feat(F-LOOP): true ReAct loop with structured Thought/Action/Observation step tracking

- app/ai/agent_loop.py: ReActStep + ReActResult dataclasses, run_react_loop()
  with LLM→Tool→Observe loop, max_steps/timeout graceful stop, ErrorCategory
  retry (TRANSIENT→retry, PERMANENT→stop, PARTIAL→continue), cost accumulation,
  agent.step hook, on_step callback
- app/plugins/builtins/automation/models.py: AgentRunStep model
- alembic/versions/0121_agent_run_steps.py: migration for agent run steps table
- app/plugins/builtins/automation/agent_runner.py: refactored to use run_react_loop(),
  saves steps to DB, updates AgentRun with cost/status/duration
- tests/test_agent_loop.py: 11 tests (all passing, mocked, no DB/LLM needed)
- PROGRESS.md: Phase F started, F-LOOP marked done

app/ai/agent_loop.py

@@ -0,0 +1,389 @@
+"""Core ReAct (Reasoning + Acting) loop for AI agents.
+
+Implements a true ReAct loop that alternates between LLM reasoning and tool
+execution. Each step records the thought (LLM content), action (tool name),
+action_input (tool arguments), and observation (tool result).
+
+The loop terminates when:
+- The LLM returns a final response without tool calls (completed)
+- max_steps is reached (stopped_max_steps)
+- timeout is exceeded (stopped_timeout)
+- A permanent error occurs (stopped_error)
+
+Error handling uses ``ErrorCategory`` from ``app.core.error_codes``:
+- TRANSIENT → retry the LLM call (up to 3 retries per step)
+- PERMANENT → stop the loop immediately
+- PARTIAL → continue with partial results
+
+Usage::
+
+    from app.ai.agent_loop import run_react_loop
+
+    result = await run_react_loop(
+        agent_definition=agent,
+        messages=[{"role": "user", "content": "Summarize recent emails"}],
+        tools=tool_schemas,
+        tool_registry=registry,
+        db=db_session,
+        tenant_id=tenant_id,
+        user_id=user_id,
+    )
+    print(result.final_content, result.total_cost_usd, result.steps_taken)
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import time
+import uuid
+from dataclasses import dataclass, field
+from datetime import UTC, datetime
+from typing import TYPE_CHECKING, Any
+
+from app.ai.llm_client import llm_complete
+from app.core.error_codes import ErrorCategory, classify_exception
+
+if TYPE_CHECKING:
+    from collections.abc import Callable
+
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+    from app.plugins.builtins.ai_assistant.tool_registry import ToolRegistry
+
+logger = logging.getLogger(__name__)
+
+# Maximum retries for transient errors per LLM step
+_MAX_TRANSIENT_RETRIES = 3
+
+
+# ──────────────────────────────────────────────────────────────────────────
+# Data structures
+# ──────────────────────────────────────────────────────────────────────────
+
+
+@dataclass
+class ReActStep:
+    """A single step in the ReAct loop (Thought → Action → Observation)."""
+
+    step_number: int
+    thought: str  # LLM content before tool calls
+    action: str | None  # Tool name (None if final response)
+    action_input: dict[str, Any] | None  # Tool arguments
+    observation: str | None  # Tool result
+    cost_usd: float
+    timestamp: str  # ISO format
+
+
+@dataclass
+class ReActResult:
+    """Final result of the ReAct loop."""
+
+    final_content: str
+    steps: list[ReActStep] = field(default_factory=list)
+    total_cost_usd: float = 0.0
+    steps_taken: int = 0
+    status: str = "completed"  # completed | stopped_max_steps | stopped_timeout | stopped_error
+    error: str | None = None
+
+
+# ──────────────────────────────────────────────────────────────────────────
+# Core loop
+# ──────────────────────────────────────────────────────────────────────────
+
+
+def _extract_tool_calls(raw_response: Any) -> list[dict[str, Any]]:
+    """Extract tool calls from a LiteLLM raw response.
+
+    Returns a list of dicts with keys: ``id``, ``name``, ``arguments``.
+    """
+    tool_calls: list[dict[str, Any]] = []
+    try:
+        msg = raw_response.choices[0].message
+        if hasattr(msg, "tool_calls") and msg.tool_calls:
+            for tc in msg.tool_calls:
+                tool_calls.append({
+                    "id": tc.id or "",
+                    "name": tc.function.name if tc.function else "",
+                    "arguments": tc.function.arguments if tc.function and tc.function.arguments else "{}",
+                })
+    except (AttributeError, IndexError, TypeError) as exc:
+        logger.debug("Failed to extract tool calls from response: %s", exc)
+    return tool_calls
+
+
+async def _execute_tool(
+    tool_registry: ToolRegistry,
+    tool_name: str,
+    arguments: dict[str, Any],
+    context: dict[str, Any],
+) -> str:
+    """Execute a single tool call via the registry.
