From 1839a536a1ebc95f5a73ec00138600b32cf24d8c Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Fri, 4 Sep 2026 01:05:36 +0900
Subject: [PATCH] perf(ci): add a no-marker fast path to extract_model_prose
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Salvaged from .github#1416 ("Bolt: extract_model_prose 빠른 반환 경로
추가"), whose branch bundled three changes against a stale base:

- extract_model_prose's fast path: still valid, applied here.
- max_tokens: REVIEW_MAX_OUTPUT_TOKENS -> 16 in the preflight request:
  superseded by ADR-0005's own escalation design already on main
  (REVIEW_PREFLIGHT_BASE_TOKENS = 16, escalating to
  REVIEW_PREFLIGHT_ESCALATED_TOKENS only when a response is truncated)
  -- not carried over.
- curl --max-time 30 -> 60 in the sidecar preflight: superseded by a
  later main-side change that removed the timeout entirely, per
  ADR-0003 ("model inference forbids a wall-clock timeout"), replacing
  it with a bounded-attempt retry loop instead -- carrying this over
  would reintroduce exactly what ADR-0003 forbids.

For the one idea that's still current: most model responses contain
neither SENTINEL_PREFIX nor CONTROL_START, so scanning every line for
both prefixes is wasted work. Bolt's own version of this fast path
(`return raw_output.strip()`) wasn't quite behavior-preserving --
splitlines()/join("\n") normalizes CRLF and other exotic line
separators the plain .strip() path leaves untouched, so a raw model
response using non-LF line endings would come back different from
before. Fixed to `"\n".join(raw_output.splitlines()).strip()`, which
is byte-for-byte identical to the slow path's output while still
skipping its per-line prefix-matching loop.

.github#1416 will be closed as fully addressed once this merges.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
---
 scripts/ci/opencode_review_surfaces.py |  8 ++++++++
 tests/test_opencode_review_surfaces.py | 25 +++++++++++++++++++++++++
 2 files changed, 33 insertions(+)

diff --git a/scripts/ci/opencode_review_surfaces.py b/scripts/ci/opencode_review_surfaces.py
index b314cddf00..55756d1ea9 100644
--- a/scripts/ci/opencode_review_surfaces.py
+++ b/scripts/ci/opencode_review_surfaces.py
@@ -329,6 +329,14 @@ def _language(value: str) -> str:
 
 def extract_model_prose(raw_output: str) -> str:
     """Return the human review body, stripping sentinel and control JSON."""
+    if "<!-- opencode-review-" not in raw_output:
+        # Neither SENTINEL_PREFIX nor CONTROL_START's shared literal prefix
+        # occurs anywhere, so the loop below is guaranteed to append every
+        # line unchanged -- skip the per-line prefix checks for the common
+        # case of a plain-prose response, without changing line-ending
+        # normalization behavior.
+        return "\n".join(raw_output.splitlines()).strip()
+
     lines: list[str] = []
     skipping_control = False
     for line in raw_output.splitlines():
diff --git a/tests/test_opencode_review_surfaces.py b/tests/test_opencode_review_surfaces.py
index 7b73dbedb2..858ca513b0 100644
--- a/tests/test_opencode_review_surfaces.py
+++ b/tests/test_opencode_review_surfaces.py
@@ -496,6 +496,31 @@ def test_extract_model_prose_strips_sentinel_and_control() -> None:
     assert "opencode-review-control-v1" not in prose
 
 
+def test_extract_model_prose_fast_path_matches_slow_path_on_plain_text() -> None:
+    """The no-marker fast path is byte-identical to the full line-scan result."""
+    raw = "line one\r\nline two\r\n\r\nline three\n"
+
+    fast_result = surfaces.extract_model_prose(raw)
+
+    lines: list[str] = []
+    skipping_control = False
+    for line in raw.splitlines():
+        stripped = line.strip()
+        if stripped.startswith(surfaces.SENTINEL_PREFIX):
+            continue
+        if stripped.startswith(surfaces.CONTROL_START):
+            skipping_control = True
+            continue
+        if skipping_control:
+            if stripped.endswith("-->"):
+                skipping_control = False
+            continue
+        lines.append(line)
+    slow_result = "\n".join(lines).strip()
+
+    assert fast_result == slow_result == "line one\nline two\n\nline three"
+
+
 def test_format_request_changes_keeps_model_prose_and_strips_fake_anchor() -> None:
     """REQUEST_CHANGES keeps the model walkthrough and never cites workflow:1."""
     body = surfaces.format_request_changes_review(