fix: preserve XML-like tags in write_file/execute_code/patch content (#72797)

agent/agent_runtime_helpers.py

@@ -800,11 +800,15 @@ def strip_think_blocks(agent, content: str) -> str:
     #     a literal <tool_call> in prose is already vanishingly rare.
     for _tc_name in ("tool_call", "tool_calls", "tool_result",
                       "function_call", "function_calls"):
+        # NOTE: intentionally does NOT use re.IGNORECASE — models emit
+        # these tags in lower case only, and case-insensitive matching
+        # corrupts data content (e.g. JS/HTML that contains `<TOOL_CALL>`
+        # literal strings). See #72797.
         content = re.sub(
             rf'<{_tc_name}\b[^>]*>.*?</{_tc_name}>',
             '',
             content,
-            flags=re.DOTALL | re.IGNORECASE,
+            flags=re.DOTALL,
         )
     # 1c. <function name="...">...</function> — Gemma-style standalone
     #     tool call. Only strip when the tag sits at a block boundary
@@ -840,11 +844,12 @@ def strip_think_blocks(agent, content: str) -> str:
     #     unterminated <function name="..."> because a truncated tail
     #     during streaming may still be valuable to the user; matches
     #     OpenClaw's intentional asymmetry.)
+    # NOTE: intentionally does NOT use re.IGNORECASE — same rationale
+    # as the block patterns above (#72797).
     content = re.sub(
         r'</(?:tool_call|tool_calls|tool_result|function_call|function_calls|function)>\s*',
         '',
         content,
-        flags=re.IGNORECASE,
     )
     return content
 

cli.py

@@ -284,13 +284,17 @@ def _strip_reasoning_tags(text: str) -> str:
             flags=re.IGNORECASE,
         )
     # Tool-call XML blocks (openclaw/openclaw#67318).
+    # NOTE: intentionally does NOT use re.IGNORECASE — models emit
+    # these tags in lower case only, and case-insensitive matching
+    # corrupts data content (e.g. JS/HTML that contains `<TOOL_CALL>`
+    # literal strings). See #72797.
     for tc_tag in ("tool_call", "tool_calls", "tool_result",
                    "function_call", "function_calls"):
         cleaned = re.sub(
             rf"<{tc_tag}\b[^>]*>.*?</{tc_tag}>\s*",
             "",
             cleaned,
-            flags=re.DOTALL | re.IGNORECASE,
+            flags=re.DOTALL,
         )
     # <function name="..."> — boundary + attribute gated to avoid prose FPs.
     cleaned = re.sub(
@@ -302,11 +306,12 @@ def _strip_reasoning_tags(text: str) -> str:
         flags=re.DOTALL | re.IGNORECASE,
     )
     # Stray tool-call close tags.
+    # NOTE: intentionally does NOT use re.IGNORECASE — same rationale
+    # as the generic tool-call tag patterns above (#72797).
     cleaned = re.sub(
         r'</(?:tool_call|tool_calls|tool_result|function_call|function_calls|function)>\s*',
         '',
         cleaned,
-        flags=re.IGNORECASE,
     )
     return cleaned.strip()
 

plugins/tool-escalator/README.md

@@ -0,0 +1,53 @@
+# tool-escalator
+
+Auto-escalate to MoA (Mixture of Agents) when consecutive tool errors are detected.
+
+## How it works
+
+The plugin watches three lifecycle hooks to detect, escalate, and de-escalate:
+
+```
+T1 (your daily model, unchanged)
+└─ on N (default 3) consecutive tool errors →
+T2 (MoA preset — user selects manually via /moa or /model)
+└─ after MoA aggregation completes → auto-de-escalate logging
+```
+
+### Three hook callbacks
+
+| Hook | What it does |
+|------|-------------|
+| `post_tool_call` | Checks every tool result for error indicators (`"error"`, `"failed"`, `Error:`-prefixed lines, non-zero exit codes). Increments a session-scoped consecutive-error counter on failure; resets to zero on success. At threshold (default 3), logs an escalation decision and sets a session flag. |
+| `pre_llm_call` | Checks the escalation flag and injects context into the user message when escalation is active, nudging the model or user to switch to MoA. Also detects ongoing MoA calls for de-escalation tracking. |
+| `post_llm_call` | Detects MoA completion (via the pre_llm_call MoA-active marker) and clears the escalation flag, logging the de-escalation. |
+
+### Error detection
+
+Pragmatic pattern matching on the string representation of every tool result:
+
+- Substring match for `"error"`, `"failed"`, `"failure"`, `"exception"`, `"traceback"`, `"timeout"`
+- `Error:` / `ERROR`-prefixed lines
+- Non-zero exit code patterns (`exit code #` where # ≠ 0)
+
+## Configuration
+
+| Key | Type | Default | Description |
+|-----|------|---------|-------------|
+| `plugins.entries.tool-escalator.config.threshold` | int | 3 | Consecutive tool errors before escalation |
+
+Example `config.yaml`:
+
+```yaml
+plugins:
+  entries:
+    tool-escalator:
+      enabled: true
+      config:
+        threshold: 5
+```
+
+## Hooks declared
+
+- `post_tool_call`
+- `pre_llm_call`
+- `post_llm_call`