refactor: pre-PR review improvements for ExecutionLoop + ReAct loop (#124)

CLAUDE.md

@@ -22,6 +22,7 @@
 - Every implementation plan must be **presented to the user** for accept/deny before coding starts
 - At **every phase** of planning and implementation, be critical — actively look for ways to improve the design in the spirit of what we're building (robustness, correctness, simplicity, future-proofing where it's free)
 - Surface improvements as suggestions, not silent changes — user decides
+- **Prioritize issues by dependency order**, not priority labels — unblocked dependencies come first
 
 ## Quick Commands
 
@@ -112,6 +113,8 @@ src/ai_company/
 - **Enforced by**: commitizen (commit-msg hook)
 - **Branches**: `<type>/<slug>` from main
 - **Pre-commit hooks**: trailing-whitespace, end-of-file-fixer, check-yaml, check-toml, check-json, check-merge-conflict, check-added-large-files, no-commit-to-branch (main), ruff check+format, gitleaks
+- **GitHub issue queries**: use `gh issue list` via Bash (not MCP tools) — MCP `list_issues` returns `null` for milestone data
+- **PR issue references**: preserve existing `Closes #NNN` references — never remove unless explicitly asked
 
 ## Post-Implementation (MANDATORY)
 

DESIGN_SPEC.md

@@ -799,6 +799,20 @@ The agent execution loop defines how an agent processes a task from start to fin
 
 > **MVP: ReAct only (Loop 1).** Plan-and-Execute and Hybrid are M4+. Auto-selection is M4+.
 
+#### ExecutionLoop Protocol
+
+All loop implementations satisfy the `ExecutionLoop` runtime-checkable protocol (defined in `engine/loop_protocol.py`):
+
+- **`get_loop_type() -> str`** — returns a unique identifier (e.g. `"react"`)
+- **`execute(...) -> ExecutionResult`** — runs the loop to completion, accepting `AgentContext`, `CompletionProvider`, optional `ToolInvoker`, optional `BudgetChecker`, and optional `CompletionConfig`
+
+Supporting models:
+
+- **`TerminationReason`** — enum: `COMPLETED`, `MAX_TURNS`, `BUDGET_EXHAUSTED`, `ERROR`
+- **`TurnRecord`** — frozen per-turn stats (tokens, cost, tool calls, finish reason)
+- **`ExecutionResult`** — frozen outcome with final context, termination reason, turn records, and optional error message (required when reason is `ERROR`)
+- **`BudgetChecker`** — callback type `Callable[[AgentContext], bool]` invoked before each LLM call
+
 #### Loop 1: ReAct (Default for Simple Tasks)
 
