ayder-cli 2.1.1__tar.gz → 2.1.4__tar.gz

{ayder_cli-2.1.1 → ayder_cli-2.1.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ayder-cli
-Version: 2.1.1
+Version: 2.1.4
 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
@@ -34,6 +34,8 @@ Provides-Extra: glm
 Requires-Dist: zhipuai; extra == 'glm'
 Provides-Extra: google
 Requires-Dist: google-genai; extra == 'google'
+Provides-Extra: mcp
+Requires-Dist: mcp; extra == 'mcp'
 Provides-Extra: qwen
 Requires-Dist: dashscope; extra == 'qwen'
 Description-Content-Type: text/markdown

{ayder_cli-2.1.1 → ayder_cli-2.1.4}/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"
@@ -62,7 +82,12 @@ verbose = false
 prompt = "STANDARD"                 # MINIMAL | STANDARD | EXTENDED
 max_background_processes = 5        # cap on concurrent background processes (e.g. test runs)
 max_output_tokens = 8192            # orchestrator response cap
-max_history_messages = 30
+# Max conversation messages kept per turn. 0 = unlimited (default): the context
+# manager's token budget + compaction are the bound, so the model uses its real
+# window. A small fixed cap (the old default was 30) thrashes a tiny history
+# window on a large-context model and wastes most of it. Set a finite number
+# only if you want a hard message cap regardless of window size.
+max_history_messages = 0
 agent_timeout = 1800               # seconds before a background agent is cancelled (30 min; reasoning+coding can be slow)
 # Tools the agents inherit. A coding harness needs file/shell/search (core),
 # tasks/notes (metadata), background process control for test suites
@@ -77,10 +102,16 @@ 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 +149,20 @@ 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-2.1.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "ayder-cli"
-version = "2.1.1"
+version = "2.1.4"
 description = "AI agent for any LLMs"
 readme = "README.md"
 requires-python = ">=3.12"
@@ -37,6 +37,7 @@ google = ["google-genai"]
 qwen = ["dashscope"]
 glm = ["zhipuai"]
 all = ["ayder-cli[anthropic,google,qwen,glm]"]
+mcp = ["mcp"]
 
 [project.urls]
 Homepage = "https://github.com/ayder/ayder-cli"

{ayder_cli-2.1.1 → ayder_cli-2.1.4}/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
 
@@ -278,7 +288,10 @@ class Config(BaseModel):
     logging_retention: str = Field(default="7 days")
     max_background_processes: int = Field(default=5)
     max_output_tokens: int = Field(default=4096)
-    max_history_messages: int = Field(default=30)
+    # 0 = unlimited: let the context manager's token budget + compaction be the
+    # bound. The old fixed 30 predates large-context models and thrashes a tiny
+    # history window on a big context.
+    max_history_messages: int = Field(default=0)
     prompt: str = Field(default="STANDARD")
     chat_protocol: str = Field(default="ollama")
     stop_sequences: list[str] = Field(default_factory=list)
@@ -309,9 +322,25 @@ 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 +388,11 @@ 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-2.1.4}/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-2.1.4}/src/ayder_cli/core/ollama_context_manager.py

@@ -237,8 +237,8 @@ class OllamaContextManager:
             cache_state = f", cache={status.state}({status.hit_ratio:.0%})"
 
         logger.info(
-            "Context: %d/%d tokens (%.0f%%), completion=%d, speed=%s%s",
-            self._real_prompt_tokens, ceiling, utilization,
+            "Context[%s]: %d/%d tokens (%.0f%%), completion=%d, speed=%s%s",
+            self._model or "?", self._real_prompt_tokens, ceiling, utilization,
             self._real_completion_tokens, prompt_speed, cache_state,
         )
 

{ayder_cli-2.1.1 → ayder_cli-2.1.4}/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-2.1.1 → ayder_cli-2.1.4}/src/ayder_cli/themes/claude.py

@@ -633,6 +633,27 @@ AgentPanel .agent-detail {
     padding: 0 0 0 4;
 }
 
+/* ThinkingPanel - Scrollable panel showing the model's reasoning stream.
+   Toggled with Ctrl+T. Taller than the tool/agent panels because reasoning is
+   long-form. Targets 'ThinkingPanel' in widgets.py. */
+ThinkingPanel {
+    layout: vertical;
+    height: auto;
+    max-height: 60%;
+    background: #0a0a18;
+    border-top: solid #333350;
+    padding: 1 1;
+    overflow-y: scroll;
+    display: none;
+}
+
+ThinkingPanel .thinking-content {
+    height: auto;
+    color: #8888aa;
+    text-style: italic;
+    padding: 0 1;
+}
+
 /* AyderApp - The root application screen. */
 Screen {
     layout: vertical;
@@ -669,6 +690,12 @@ Screen:inline {
     width: 100%;
 }
 
+#thinking-panel {
+    min-height: 0;
+    max-height: 60%;
+    width: 100%;
+}
+
 #input-bar {
     height: auto;
     min-height: 1;

