fix(compression): autoraise gpt-5.3-codex-spark threshold to 70% (#48621)

gpt-5.3-codex-spark has a native 128K context window but the default
50% compaction trigger fires at ~64K, wasting half the usable window
before the session has accumulated enough turns to summarize
meaningfully.  This raises the trigger to 70% (~90K) on the Codex OAuth
route only, leaving ~38K headroom for the summary and continued
conversation before the 128K hard limit.

The override is not gated by allow_codex_gpt55_autoraise because 128K
is the model's native window (unlike gpt-5.5's artificial 272K Codex
cap).  Non-Codex routes are unaffected.

Also adds a boundary regression test verifying the short-session
scenario from the issue always yields a non-empty compressible window
(no silent context wipe).

agent/auxiliary_client.py

@@ -325,6 +325,15 @@ def _is_arcee_trinity_thinking(model: Optional[str]) -> bool:
 # gpt-5.5 sessions use the window they actually have.
 _CODEX_GPT54_GPT55_COMPACTION_THRESHOLD = 0.85
 
+# gpt-5.3-codex-spark is Codex-OAuth-only (ChatGPT Pro entitlement) with a
+# native 128K context window.  The default 50% compaction trigger fires at
+# ~64K — wasting half the usable window, often before the session has enough
+# turns to summarize meaningfully.  We raise the trigger to 70% (~90K) so
+# spark sessions use more of the window before summarization, while still
+# leaving ~38K headroom for the summary and continued conversation before
+# the 128K hard limit.
+_CODEX_SPARK_COMPACTION_THRESHOLD = 0.70
+
 
 def _is_codex_gpt54_or_gpt55(model: Optional[str], provider: Optional[str] = None) -> bool:
     """True for gpt-5.4 / gpt-5.5 on the ChatGPT Codex OAuth backend.
@@ -350,6 +359,21 @@ def _is_codex_gpt54_or_gpt55(model: Optional[str], provider: Optional[str] = Non
     )
 
 
+def _is_codex_spark(model: Optional[str], provider: Optional[str] = None) -> bool:
+    """True for ``gpt-5.3-codex-spark`` on the ChatGPT Codex OAuth backend.
+
+    The model is Codex-OAuth-only (ChatGPT Pro entitlement) with a native
+    128K context window.  Only the Codex OAuth route (provider
+    ``openai-codex``) is matched — the slug is not available on other
+    routes.
+    """
+    prov = (provider or "").strip().lower()
+    if prov != "openai-codex":
+        return False
+    bare = (model or "").strip().lower().rsplit("/", 1)[-1]
+    return bare == "gpt-5.3-codex-spark"
+
+
 def _fixed_temperature_for_model(
     model: Optional[str],
     base_url: Optional[str] = None,
@@ -391,6 +415,9 @@ def _compression_threshold_for_model(
         ~136K. Gated by ``allow_codex_gpt55_autoraise`` (historical config-key
         name kept for backward compatibility) so the user can opt back down to
         the global default (the caller passes the config flag through here).
+      - gpt-5.3-codex-spark on the Codex OAuth route → 0.70, because the model
+        has a native 128K window and the default 50% trigger would compact at
+        ~64K — wasting half the usable context. Gated by the same flag.
 
     Returns a float in (0, 1] to override the global ``compression.threshold``
     config value, or ``None`` to leave the user's config value unchanged.
@@ -399,6 +426,8 @@ def _compression_threshold_for_model(
         return 0.75
     if allow_codex_gpt55_autoraise and _is_codex_gpt54_or_gpt55(model, provider):
         return _CODEX_GPT54_GPT55_COMPACTION_THRESHOLD
+    if allow_codex_gpt55_autoraise and _is_codex_spark(model, provider):
+        return _CODEX_SPARK_COMPACTION_THRESHOLD
     return None
 
 # Default auxiliary models for direct API-key providers (cheap/fast for side tasks)

tests/agent/test_arcee_trinity_overrides.py

@@ -18,6 +18,7 @@ from agent.auxiliary_client import (
     _fixed_temperature_for_model,
     _is_arcee_trinity_thinking,
     _is_codex_gpt54_or_gpt55,
+    _is_codex_spark,
 )
 
