ayder-cli 2.3.1__tar.gz → 2.4.1__tar.gz

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/AGENTS.md

@@ -87,6 +87,7 @@ ayder-cli/
 │   ├── providers/              # LLM provider implementations
 │   │   ├── base.py             # AIProvider protocol, NormalizedStreamChunk
 │   │   ├── orchestrator.py  # Factory: create(config) → provider
+│   │   ├── evren.py            # evren gateway terms-of-use acceptance (/model)
 │   │   └── impl/               # Individual provider drivers
 │   │       ├── openai.py       # OpenAIProvider (base for most drivers)
 │   │       ├── ollama.py       # OllamaProvider (native + XML fallback)
@@ -385,6 +386,22 @@ uv run poe test-cov      # With coverage report
 - **Coverage**: Maintain existing coverage levels
 - **Timeout**: Use `--timeout=30` to prevent hanging tests
 
+### Test Definitions
+
+Plans and reviews use these values.
+
+- **Kinds**
+  - `baseline regression` — fails on the parent commit for a named reason, passes on the fix.
+  - `new interface` — fails at its assertion once a minimal callable version exists (never on an `ImportError`).
+  - `preservation` — passes before and after the change; its assertion depends on the behaviour it preserves.
+- **Layers**
+  - `unit` — a function or class, without a Textual app.
+  - `tui` — a Textual app driven under `App.run_test()`.
+- **Environment**
+  - `local-faked` — the only one. No network: HTTP is faked with `httpx.MockTransport` passed to the SDK client; other I/O uses `unittest.mock`. No test reaches a real provider.
+
+The gate procedure lives in [docs/GATE.md](docs/GATE.md).
+
 ### Writing Tests
 
 ```python

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: ayder-cli
-Version: 2.3.1
+Version: 2.4.1
 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
@@ -155,6 +155,8 @@ in the core install.
 
 ayder-cli uses a flexible profile-based configuration system. On the first run, it creates a config file at `~/.ayder/config.toml`.
 
+The generated file includes an `[llm.evren]` profile for the evren gateway (`https://evren-llmapi.ssyz.org.tr/v1`) with an empty `api_key`. Fill in your key and switch to it with `/provider evren`.
+
 **Key Concepts:**
 - **Profile Name:** A custom named section (e.g., `[llm.my_ollama]`). You can define as many profiles as you want.
 - **Driver:** The underlying native SDK or adapter used by the profile (`ollama`, `openai`, `anthropic`, `google`, `deepseek`, `qwen`, or `glm`). Each driver guarantees full support for native tool calling and streaming.
@@ -202,7 +204,7 @@ editor = "vim"                # Editor for /task-edit command
 verbose = false               # Show file contents after write + LLM debug
 max_background_processes = 5  # Max concurrent background processes (1-20)
 max_iterations = 10           # Max agentic iterations per message (1-100)
-max_output_tokens = 4096      # Max tokens in LLM response
+max_output_tokens = 32768     # Max tokens in LLM response
 max_history_messages = 30     # Messages kept in history
 prompt = "STANDARD"           # System prompt tier: MINIMAL, STANDARD, EXTENDED
 tool_tags = ["core", "metadata"]  # Enabled tool tags (see /plugin)
@@ -319,6 +321,10 @@ Once a profile is active, use `/model` in the TUI to swap models on the fly:
 - **Interactive Picker:** `/model` with no arguments queries the active driver for available models and opens a picker.
 - **Direct Switch:** `/model <model-name>` immediately switches to that model.
 
+If the provider rejects the listing request, `/model` shows the server's error.
+
+For the evren gateway (`evren-llmapi.ssyz.org.tr`), if your account has not accepted the current terms of use, `/model` shows them and asks you to Approve or Deny; Approve accepts them on the server.
+
 Changes made with `/model` apply to the current session only. To make a model the permanent default, update `model = "..."` in your `config.toml`.
 
 ## Usage
@@ -408,12 +414,13 @@ ayder-cli includes a context manager that automatically keeps your conversation
 - **Real token counts**: Uses Ollama's `prompt_eval_count` and `eval_count` from native API responses instead of estimates. Compaction decisions are based on actual usage, not heuristics.
 - **Cache monitoring**: Tracks `prompt_eval_duration / prompt_eval_count` ratio to detect KV-cache hits (hot/warm/miss). When the cache is hot, compaction is delayed (threshold raised from 70% to 90%) to avoid invalidating it.
 - **Auto-detected context length**: On startup, queries `/api/show` to get the model's real context length instead of relying on `num_ctx`. Falls back to `num_ctx` if detection fails.
-- **Heuristic compaction**: When the context fills up, oldest message units are replaced with a text summary — no LLM call needed.
+- **Heuristic compaction**: When the context fills up, the oldest half of the message units is replaced with a text summary — no LLM call needed. The newest unit is never folded. The summary persists: every later request reuses the same summary message, so each request extends the previous one until the next compaction. A new compaction folds the previous summary into one capped summary (8,000 characters).
+- **Prefix contract**: Message-level prefix stability holds while the model, tools and `max_history = 0` stay unchanged and the history is append-only. Rewriting history (`/clear`, a compact, loading a session) starts a new generation: the summary is dropped and compaction waits for a fresh token count. This guarantees a stable message prefix; whether the server reuses its cache is up to the server.
 - **Native SDK**: Uses `ollama.AsyncClient` directly (not the OpenAI compatibility layer) for full timing data, native thinking support, and `keep_alive=-1` to prevent model unloading.
 
 **General-purpose manager** (Anthropic, OpenAI, Gemini, and other drivers):
 
-Every message is assigned an importance tier (system > recent user > recent assistant > tool results > old history). When the conversation approaches the token budget, old tool results are compressed (JSON structurally summarized, large text head/tail truncated) and the lowest-priority messages are pruned. Tool call + result pairs are kept as atomic units.
+Every message is assigned an importance tier (system > recent user > recent assistant > tool results > old history). When the conversation approaches the token budget, old tool results are compressed (JSON structurally summarized, large text head/tail truncated) and the lowest-priority messages are pruned. Tool call + result pairs are kept as atomic units. The newest unit is always kept, even when it alone exceeds the budget, so the model always sees the request it is answering.
 
 The context manager is configured under `[context_manager]` in `config.toml`:
 
@@ -539,6 +546,8 @@ ayder-cli includes a built-in task system:
 
 Tasks are stored as markdown files in `.ayder/tasks/` using slug-based filenames (e.g., `TASK-001-add-auth-middleware.md`).
 
+The `task` tool resolves identifiers (a path, filename, ID or slug) only to files inside `.ayder/tasks/`. An identifier that points elsewhere, including through a symlink, is not resolved. If `.ayder/tasks/`, `.ayder/context/` or `.ayder/notes/` itself resolves outside the project, those tools return a security error.
+
 ```
 > /tasks            # Interactive task browser
 > /task-edit 1      # Open TASK-001 in the in-app editor

