retoorandClaude Sonnet 5
569f1dcc64
AI gateway: - Add a generic, admin-selectable `client_profile` field on gateway_providers (e.g. "opencode") so a provider needing special request headers (OpenCode Zen's client-identity spoofing) is configured like any other provider, not hardcoded by name. - Track per-(provider, model) reliability/speed/latency health in memory, seeded from the existing gateway_usage_ledger at startup - purely observational, never influences routing. - New Stats tab on /admin/gateway: request volume, latency, per-model breakdowns, and reliability weight, charted with a vendored Chart.js and devplace's own theme tokens. - Record which model a failed request actually fell back to (fallback_used_route), surfaced in the Recent Failures table. - Stop excluding context_length errors from fallback, and skip a primary attempt outright when its known context window is already too small for the estimated request size, going straight to the fallback. - gateway_usage_ledger's provider/fallback_used_route columns and indexes are ensured centrally in database/schema.py's init_db(), the single point of truth for this table's schema. - Non-OpenAI upstream routing and client-model passthrough; trust only the upstream's own X-Gateway-Model header for served-model attribution. Devii agent: - Fix a real lockup: plan/verify tools could be individually disabled via the admin tool toggles while still being required by the protocol gate, permanently bricking any task that needed tools. They can no longer be disabled, and the gate now also checks the tool is actually offered. - Fix compaction being silently calibrated for a 1M-token model while running a much smaller one: context budget is now percentage-based and the summarizer's own request is sized to fit the real model. - Give a specific, actionable retry message when plan()'s own arguments get cut off by the output limit, and tighten its schema to discourage overlong plans. Other: - Backup service: offload completed backups to a remote Hetzner Storage Box. - Container manager: fix orphan blob leaks from sync races, add a two-phase plan/execute `system prune` CLI command. - Admin gateway UI: replace the JS-rendered model/provider tables with server-rendered forms and pages. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DhmEkvutuwtzFVcLbTrhdo
155 lines
5.0 KiB
Python
155 lines
5.0 KiB
Python
# retoor <retoor@molodetz.nl>
|
|
|
|
import json
|
|
|
|
from devplacepy.services.devii.agentic.compaction import (
|
|
compact_messages,
|
|
find_compaction_split,
|
|
is_context_length_error,
|
|
)
|
|
from devplacepy.services.devii.errors import LLMError
|
|
from tests.conftest import run_async
|
|
|
|
|
|
def _error(status, body):
|
|
return LLMError("Model endpoint returned error", status=status, body=body)
|
|
|
|
|
|
def test_openrouter_style_message_detected():
|
|
body = json.dumps(
|
|
{
|
|
"error": {
|
|
"message": (
|
|
"This endpoint's maximum context length is 131072 tokens. "
|
|
"However, you requested about 403355 tokens (349784 of text "
|
|
"input, 53571 of tool input). Please reduce the length of "
|
|
"either one, or use the context-compression plugin."
