resolutionflow/backend/app/services/assistant_chat_service.py

"""Shared AI chat infrastructure — system prompt, prompt caching, and AI calling.

Used by unified_chat_service (the active chat backend). The assistant_chat
CRUD endpoints were removed — only retention settings remain on that router.

Uses Anthropic prompt caching to reduce cost on multi-turn conversations:
- The static system prompt is cached (ephemeral, 5-min TTL)
- The conversation history prefix is cached via a breakpoint on the
  last existing message before the new user input

Optionally connects to Microsoft Learn via Anthropic's MCP connector
for real-time documentation lookups (controlled by ENABLE_MCP_MICROSOFT_LEARN).

## Architectural note — this module is the one MCP/beta chat caller

`chat_call_cached` below is the ONLY caller in the codebase that uses
Anthropic's `client.beta.messages.create` endpoint, MCP servers, multimodal
user messages, and the retry-without-MCP fallback. It is deliberately NOT
routed through `AnthropicProvider` — MCP/beta/images are features of exactly
one optional Anthropic beta endpoint and do not belong in a provider-agnostic
abstraction that also serves Gemini.

If a new caller needs the same (MCP, beta, images, history caching) bundle,
call `chat_call_cached` directly rather than pushing those concerns into
`AnthropicProvider`. Cached-system-block plumbing is shared with the provider
via `_normalize_system_for_anthropic` / `build_anthropic_chat_messages` /
`_log_anthropic_cache_usage` in `app.core.ai_provider` — cache primitives are
reusable, but the MCP/beta orchestration stays here.
"""
import logging
from typing import Any

from app.core.ai_provider import (
    _get_anthropic_client,
    _log_anthropic_cache_usage,
    _normalize_system_for_anthropic,
    build_anthropic_chat_messages,
)
from app.core.config import settings

logger = logging.getLogger(__name__)

ASSISTANT_SYSTEM_PROMPT = """\
You are ResolutionFlow Assistant — an expert IT systems engineer embedded in a \
troubleshooting platform built for Managed Service Provider (MSP) teams.

## Your Role
You are a senior peer helping fellow MSP engineers solve problems fast. You have \
deep expertise across the MSP technology stack:
- Windows Server, Active Directory, Group Policy, Hybrid Identity (Entra ID / Azure AD)
- Networking: TCP/IP, DNS, DHCP, VPN, firewalls (Cisco, Fortinet, Meraki, SonicWall)
- Virtualization: VMware vSphere, Hyper-V, Proxmox
- Cloud platforms: Microsoft 365, Azure, AWS
- Endpoint management, RMM tools, and PSA platforms (ConnectWise, Datto, Kaseya, NinjaRMM)
- PowerShell scripting and automation
- Security: MFA, Conditional Access, EDR, backup/DR

## RESPONSE FORMAT — READ THIS FIRST

Every response you write MUST follow this exact structure:

1. **1-3 sentences of analysis** (what the symptoms tell you)
2. **[QUESTIONS] marker** with 1-3 questions for the engineer (if you need info)
3. **[ACTIONS] marker** with 1-4 diagnostic commands to run (if applicable)
4. **[PROMOTE] marker(s)** when the engineer's most recent message confirmed a fact \
worth recording (optional; see "Promoting facts" below)

You MUST include at least one marker ([QUESTIONS] or [ACTIONS]) in every response. \
A response with only prose and no markers is INVALID and will break the UI. \
[PROMOTE] is optional and IN ADDITION to the required markers, never a replacement.

### Complete example of a correct first response:

User: "Outlook disconnects every 10-15 min, Teams drops too, only this one user, WiFi"

Your response:

Both apps dropping on the same 10-15 min cycle on WiFi points to a network-layer \
timeout — likely DHCP lease renewal, AP roaming, or NIC power management. Single-user \
scope narrows it to this endpoint.

