ayder-cli 2.1.1__tar.gz → 2.1.2__tar.gz

{ayder_cli-2.1.1 → ayder_cli-2.1.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ayder-cli
-Version: 2.1.1
+Version: 2.1.2
 Summary: AI agent for any LLMs
 Project-URL: Homepage, https://github.com/ayder/ayder-cli
 Project-URL: Repository, https://github.com/ayder/ayder-cli.git

{ayder_cli-2.1.1 → ayder_cli-2.1.2}/docs/config.toml.example

@@ -48,9 +48,29 @@
 #      `export OLLAMA_API_KEY="<key>"`. The api_key config field is INERT on the
 #      native path, so a missing env var ⇒ 401 Unauthorized.
 #
-# NOTE on `think`: value is model/driver-dependent. Recent Ollama accepts
-#       true/false for most models and "low"|"medium"|"high" for reasoning
-#       models like deepseek-v4-pro. Adjust per model if a value is rejected.
+# NOTE on `think`: on the native `ollama` driver, ayder hardcodes thinking ON
+#       (think=True) and does NOT forward the configured level — so a
+#       `think = "high"` below documents intent, but the "low/medium/high"
+#       granularity is not applied. The `openai` driver sends no reasoning-effort
+#       parameter at all.
+#
+# NOTE on tuning knobs — which actually take effect depends on the DRIVER. These
+#       agents run on the native `ollama` driver, which ignores most of them:
+#         • temperature       — forwarded by NO driver yet ⇒ no effect anywhere.
+#         • max_output_tokens — NOT sent on the ollama path (Ollama sizes
+#                               generation itself); honored only by the
+#                               openai-family drivers as max_tokens. Default 4096.
+#         • think             — hardcoded ON for ollama (level ignored); openai
+#                               sends none.
+#         • num_ctx           — on ollama it is provisional only (the real window
+#                               is auto-detected from /api/show); on the openai
+#                               path it sizes the context-manager budget.
+#                               Default 65536.
+#       The per-agent temperature / max_output_tokens / think values below are
+#       therefore largely INERT on this Ollama pipeline — they are kept to
+#       document intent and to carry over if you switch an agent to an `openai`
+#       profile. To actually tune these, use the `openai` driver — see
+#       [llm.openai] below.
 # =============================================================================
 
 config_version = "2.0"
@@ -77,10 +97,16 @@ retention = "7 days"
 
 [context_manager]
 enabled = true
-# On the native Ollama path, OllamaContextManager auto-detects the model's real
-# window from /api/show; max_context_tokens is ignored there and only matters on
-# the openai/non-ollama path. reserve_ratio + compaction_threshold ARE read.
-max_context_tokens = 131072
+# Destructive tool-result compression is OPT-IN and OFF by default. It summarizes
+# old tool results in place (lossy) and thrashes the KV cache, so leave it off
+# unless you specifically want it; eviction will replace it in a later release.
+enable_compression = false
+# max_context_tokens is OPTIONAL. When unset (recommended), the budget is derived
+# from the active model's num_ctx. It is read only on the openai/non-ollama path
+# (DefaultContextManager); the native Ollama path auto-detects the real window
+# from /api/show and ignores it. reserve_ratio + compaction_threshold ARE read on
+# both paths.
+# max_context_tokens = 131072
 reserve_ratio = 0.30
 compaction_threshold = 0.80
 
@@ -118,6 +144,20 @@ num_ctx = 32768                     # keep modest: local VRAM is the constraint
 # model = "gemma4:31b-cloud"
 # num_ctx = 131072
 