|
|
),
|
|
"code": 400,
|
|
"metadata": {"provider_name": None},
|
|
}
|
|
}
|
|
)
|
|
assert is_context_length_error(_error(400, body)) is True
|
|
|
|
|
|
def test_openai_style_code_detected():
|
|
body = json.dumps(
|
|
{
|
|
"error": {
|
|
"message": "This model's maximum context length is 128000 tokens.",
|
|
"type": "invalid_request_error",
|
|
"param": None,
|
|
"code": "context_length_exceeded",
|
|
}
|
|
}
|
|
)
|
|
assert is_context_length_error(_error(400, body)) is True
|
|
|
|
|
|
def test_unrelated_400_not_detected():
|
|
body = json.dumps({"error": {"message": "Invalid API key.", "code": 400}})
|
|
assert is_context_length_error(_error(400, body)) is False
|
|
|
|
|
|
def test_non_400_status_not_detected_even_with_matching_text():
|
|
body = json.dumps(
|
|
{"error": {"message": "maximum context length is 131072 tokens"}}
|
|
)
|
|
assert is_context_length_error(_error(429, body)) is False
|
|
|
|
|
|
def test_malformed_body_falls_back_to_phrase_match():
|
|
truncated = "maximum context length is 131072 tokens, please reduce the length"
|
|
assert is_context_length_error(_error(400, truncated)) is True
|
|
|
|
|
|
def test_malformed_body_with_no_match_is_false():
|
|
assert is_context_length_error(_error(400, "not valid json at all")) is False
|
|
|
|
|
|
class _StubLlm:
|
|
def __init__(self, summary="a concise summary of the earlier turns"):
|
|
self.summary = summary
|
|
self.calls = 0
|
|
|
|
async def summarize(self, prompt):
|
|
self.calls += 1
|
|
return self.summary
|
|
|
|
|
|
def _tool_call_message(name="run_tool"):
|
|
return {
|
|
"role": "assistant",
|
|
"content": "",
|
|
"tool_calls": [
|
|
{"id": "c1", "function": {"name": name, "arguments": "{}"}}
|
|
],
|
|
}
|
|
|
|
|
|
def _tool_result_message(content="result"):
|
|
return {"role": "tool", "tool_call_id": "c1", "name": "run_tool", "content": content}
|
|
|
|
|
|
def _long_tool_heavy_conversation(rounds=20):
|
|
messages = [
|
|
{"role": "system", "content": "sys"},
|
|
{"role": "user", "content": "start the long task"},
|
|
]
|
|
for i in range(rounds):
|
|
messages.append(_tool_call_message())
|
|
messages.append(_tool_result_message(f"result {i}" * 200))
|
|
return messages
|
|
|
|
|
|
def test_find_compaction_split_prefers_a_user_message():
|
|
messages = [
|
|
{"role": "system", "content": "sys"},
|
|
{"role": "user", "content": "first"},
|
|
{"role": "assistant", "content": "reply"},
|
|
{"role": "user", "content": "second"},
|
|
{"role": "assistant", "content": "reply2"},
|
|
{"role": "user", "content": "third"},
|
|
{"role": "assistant", "content": "reply3"},
|
|
]
|
|
split = find_compaction_split(messages, keep_tail=2)
|
|
assert messages[split]["role"] == "user"
|
|
|
|
|
|
def test_find_compaction_split_falls_back_to_a_non_tool_boundary_without_a_recent_user_message():
|
|
messages = _long_tool_heavy_conversation(rounds=20)
|
|
split = find_compaction_split(messages, keep_tail=4)
|
|
assert split > 1
|
|
assert messages[split].get("role") != "tool"
|
|
|
|
|
|
def test_find_compaction_split_never_lands_inside_a_tool_result_run():
|
|
messages = _long_tool_heavy_conversation(rounds=30)
|
|
for keep_tail in (2, 3, 4, 5, 8, 10, 15):
|
|
split = find_compaction_split(messages, keep_tail)
|
|
assert messages[split].get("role") != "tool", (
|
|
f"keep_tail={keep_tail} split at a tool message, orphaning its tool_calls"
|
|
)
|
|
|
|
|
|
def test_compact_messages_shrinks_a_tool_heavy_conversation_with_no_recent_user_message():
|
|
messages = _long_tool_heavy_conversation(rounds=20)
|
|
original_len = len(messages)
|
|
llm = _StubLlm()
|
|
compacted = run_async(compact_messages(llm, messages, keep_tail=4))
|
|
assert llm.calls == 1
|
|
assert len(compacted) < original_len
|
|
assert compacted[0]["role"] == "system"
|
|
assert "[compacted earlier turns]" in compacted[1]["content"]
|
|
assert compacted[-1] == messages[-1]
|
|
|
|
|
|
def test_compact_messages_tail_never_starts_with_a_dangling_tool_result():
|
|
messages = _long_tool_heavy_conversation(rounds=25)
|
|
llm = _StubLlm()
|
|
compacted = run_async(compact_messages(llm, messages, keep_tail=6))
|
|
tail = compacted[2:]
|
|
assert tail
|
|
assert tail[0].get("role") != "tool"
|