fix(reasoning): keep gpt-5.x summary parts as separate blocks on the chat wire

Reasoning-summary models emit one reasoning_content delta per completed
summary part, each a self-contained bold heading. The Responses API delimits
those parts with summary_index; the OpenAI chat wire carries no such field —
verified live against Nous Portal, whose reasoning chunks contain nothing but
delta.reasoning_content — so concatenating them glued every part into one
unspaced, half-bold paragraph.

Re-derive the boundary from the signal the wire does carry: a delta opening a
closed bold heading against a mid-line tail. This matches Hermes own Responses
adapter, which already joins its summary parts with a blank line.

agent/chat_completion_helpers.py

@@ -39,6 +39,7 @@ from agent.message_sanitization import (
     _sanitize_surrogates,
     _repair_tool_call_arguments,
 )
+from agent.reasoning_summaries import separate_glued_reasoning_blocks
 from agent.stream_single_writer import claim_stream_writer, stream_writer_is_current
 from tools.terminal_tool import is_persistent_env
 from utils import base_url_host_matches, base_url_hostname, env_float, env_int
@@ -3256,6 +3257,14 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
             # Accumulate reasoning content
             reasoning_text = getattr(delta, "reasoning_content", None) or getattr(delta, "reasoning", None)
             if reasoning_text:
+                # Summary-part models (gpt-5.x and other Responses relays) send
+                # one complete markdown block per delta with no separator, so
+                # the parts glue into a single unreadable run. Only the tail of
+                # what's accumulated matters. See agent/reasoning_summaries.py.
+                reasoning_text = separate_glued_reasoning_blocks(
+                    reasoning_parts[-1] if reasoning_parts else "",
+                    reasoning_text,
+                )
                 reasoning_parts.append(reasoning_text)
                 _fire_first_delta()
                 agent._fire_reasoning_delta(reasoning_text)

agent/reasoning_summaries.py

@@ -0,0 +1,67 @@
+"""Boundary repair for providers that stream reasoning as discrete summary parts.
+
+Reasoning-summary models (OpenAI's gpt-5.x family, and anything relaying the
+Responses API onto the OpenAI chat wire) do not stream a chain of thought token
+by token. They emit one ``reasoning_content`` delta per *completed* summary
+part, each opening with a bold markdown heading::
+
+    {"delta": {"reasoning_content": "**Investigating likely culprit PRs**"}}
+    {"delta": {"reasoning_content": "**Inspecting message schema**"}}
+
+On the Responses API those parts are delimited by ``summary_index``
+(``response.reasoning_summary_part.added`` / ``.done``). The OpenAI chat wire
+carries no such field — verified live against Nous Portal's
+``openai/gpt-5.6-sol``, whose reasoning chunks contain nothing but
+``delta.reasoning_content`` — so the boundary cannot be recovered from
+metadata, and consumers that concatenate deltas glue the parts together:
+
+    **Investigating likely culprit PRs****Inspecting message schema**
+
+That ``****`` run is neither a bold close nor a bold open to a markdown parser,
+so the whole trace renders as one unbroken, unspaced, half-bold paragraph.
+
+The AI SDK hit exactly this (vercel/ai#6742) and fixed it upstream by starting
+a new reasoning part per ``summary_index``. That route needs the index, which
+this wire does not give us, so we re-derive the boundary from the one signal it
+does carry: a delta opening a bold heading. Hermes' own Responses adapter
+already joins its summary parts with a blank line
+(``agent/codex_responses_adapter.py``), so this brings the chat-completions
+stream in line with the path that keeps the structure.
+"""
+
+from __future__ import annotations
+
+__all__ = ["separate_glued_reasoning_blocks"]
+
+
+def separate_glued_reasoning_blocks(previous: str, delta: str) -> str:
+    """Return *delta*, prefixed with a paragraph break when it glues onto *previous*.
+
+    *previous* is the reasoning text accumulated so far (only its tail matters).
+    A break is inserted when *delta* opens a bold heading and *previous* is
+    mid-line, which is the summary-part boundary the chat wire drops. Both
+    shapes the upstream issue reports are covered: a heading-only part butting
+    against the next heading (``**One****Two**``), and a part whose prose body
+    butts against the next heading (``...interaction!**Next**``).
+
+    Token-streamed reasoning is left alone: its deltas carry their own leading
+    whitespace, so *previous* ends mid-line only when the model really did run
+    two parts together.
+    """
+    if not previous or not delta:
+        return delta
+
+    if not delta.startswith("**"):
+        return delta
+
+    # Already separated — the provider (or an earlier part) ended the line.
+    if previous[-1].isspace():
+        return delta
+
+    # Require a *closed* heading. A token-streamed fragment that merely opens
+    # emphasis ("**" then "bold" then "**" across three deltas) is not a part
+    # boundary; a summary part always carries its whole heading in one delta.
+    if "**" not in delta[2:]:
+        return delta
+
+    return f"\n\n{delta}"