 A single interleaved loop: the agent reasons about the current state, selects an action (tool call or response), observes the result, and repeats until done or `max_turns` is reached.
@@ -814,7 +828,7 @@ A single interleaved loop: the agent reasons about the current state, selects an
 │       └─────────────────────────┘        │
 │                                          │
 │  Terminate when: task complete, max      │
-│  turns, budget exhausted, or blocked     │
+│  turns, budget exhausted, or error       │
 └──────────────────────────────────────────┘
 ```
 
@@ -2106,7 +2120,9 @@ ai-company/
 │       │   ├── prompt_template.py  # System prompt Jinja2 templates
 │       │   ├── task_execution.py   # TaskExecution + StatusTransition
 │       │   ├── context.py          # AgentContext + AgentContextSnapshot
-│       │   ├── agent_engine.py     # Agent execution loop (M3)
+│       │   ├── loop_protocol.py    # ExecutionLoop protocol + result models
+│       │   ├── react_loop.py       # ReAct loop implementation
+│       │   ├── agent_engine.py     # Agent execution engine (M3)
 │       │   ├── task_engine.py      # Task routing & scheduling (M3-M4)
 │       │   ├── workflow_engine.py  # Workflow orchestration (M4)
 │       │   ├── meeting_engine.py   # Meeting coordination (M4)

src/ai_company/engine/__init__.py

@@ -1,7 +1,7 @@
 """Agent execution engine.
 
 Re-exports the public API for system prompt construction,
-runtime execution state, and engine errors.
+runtime execution state, execution loops, and engine errors.
 """
 
 from ai_company.engine.context import (
@@ -10,17 +10,27 @@
     AgentContextSnapshot,
 )
 from ai_company.engine.errors import (
+    BudgetExhaustedError,
     EngineError,
     ExecutionStateError,
+    LoopExecutionError,
     MaxTurnsExceededError,
     PromptBuildError,
 )
+from ai_company.engine.loop_protocol import (
+    BudgetChecker,
+    ExecutionLoop,
+    ExecutionResult,
+    TerminationReason,
+    TurnRecord,
+)
 from ai_company.engine.prompt import (
     DefaultTokenEstimator,
     PromptTokenEstimator,
     SystemPrompt,
     build_system_prompt,
 )
+from ai_company.engine.react_loop import ReactLoop
 from ai_company.engine.task_execution import StatusTransition, TaskExecution
 from ai_company.providers.models import ZERO_TOKEN_USAGE, add_token_usage
 
@@ -29,15 +39,23 @@
     "ZERO_TOKEN_USAGE",
     "AgentContext",
     "AgentContextSnapshot",
+    "BudgetChecker",
+    "BudgetExhaustedError",
     "DefaultTokenEstimator",
     "EngineError",
+    "ExecutionLoop",
+    "ExecutionResult",
     "ExecutionStateError",
+    "LoopExecutionError",
     "MaxTurnsExceededError",
     "PromptBuildError",
     "PromptTokenEstimator",
+    "ReactLoop",
     "StatusTransition",
     "SystemPrompt",
     "TaskExecution",
+    "TerminationReason",
+    "TurnRecord",
     "add_token_usage",
     "build_system_prompt",
 ]

src/ai_company/engine/errors.py

@@ -19,3 +19,21 @@ class MaxTurnsExceededError(EngineError):
     Enforced by ``AgentContext.with_turn_completed`` when the hard turn
     limit has been reached.
     """
+
+
+class BudgetExhaustedError(EngineError):
+    """Budget exhaustion signal for the engine layer.
+
+    The execution loop returns ``TerminationReason.BUDGET_EXHAUSTED``
+    internally.  This exception is available for the engine layer above
+    the loop to convert that result into a raised error when appropriate.
+    """
+
+
+class LoopExecutionError(EngineError):
+    """Non-recoverable execution loop error for the engine layer.
+
+    The execution loop returns ``TerminationReason.ERROR`` internally.
+    This exception is available for the engine layer above the loop to
+    convert that result into a raised error when appropriate.
+    """

src/ai_company/engine/loop_protocol.py

@@ -0,0 +1,160 @@
+"""Execution loop protocol and supporting models.
+
+Defines the ``ExecutionLoop`` protocol that the agent engine calls to
+run a task, along with ``ExecutionResult``, ``TurnRecord``,
+``TerminationReason``, and ``BudgetChecker``.
+"""
+
+from collections.abc import Callable
+from enum import StrEnum
+from typing import TYPE_CHECKING, Any, Protocol, Self, runtime_checkable
+
+from pydantic import BaseModel, ConfigDict, Field, computed_field, model_validator
+
+from ai_company.engine.context import AgentContext
+from ai_company.providers.enums import FinishReason  # noqa: TC001
+
+if TYPE_CHECKING:
+    from ai_company.providers.models import CompletionConfig
+    from ai_company.providers.protocol import CompletionProvider
+    from ai_company.tools.invoker import ToolInvoker
+
+
+class TerminationReason(StrEnum):
+    """Why the execution loop terminated."""
+
+    COMPLETED = "completed"
+    MAX_TURNS = "max_turns"
+    BUDGET_EXHAUSTED = "budget_exhausted"
+    ERROR = "error"
+
+
+class TurnRecord(BaseModel):