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/README.md

@@ -112,6 +112,8 @@ in the core install.
 
 ayder-cli uses a flexible profile-based configuration system. On the first run, it creates a config file at `~/.ayder/config.toml`.
 
+The generated file includes an `[llm.evren]` profile for the evren gateway (`https://evren-llmapi.ssyz.org.tr/v1`) with an empty `api_key`. Fill in your key and switch to it with `/provider evren`.
+
 **Key Concepts:**
 - **Profile Name:** A custom named section (e.g., `[llm.my_ollama]`). You can define as many profiles as you want.
 - **Driver:** The underlying native SDK or adapter used by the profile (`ollama`, `openai`, `anthropic`, `google`, `deepseek`, `qwen`, or `glm`). Each driver guarantees full support for native tool calling and streaming.
@@ -159,7 +161,7 @@ editor = "vim"                # Editor for /task-edit command
 verbose = false               # Show file contents after write + LLM debug
 max_background_processes = 5  # Max concurrent background processes (1-20)
 max_iterations = 10           # Max agentic iterations per message (1-100)
-max_output_tokens = 4096      # Max tokens in LLM response
+max_output_tokens = 32768     # Max tokens in LLM response
 max_history_messages = 30     # Messages kept in history
 prompt = "STANDARD"           # System prompt tier: MINIMAL, STANDARD, EXTENDED
 tool_tags = ["core", "metadata"]  # Enabled tool tags (see /plugin)
@@ -276,6 +278,10 @@ Once a profile is active, use `/model` in the TUI to swap models on the fly:
 - **Interactive Picker:** `/model` with no arguments queries the active driver for available models and opens a picker.
 - **Direct Switch:** `/model <model-name>` immediately switches to that model.
 