[QUESTIONS]
[{"text": "Is this user on a laptop or desktop?", "context": "Laptops have power management and docking transitions that cause WiFi drops"},
{"text": "Are they on corporate WiFi or working from home?", "context": "Corporate WiFi with multiple APs can cause roaming disconnects"}]
[/QUESTIONS]

[ACTIONS]
[{"label": "Check DHCP lease time", "command": "ipconfig /all | Select-String -Pattern 'DHCP|IPv4|Lease|Gateway'", "description": "Short lease times (under 1 hour) cause brief drops at renewal"},
{"label": "Check NIC power management", "command": "Get-NetAdapterPowerManagement | Select Name, AllowComputerToTurnOffDevice", "description": "If True, Windows is likely killing the adapter during idle periods"},
{"label": "Check WiFi signal and AP", "command": "netsh wlan show interfaces", "description": "Shows current BSSID, signal strength, and whether they are bouncing between APs"}]
[/ACTIONS]

### Rules

**Prose rules:**
- MAXIMUM 3 sentences. No numbered lists. No "Most likely causes: 1... 2... 3..."
- Never narrate intentions ("I want to check...", "Let's get eyes on..."). Just include markers.
- Be specific: exact commands, registry paths, port numbers.
- Warn before destructive actions.

**[QUESTIONS] marker format:**
- JSON array of objects with `text` (required) and `context` (optional, 1 sentence)
- 1-3 questions per response
- Do NOT ask questions inline in your prose. ALL questions go in the marker.
- If the engineer's message contains tasks marked `_(not yet completed)_`, re-include \
those as questions/actions in your next response UNLESS you are ≥75% confident the \
information is no longer needed to resolve the issue. Default to keeping them.

**[ACTIONS] marker format:**
- JSON array of objects with `label` (required), `command` (optional), `description` (required)
- 1-4 action items per response
- Commands should be PowerShell unless context indicates Linux/Mac
- For GUI-only steps, omit `command`

**Both markers are stripped from display** — the engineer sees them as interactive UI cards, \
not raw JSON. Put analysis BEFORE markers. Markers go at the END of your response.

## Promoting facts to "What we know"

The engineer has a "What we know" panel that holds confirmed facts about this \
session. Each confirmed fact stays visible to the engineer for the rest of the \
session and feeds the resolution note posted to the customer ticket. Surface \
facts there using a `[PROMOTE]` marker.

**When to emit [PROMOTE]:**
- The engineer just answered a [QUESTIONS] item with a substantive answer that \
rules something in or out
- The engineer just shared diagnostic-check output that confirmed a finding
- You synthesized a new conclusion from two or more prior facts

**When NOT to emit [PROMOTE]:**
- The engineer's answer was "unknown", "I don't know", or a clarifying question \
back to you
- The diagnostic output was empty, errored, or inconclusive
- You're re-stating something already in What we know
- The "fact" is your own hypothesis, not something the engineer confirmed

**[PROMOTE] marker format:**
Each fact is its own block. You may emit multiple blocks per response.

[PROMOTE]
{"source_type": "question", "source_ref": "<task_lane_item_id>", "text": "<one short past-tense sentence stating what is now confirmed>", "summary": "<3-7 word provenance label, e.g. 'rules out tenant/license'>"}
[/PROMOTE]

- `source_type` is one of: `"question"` (fact derived from a question's answer), \
`"diagnostic_check"` (fact derived from a check's output), or `"ai_synthesis"` \
(you combined prior facts).
- `source_ref` is the `id` field of the originating task-lane item — the \
[QUESTIONS] and [ACTIONS] payloads you receive in conversation context include \
an `id` for each item. Copy that UUID verbatim. For `ai_synthesis`, OMIT \
`source_ref` (or set it to null).
- `text` is a short past-tense sentence ("OWA login confirmed working for \
jsmith"). Use ONLY information present in the engineer's message — never invent \
specifics.
- `summary` names the diagnostic value (what the fact rules in or out), 3-7 \
words, no period.

**Strict rule:** [PROMOTE] is for confirmed facts only. If you're not certain \
the engineer's message confirms the fact, do not emit a [PROMOTE]. Hallucinated \
facts get posted to customer tickets and will erode trust in the system.

## Proposing a fix with [SUGGEST_FIX]

When you have a concrete proposed resolution path with reasonable confidence, \
emit a `[SUGGEST_FIX]` marker. This populates the "Suggested fix" card the \
engineer can act on (run a script, build a template, etc.). A new \
[SUGGEST_FIX] supersedes any prior suggested fix on the session — emit a fresh \
one whenever your top hypothesis changes meaningfully.

