: GOOGLE

• Received: GOGLE (missing O)

Root Cause

Leading whitespace mismatch:

1. Pi embedded (normalizeDirectiveWhitespace): ✅ Trims leading/trailing whitespace

   // src/utils/directive-tags.ts:40-65
   const normalized = masked
     .replace(/^\n+/, "")
     .replace(/^[ \t](?=\S)/, "")  // Removes leading space
     .trimEnd();

2. Gateway (stripInlineDirectiveTagsForDisplay in emitChatDelta): ❌ Does NOT trim

   // src/gateway/server-chat.ts:691-693
   const cleanedText = stripInlineDirectiveTagsForDisplay(text).text;  // ❌ No .trim()
   const cleanedDelta = typeof delta === "string" 
     ? stripInlineDirectiveTagsForDisplay(delta).text  // ❌ No .trim()
     : "";

3. Consequence: previousText.startsWith(nextText) fails → appendUniqueSuffix triggers false overlap detection → character truncation

Corruption Mechanism

LLM output: "[[reply_to_current]] Domain www.6.xumum.xyz"
  ↓
Pi embedded sends (trimmed): "Domain www.6.xumum.xyz" (no leading space)
  ↓
Gateway caches (not trimmed): " Domain www.6.xumum.xyz" (has leading space)
  ↓
Next delta: startsWith() check fails (leading space mismatch)
  ↓
appendUniqueSuffix: detects false overlap "um" at boundary → truncates
  ↓
Corruption: "www.6.xum.xyz" ❌

Impact

• ❌ WebSocket delta messages corrupted
• ❌ WebSocket final messages corrupted (reads from corrupted buffer)
• ✅ Session files correct (different code path already uses .trim())

Trigger Conditions

1. Agent config: replyToMode: "all" or "threaded" (default)
2. LLM outputs inline directive tag at beginning (e.g., [[reply_to_current]], [[audio_as_voice]])
3. Client consumes WebSocket delta/final messages
4. Content contains repeated character sequences (triggers false overlap detection)

Evidence

Session file (correct data)

Remote session file on production Gateway contains correct content:

"content": "...www.6.xumum.xyz...GOOGLE..."

Java client WebSocket log (corrupted data)

{
  "type": "event",
  "event": "chat",
  "payload": {
    "state": "final",
    "message": {
      "content": [{"type": "text", "text": "...www.6.xum.xyz...GOGLE..."}]
    }
  }
}

Code comparison

✅ Correct usage (already using .trim()):

• src/gateway/session-utils.fs.ts:256 - .text.trim()
• src/gateway/session-utils.fs.ts:267 - .text.trim()
• extensions/discord/src/voice/sanitize.ts:20 - .text.trim()

❌ Missing .trim() (bug location):

• src/gateway/server-chat.ts:691 - emitChatDelta cleanedText
• src/gateway/server-chat.ts:693 - emitChatDelta cleanedDelta

Proposed Fix

Add .trim() to match other code paths:

--- a/src/gateway/server-chat.ts
+++ b/src/gateway/server-chat.ts
@@ -688,9 +688,9 @@ export function createAgentEventHandler({
     text: string,
     delta?: unknown,
   ) => {
-    const cleanedText = stripInlineDirectiveTagsForDisplay(text).text;
+    const cleanedText = stripInlineDirectiveTagsForDisplay(text).text.trim();
     const cleanedDelta =
-      typeof delta === "string" ? stripInlineDirectiveTagsForDisplay(delta).text : "";
+      typeof delta === "string" ? stripInlineDirectiveTagsForDisplay(delta).text.trim() : "";
     const previousRawText = chatRunState.rawBuffers.get(clientRunId) ?? "";

Benefits:

• ✅ Aligns with session-utils.fs.ts and Discord extension
• ✅ Fixes delta and final message corruption
• ✅ Single-point change, minimal impact
• ✅ Backward compatible (trim doesn't change semantics)

Reproduction Steps

1. Configure agent with replyToMode: "all"
2. Send message: "Please respond with: domain www.6.xumum.xyz, search engine GOOGLE"
3. Monitor WebSocket event: "chat" with state: "final"
4. Observe character loss: xumum → xum, GOOGLE → GOGLE

Environment

• OpenClaw version: Latest main branch (commit 82364e901a as of 2026-04-14)
• Gateway: Production instance
• Client: Java WebSocket client monitoring delta/final events
• Agent: Site-builder with default replyToMode: "all"

Why This Bug Is Rare

1. Requires LLM to output inline directive tag at beginning
2. Requires client to consume WebSocket delta/final (most clients read session files, which are correct)
3. Requires repeated character sequences to trigger false overlap detection
4. Test coverage gap for inline directive + delta combination

Related Files

• Bug location: src/gateway/server-chat.ts:691-693
• Related: src/utils/directive-tags.ts (trim behavior)
• Correct examples: src/gateway/session-utils.fs.ts:256, extensions/discord/src/voice/sanitize.ts:20