Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

strata connect: scale the app configs' compaction point with the server's context

demetree7a03fc941425
diff --git a/strata_cli/launch/harness.py b/strata_cli/launch/harness.py
index 5f9379c10..dc4e2de9b 100644
--- a/strata_cli/launch/harness.py
+++ b/strata_cli/launch/harness.py
@@ -21,7 +21,8 @@ from typing import Callable
PLACEHOLDER_KEY = "strata-local" # the tested configs' placeholder when the server has no key (not a secret)
OUTPUT_CAP = 4096 # the output cap every tested config used
-COMPACT_AT = 15000 # where the tested configs make the app compact its conversation
+TESTED_CONTEXT = 32768 # the window every config below was tested on
+COMPACT_AT = 15000 # and where those tested configs make the app compact its conversation
@dataclass
@@ -42,6 +43,17 @@ class Target:
return max(1024, self.context - OUTPUT_CAP)
+def compact_at(t: Target) -> int:
+ """Where the app should compact its conversation: the share of the input limit the tested configs used, so that the
+ summarising request (which carries the whole conversation) still fits the window.
+
+ A flat 15,000 summarises a wider server every 15,000 tokens - 6% of a 256K window - while the tested configs compact
+ a 32K server at 52% of its input limit. The share is what carries over, not the number: a 32K server keeps exactly
+ the value it was tested at."""
+ tested_limit = TESTED_CONTEXT - OUTPUT_CAP
+ return max(COMPACT_AT, int(t.input_limit * COMPACT_AT / tested_limit))
+
+
@dataclass
class FileSpec:
path: str # may start with ~
@@ -80,7 +92,7 @@ def _j(obj) -> str:
def opencode_file(t: Target) -> str:
return _j({"$schema": "https://opencode.ai/config.json", "model": f"strata/{t.model}", "small_model": f"strata/{t.model}",
"autoupdate": False,
- "compaction": {"auto": True, "buffer": max(1024, t.input_limit - COMPACT_AT)},
+ "compaction": {"auto": True, "buffer": max(1024, t.input_limit - compact_at(t))},
"share": "disabled",
"provider": {"strata": {"npm": "@ai-sdk/openai-compatible", "name": "Strata (local)",
"options": {"baseURL": t.base, "apiKey": t.key},
@@ -90,7 +102,7 @@ def opencode_file(t: Target) -> str:
def codex_file(t: Target) -> str:
return (f'model = "{t.model}"\nmodel_provider = "strata"\nmodel_context_window = {t.context}\n'
- f"model_auto_compact_token_limit = {COMPACT_AT}\nmodel_reasoning_summary = \"auto\"\napproval_policy = \"never\"\n"
+ f"model_auto_compact_token_limit = {compact_at(t)}\nmodel_reasoning_summary = \"auto\"\napproval_policy = \"never\"\n"
'sandbox_mode = "workspace-write"\ncheck_for_update_on_startup = false\n\n[model_providers.strata]\n'
f'name = "Strata (local)"\nbase_url = "{t.base}"\nwire_api = "responses"\nenv_key = "STRATA_API_KEY"\n'
"stream_idle_timeout_ms = 600000\n")
@@ -118,7 +130,7 @@ def omp_models(t: Target) -> str:
def omp_config(t: Target) -> str:
return ("modelRoles:\n default: strata/" + t.model + "\ndefaultThinkingLevel: medium\ncompaction:\n enabled: true\n thresholdTokens: "
- + str(COMPACT_AT) + "\n keepRecentTokens: 4000\n methodOrder:\n - soft\n idleEnabled: false\nsetupVersion: 2\n")
+ + str(compact_at(t)) + "\n keepRecentTokens: 4000\n methodOrder:\n - soft\n idleEnabled: false\nsetupVersion: 2\n")
def pi_models(t: Target) -> str:
@@ -131,12 +143,12 @@ def pi_models(t: Target) -> str:
def pi_settings(t: Target) -> str:
return _j({"defaultProvider": "strata", "defaultModel": t.model, "defaultThinkingLevel": "medium",
- "compaction": {"enabled": True, "reserveTokens": max(1024, t.context - COMPACT_AT), "keepRecentTokens": 4000}})