+# OpenAI-compatible driver. Unlike the `ollama` driver above, this one HONORS the
+# tuning knobs (defaults shown). Use it for api.openai.com, Azure, vLLM,
+# llama.cpp, LM Studio, or Ollama's own `/v1` endpoint. Point an agent at it with
+# `provider = "openai"`.
+# [llm.openai]
+# driver = "openai"
+# base_url = "https://api.openai.com/v1"   # OpenAI-compatible endpoint
+# api_key = "sk-..."                        # your key (or "ollama"/"none" for local servers)
+# model = "gpt-5.5"
+# num_ctx = 65536            # sizes the context-manager budget (default 65536); NOT sent to the API
+# max_output_tokens = 8192   # -> max_tokens / max_completion_tokens (default 4096); keep WELL below num_ctx
+# stop_sequences = []        # -> stop (default [])
+# # temperature / think are NOT applied by the openai driver today (no-ops) — omitted on purpose.
+
 # =============================================================================
 # Agents
 # =============================================================================

{ayder_cli-2.1.1 → ayder_cli-2.1.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "ayder-cli"
-version = "2.1.1"
+version = "2.1.2"
 description = "AI agent for any LLMs"
 readme = "README.md"
 requires-python = ">=3.12"

{ayder_cli-2.1.1 → ayder_cli-2.1.2}/src/ayder_cli/core/config.py

@@ -1,8 +1,11 @@
+import logging
 import tomllib
 from pathlib import Path
 from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
 from typing import Any, Callable, Dict
 
+logger = logging.getLogger(__name__)
+
 CONFIG_DIR = Path("~/.ayder").expanduser()
 CONFIG_PATH = CONFIG_DIR / "config.toml"
 
@@ -174,12 +177,19 @@ class ContextManagerConfigSection(BaseModel):
     model_config = ConfigDict(frozen=True)
 
     enabled: bool = Field(default=True)
-    max_context_tokens: int = Field(default=8192)
+    # Optional: when unset, the context manager derives the budget from the
+    # model's num_ctx. The old fixed 8192 default predates large-context models.
+    max_context_tokens: int | None = Field(default=None)
     reserve_ratio: float = Field(default=0.30)
     compaction_threshold: float = Field(default=0.70)
     tool_result_compress_age: int = Field(default=5)
     max_tool_result_length: int = Field(default=2048)
     compress_tool_results: bool = Field(default=True)
+    # Master switch for destructive tool-result compression. Compression is
+    # known to be too aggressive (it summarizes old tool results in place and
+    # thrashes the KV cache), so it is OPT-IN: disabled unless explicitly set.
+    # Eviction will replace it in a later phase.
+    enable_compression: bool = Field(default=False)
 
     @field_validator("reserve_ratio", "compaction_threshold")
     @classmethod
@@ -190,8 +200,8 @@ class ContextManagerConfigSection(BaseModel):
 
     @field_validator("max_context_tokens")
     @classmethod
-    def validate_positive(cls, v: int) -> int:
-        if v <= 0:
+    def validate_positive(cls, v: int | None) -> int | None:
+        if v is not None and v <= 0:
             raise ValueError("max_context_tokens must be positive")
         return v
 
@@ -309,9 +319,25 @@ class Config(BaseModel):
         if isinstance(llm_section, dict):
             profile = llm_section.get(provider)
             if not isinstance(profile, dict):
+                available = ", ".join(sorted(llm_section)) or "<none>"
                 default_profile = llm_section.get(DEFAULTS["provider"])
                 if isinstance(default_profile, dict):
                     profile = default_profile
+                    logger.error(
+                        "Provider %r has no [llm.%s] profile; falling back to the "
+                        "default [llm.%s] profile. Available profiles: %s. Fix the "
+                        "provider name or add an [llm.%s] section.",
+                        provider, provider, DEFAULTS["provider"], available, provider,
+                    )
+                else:
+                    logger.error(
+                        "Provider %r has no [llm.%s] profile and no default "
+                        "[llm.%s] profile exists; the driver will default to "
+                        "'openai' with built-in endpoint defaults. Available "
+                        "profiles: %s. Fix the provider name or add an [llm.%s] "
+                        "section.",
+                        provider, provider, DEFAULTS["provider"], available, provider,
+                    )
             if isinstance(profile, dict):
                 new_data.update(profile)
             new_data.pop("llm", None)
@@ -359,6 +385,11 @@ class Config(BaseModel):
             "qwen", "dashscope", "glm", "zhipu",
         )
         if v not in valid:
+            logger.error(
+                "Unsupported driver %r; must be one of: %s. Fix the driver in "
+                "your [llm.<provider>] profile.",
+                v, ", ".join(valid),
+            )
             raise ValueError(f"driver must be one of: {', '.join(valid)}")
         return v
 