{ayder_cli-2.1.1 → ayder_cli-2.1.4}/src/ayder_cli/themes/original.py

@@ -514,6 +514,27 @@ ToolPanel .tool-item.completed {
     color: $success;
 }
 
+/* ThinkingPanel - Scrollable reasoning stream, toggled with Ctrl+T.
+   Taller than the tool/agent panels because reasoning is long-form. */
+ThinkingPanel {
+    layout: vertical;
+    height: auto;
+    min-height: 6;
+    max-height: 24;
+    background: $surface-darken-1;
+    border-top: heavy $primary;
+    padding: 1 1;
+    overflow-y: scroll;
+    display: none;
+}
+
+ThinkingPanel .thinking-content {
+    height: auto;
+    color: $text-muted;
+    text-style: italic;
+    padding: 0 1;
+}
+
 /* AyderApp - Main application */
 Screen {
     layout: vertical;

{ayder_cli-2.1.1 → ayder_cli-2.1.4}/src/ayder_cli/tools/definition.py

@@ -168,7 +168,15 @@ def _discover_definitions() -> Tuple["ToolDefinition", ...]:
 
 _BUILTIN_DEFINITIONS: Tuple[ToolDefinition, ...] = _discover_definitions()
 
-# Phase 1: Merge global plugins at import time
+# Phase 1: Merge global plugins at import time.
+#
+# Discovery runs *inside this module's own body*, so a plugin whose definitions
+# module reads ``TOOL_DEFINITIONS`` at import time (e.g. to avoid tool-name
+# collisions, as mcp-tool does) would otherwise hit a circular import — the name
+# isn't bound yet. Bind it provisionally to the builtins first; those are the
+# only definitions that exist during discovery anyway. It is reassigned to the
+# full set (builtins + plugins) immediately after discovery completes.
+TOOL_DEFINITIONS: Tuple[ToolDefinition, ...] = _BUILTIN_DEFINITIONS
 _GLOBAL_PLUGIN_DEFS: Tuple[ToolDefinition, ...] = ()
 _PLUGIN_HANDLERS: Dict[str, Any] = {}
 
@@ -179,9 +187,7 @@ try:
 except Exception as e:
     logger.warning(f"Failed to load global plugins: {e}")
 
-TOOL_DEFINITIONS: Tuple[ToolDefinition, ...] = (
-    _BUILTIN_DEFINITIONS + _GLOBAL_PLUGIN_DEFS
-)
+TOOL_DEFINITIONS = _BUILTIN_DEFINITIONS + _GLOBAL_PLUGIN_DEFS
 TOOL_DEFINITIONS_BY_NAME: Dict[str, ToolDefinition] = {
     td.name: td for td in TOOL_DEFINITIONS
 }

{ayder_cli-2.1.1 → ayder_cli-2.1.4}/src/ayder_cli/tools/plugin_manager.py

@@ -147,10 +147,15 @@ def install_plugin_from_local(
             "Use --force to overwrite."
         )
 
-    # Copy plugin files
-    if target_dir.exists():
-        shutil.rmtree(target_dir)
-    shutil.copytree(source_dir, target_dir)
+    # Copy plugin files. When the source IS the install target — e.g. reinstalling
+    # an already-installed plugin from its own directory with --force, which is
+    # exactly what the missing-dependency warning recommends — the files are
+    # already in place. Skip the copy so we don't rmtree the source out from under
+    # ourselves (the dependency reinstall + registry refresh below still run).
+    if source_dir.resolve() != target_dir.resolve():
+        if target_dir.exists():
+            shutil.rmtree(target_dir)
+        shutil.copytree(source_dir, target_dir)
 
     # Update plugins.json
     data = read_plugins_json(plugins_dir)

ayder_cli-2.1.4/src/ayder_cli/tools/plugin_status.py

@@ -0,0 +1,33 @@
+"""Plugin status-badge registry.
+
+A lightweight, process-global registry that plugins use to publish a status
+badge (a short label plus a colour) for display in the TUI status bar. The
+mcp-tool plugin, for example, calls ``set_status("mcp", "MCP: news-digest",
+"green")`` once it connects to its MCP servers, and ``set_status("mcp", "MCP",
+"red")`` when configured servers are unreachable.
+
+The TUI ``StatusBar`` reads :func:`get_all` to render the badges. Keeping this
+in ayder core (rather than inside a plugin) means the badge survives even when
+the plugin's own modules are loaded under throwaway names during discovery, and
+gives the plugin a stable import target instead of a no-op fallback.
+"""
+
+from __future__ import annotations
+
+# plugin name -> (label, colour). Colour is a Rich colour/style name ("green").
+_status: dict[str, tuple[str, str]] = {}
+
+
+def set_status(name: str, label: str, color: str) -> None:
+    """Register or update a plugin's status badge."""
+    _status[name] = (label, color)
+
+
+def get_all() -> dict[str, tuple[str, str]]:
+    """Return a copy of all registered status badges (name -> (label, color))."""
+    return dict(_status)
+
+
+def clear() -> None:
+    """Remove all registered badges (primarily for tests)."""
+    _status.clear()