+    """Per-turn metadata recorded during execution.
+
+    Attributes:
+        turn_number: 1-indexed turn number.
+        input_tokens: Input tokens consumed this turn.
+        output_tokens: Output tokens generated this turn.
+        total_tokens: Sum of input and output tokens (computed).
+        cost_usd: Cost in USD for this turn.
+        tool_calls_made: Names of tools invoked this turn.
+        finish_reason: LLM finish reason for this turn.
+    """
+
+    model_config = ConfigDict(frozen=True)
+
+    turn_number: int = Field(gt=0, description="1-indexed turn number")
+    input_tokens: int = Field(ge=0, description="Input tokens this turn")
+    output_tokens: int = Field(ge=0, description="Output tokens this turn")
+    cost_usd: float = Field(ge=0.0, description="Cost in USD this turn")
+    tool_calls_made: tuple[str, ...] = Field(
+        default=(),
+        description="Tool names invoked this turn",
+    )
+    finish_reason: FinishReason = Field(
+        description="LLM finish reason this turn",
+    )
+
+    @computed_field(description="Total token count")  # type: ignore[prop-decorator]
+    @property
+    def total_tokens(self) -> int:
+        """Sum of input and output tokens."""
+        return self.input_tokens + self.output_tokens
+
+
+class ExecutionResult(BaseModel):
+    """Result returned by an execution loop.
+
+    Attributes:
+        context: Final agent context after execution.
+        termination_reason: Why the loop stopped.
+        turns: Per-turn metadata records.
+        total_tool_calls: Total tool calls across all turns (computed).
+        error_message: Error description when termination_reason is ERROR.
+        metadata: Forward-compatible dict for future loop types.
+            Note: ``frozen=True`` prevents field reassignment but not
+            in-place mutation of the dict contents; deep-copy at
+            system boundaries per project conventions.
+    """
+
+    model_config = ConfigDict(frozen=True)
+
+    context: AgentContext = Field(description="Final agent context")
+    termination_reason: TerminationReason = Field(
+        description="Why the loop stopped",
+    )
+    turns: tuple[TurnRecord, ...] = Field(
+        default=(),
+        description="Per-turn metadata",
+    )
+    error_message: str | None = Field(
+        default=None,
+        description="Error description (when reason is ERROR)",
+    )
+    metadata: dict[str, Any] = Field(
+        default_factory=dict,
+        description="Forward-compatible metadata for future loop types",
+    )
+
+    @computed_field(  # type: ignore[prop-decorator]
+        description="Total tool calls across all turns",
+    )
+    @property
+    def total_tool_calls(self) -> int:
+        """Sum of tool calls from all turn records."""
+        return sum(len(t.tool_calls_made) for t in self.turns)
+
+    @model_validator(mode="after")
+    def _validate_error_message(self) -> Self:
+        if self.termination_reason == TerminationReason.ERROR:
+            if self.error_message is None:
+                msg = "error_message is required when termination_reason is ERROR"
+                raise ValueError(msg)
+        elif self.error_message is not None:
+            msg = "error_message must be None when termination_reason is not ERROR"
+            raise ValueError(msg)
+        return self
+
+
+BudgetChecker = Callable[[AgentContext], bool]
+"""Callback that returns ``True`` when the budget is exhausted."""
+
+
+@runtime_checkable
+class ExecutionLoop(Protocol):
+    """Protocol for agent execution loops.
+
+    The agent engine calls ``execute`` to run a task through the loop.
+    Implementations decide the control flow (ReAct, Plan-and-Execute, etc.)
+    but all return an ``ExecutionResult`` with a ``TerminationReason``.
+    """
+
+    async def execute(
+        self,
+        *,
+        context: AgentContext,
+        provider: CompletionProvider,
+        tool_invoker: ToolInvoker | None = None,
+        budget_checker: BudgetChecker | None = None,
+        completion_config: CompletionConfig | None = None,
+    ) -> ExecutionResult:
+        """Run the execution loop.
+
+        Args:
+            context: Initial agent context with conversation and identity.
+            provider: LLM completion provider.
+            tool_invoker: Optional tool invoker for tool execution.
+            budget_checker: Optional callback; returns ``True`` when
+                budget is exhausted.
+            completion_config: Optional per-execution override for
+                temperature/max_tokens (defaults to identity's model config).
+
+        Returns:
+            Execution result with final context and termination reason.
+        """
+        ...
+
+    def get_loop_type(self) -> str:
+        """Return the loop type identifier (e.g. ``"react"``)."""
+        ...