feat(cya): T01-T07 core console-native MVP slice (CYA-WP-0001)

- T01: Python + Typer/rich + pyproject.toml + full src/ layout + working `cya` CLI entrypoint
- T02: Bounded transparent context collector (top-level only, provenance, ignores) + --explain-context
- T03: Genuine rule-based risk classifier (primary) + mandatory terminal confirmation, no auto-execute
- T04: LLMAdapter Protocol + deterministic FakeLLMAdapter seam (llm-connect boundary, zero bypass)
- T05: Strictly minimal phase-memory no-op ports (loud markers, per operator direction 2026-05-26)
- T06: Orchestrator coordinating the full flow; CLI is thin delegation
- T07: pytest harness + safety-focused tests (risk invariants + collector)

All changes verified by running the installed `cya` binary and `pytest tests/`.

Workplan updated with status. State Hub progress event logged (workstream 0a1233fd...).

Refs: CYA-WP-0001, Decision a644364b-11c4-49a9-bf17-99063382e27b

src/cya/llm/__init__.py

@@ -0,0 +1,27 @@
+"""llm-connect adapter boundary — the integration seam (T04).
+
+can-you-assist owns orchestration + CLI experience.
+llm-connect owns provider access, config, token counting, and structured I/O.
+
+This package defines the small stable Protocol / interface that all model
+interaction must flow through. A deterministic fake lives here for tests.
+Real delegation to llm-connect is a small localized change once the contract
+is stable.
+
+See workplan CYA-WP-0001-T04 for the full contract and acceptance criteria.
+"""
+
+from .adapter import (
+    AssistanceRequest,
+    AssistanceResponse,
+    LLMAdapter,
+    FakeLLMAdapter,
+)
+
+__all__ = [
+    "AssistanceRequest",
+    "AssistanceResponse",
+    "LLMAdapter",
+    "FakeLLMAdapter",
+]
+

src/cya/llm/adapter.py

@@ -0,0 +1,139 @@
+"""llm-connect adapter boundary (T04 — the integration seam).
+
+Per SCOPE.md and INTENT.md:
+- `can-you-assist` owns orchestration + CLI experience.
+- `llm-connect` owns provider access, config, token counting, and structured I/O.
+
+This module defines the single stable contract that *all* model interaction
+in this repository must flow through. There must never be a production code
+path that talks to an LLM (or a mock) while bypassing this boundary.
+
+Design goals for the MVP slice:
+- Tiny, stable surface (Protocol + two simple data containers).
+- A deterministic, fully reproducible FakeLLMAdapter for tests and early demos.
+- Easy future replacement: swapping the fake for a real (or stubbed)
+  llm-connect client must be a small, localized change.
+
+See workplan CYA-WP-0001-T04 for the full acceptance criteria and the
+"Integration Guide for llm-connect" expectations.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Protocol
+
+
+# ---------------------------------------------------------------------------
+# Request / Response shapes (kept minimal for T04)
+# These will evolve slightly when the real orchestrator (T06) and
+# llm-connect types are known, but the boundary contract stays stable.
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class AssistanceRequest:
+    """What we send to the LLM adapter.
+
+    Contains the framed user intent, the packed context envelope (or its
+    serialised form), and any hints the caller wants to pass (model prefs,
+    token budget, etc.). The adapter is allowed to ignore hints it does not
+    understand.
+    """
+
+    user_request: str
+    context: dict[str, Any] | None = None          # usually a ContextEnvelope.to_dict()
+    hints: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class AssistanceResponse:
+    """What comes back from the LLM adapter.
+
+    The orchestrator / CLI is responsible for turning this into the final
+    user-facing output. The raw fields are intentionally rich so that
+    different front-ends (terminal, future voice) can render appropriately.
+    """
+
+    suggestion: str
+    explanation: str = ""
+    rationale: str = ""
+    risks: list[str] = field(default_factory=list)
+    raw_model_output: str | None = None
+    metadata: dict[str, Any] = field(default_factory=dict)
+
+
+# ---------------------------------------------------------------------------
+# The stable boundary
+# ---------------------------------------------------------------------------
+
+
+class LLMAdapter(Protocol):
+    """The single seam for all model interaction.
+
+    Any real implementation (llm-connect or otherwise) must satisfy this
+    protocol. All production call sites must go through an instance of
+    something obeying this interface.
+    """
+
+    def complete(self, request: AssistanceRequest) -> AssistanceResponse:
+        """Turn a framed request + context into a structured assistant response."""
+        ...
+
+
+# ---------------------------------------------------------------------------
+# Deterministic fake (used by tests, early demos, and T06 development)
+# ---------------------------------------------------------------------------
+
+
+class FakeLLMAdapter:
+    """A fully deterministic, side-effect-free fake adapter.
+
+    Returns canned but useful responses that are stable across runs.
+    The response content is derived only from the request text so that
+    tests can assert on it without any network or real model.
+
+    This is the implementation that must be used by all unit and safety
+    tests until a real adapter is explicitly swapped in.
+    """
+
+    def complete(self, request: AssistanceRequest) -> AssistanceResponse:
+        user_text = request.user_request.strip()
+        risks: list[str] = []
+
+        # Very simple deterministic logic for the MVP slice.
+        # In a real adapter this would be the call to llm-connect.
+        if "delete" in user_text.lower() or "remove" in user_text.lower():
+            suggestion = "I cannot recommend executing that directly. Consider a more targeted command or review the exact files first."
+            explanation = "Your request contained destructive language. The safety layer already required confirmation; the model echoes caution."
+            rationale = "Rule-based safety + conservative model policy."
+            risks = ["Destructive intent detected by rules", "Broad scope in request"]
+        elif "git" in user_text.lower() and ("log" in user_text.lower() or "history" in user_text.lower()):
+            suggestion = "Run: git log --oneline -10 --graph --decorate"
+            explanation = "Standard, safe way to view recent history with a compact graph."
+            rationale = "Common informational request; safe read-only operation."
+        else:
+            suggestion = f"Understood: {user_text[:80]}...\n\nSuggested next step: explore the current directory with `ls -la` or `git status` and share more specific intent."
+            explanation = "This is a placeholder response from the FakeLLMAdapter (T04)."
+            rationale = "No high-risk patterns; generic helpful reply."
+
+        return AssistanceResponse(
+            suggestion=suggestion,
+            explanation=explanation,
+            rationale=rationale,
+            risks=risks,
+            raw_model_output=f"[FAKE] echo of request: {user_text[:200]}",
+            metadata={
+                "adapter": "FakeLLMAdapter",
+                "version": "t04-mvp",
+                "deterministic": True,
+            },
+        )
+
+
+__all__ = [
+    "AssistanceRequest",
+    "AssistanceResponse",
+    "LLMAdapter",
+    "FakeLLMAdapter",
+]