+If the provider rejects the listing request, `/model` shows the server's error.
+
+For the evren gateway (`evren-llmapi.ssyz.org.tr`), if your account has not accepted the current terms of use, `/model` shows them and asks you to Approve or Deny; Approve accepts them on the server.
+
 Changes made with `/model` apply to the current session only. To make a model the permanent default, update `model = "..."` in your `config.toml`.
 
 ## Usage
@@ -365,12 +371,13 @@ ayder-cli includes a context manager that automatically keeps your conversation
 - **Real token counts**: Uses Ollama's `prompt_eval_count` and `eval_count` from native API responses instead of estimates. Compaction decisions are based on actual usage, not heuristics.
 - **Cache monitoring**: Tracks `prompt_eval_duration / prompt_eval_count` ratio to detect KV-cache hits (hot/warm/miss). When the cache is hot, compaction is delayed (threshold raised from 70% to 90%) to avoid invalidating it.
 - **Auto-detected context length**: On startup, queries `/api/show` to get the model's real context length instead of relying on `num_ctx`. Falls back to `num_ctx` if detection fails.
-- **Heuristic compaction**: When the context fills up, oldest message units are replaced with a text summary — no LLM call needed.
+- **Heuristic compaction**: When the context fills up, the oldest half of the message units is replaced with a text summary — no LLM call needed. The newest unit is never folded. The summary persists: every later request reuses the same summary message, so each request extends the previous one until the next compaction. A new compaction folds the previous summary into one capped summary (8,000 characters).
+- **Prefix contract**: Message-level prefix stability holds while the model, tools and `max_history = 0` stay unchanged and the history is append-only. Rewriting history (`/clear`, a compact, loading a session) starts a new generation: the summary is dropped and compaction waits for a fresh token count. This guarantees a stable message prefix; whether the server reuses its cache is up to the server.
 - **Native SDK**: Uses `ollama.AsyncClient` directly (not the OpenAI compatibility layer) for full timing data, native thinking support, and `keep_alive=-1` to prevent model unloading.
 
 **General-purpose manager** (Anthropic, OpenAI, Gemini, and other drivers):
 
-Every message is assigned an importance tier (system > recent user > recent assistant > tool results > old history). When the conversation approaches the token budget, old tool results are compressed (JSON structurally summarized, large text head/tail truncated) and the lowest-priority messages are pruned. Tool call + result pairs are kept as atomic units.
+Every message is assigned an importance tier (system > recent user > recent assistant > tool results > old history). When the conversation approaches the token budget, old tool results are compressed (JSON structurally summarized, large text head/tail truncated) and the lowest-priority messages are pruned. Tool call + result pairs are kept as atomic units. The newest unit is always kept, even when it alone exceeds the budget, so the model always sees the request it is answering.
 
 The context manager is configured under `[context_manager]` in `config.toml`:
 
@@ -496,6 +503,8 @@ ayder-cli includes a built-in task system:
 
 Tasks are stored as markdown files in `.ayder/tasks/` using slug-based filenames (e.g., `TASK-001-add-auth-middleware.md`).
 