 
@@ -175,3 +176,75 @@ def test_compression_threshold_opt_out_does_not_disable_trinity() -> None:
         )
         == 0.75
     )
+
+
+# ---------------------------------------------------------------------------
+# Codex gpt-5.3-codex-spark compaction-threshold autoraise
+#
+# gpt-5.3-codex-spark is Codex-OAuth-only (ChatGPT Pro entitlement) with a
+# native 128K context window.  The default 50% compaction trigger would fire
+# at ~64K — wasting half the usable window, often before the session has
+# accumulated enough turns to summarize meaningfully.  This route raises the
+# trigger to 70% (~90K) to preserve more raw context while leaving ~38K
+# headroom before the 128K hard limit.
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    "model",
+    [
+        "gpt-5.3-codex-spark",
+        "openai/gpt-5.3-codex-spark",  # aggregator-prefixed (still on the codex route)
+        "GPT-5.3-CODEX-SPARK",  # case-insensitive
+        "  gpt-5.3-codex-spark  ",  # whitespace tolerant
+    ],
+)
+def test_is_codex_spark_matches_on_codex_provider(model: str) -> None:
+    assert _is_codex_spark(model, "openai-codex") is True
+
+
+@pytest.mark.parametrize(
+    "provider",
+    ["openrouter", "openai", "copilot", "openai-api", "", None],
+)
+def test_is_codex_spark_rejects_non_codex_providers(provider) -> None:
+    # spark on any non-Codex route is not a real slug — no override.
+    assert _is_codex_spark("gpt-5.3-codex-spark", provider) is False
+
+
+@pytest.mark.parametrize(
+    "model",
+    [
+        "gpt-5.5",  # different family
+        "gpt-5.3-codex",  # sibling, not spark
+        "gpt-5.3",  # bare 5.3, not spark
+        "gpt-5.3-codex-spark-mini",  # hypothetical variant — not matched yet
+        "", None,
+    ],
+)
+def test_is_codex_spark_rejects_non_spark_models(model) -> None:
+    assert _is_codex_spark(model, "openai-codex") is False
+
+
+def test_compression_threshold_for_codex_spark() -> None:
+    assert _compression_threshold_for_model("gpt-5.3-codex-spark", "openai-codex") == 0.70
+    assert _compression_threshold_for_model("openai/gpt-5.3-codex-spark", "openai-codex") == 0.70
+
+
+def test_compression_threshold_codex_spark_other_routes_unaffected() -> None:
+    # Same slug, different route → no override (keep the user's config value).
+    assert _compression_threshold_for_model("gpt-5.3-codex-spark", "openrouter") is None
+    assert _compression_threshold_for_model("gpt-5.3-codex-spark", "openai") is None
+    assert _compression_threshold_for_model("gpt-5.3-codex-spark") is None  # no provider
+
+
+def test_compression_threshold_codex_spark_not_gated_by_gpt55_optout() -> None:
+    # The spark autoraise is independent of the gpt-5.5 opt-out flag — 128K is
+    # the model's native window, so 70% is unambiguously correct regardless of
+    # whether the user opted out of the (artificial-cap) gpt-5.5 autoraise.
+    assert (
+        _compression_threshold_for_model(
+            "gpt-5.3-codex-spark", "openai-codex", allow_codex_gpt55_autoraise=False
+        )
+        == 0.70
+    )

tests/run_agent/test_infinite_compaction_loop.py

@@ -283,3 +283,58 @@ class TestCooldownGuard:
         comp.last_prompt_tokens = 65_000
         assert comp._summary_failure_cooldown_until == 0.0
         assert comp.should_compress(65_000)
+
+
+# ---------------------------------------------------------------------------
+# Test: #48621 — gpt-5.3-codex-spark short-session boundary
+#
+# Issue #48621 Bug 2 claims that a short high-token session (15-20 messages,
+# ~90k tokens on a 128k model with protect_last_n=20) hits
+# compress_start >= compress_end, causing a silent context wipe.  The
+# raw-budget fallback added in the #40803 fix already mitigates this: the
+# boundary logic always exposes a minimal compressible window.  This test
+# locks that behavior in for the exact #48621 parameters.
+# ---------------------------------------------------------------------------
+
+class TestCodexSparkShortSessionBoundary:
+    """Verify that gpt-5.3-codex-spark's short-session scenario always yields
+    a non-empty compressible window (no silent wipe)."""
+
+    def test_short_high_token_session_has_compressible_window(self):
+        """16 messages with large tool outputs on a 128k model must leave
+        a compressible middle (compress_start < compress_end)."""
+        comp = _make_compressor(
+            model="gpt-5.3-codex-spark",
+            threshold_percent=0.70,
+            protect_first_n=3,
+            protect_last_n=20,
+            config_context_length=128000,
+        )
+        # Build system + 3 head pairs + 3 tool groups (large outputs) + tail pair
+        big_tool = "x" * 20000  # ~5k tokens each
+        messages = [{"role": "system", "content": "You are a helpful agent."}]
+        for i in range(3):
+            messages.append({"role": "user", "content": f"Question {i}"})
+            messages.append({"role": "assistant", "content": f"Answer {i}"})
+        for i in range(3):
+            messages.append({"role": "user", "content": f"Run command {i}"})
+            messages.append({
+                "role": "assistant", "content": "",
+                "tool_calls": [{
+                    "id": f"tc{i}", "type": "function",
+                    "function": {"name": "terminal", "arguments": "{}"},
+                }],
+            })
+            messages.append({"role": "tool", "tool_call_id": f"tc{i}", "content": big_tool})
+        messages.append({"role": "user", "content": "Final question"})
+        messages.append({"role": "assistant", "content": "Final answer"})
+
+        head = comp._protect_head_size(messages)
+        compress_start = comp._align_boundary_forward(messages, head)
+        compress_end = comp._find_tail_cut_by_tokens(messages, compress_start)
+
+        assert compress_start < compress_end, (
+            f"No compressible window: start={compress_start}, end={compress_end}. "
+            f"This would cause the silent context wipe described in #48621."
+        )
+        assert comp.has_content_to_compress(messages) is True