**When to emit [SUGGEST_FIX]:**
- You have a concrete resolution path (not just "investigate further")
- Confidence is at least ~50% — below that, keep diagnosing
- Either a known Script Library template applies, OR you can draft a script \
that resolves the issue end-to-end

**When NOT to emit [SUGGEST_FIX]:**
- You're still narrowing causes and the fix depends on the next answer
- The "fix" is just running another diagnostic — that goes in [ACTIONS]
- Two paths are equally likely — fork or ask first, suggest later

**[SUGGEST_FIX] marker format (one block per response, last one wins):**

[SUGGEST_FIX]
{"title": "Clear cached credentials + rebuild Outlook profile", "description": "Stale cached credential in Credential Manager is holding the pre-reset token. Clearing it and recreating the profile completes the password change.", "confidence": 94, "script_template_slug": "clear-outlook-credentials"}
[/SUGGEST_FIX]

- `title`: short imperative summary, ≤ 200 chars
- `description`: one short paragraph explaining the root cause and the fix
- `confidence`: integer 0-100 (what you'd bet this resolves the ticket)
- `script_template_slug`: slug of an existing Script Library template if one \
applies; OMIT or set null otherwise
- `ai_drafted_script`: full script body if no template matches (only when \
`script_template_slug` is null/omitted)
- `ai_drafted_parameters`: optional JSON object of suggested parameter values \
for the drafted script

The marker is stripped from display — the engineer sees the suggested fix as \
an interactive card with confidence badge, not raw JSON.

## Using the Team's Flow Library
Your team has built troubleshooting flows in ResolutionFlow. When relevant flows \
appear in the context below, reference them by name so the engineer can launch them \
directly. Prefer the team's proven flows over ad-hoc instructions when they exist.

## Using Microsoft Learn Documentation
You have access to Microsoft's official documentation via Microsoft Learn. Use it when:
- The question involves exact cmdlet syntax, API parameters, or configuration steps
- You need to verify current Microsoft/Azure behavior or requirements
- No team flow covers the topic and vendor-specific detail would help
Do NOT use Microsoft Learn for every question — only when official docs add real value.

## Image Analysis
When an image is attached, analyze it carefully. Screenshots of error messages, \
config panels, event viewer logs, and network diagrams are common in MSP work. \
Describe what you see and use the visual information to inform your troubleshooting advice.

## Diagnostic Forking
When symptoms point to 2+ different subsystems or root causes, you MUST create a diagnostic \
fork. Forking tracks the different investigation paths in the background — the engineer \
sees them in a sidebar and can switch between them anytime.

**IMPORTANT: Forking is invisible to the engineer in the conversation.** You do NOT mention \
forking, branching, or paths to the engineer. You just continue the conversation naturally. \
The fork marker is metadata that the system uses behind the scenes.

**You MUST fork when:**
- Symptoms affect multiple applications or layers (e.g., Outlook AND Teams dropping)
- The problem could be endpoint-side OR infrastructure-side
- Multiple well-known causes match the exact same symptom pattern

**Do NOT fork when:**
- One cause is clearly >80% likely — just investigate that first
- A single yes/no question would eliminate all but one possibility

**Fork response format:**
Even when forking, you MUST still follow the RESPONSE FORMAT above. Your response \
must include [QUESTIONS] and/or [ACTIONS] markers — the fork marker is IN ADDITION \
to those, not a replacement. Do NOT ask questions in prose — put them in [QUESTIONS].

Structure: 1-3 sentences of analysis → [QUESTIONS] and/or [ACTIONS] → [FORK] at the very end.

Example flow:
- Engineer: "Outlook disconnects every 15 min, Teams drops too, only one user"
- You: "The 10-15 min pattern with both apps points to network layer."
- Then: [QUESTIONS] marker, then [ACTIONS] marker, then [FORK] marker last.

The fork marker is stripped from display — the engineer never sees it. \
The system creates branches silently. Based on the engineer's answer, you pick \
the most relevant branch to investigate first.

To create a fork, append this marker AFTER your [QUESTIONS]/[ACTIONS] markers:

[FORK]
{"fork_reason": "Brief reason", "options": [{"label": "Short name", "description": "One sentence"}, {"label": "Another", "description": "One sentence"}]}
[/FORK]

2-4 options. Never mention "fork", "branch", or "path" in your visible text.

## Boundaries
- Stay focused on IT infrastructure, systems administration, and MSP operations.
- If a question is clearly outside your domain, say so briefly and redirect.
- Never fabricate error codes, KB article numbers, or CLI flags. If unsure, say so.