+The `task` tool resolves identifiers (a path, filename, ID or slug) only to files inside `.ayder/tasks/`. An identifier that points elsewhere, including through a symlink, is not resolved. If `.ayder/tasks/`, `.ayder/context/` or `.ayder/notes/` itself resolves outside the project, those tools return a security error.
+
 ```
 > /tasks            # Interactive task browser
 > /task-edit 1      # Open TASK-001 in the in-app editor

ayder_cli-2.4.1/docs/GATE.md

@@ -0,0 +1,28 @@
+# Gate
+
+The procedure an agent follows to gate a candidate commit. Follow the steps in order and do
+not interpret the results. Never edit a file, fix a failure, skip a step or create a worktree.
+Test definitions live in [AGENTS.md](../AGENTS.md#test-definitions).
+
+Input: the candidate SHA.
+
+1. `git rev-parse HEAD` → expected: exactly the candidate SHA.
+2. `git status --porcelain` → expected: empty output.
+3. `uv run poe check-all` → expected: exit code 0. This runs lint, typecheck, the logging
+   gates and the test suite, and stops at the first failing task.
+4. Copy the final pytest summary line (`=== N passed, M skipped in …s ===`) verbatim.
+
+Last measured: `2606 passed, 2 skipped` on `fix/max-output-tokens-default` (2026-10-16).
+
+## Report format
+
+```text
+SHA: <candidate SHA>
+1 git rev-parse HEAD: <output> (match|MISMATCH)
+2 git status --porcelain: <empty|output>
+3 uv run poe check-all: exit <code>
+4 summary: <pytest summary line verbatim>
+GATE COMPLETE | GATE FAILED at step <n>
+```
+
+`GATE COMPLETE` only when steps 1–3 meet their expected results.

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/docs/PROJECT_STRUCTURE.md

@@ -220,7 +220,7 @@ cli.py:main()
 | `tui/adapter.py` | Adapter glue between `AyderApp` and `ChatLoop` | callback wiring helpers |
 | `tui/commands.py` | Slash command handlers | `COMMAND_MAP`, `handle_*()` |
 | `tui/widgets.py` | Custom widgets | `ChatView`, `ToolPanel`, `CLIInputBar`, `StatusBar`, `AutoCompleteInput` |
-| `tui/screens.py` | Modal screens | `AgentListScreen`, `CLIConfirmScreen`, `CLIHelpScreen`, `CLIMultiSelectScreen`, `CLIPermissionScreen`, `CLISafeModeScreen`, `CLISelectScreen`, `TaskEditScreen` |
+| `tui/screens.py` | Modal screens | `AgentListScreen`, `CLIConfirmScreen`, `CLIHelpScreen`, `CLIMultiSelectScreen`, `CLIPermissionScreen`, `CLISafeModeScreen`, `CLISelectScreen`, `EvrenTermsScreen`, `TaskEditScreen` |
 | `tui/parser.py` | TUI-specific parsing | `content_processor()` |
 | `tui/helpers.py` | UI helpers | `create_tui_banner()` |
 | `tui/keybindings.py` | Keybinding declarations | `BINDINGS` |
@@ -242,6 +242,7 @@ All LLM provider implementations live here. Previously some sat under `services/
 | `providers/__init__.py` | Re-exports | `AIProvider`, `NormalizedStreamChunk`, `ToolCallDef`, `provider_orchestrator` |
 | `providers/base.py` | Provider protocol + shared DTOs | `AIProvider`, `NormalizedStreamChunk`, `ToolCallDef`, `_ToolCall`, `_FunctionCall` |
 | `providers/orchestrator.py` | Driver-keyed provider factory | `ProviderOrchestrator`, `provider_orchestrator` |
+| `providers/evren.py` | evren gateway terms-of-use acceptance (used by `/model`) | `EVREN_HOST`, `EvrenTerms`, `is_evren_terms_error`, `fetch_terms`, `accept_terms` |
 | `providers/impl/ollama.py` | Native Ollama provider (ollama SDK) | `OllamaProvider` |
 | `providers/impl/ollama_inspector.py` | Ollama model introspection | `OllamaInspector`, `ModelInfo`, `RuntimeState` |
 | `providers/impl/openai.py` | OpenAI / OpenAI-compatible backend | `OpenAIProvider` |
@@ -604,9 +605,11 @@ previously-split `AgentLoopBase` and `LoopConfig` were merged back in):
 - Routes *every* LLM response through the injected `ContextManagerProtocol` (`prepare_messages()` / `update_from_response()`); tool outputs are truncated at insertion time via `truncate_tool_result()` so the KV-cache prefix remains stable
 
 Key features:
-- Auto-approved tools run in parallel (`asyncio.gather`); confirmation-required tools run sequentially
+- Auto-approved tools run in parallel (`asyncio.as_completed`); confirmation-required tools run sequentially
 - Confirmation gate applies to **both** OpenAI-format and XML/JSON custom tool calls
-- `ExecutionPolicy.execute_with_registry()` is the single execution entry (validate → permission → execute)
+- `ExecutionPolicy.execute_with_registry()` is the single execution entry (object check → `canonicalize_arguments()` → validate → permission → execute). Canonicalization applies aliases and `"123"`/`"true"`/`"false"` coercion; the loop's required-argument gate uses the same function