{ayder_cli-2.1.1 → ayder_cli-2.1.2}/src/ayder_cli/core/default_context_manager.py

@@ -19,6 +19,10 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
+# Fallback context budget used only when neither an explicit max_context_tokens
+# nor a model num_ctx is available (e.g. a bare config section in tests).
+DEFAULT_MAX_CONTEXT_TOKENS = 65536
+
 
 class MessageTier(Enum):
     """Importance tiers for messages."""
@@ -147,10 +151,34 @@ class DefaultContextManager:
         elif isinstance(config, ContextManagerConfigSection):
             self._config = config
         else:
-            max_tokens = getattr(config, "num_ctx", 8192)
-            if max_tokens <= 0:
-                max_tokens = 8192
-            self._config = ContextManagerConfigSection(max_context_tokens=max_tokens)
+            self._config = ContextManagerConfigSection()
+
+        # Resolve the effective context budget. An explicit max_context_tokens
+        # wins; otherwise derive it from the model's num_ctx so the budget tracks
+        # the real window instead of the legacy fixed default.
+        configured = self._config.max_context_tokens
+        num_ctx = getattr(config, "num_ctx", None)
+        if isinstance(configured, int) and configured > 0:
+            self._max_context_tokens = configured
+            logger.info(
+                "Context budget: %d tokens (explicit max_context_tokens; num_ctx=%s)",
+                configured, num_ctx,
+            )
+        elif isinstance(num_ctx, int) and num_ctx > 0:
+            self._max_context_tokens = num_ctx
+            logger.info(
+                "Context budget: %d tokens (derived from num_ctx; "
+                "max_context_tokens unset)",
+                num_ctx,
+            )
+        else:
+            self._max_context_tokens = DEFAULT_MAX_CONTEXT_TOKENS
+            logger.warning(
+                "Context budget: could not read num_ctx or max_context_tokens; "
+                "falling back to default %d tokens. Recommend setting num_ctx "
+                "(or [context_manager] max_context_tokens) in your config.",
+                DEFAULT_MAX_CONTEXT_TOKENS,
+            )
 
         provider = getattr(config, "provider", "openai")
         self._counter = TokenCounter(model=model, provider=provider)
@@ -214,7 +242,7 @@ class DefaultContextManager:
         response arrives. See opus47.md finding #6.
         """
         if self._last_prompt_tokens:
-            ceiling = self._config.max_context_tokens
+            ceiling = self._max_context_tokens
             return self._last_prompt_tokens >= ceiling * self._config.compaction_threshold
         return self.should_trigger_summarization()
 
@@ -227,8 +255,8 @@ class DefaultContextManager:
             return self._stats_cache
 
         total = self._calculate_total_tokens()
-        available = max(0, self._config.max_context_tokens - total)
-        utilization = (total / self._config.max_context_tokens) * 100
+        available = max(0, self._max_context_tokens - total)
+        utilization = (total / self._max_context_tokens) * 100
 
         self._stats_cache = ContextStats(
             total_tokens=total,
@@ -245,7 +273,7 @@ class DefaultContextManager:
 
