fix(streaming): filter <think> blocks from gateway stream consumer

Models like MiniMax emit inline <think>...</think> reasoning blocks in
their content field. The CLI already suppresses these via a state machine
in _stream_delta, but the gateway's GatewayStreamConsumer had no
equivalent filtering — raw think blocks were streamed directly to
Discord/Telegram/Slack.

The fix adds a _filter_and_accumulate() method that mirrors the CLI's
approach: a state machine tracks whether we're inside a reasoning block
and silently discards the content. Includes the same block-boundary
check (tag must appear at line start or after whitespace-only prefix)
to avoid false positives when models mention <think> in prose.

Handles all tag variants: <think>, <thinking>, <THINKING>, <thought>,
<reasoning>, <REASONING_SCRATCHPAD>.

Also handles edge cases:
- Tags split across streaming deltas (partial tag buffering)
- Unclosed blocks (content suppressed until stream ends)
- Multiple consecutive blocks
- _flush_think_buffer on stream end for held-back partial tags

Adds 22 unit tests + 1 integration test covering all scenarios.

gateway/stream_consumer.py

@@ -64,6 +64,18 @@ class GatewayStreamConsumer:
     # progressive edits for the remainder of the stream.
     _MAX_FLOOD_STRIKES = 3
 
+    # Reasoning/thinking tags that models emit inline in content.
+    # Must stay in sync with cli.py _OPEN_TAGS/_CLOSE_TAGS and
+    # run_agent.py _strip_think_blocks() tag variants.
+    _OPEN_THINK_TAGS = (
+        "<REASONING_SCRATCHPAD>", "<think>", "<reasoning>",
+        "<THINKING>", "<thinking>", "<thought>",
+    )
+    _CLOSE_THINK_TAGS = (
+        "</REASONING_SCRATCHPAD>", "</think>", "</reasoning>",
+        "</THINKING>", "</thinking>", "</thought>",
+    )
+
     def __init__(
         self,
         adapter: Any,
@@ -88,6 +100,10 @@ class GatewayStreamConsumer:
         self._current_edit_interval = self.cfg.edit_interval  # Adaptive backoff
         self._final_response_sent = False
 
+        # Think-block filter state (mirrors CLI's _stream_delta tag suppression)
+        self._in_think_block = False
+        self._think_buffer = ""
+
     @property
     def already_sent(self) -> bool:
         """True if at least one message was sent or edited during the run."""
@@ -132,6 +148,112 @@ class GatewayStreamConsumer:
         """Signal that the stream is complete."""
         self._queue.put(_DONE)
 
+    # ── Think-block filtering ────────────────────────────────────────
+    # Models like MiniMax emit inline <think>...</think> blocks in their
+    # content.  The CLI's _stream_delta suppresses these via a state
+    # machine; we do the same here so gateway users never see raw
+    # reasoning tags.  The agent also strips them from the final
+    # response (run_agent.py _strip_think_blocks), but the stream
+    # consumer sends intermediate edits before that stripping happens.
+
+    def _filter_and_accumulate(self, text: str) -> None:
+        """Add a text delta to the accumulated buffer, suppressing think blocks.
+
+        Uses a state machine that tracks whether we are inside a
+        reasoning/thinking block.  Text inside such blocks is silently
+        discarded.  Partial tags at buffer boundaries are held back in
+        ``_think_buffer`` until enough characters arrive to decide.
+        """
+        buf = self._think_buffer + text
+        self._think_buffer = ""
+
+        while buf:
+            if self._in_think_block:
+                # Look for the earliest closing tag
+                best_idx = -1
+                best_len = 0
+                for tag in self._CLOSE_THINK_TAGS:
+                    idx = buf.find(tag)
+                    if idx != -1 and (best_idx == -1 or idx < best_idx):
+                        best_idx = idx
+                        best_len = len(tag)
+
+                if best_len:
+                    # Found closing tag — discard block, process remainder
+                    self._in_think_block = False
+                    buf = buf[best_idx + best_len:]
+                else:
+                    # No closing tag yet — hold tail that could be a
+                    # partial closing tag prefix, discard the rest.
+                    max_tag = max(len(t) for t in self._CLOSE_THINK_TAGS)
+                    self._think_buffer = buf[-max_tag:] if len(buf) > max_tag else buf
+                    return
+            else:
+                # Look for earliest opening tag at a block boundary
+                # (start of text / preceded by newline + optional whitespace).
+                # This prevents false positives when models *mention* tags
+                # in prose (e.g. "the <think> tag is used for…").
+                best_idx = -1
+                best_len = 0
+                for tag in self._OPEN_THINK_TAGS:
+                    search_start = 0
+                    while True:
+                        idx = buf.find(tag, search_start)
+                        if idx == -1:
+                            break
+                        # Block-boundary check (mirrors cli.py logic)
+                        if idx == 0:
+                            is_boundary = (
+                                not self._accumulated
+                                or self._accumulated.endswith("\n")
+                            )
+                        else:
+                            preceding = buf[:idx]
+                            last_nl = preceding.rfind("\n")
+                            if last_nl == -1:
+                                is_boundary = (
+                                    (not self._accumulated
+                                     or self._accumulated.endswith("\n"))
+                                    and preceding.strip() == ""
+                                )
+                            else:
+                                is_boundary = preceding[last_nl + 1:].strip() == ""
+
+                        if is_boundary and (best_idx == -1 or idx < best_idx):
+                            best_idx = idx
+                            best_len = len(tag)
+                            break  # first boundary hit for this tag is enough
+                        search_start = idx + 1
+
+                if best_len:
+                    # Emit text before the tag, enter think block
+                    self._accumulated += buf[:best_idx]
+                    self._in_think_block = True
+                    buf = buf[best_idx + best_len:]
+                else:
+                    # No opening tag — check for a partial tag at the tail
+                    held_back = 0
+                    for tag in self._OPEN_THINK_TAGS:
+                        for i in range(1, len(tag)):
+                            if buf.endswith(tag[:i]) and i > held_back:
+                                held_back = i
+                    if held_back:
+                        self._accumulated += buf[:-held_back]
+                        self._think_buffer = buf[-held_back:]
+                    else:
+                        self._accumulated += buf
+                    return
+
+    def _flush_think_buffer(self) -> None:
+        """Flush any held-back partial-tag buffer into accumulated text.
+
+        Called when the stream ends (got_done) so that partial text that
+        was held back waiting for a possible opening tag is not lost.
+        """
+        if self._think_buffer and not self._in_think_block:
+            self._accumulated += self._think_buffer
+            self._think_buffer = ""
+
     async def run(self) -> None:
         """Async task that drains the queue and edits the platform message."""
         # Platform message length limit — leave room for cursor + formatting
@@ -156,10 +278,16 @@ class GatewayStreamConsumer:
                         if isinstance(item, tuple) and len(item) == 2 and item[0] is _COMMENTARY:
                             commentary_text = item[1]
                             break
-                        self._accumulated += item
+                        self._filter_and_accumulate(item)
                     except queue.Empty:
                         break
 
+                # Flush any held-back partial-tag buffer on stream end
+                # so trailing text that was waiting for a potential open
+                # tag is not lost.
+                if got_done:
+                    self._flush_think_buffer()
+
                 # Decide whether to flush an edit
                 now = time.monotonic()
                 elapsed = now - self._last_edit_time