tests/agent/test_reasoning_summaries.py

@@ -0,0 +1,73 @@
+"""Reasoning summary-part boundary repair (agent/reasoning_summaries.py)."""
+
+from agent.reasoning_summaries import separate_glued_reasoning_blocks
+
+
+def _stream(deltas):
+    """Accumulate *deltas* the way the chat-completions stream loop does."""
+    parts: list[str] = []
+    for delta in deltas:
+        parts.append(
+            separate_glued_reasoning_blocks(parts[-1] if parts else "", delta)
+        )
+    return "".join(parts)
+
+
+def test_heading_only_parts_do_not_glue_into_one_run():
+    # The shape observed live on Nous Portal's openai/gpt-5.6-sol: each delta
+    # is a bare heading, so consecutive parts produce a `****` run.
+    text = _stream(
+        [
+            "**Investigating likely culprit PRs**",
+            "**Inspecting message schema and tool_calls content**",
+            "**Analyzing interrupted tool call impact**",
+        ]
+    )
+
+    assert "****" not in text
+    assert text.splitlines() == [
+        "**Investigating likely culprit PRs**",
+        "",
+        "**Inspecting message schema and tool_calls content**",
+        "",
+        "**Analyzing interrupted tool call impact**",
+    ]
+
+
+def test_prose_body_does_not_glue_onto_the_next_heading():
+    # vercel/ai#6742's repro: a part ends in prose and the next heading butts
+    # straight onto it, with no `****` run to key off.
+    text = _stream(
+        [
+            "**Simulating a greeting stream**\n\nIt feels like a streaming interaction!",
+            "**Simulating a greeting stream**\n\nI want to meet the request.",
+        ]
+    )
+
+    assert "interaction!**" not in text
+    assert "interaction!\n\n**Simulating" in text
+
+
+def test_token_streamed_reasoning_is_untouched():
+    deltas = ["Looking at", " the session", " logs, I see", " one bold word."]
+
+    assert _stream(deltas) == "".join(deltas)
+
+
+def test_bold_word_mid_sentence_is_not_a_boundary():
+    # Emphasis inside token-streamed prose arrives after a space.
+    assert separate_glued_reasoning_blocks("I see the ", "**signature**") == "**signature**"
+
+
+def test_unclosed_emphasis_fragment_is_not_a_boundary():
+    # A token stream splitting emphasis across deltas opens but never closes.
+    assert separate_glued_reasoning_blocks("weighing", "**") == "**"
+
+
+def test_boundary_needs_a_bold_opener():
+    assert separate_glued_reasoning_blocks("**Closing**", "plain head") == "plain head"
+
+
+def test_empty_operands_pass_through():
+    assert separate_glued_reasoning_blocks("", "**first**") == "**first**"
+    assert separate_glued_reasoning_blocks("**first**", "") == ""