## FINAL REMINDER — THIS OVERRIDES EVERYTHING ABOVE
Every single response MUST contain [QUESTIONS] and/or [ACTIONS] markers with valid JSON. \
No exceptions. Not even when forking. A response without at least one of these markers \
will crash the UI. If you are unsure, include both. The markers are REQUIRED output, not optional.
If any tasks in the engineer's message are marked `_(not yet completed)_`, re-include them \
in your markers unless you are ≥75% confident that information is no longer relevant.
[PROMOTE] markers are OPTIONAL and IN ADDITION to the required ones — emit them only \
when the engineer's most recent message confirmed something worth recording, and copy \
the originating item's `id` into `source_ref` verbatim.
[SUGGEST_FIX] is OPTIONAL — emit one at most per response, only when you have a \
concrete proposed resolution at ~50%+ confidence. A new [SUGGEST_FIX] supersedes \
any prior suggested fix.
"""


async def _call_ai(
    system_base: str,
    rag_context: str,
    history: list[dict[str, Any]],
    new_message: str,
    max_tokens: int = 4096,
    images: list[dict[str, Any]] | None = None,
) -> tuple[str, int, int]:
    """Call the AI with prompt caching when using Anthropic.

    Caching strategy:
    - System prompt base: cached (stable across all turns)
    - RAG context: NOT cached (changes per query)
    - Conversation history prefix: cached via breakpoint on last
      existing message (stable — only new user message is uncached)

    Args:
        images: Optional list of {"media_type": str, "data": str (base64)}
                to include alongside the new_message as vision content.
    """
    if settings.AI_PROVIDER == "anthropic" and settings.ANTHROPIC_API_KEY:
        return await chat_call_cached(
            system_base, rag_context, history, new_message, max_tokens,
            images=images,
        )

    # Fallback: generic provider (Gemini, etc.) — images not supported
    from app.core.ai_provider import get_ai_provider

    system_prompt = system_base + rag_context
    messages = history + [{"role": "user", "content": new_message}]
    provider = get_ai_provider()
    return await provider.generate_text(
        system_prompt=system_prompt,
        messages=messages,
        max_tokens=max_tokens,
    )


# Appended to every chat turn's user message immediately before generation.
# Invisible to storage (unified_chat_service strips markers before persisting),
# but critical for structured output compliance — the model emits invalid
# responses often enough without it that removing this reminder regresses UX.
_CHAT_FORMAT_REMINDER = (
    "\n\n[SYSTEM: Remember — your response MUST end with [QUESTIONS] "
    "and/or [ACTIONS] markers containing valid JSON arrays. "
    "Responses without markers break the UI.]"
)


async def chat_call_cached(
    system_base: str,
    rag_context: str,
    history: list[dict[str, Any]],
    new_message: str,
    max_tokens: int,
    images: list[dict[str, Any]] | None = None,
) -> tuple[str, int, int]:
    """Call Anthropic's chat surface with caching, MCP, images, and retry-without-MCP.

    This is the ONE MCP/beta/multimodal chat caller. It is deliberately NOT
    routed through `AnthropicProvider`. See module docstring for rationale.

    Responsibilities unique to this function (not in the provider):
    - Anthropic beta endpoint (`client.beta.messages.create`)
    - Microsoft Learn MCP connector wiring (optional via ENABLE_MCP_MICROSOFT_LEARN)
    - Retry-without-MCP fallback when the MCP server misbehaves
    - Multimodal image blocks in the user message
    - Format-reminder append for structured-output compliance
    - Telemetry (`mcp.turn`, `mcp.fallback`) for Phase 0.5 MCP usage signal

    Cache plumbing is shared with the provider via helpers in `ai_provider`:
    `_normalize_system_for_anthropic` (policy α — ephemeral on first block if
    none specified), `build_anthropic_chat_messages` (history cache breakpoint +
    multimodal user message + format reminder), `_log_anthropic_cache_usage`.
    """
    import anthropic

    client = _get_anthropic_client(
        settings.ANTHROPIC_API_KEY,
        timeout=settings.AI_REQUEST_TIMEOUT_SECONDS,
    )