     @property
     def reserve(self) -> int:
-        return int(self._config.max_context_tokens * self._config.reserve_ratio)
+        return int(self._max_context_tokens * self._config.reserve_ratio)
 
     def count_tokens(self, text: str) -> int:
         """Legacy support."""
@@ -362,18 +390,24 @@ class DefaultContextManager:
             return False
 
         total = self._calculate_total_tokens()
-        utilization = total / self._config.max_context_tokens
+        utilization = total / self._max_context_tokens
         return utilization >= self._config.compaction_threshold
 
     def should_trigger_compression(self) -> bool:
+        if not getattr(self._config, "enable_compression", False):
+            return False
         if not self._config.enabled or not self._config.compress_tool_results:
             return False
 
         total = self._calculate_total_tokens()
-        utilization = total / self._config.max_context_tokens
+        utilization = total / self._max_context_tokens
         return utilization >= self._config.compaction_threshold
 
     def compress_old_results(self, age_threshold: int | None = None) -> int:
+        # Compression is opt-in. Disabled by default because it destructively
+        # summarizes old tool results in place and thrashes the KV cache.
+        if not getattr(self._config, "enable_compression", False):
+            return 0
         threshold = (
             age_threshold
             if age_threshold is not None
@@ -449,10 +483,12 @@ class DefaultContextManager:
         return overhead
 
     def _available_budget(self, overhead: int) -> int:
-        reserve = int(self._config.max_context_tokens * self._config.reserve_ratio)
-        return self._config.max_context_tokens - overhead - reserve
+        reserve = int(self._max_context_tokens * self._config.reserve_ratio)
+        return self._max_context_tokens - overhead - reserve
 
     def _auto_compress(self) -> None:
+        if not getattr(self._config, "enable_compression", False):
+            return
         if not self._config.compress_tool_results:
             return
         if self.should_trigger_compression():

{ayder_cli-2.1.1 → ayder_cli-2.1.2}/src/ayder_cli/providers/impl/ollama.py

@@ -146,8 +146,8 @@ class OllamaProvider(AIProvider):
         tools: list[dict[str, Any]] | None,
         options: dict[str, Any] | None,
     ) -> AsyncGenerator[NormalizedStreamChunk, None]:
-        opts = options or {}
-        num_ctx = opts.get("num_ctx", 65536)
+        # Let Ollama size the context window from the model itself (recent
+        # versions do this automatically); do not force num_ctx.
         ollama_tools = self._convert_tools(tools) if tools else None
         ollama_messages = self._convert_messages(messages)
 
@@ -155,7 +155,6 @@ class OllamaProvider(AIProvider):
             model=model,
             messages=ollama_messages,
             tools=ollama_tools,
-            options={"num_ctx": num_ctx},
             keep_alive=-1,
             think=True,
             stream=True,
@@ -204,8 +203,8 @@ class OllamaProvider(AIProvider):
         tools: list[dict[str, Any]] | None,
         options: dict[str, Any] | None,
     ) -> AsyncGenerator[NormalizedStreamChunk, None]:
-        opts = options or {}
-        num_ctx = opts.get("num_ctx", 65536)
+        # Let Ollama size the context window from the model itself; do not
+        # force num_ctx.
         formatted_messages = driver.render_tools_into_messages(messages, tools or [])
         ollama_messages = self._convert_messages(formatted_messages)
 
@@ -213,7 +212,6 @@ class OllamaProvider(AIProvider):
             model=model,
             messages=ollama_messages,
             tools=None,
-            options={"num_ctx": num_ctx},
             keep_alive=-1,
             think=True,
             stream=True,