+ "compaction": {"enabled": True, "reserveTokens": max(1024, t.context - compact_at(t)), "keepRecentTokens": 4000}})
def hermes_file(t: Target) -> str:
return ("model:\n default: " + t.model + "\n provider: lmstudio\n base_url: " + t.base + "\n context_length: " + str(t.context) +
- "\ncompression:\n enabled: true\n threshold_tokens: " + str(COMPACT_AT) + "\nauxiliary:\n compression:\n context_length: 65536\n")
+ "\ncompression:\n enabled: true\n threshold_tokens: " + str(compact_at(t)) + "\nauxiliary:\n compression:\n context_length: 65536\n")
def openclaw_file(t: Target) -> str:
@@ -181,14 +193,15 @@ def _limits_omp(t: Target) -> list[str]:
def _limits_pi(t: Target) -> list[str]:
return ["pi has the smallest prompt of the eight (1.5K tokens): the most room for the conversation.",
- "It compacts after the turn that crosses the line, so one turn can reach about 15K tokens before it summarises."]
+ "It compacts after the turn that crosses the line, so one turn can reach about %dK tokens before it summarises."
+ % round(compact_at(t) / 1000)]
def _limits_hermes(t: Target) -> list[str]:
out = ["Hermes refuses a window under 64,000 tokens unless the provider is lmstudio with an explicit context_length: the config uses "
"that, and auxiliary.compression.context_length 65536 for its summariser (the same floor).",
"Its system prompt and 20 tools take about 12.9K tokens (40% of a 32K window): about 19K is left for the conversation.",
- "Its compaction threshold is floored at its own 64K minimum, so it compacts later than 15K (around 20K); it still fits 32K."]
+ "Its own 64K minimum floors the compaction threshold, so it compacts later than the value above (around 20K on a 32K window); it still fits 32K."]
if t.context < 65536:
out.insert(0, "WARNING: Hermes is built for 64K or more. At %dK it works only through the workaround above." % (t.context // 1024))
return out
diff --git a/tests/terminal/test_launch.py b/tests/terminal/test_launch.py
index 8093ef4f0..ff217d2d6 100644
--- a/tests/terminal/test_launch.py
+++ b/tests/terminal/test_launch.py
@@ -1236,7 +1236,19 @@ class ConnectCommand(unittest.TestCase):
def test_context_flag_reaches_the_config(self):
code, out, _ = run_cli(["connect", "pi", "--url", "http://127.0.0.1:8080", "--context", "64k", "--show-key"])
self.assertIn('"contextWindow": 65536', out)
- self.assertIn('"reserveTokens": 50536', out)
+ self.assertIn('"reserveTokens": 33394', out) # 65536 - 32142: the tested 52.3% share, not a flat 15K
+
+ def test_the_compaction_point_scales_with_the_window(self):
+ """A wider server must not be summarised every 15,000 tokens: the share the tested 32K configs used carries over."""
+ wide = harness.target("http://127.0.0.1:18900", None, "qwen3.8-flash-next-q2_0", 262144)
+ self.assertEqual(harness.compact_at(wide), 135000) # 52.3% of the 258048 input limit
+ cfg = json.loads(self.render("opencode", context=262144)["files"][0]["text"])
+ self.assertEqual(wide.input_limit - cfg["compaction"]["buffer"], 135000)
+ self.assertIn("model_auto_compact_token_limit = 135000", self.render("codex", context=262144)["files"][0]["text"])
+ self.assertIn("thresholdTokens: 135000", self.render("omp", context=262144)["files"][1]["text"])
+ # the window every config was tested on keeps exactly the value it was tested at
+ tested = harness.target("http://127.0.0.1:18900", None, "qwen3.8-flash-next-q2_0", 32768)
+ self.assertEqual(harness.compact_at(tested), 15000)
class GpuRolesFlag(unittest.TestCase):