+- Argument integrity: `_parse_tool_arguments()` reports `ok`, `repaired` or `invalid`. Non-object arguments never run. Arguments repaired after being cut off (output token limit) run only for tools registered explicitly as `r`; other tools get an error asking the model to resend in smaller calls
+- Cancellation: each batch owns a `_ToolBatch`. Workers refuse to start once the batch is cancelled and record their own result and `tool_call` event. When a batch ends early (Ctrl+C or a raising callback), a `finally` appends exactly one tool message per unanswered call: its result, an "interrupted, may have partially applied" notice, or a "cancelled before it ran" notice. Running workers are not awaited, so Ctrl+C returns at once
 - Token usage tracking and iteration limiting
 - Pre-iteration hook (`pre_iteration_hook`) used by the agent system to inject summaries
 - Graceful cancellation via `is_cancelled()`

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/docs/config.toml.example

@@ -201,11 +201,20 @@ think = true                        # default; set false if the selected model r
 # 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
+# max_output_tokens = 8192   # -> max_tokens / max_completion_tokens (default 32768); keep WELL below num_ctx
 # stop_sequences = []        # -> stop (default [])
 # reasoning_effort = "medium" # choose a value supported by the selected model
 # # temperature / think are NOT applied by the openai driver.
 
+# evren gateway (OpenAI-compatible). A fresh config.toml ships this profile with
+# an empty api_key. The first `/model` asks you to accept evren's terms of use.
+# [llm.evren]
+# driver = "openai"
+# base_url = "https://evren-llmapi.ssyz.org.tr/v1"
+# api_key = ""               # your evren key
+# model = "deepseek-v4-flash"
+# num_ctx = 131072
+
 # =============================================================================
 # Agents
 # =============================================================================

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "ayder-cli"
-version = "2.3.1"
+version = "2.4.1"
 description = "AI agent for any LLMs"
 readme = "README.md"
 requires-python = ">=3.12"

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/scripts/check_exception_handlers.py

@@ -24,7 +24,7 @@ DEFAULT_BASELINE = REPO / "scripts" / "exception_baseline.txt"
 
 # Frozen census (§C10). A different number means the gate is looking at the
 # wrong tree, or the census moved without the plan being updated.
-EXPECTED_BROAD = 138
+EXPECTED_BROAD = 137
 BRANCHING = (ast.If, ast.Try, ast.For, ast.While, ast.With, ast.Match)
 TERMINAL = (ast.Return, ast.Break, ast.Continue, ast.Pass)
 

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/scripts/check_log_arguments.py

@@ -203,13 +203,13 @@ FROZEN_CHAIN_SITES: dict[tuple[str, str, str, str], tuple[str, int]] = {
     ("log.py", "emit_event", "bind", "**"): ("Name(id='fields', ctx=Load())", 1),
 }
 
-# The 100 `exception=True` inputs are homogeneous, so they are governed as one
-# package-wide shape/count rule rather than 100 site identities - their per-site
+# The 101 `exception=True` inputs are homogeneous, so they are governed as one
+# package-wide shape/count rule rather than 101 site identities - their per-site
 # discipline is C10's.
-FROZEN_EXCEPTION_TRUE = 100
+FROZEN_EXCEPTION_TRUE = 101
 
 # Live chain inputs, counted per method. Printed on every run.
-CHAIN_TALLY: dict[str, int] = {"bind": 6, "opt": 104, "patch": 0}
+CHAIN_TALLY: dict[str, int] = {"bind": 6, "opt": 105, "patch": 0}
 
 # The reviewed composition of THE COMMITTED baseline, frozen by the Phase 5
 # census: 252 retained non-path + 22 paths + 2 deferred interpolated rows, plus

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/src/ayder_cli/application/execution_policy.py

@@ -158,19 +158,40 @@ class ExecutionPolicy:
         *,
         pre_approved: bool = False,
     ) -> ExecutionResult:
-        """Single shared execution path: validate → check permission → execute.
+        """Single shared execution path: canonicalize → validate → permission → execute.
 
         Both CLI and TUI route through this method so registry.execute is never
         called directly from interface-specific code.
 
+        Order:
+        1. Arguments that are not a JSON object are a validation error.
+        2. ``canonicalize_arguments``: aliases, integer and boolean coercion.
+        3. ``ValidationAuthority`` on the canonical arguments.
