"""Chat transcripts, one file per thread, in the per-student dataset storage.

Layout, alongside the profile that's already there:

    students/{student_id}/chats/index.json          — the thread list
    students/{student_id}/chats/{thread_id}.json    — one conversation

Every write is a git commit in the dataset repo (`server/storage.py`), so the
budget is **one write per completed turn** — never per token. The index is only
rewritten when the sidebar's ordering would actually change, which for a normal
conversation means once, when it starts.

This is also the shape the planned "My Story" feature needs: a durable,
timestamped record of what a student asked and what they were pointed toward.
"""
from __future__ import annotations

import logging
import re
import secrets
from datetime import datetime, timezone

from .. import storage

log = logging.getLogger("foresight.agent")

MAX_THREADS_LISTED = 50


def _now() -> str:
    return datetime.now(timezone.utc).isoformat(timespec="seconds")


def new_thread_id() -> str:
    return "t_" + secrets.token_hex(6)


def _valid(thread_id: str) -> bool:
    """Thread ids become path segments — never let one traverse."""
    return bool(re.fullmatch(r"t_[0-9a-f]{12}", thread_id or ""))


def _path(student_id: str, name: str) -> str:
    return storage.student_path(student_id, f"chats/{name}")


def title_from(question: str) -> str:
    """A thread title from the first question. Cheap on purpose — a second model
    call to name a conversation isn't worth the latency or the tokens."""
    words = (question or "").strip().split()
    title = " ".join(words[:7])
    if len(words) > 7:
        title += "…"
    return title[:80] or "New conversation"


# --- read -------------------------------------------------------------------
def load(student_id: str, thread_id: str) -> dict | None:
    """The transcript, or None if there isn't one to show.

    A deleted thread reads back as its tombstone (see `delete`), which is not a
    conversation — returning it would 200 an empty transcript to the client and
    let a new turn resurrect the thread the student threw away.
    """
    if not _valid(thread_id):
        return None
    data = storage.read_json(_path(student_id, f"{thread_id}.json"))
    if not data or data.get("deleted_at"):
        return None
    return data


def index(student_id: str) -> list[dict]:
    data = storage.read_json(_path(student_id, "index.json")) or {}
    threads = data.get("threads") or []
    threads.sort(key=lambda t: t.get("updated_at") or "", reverse=True)
    return threads[:MAX_THREADS_LISTED]


def most_recent_id(student_id: str) -> str | None:
    """The conversation to fall back to when a session doesn't know where it was
    — a fresh sign-in, or the same student on a second device."""
    listed = index(student_id)
    return listed[0].get("thread_id") if listed else None


# --- write ------------------------------------------------------------------
def append_turn(student_id: str, thread_id: str | None, question: str,
                answer: dict, partial: bool = False) -> dict:
    """Persist one exchange and return the saved thread.

    `answer` is the runner's final event: text, sources, suggestion, tools.

    `partial` marks an answer the student walked out on — they reloaded or closed
    the tab while it was still streaming, so what we have is however far it got.
    It's saved rather than dropped: the exchange was on their screen, and finding
    it missing when they come back is worse than finding it cut short.
    """
    created = False
    if not thread_id or not _valid(thread_id):
        thread_id, created = new_thread_id(), True

    thread = load(student_id, thread_id) if not created else None
    if thread is None:
        thread = {"thread_id": thread_id, "created_at": _now(),
                  "title": title_from(question), "messages": []}
        created = True

    stamp = _now()
    thread["messages"].append({"role": "user", "text": question, "at": stamp})
    reply = {
        "role": "assistant",
        "text": answer.get("text") or "",
        "at": stamp,
        "sources": answer.get("sources") or [],
        "suggestion": answer.get("suggestion"),
        "tools": answer.get("tools") or [],
    }
    if partial:
        reply["partial"] = True
    thread["messages"].append(reply)
    thread["updated_at"] = stamp

    storage.write_json(_path(student_id, f"{thread_id}.json"), thread,
                       message=f"chat: turn in {thread_id}")
    _touch_index(student_id, thread)
    return thread


def _touch_index(student_id: str, thread: dict) -> None:
    """Keep the index ordered by last activity, for as few commits as possible.

    An index write is a second commit per turn, so we only pay for one when the
    ordering would actually change: a brand-new thread, or a turn in a thread
    that isn't already at the top. Continuing the conversation you're already in
    — the common case by far — still costs one commit per turn, which is what the
    write budget here has always been.

    The consequence is that an entry's `updated_at` is not "last message"; it's
    "last time this thread moved to the top", which is exactly what ordering
    needs and nothing more. Don't render it as a last-activity time. `created_at`
    is carried alongside it precisely so the sidebar has an honest stamp to show
    ("Started …"), and `load()` always has the real per-message timestamps.
    """
    data = storage.read_json(_path(student_id, "index.json")) or {"threads": []}
    threads = data.get("threads") or []
    if threads and threads[0].get("thread_id") == thread["thread_id"]:
        return                      # already first; a rewrite would change nothing

    # `preview` is the opening question and `answer` the first reply — both are
    # the *first* of their kind, never the latest, so an entry stays true however
    # long the conversation runs and however rarely this function rewrites it.
    # The title is the question too, so a sidebar showing both lines wants the
    # answer on the second one; `preview` stays for entries written before that.
    preview = answer = ""
    for m in thread.get("messages", []):
        if not preview and m.get("role") == "user":
            preview = (m.get("text") or "")[:120]
        elif not answer and m.get("role") == "assistant":
            answer = " ".join((m.get("text") or "").split())[:140]
        if preview and answer:
            break
    rest = [t for t in threads if t.get("thread_id") != thread["thread_id"]]
    rest.insert(0, {"thread_id": thread["thread_id"], "title": thread["title"],
                    "created_at": thread.get("created_at"),
                    "updated_at": thread.get("updated_at"),
                    "preview": preview, "answer": answer})
    data["threads"] = rest[:MAX_THREADS_LISTED]
    try:
        storage.write_json(_path(student_id, "index.json"), data,
                           message=f"chat: index {thread['thread_id']}")
    except Exception:
        # The transcript is already saved; a missing index entry is recoverable
        # and shouldn't fail the student's turn.
        log.exception("chat: failed to update the thread index")


def delete(student_id: str, thread_id: str) -> bool:
    """Remove a thread from the index and blank its transcript.

    `storage` has no delete — it's an upload-only wrapper over the dataset repo —
    so the transcript is overwritten with a tombstone rather than removed. The
    repo's history keeps the old content either way; this is "gone from the app",
    not "erased from git".
    """
    if not _valid(thread_id):
        return False
    data = storage.read_json(_path(student_id, "index.json")) or {"threads": []}
    before = len(data.get("threads", []))
    data["threads"] = [t for t in data.get("threads", []) if t.get("thread_id") != thread_id]
    storage.write_json(_path(student_id, "index.json"), data,
                       message=f"chat: delete {thread_id}")
    storage.write_json(_path(student_id, f"{thread_id}.json"),
                       {"thread_id": thread_id, "deleted_at": _now(), "messages": []},
                       message=f"chat: tombstone {thread_id}")
    return len(data["threads"]) < before