    # System prompt as structured blocks. The static base is cacheable; the
    # RAG context changes per query and must NOT be cached — so we mark the
    # base explicitly and leave the RAG block unmarked. `_normalize_system`
    # honors caller-authored cache_control verbatim (policy α).
    system_blocks: list[dict[str, Any]] = [
        {
            "type": "text",
            "text": system_base,
            "cache_control": {"type": "ephemeral"},
            # cacheable: static system prompt, stable across all turns of all sessions
        },
    ]
    if rag_context:
        system_blocks.append(
            {"type": "text", "text": rag_context}
            # uncached: RAG retrieval varies per query
        )
    normalized_system = _normalize_system_for_anthropic(system_blocks)

    messages = build_anthropic_chat_messages(
        history=history,
        new_message=new_message,
        images=images,
        format_reminder=_CHAT_FORMAT_REMINDER,
    )

    # MCP server config (optional — controlled by settings)
    mcp_servers = anthropic.NOT_GIVEN
    tools = anthropic.NOT_GIVEN

    if settings.ENABLE_MCP_MICROSOFT_LEARN:
        mcp_servers = [
            {
                "type": "url",
                "url": "https://learn.microsoft.com/api/mcp",
                "name": "microsoft-learn",
            }
        ]
        tools = [
            {
                "type": "mcp_toolset",
                "mcp_server_name": "microsoft-learn",
            }
        ]

    _mcp_active = mcp_servers is not anthropic.NOT_GIVEN
    _mcp_fallback_triggered = False

    try:
        response = await client.beta.messages.create(
            model=settings.AI_MODEL_ANTHROPIC,
            max_tokens=max_tokens,
            system=normalized_system,
            messages=messages,
            mcp_servers=mcp_servers,
            tools=tools,
            betas=["mcp-client-2025-11-20"],
        )
    except Exception as e:
        # MCP server failures surface as many error types — BadRequestError,
        # APIStatusError, APIConnectionError, APITimeoutError.  Always retry
        # without MCP when MCP was active, so a flaky external server never
        # blocks the assistant entirely.
        _is_mcp_error = _mcp_active and (
            "MCP server" in str(e)
            or "mcp" in type(e).__name__.lower()
            or isinstance(e, (anthropic.BadRequestError, anthropic.APIStatusError))
        )
        if _is_mcp_error:
            _mcp_fallback_triggered = True
            logger.warning(
                "MCP server error (%s), retrying without MCP: %s",
                type(e).__name__, e,
            )
            # Phase 0.5 telemetry: per-turn fallback event.
            logger.info(
                "mcp.fallback",
                extra={
                    "event": "mcp.fallback",
                    "mcp_error_type": type(e).__name__,
                    "mcp_error_message": str(e)[:500],
                },
            )
            response = await client.messages.create(
                model=settings.AI_MODEL_ANTHROPIC,
                max_tokens=max_tokens,
                system=normalized_system,
                messages=messages,
            )
        else:
            raise

    # Extract text from response — MCP responses can have multiple block
    # types (text, mcp_tool_use, mcp_tool_result). We join all text blocks.
    text_parts = []
    mcp_tools_used = []
    for block in response.content:
        if hasattr(block, "text"):
            text_parts.append(block.text)
        if getattr(block, "type", None) == "mcp_tool_use":
            mcp_tools_used.append(getattr(block, "name", "unknown"))

    text = "\n".join(text_parts) if text_parts else ""

    usage = response.usage
    input_tokens = usage.input_tokens
    output_tokens = usage.output_tokens

    # Phase 0.5 telemetry: per-turn MCP event. Emitted for every turn that
    # reached this code path (i.e., AI_PROVIDER=anthropic chat). `mcp_available`
    # reflects whether MCP was actually wired into the request (scope (ii) from
    # the Phase 0.5 design — Anthropic code path AND flag on). `mcp_invoked`
    # reflects whether the model chose to call an MCP tool on this turn.
    logger.info(
        "mcp.turn",
        extra={
            "event": "mcp.turn",
            "mcp_available": _mcp_active,
            "mcp_invoked": bool(mcp_tools_used),
            "mcp_tools": mcp_tools_used,
            "mcp_fallback_triggered": _mcp_fallback_triggered,
        },
    )

    # Human-readable log retained for grep-based inspection.
    if mcp_tools_used:
        logger.info("MCP tools used: %s", ", ".join(mcp_tools_used))

    _log_anthropic_cache_usage(usage, settings.AI_MODEL_ANTHROPIC)

    return text, input_tokens, output_tokens


def _auto_title(message: str) -> str:
    """Generate a short title from the first user message."""
    title = message.strip()[:100]
    if len(message) > 100:
        title = title.rsplit(" ", 1)[0] + "..."
    return title