+        4. Permission check (unless pre-approved), then ``registry.execute``
+           with the canonical arguments.
+
         Args:
             pre_approved: If True, skip the permission check (user already
                 confirmed via the confirmation dialog).
         """
         from ayder_cli.application.validation import ValidationAuthority
+        from ayder_cli.tools.normalization import canonicalize_arguments
+
+        if not isinstance(request.arguments, dict):
+            return ExecutionResult(
+                success=False,
+                error=ValidationError(
+                    tool_name=request.name,
+                    message=(
+                        "arguments must be a JSON object, got "
+                        f"{type(request.arguments).__name__}"
+                    ),
+                ),
+            )
+        canonical = canonicalize_arguments(request.name, request.arguments)
 
         authority = ValidationAuthority()
-        v_request = ToolRequest(name=request.name, arguments=request.arguments)
+        v_request = ToolRequest(name=request.name, arguments=canonical)
         valid, err = authority.validate(v_request)
         if not valid:
             return ExecutionResult(success=False, error=err)
@@ -180,7 +201,7 @@ class ExecutionPolicy:
             if perm_err is not None:
                 return ExecutionResult(success=False, error=perm_err)
 
-        raw = registry.execute(request.name, request.arguments)
+        raw = registry.execute(request.name, v_request.arguments)
         if isinstance(raw, ToolError):
             # A tool that ran and failed is NOT a successful execution.
             return ExecutionResult(success=False, was_confirmed=pre_approved, error=raw)

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/src/ayder_cli/application/validation.py

@@ -101,6 +101,12 @@ class SchemaValidator:
                     field=param_name,
                     message=f"'{param_name}' must be integer, got {type(value).__name__}",
                 )
+            if expected_type == "boolean" and not isinstance(value, bool):
+                return False, ValidationError(
+                    tool_name=request.name,
+                    field=param_name,
+                    message=f"'{param_name}' must be boolean, got {type(value).__name__}",
+                )
             if expected_type == "string" and not isinstance(value, str):
                 return False, ValidationError(
                     tool_name=request.name,

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/src/ayder_cli/core/config.py

@@ -45,6 +45,16 @@ DEFAULTS: Dict[str, Any] = {
         "model": "gemini-3-flash",
         "num_ctx": 65536,
     },
+    # Extra [llm.<name>] profiles written into a freshly created config.toml.
+    "extra_llm_profiles": {
+        "evren": {
+            "driver": "openai",
+            "base_url": "https://evren-llmapi.ssyz.org.tr/v1",
+            "api_key": "",
+            "model": "deepseek-v4-flash",
+            "num_ctx": 131072,
+        },
+    },
     "editor": "vim",
     "verbose": False,
     "logging": {
@@ -342,7 +352,7 @@ class Config(BaseModel):
     max_background_processes: int = Field(default=5)
     # Local Unix-socket inbox that accepts prompts from peer processes.
     messaging_enabled: bool = Field(default=True)
-    max_output_tokens: int = Field(default=4096)
+    max_output_tokens: int = Field(default=32768)
     # 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.

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/src/ayder_cli/core/config_migration.py

@@ -74,6 +74,8 @@ def render_v2_config(
         section = dict(defaults.get(provider, {}))
         section.setdefault("driver", _DRIVER_BY_PROVIDER[provider])
         llm_profiles[provider] = section
+    for name, section in defaults.get("extra_llm_profiles", {}).items():
+        llm_profiles[name] = dict(section)
     if llm_overrides:
         for name, section in llm_overrides.items():
             merged = dict(llm_profiles.get(name, {}))

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/src/ayder_cli/core/context.py

@@ -44,6 +44,39 @@ class ProjectContext:
 
         return target_path
 
+    def harness_dir(self, *parts: str) -> Path:
+        """Return ``<root>/.ayder/<parts>``, refusing a location outside the root.
+
+        ``validate_path`` resolves symlinks, so a ``.ayder/<parts>`` symlink that
+        points outside the project raises instead of redirecting writes.