+
+    Returns the tool result as a string, or an error message.
+    """
+    tool = tool_registry.get(tool_name)
+    if tool is None:
+        return f"Error: Tool '{tool_name}' not found"
+
+    try:
+        result = await tool.handler(arguments=arguments, context=context)
+        return result if isinstance(result, str) else json.dumps(result)
+    except Exception as exc:
+        logger.exception("Tool '%s' execution failed", tool_name)
+        return f"Error: {exc}"
+
+
+async def run_react_loop(
+    agent_definition: Any,  # AgentDefinition from automation models
+    messages: list[dict[str, Any]],
+    tools: list[dict[str, Any]],
+    tool_registry: ToolRegistry,
+    db: AsyncSession,
+    tenant_id: uuid.UUID,
+    user_id: uuid.UUID,
+    agent_run_id: uuid.UUID | None = None,
+    max_steps: int = 20,
+    timeout_seconds: int = 300,
+    trace_id: str | None = None,
+    on_step: Callable | None = None,
+) -> ReActResult:
+    """Execute a ReAct loop: LLM reasoning → tool execution → repeat.
+
+    Args:
+        agent_definition: AgentDefinition with llm_model, system_prompt, etc.
+        messages: Initial chat messages (without system prompt).
+        tools: OpenAI-format tool schemas for function calling.
+        tool_registry: ToolRegistry instance for tool execution.
+        db: Async DB session.
+        tenant_id: Tenant ID for multi-tenancy.
+        user_id: User ID for permission context.
+        agent_run_id: Optional AgentRun ID for step persistence.
+        max_steps: Maximum loop iterations (default 20).
+        timeout_seconds: Overall timeout (default 300).
+        trace_id: Optional trace ID for correlation.
+        on_step: Optional async callback fired after each step.
+
+    Returns:
+        ReActResult with final content, steps, cost, and status.
+    """
+    from app.core.hooks import do_action
+
+    result = ReActResult(final_content="", status="completed")
+    start_time = time.monotonic()
+
+    # Build LLM parameters from agent definition
+    litellm_model = getattr(agent_definition, "llm_model", None) or "gpt-4o"
+    system_prompt = getattr(agent_definition, "system_prompt", "") or "You are a helpful AI assistant."
+    api_key = getattr(agent_definition, "api_key", None)
+    api_base = getattr(agent_definition, "api_base", None)
+    provider = getattr(agent_definition, "provider", None)
+    max_tokens = getattr(agent_definition, "max_tokens", None) or 1000
+
+    # Build the full message list with system prompt prepended
+    full_messages: list[dict[str, Any]] = [
+        {"role": "system", "content": system_prompt},
+        *messages,
+    ]
+
+    tool_context: dict[str, Any] = {
+        "tenant_id": str(tenant_id),
+        "user_id": str(user_id),
+        "db": db,
+    }
+
+    for step_num in range(1, max_steps + 1):
+        # ── Timeout check ──
+        elapsed = time.monotonic() - start_time
+        if elapsed >= timeout_seconds:
+            result.status = "stopped_timeout"
+            result.error = f"Timeout after {elapsed:.1f}s (limit {timeout_seconds}s)"
+            logger.warning("ReAct loop timed out at step %d: %s", step_num, result.error)
+            break
+
+        # ── LLM call with transient retry ──
+        llm_result: dict[str, Any] | None = None
+        last_error: str | None = None
+
+        for retry in range(_MAX_TRANSIENT_RETRIES + 1):
+            try:
+                llm_result = await llm_complete(
+                    model=litellm_model,
+                    messages=full_messages,
+                    tools=tools if tools else None,
+                    temperature=0.3,
+                    max_tokens=max_tokens,
+                    api_key=api_key,
+                    api_base=api_base,
+                    provider=provider,
+                    trace_id=trace_id,
+                    tenant_id=tenant_id,
+                    db=db,
+                )
+                break
+            except Exception as exc:
+                last_error = str(exc)
+                category = classify_exception(exc)
+
+                if category == ErrorCategory.PERMANENT:
+                    result.status = "stopped_error"
+                    result.error = f"Permanent error at step {step_num}: {exc}"
+                    logger.error("ReAct loop permanent error: %s", result.error)
+                    return result
+
+                if category == ErrorCategory.TRANSIENT and retry < _MAX_TRANSIENT_RETRIES:
+                    backoff = 2 ** retry
+                    logger.warning(
+                        "Transient error at step %d (retry %d/%d): %s — retrying in %ds",
+                        step_num, retry + 1, _MAX_TRANSIENT_RETRIES, exc, backoff,
+                    )
+                    await asyncio.sleep(backoff)
+                    continue
+
+                # PARTIAL or exhausted retries
+                if category == ErrorCategory.PARTIAL:
+                    logger.warning("Partial error at step %d: %s — continuing", step_num, exc)
+                    last_error = str(exc)
+                    break