ayder_cli-2.1.2/tests/core/test_config_provider_logging.py

@@ -0,0 +1,46 @@
+"""Config should log an actionable error when a provider/driver is misconfigured."""
+import logging
+
+import pytest
+from pydantic import ValidationError
+
+from ayder_cli.core.config import Config
+
+
+def test_missing_provider_profile_logs_error(caplog):
+    """A provider with no matching [llm.<provider>] profile logs an error and
+    still falls back to the openai driver (so the run does not silently misroute
+    without any trace)."""
+    data = {
+        "app": {"provider": "ollama_cloud"},
+        "llm": {"ollama": {"driver": "ollama", "model": "m"}},
+    }
+    with caplog.at_level(logging.ERROR):
+        cfg = Config(**data)
+
+    assert cfg.driver == "openai"  # silent fallback — now logged
+    text = caplog.text
+    assert "ollama_cloud" in text
+    assert "ollama" in text  # names the available profile / section to add
+
+
+def test_valid_provider_profile_logs_no_error(caplog):
+    """A provider that has a matching profile must not log an error."""
+    data = {
+        "app": {"provider": "ollama"},
+        "llm": {"ollama": {"driver": "ollama", "model": "m"}},
+    }
+    with caplog.at_level(logging.ERROR):
+        cfg = Config(**data)
+
+    assert cfg.driver == "ollama"
+    assert caplog.text == ""
+
+
+def test_invalid_driver_logs_error(caplog):
+    """An unsupported driver name logs an error before raising."""
+    with caplog.at_level(logging.ERROR):
+        with pytest.raises(ValidationError):
+            Config(driver="not_a_real_driver", provider="openai")
+
+    assert "not_a_real_driver" in caplog.text

{ayder_cli-2.1.1 → ayder_cli-2.1.2}/tests/core/test_default_context_manager.py

@@ -178,6 +178,139 @@ def test_group_into_units():
     assert units[2] == [messages[3]]
 
 
+def _make_old_long_tool_result_manager(config):
+    """Manager holding one old, long tool result eligible for compression."""
+    mgr = DefaultContextManager(config)
+    long_content = "x" * 5000  # exceeds max_tool_result_length (2048)
+    mgr.add_message({"role": "user", "content": "go"})
+    mgr.add_message(
+        {
+            "role": "assistant",
+            "content": "",
+            "tool_calls": [
+                {"id": "1", "function": {"name": "read_file", "arguments": "{}"}}
+            ],
+        }
+    )
+    mgr.add_message(
+        {"role": "tool", "tool_call_id": "1", "name": "read_file", "content": long_content}
+    )
+    # Filler so the tool result is "old" (age > tool_result_compress_age=5).
+    for i in range(10):
+        mgr.add_message({"role": "user", "content": f"m{i}"})
+    return mgr, long_content
+
+
+def test_max_context_tokens_defaults_to_none():
+    """max_context_tokens is now optional — unset means 'derive from num_ctx'."""
+    from ayder_cli.core.config import ContextManagerConfigSection
+
+    assert ContextManagerConfigSection().max_context_tokens is None
+
+
+def test_budget_syncs_from_num_ctx_when_max_context_tokens_unset():
+    """When max_context_tokens is unset, the budget tracks the model's num_ctx."""
+    from types import SimpleNamespace
+    from ayder_cli.core.config import ContextManagerConfigSection
+
+    cfg = SimpleNamespace(
+        context_manager=ContextManagerConfigSection(),  # max_context_tokens=None
+        num_ctx=131072,
+        provider="openai",
+        model="x",
+    )
+    mgr = DefaultContextManager(cfg)
+
+    assert mgr._max_context_tokens == 131072
+    assert mgr.reserve == int(131072 * 0.30)
+
+
+def test_budget_fallback_to_default_logs_warning(caplog):
+    """When neither max_context_tokens nor num_ctx is readable, warn + use default."""
+    import logging
+    from ayder_cli.core.config import ContextManagerConfigSection
+    from ayder_cli.core.default_context_manager import DEFAULT_MAX_CONTEXT_TOKENS
+
+    config = ContextManagerConfigSection()  # max_context_tokens=None, no num_ctx attr
+    with caplog.at_level(logging.WARNING):
+        mgr = DefaultContextManager(config)
+
+    assert mgr._max_context_tokens == DEFAULT_MAX_CONTEXT_TOKENS
+    assert "num_ctx" in caplog.text
+    assert str(DEFAULT_MAX_CONTEXT_TOKENS) in caplog.text
+
+
+def test_budget_derived_from_num_ctx_logs_info(caplog):
+    """Deriving the budget from num_ctx is reported at INFO with both values."""
+    import logging
+    from types import SimpleNamespace
+    from ayder_cli.core.config import ContextManagerConfigSection
+
+    cfg = SimpleNamespace(
+        context_manager=ContextManagerConfigSection(),
+        num_ctx=131072,
+        provider="openai",
+        model="x",
+    )
+    with caplog.at_level(logging.INFO):
+        DefaultContextManager(cfg)
+
+    assert "131072" in caplog.text
+    assert "num_ctx" in caplog.text
+
+
+def test_explicit_max_context_tokens_overrides_num_ctx():
+    """An explicit max_context_tokens wins over num_ctx."""
+    from types import SimpleNamespace
+    from ayder_cli.core.config import ContextManagerConfigSection
+
+    cfg = SimpleNamespace(
+        context_manager=ContextManagerConfigSection(max_context_tokens=4096),
+        num_ctx=131072,
+        provider="openai",
+        model="x",
+    )
+    mgr = DefaultContextManager(cfg)
+
+    assert mgr._max_context_tokens == 4096
+
+
+def test_enable_compression_defaults_off():
+    """Compression must be opt-in: the global flag defaults to False."""
+    from ayder_cli.core.config import ContextManagerConfigSection
+
+    assert ContextManagerConfigSection().enable_compression is False
+
+
+def test_compress_old_results_noop_when_compression_disabled():
+    """With the default (disabled) flag, old long tool results are NOT compressed."""
+    from ayder_cli.core.config import ContextManagerConfigSection
+
+    config = ContextManagerConfigSection(max_context_tokens=1000)
+    mgr, original = _make_old_long_tool_result_manager(config)
+
+    compressed = mgr.compress_old_results()
+
+    assert compressed == 0
+    assert mgr._messages[2]["content"] == original  # untouched
+
+
+def test_compress_old_results_runs_when_compression_enabled():
+    """When enable_compression=True, eligible old long tool results are compressed."""
+    from ayder_cli.core.config import ContextManagerConfigSection
+
+    config = ContextManagerConfigSection(
+        max_context_tokens=1000, enable_compression=True
+    )
+    mgr, original = _make_old_long_tool_result_manager(config)
+
+    compressed = mgr.compress_old_results()
+
+    assert compressed == 1
+    assert mgr._messages[2]["content"] != original
+    assert "truncated" in mgr._messages[2]["content"].lower()
+
+
 def test_assign_tier_user_messages_never_old_assistant():
     """_assign_tier must not assign OLD_ASSISTANT to old user messages."""
     from ayder_cli.core.config import ContextManagerConfigSection