+        The directory is not created.
+
+        Raises:
+            ValueError: If the resolved directory is outside the project root.
+        """
+        return self.validate_path(str(Path(".ayder", *parts)))
+
+    @staticmethod
+    def within(directory: Path, path: Path | str) -> Path:
+        """Return ``path`` unchanged if it resolves inside ``directory``.
+
+        The resolved form is only checked, never returned, so names and
+        relative paths stay stable when an in-directory symlink is followed.
+        ``os.path.realpath`` is used rather than ``Path.resolve``, which also
+        stats the final target. Resolving still lstats and reads links along
+        the path, including outside components; what it never does is read or
+        follow-stat the target's content, so a refused entry stays unread.
+
+        Raises:
+            ValueError: If ``path`` resolves outside ``directory``.
+        """
+        resolved = Path(os.path.realpath(path))
+        if not resolved.is_relative_to(Path(os.path.realpath(directory))):
+            raise ValueError(
+                f"Security Alert: Path '{resolved}' is outside '{directory}'"
+            )
+        return Path(path)
+
     def to_relative(self, absolute_path: Path) -> str:
         """Convert an absolute path to relative path for LLM output.
 

{ayder_cli-2.3.1 → ayder_cli-2.4.1}/src/ayder_cli/core/default_context_manager.py

@@ -371,39 +371,45 @@ class DefaultContextManager:
         overhead = effective_system_tokens + effective_schema_tokens
         available = self._available_budget(overhead)
 
-        if available <= 0:
-            # Overhead ate the whole budget: every history message is dropped.
-            kept = 1 if system_msg else 0
-            self._emit_trim(len(self._messages), kept, "budget_exhausted")
-            return [system_msg] if system_msg else []
-
         start_idx = 1 if system_msg else 0
         history_msgs = self._messages[start_idx:]
 
         units = self._group_into_units(history_msgs)
 
         result: List[Dict[str, Any]] = []
-        used = 0
-        messages_count = 0
-        reason: str | None = None
-        for unit in reversed(units):
-            if max_history > 0 and messages_count + len(unit) > max_history:
-                reason = "max_history"
-                break
-
-            cost = self._counter.count_messages(unit)
-            if used + cost > available:
-                reason = "token_budget"
-                break
-
-            result = unit + result
-            used += cost
-            messages_count += len(unit)
-
-        # Emitted BEFORE the truncation marker below: that marker ADDS a
-        # message, and counting after it would under-report `dropped` by one.
-        if reason is not None:
-            self._emit_trim(len(history_msgs), len(result), reason)
+        if available <= 0:
+            # Overhead ate the whole budget: every history message is dropped
+            # except the newest unit, the request the model must see (F8).
+            if units:
+                result = list(units[-1])
+            kept = (1 if system_msg else 0) + len(result)
+            self._emit_trim(len(self._messages), kept, "budget_exhausted")
+        else:
+            used = 0
+            messages_count = 0
+            reason: str | None = None
+            for unit in reversed(units):
+                # The newest unit is always kept, even when it alone exceeds
+                # the budget or max_history: dropping it would send a request
+                # the model never saw (F8). If it truly exceeds the model's
+                # window, the provider's error reaches the user.
+                if result and max_history > 0 and messages_count + len(unit) > max_history:
+                    reason = "max_history"
+                    break
+
+                cost = self._counter.count_messages(unit)
+                if result and used + cost > available:
+                    reason = "token_budget"
+                    break
+
+                result = unit + result
+                used += cost
+                messages_count += len(unit)
+
+            # Emitted BEFORE the truncation marker below: that marker ADDS a
+            # message, and counting after it would under-report `dropped` by one.
+            if reason is not None:
+                self._emit_trim(len(history_msgs), len(result), reason)
 
         if result and result[0].get("role") != "user":
             result.insert(