+
+                # Exhausted transient retries
+                result.status = "stopped_error"
+                result.error = f"Error after {retry + 1} retries at step {step_num}: {exc}"
+                logger.error("ReAct loop error: %s", result.error)
+                return result
+
+        if llm_result is None:
+            result.status = "stopped_error"
+            result.error = f"LLM call failed at step {step_num}: {last_error}"
+            return result
+
+        # ── Extract response data ──
+        content = llm_result.get("content", "")
+        cost_usd = llm_result.get("cost_usd", 0.0)
+        result.total_cost_usd += cost_usd
+
+        tool_calls = _extract_tool_calls(llm_result.get("raw_response"))
+
+        # ── No tool calls → final response ──
+        if not tool_calls:
+            step = ReActStep(
+                step_number=step_num,
+                thought=content,
+                action=None,
+                action_input=None,
+                observation=None,
+                cost_usd=cost_usd,
+                timestamp=datetime.now(UTC).isoformat(),
+            )
+            result.steps.append(step)
+            result.final_content = content
+            result.steps_taken = step_num
+
+            # Fire hook
+            await do_action(
+                "agent.step",
+                agent_id=str(getattr(agent_definition, "id", "")),
+                step_number=step_num,
+                thought=content,
+                action=None,
+                observation=None,
+                cost_usd=cost_usd,
+                agent_run_id=str(agent_run_id) if agent_run_id else None,
+                trace_id=trace_id,
+            )
+
+            # Callback
+            if on_step:
+                try:
+                    await on_step(step)
+                except Exception:
+                    logger.debug("on_step callback failed", exc_info=True)
+
+            break
+
+        # ── Execute tool calls ──
+        # Append assistant message with tool calls to conversation
+        full_messages.append({
+            "role": "assistant",
+            "content": content,
+            "tool_calls": [
+                {
+                    "id": tc["id"],
+                    "type": "function",
+                    "function": {"name": tc["name"], "arguments": tc["arguments"]},
+                }
+                for tc in tool_calls
+            ],
+        })
+
+        # Execute each tool call and collect observations
+        observations: list[str] = []
+        for tc in tool_calls:
+            tool_name = tc["name"]
+            try:
+                args = json.loads(tc["arguments"]) if tc["arguments"] else {}
+            except json.JSONDecodeError:
+                args = {}
+                logger.warning("Invalid JSON arguments for tool '%s': %s", tool_name, tc["arguments"])
+
+            observation = await _execute_tool(tool_registry, tool_name, args, tool_context)
+            observations.append(observation)
+
+            # Feed tool result back into conversation
+            full_messages.append({
+                "role": "tool",
+                "tool_call_id": tc["id"],
+                "content": observation,
+            })
+
+            # Record step
+            step = ReActStep(
+                step_number=step_num,
+                thought=content,
+                action=tool_name,
+                action_input=args,
+                observation=observation,
+                cost_usd=cost_usd / len(tool_calls) if tool_calls else cost_usd,
+                timestamp=datetime.now(UTC).isoformat(),
+            )
+            result.steps.append(step)
+
+            # Fire hook
+            await do_action(
+                "agent.step",
+                agent_id=str(getattr(agent_definition, "id", "")),
+                step_number=step_num,
+                thought=content,
+                action=tool_name,
+                observation=observation,
+                cost_usd=cost_usd,
+                agent_run_id=str(agent_run_id) if agent_run_id else None,
+                trace_id=trace_id,
+            )
+
+            # Callback
+            if on_step:
+                try:
+                    await on_step(step)
+                except Exception:
+                    logger.debug("on_step callback failed", exc_info=True)
+
+        result.steps_taken = step_num
+
+        # If this was the last allowed step, stop gracefully
+        if step_num >= max_steps:
+            result.status = "stopped_max_steps"
+            result.error = f"Reached max_steps limit ({max_steps})"
+            result.final_content = content
+            logger.warning("ReAct loop stopped at max_steps=%d", max_steps)
+            break
+
+    # If loop completed without a final response (e.g. all steps had tool calls)
+    if not result.final_content and result.steps:
+        result.final_content = result.steps[-1].thought or ""
+
+    if result.status == "completed" and not result.final_content:
+        result.final_content = ""
+
+    return result