{ayder_cli-2.1.1 → ayder_cli-2.1.2}/tests/providers/test_ollama_native.py

@@ -204,6 +204,39 @@ async def test_keep_alive_set_to_infinite():
         assert call_kwargs.kwargs.get("keep_alive") == -1 or call_kwargs[1].get("keep_alive") == -1
 
 
+@pytest.mark.asyncio
+async def test_native_stream_does_not_force_num_ctx():
+    """Provider must let Ollama derive context from the model, not force num_ctx.
+
+    Recent Ollama versions size the context window from the model itself, so the
+    provider should not override it via options.
+    """
+    cfg = make_config()
+
+    async def mock_stream(*args, **kwargs):
+        yield make_chat_response(content="ok", done=True,
+                                prompt_eval_count=1, eval_count=1,
+                                prompt_eval_duration=10, eval_duration=10,
+                                load_duration=0)
+
+    with patch("ayder_cli.providers.impl.ollama.AsyncClient") as MockClient:
+        instance = AsyncMock()
+        instance.chat.return_value = mock_stream()
+        MockClient.return_value = instance
+
+        provider = OllamaProvider(cfg)
+        async for _ in provider.stream_with_tools(
+            messages=[{"role": "user", "content": "test"}],
+            model="test:latest",
+            tools=None,
+            options={"num_ctx": 4096},
+        ):
+            pass
+
+        sent_options = instance.chat.call_args.kwargs.get("options")
+        assert not (sent_options and "num_ctx" in sent_options)
+
+
 # =============================================================================
 # SDK Compliance Regression Tests
 # =============================================================================