MeshTerm/meshterm/tools/trace.py

"""The trace tools: path traces, live in the menu, one-shot on the CLI.

A trace is one walked path — the protocol has no destination field, so "target" is
purely a UX notion — and the menu splits the feature along exactly that line:

* **Trace target** (``trace``): *can I reach this node?* Pick a target from a
  recency-ordered list; the live screen composes/explores symmetric routes that turn
  at the target and come home over the mirrored hops. The list stays pushed underneath
  for the whole visit, so Esc from a walk lands back on the row that launched it —
  freshly re-sorted by what was just traced — and the next node is one Enter away.
* **Trace path** (``trace-path``): *how far can a route I build carry?* No target and
  no picker — the whole walk is composed hop by hop and only has to end within our
  own earshot. Traces record under the ``(path)`` sentinel, keeping composed walks
  out of the target picker's history.

Both open the live trace screen (:mod:`meshterm.ui.trace_screen`) armed but idle:
nothing transmits until Trace is committed, which runs the chosen sample count
(1–8 traces, paced between transmissions — repeaters penalize, and can blacklist,
nodes that burst traffic, so multi-trace runs always sleep a cooldown between sends).
On the CLI each stays a scriptable one-shot: run one trace, print the route and
per-hop SNR, exit.

Both front ends persist identically: one ``runs`` row per trace, recorded under it, so
stored history reads the same no matter where it came from.
"""

from __future__ import annotations

from datetime import datetime
from typing import TYPE_CHECKING, Any

import typer

from ..context import AppContext
from ..core.models import LOCAL_DEVICE_LABEL, PATH_TRACE_TARGET, Contact, TraceStats
from ..services import trace_runner
from ..ui.widgets import stats_panel
from .base import Tool, ToolResult, register

if TYPE_CHECKING:  # pragma: no cover - typing only
    from ..ui.contactlist import ContactListScreen, ContactRow

#: The characters a stored trace target must consist of to be treated as a hex key
#: prefix when folding it back to a contact name in the target picker.
_HEX_DIGITS = frozenset("0123456789abcdef")


@register
class TraceTool(Tool):
    """Trace the path to a target and watch per-hop SNR, live or scripted."""

    name = "trace"
    title = "Trace target"
    icon = "🎯"
    help = "Trace a target and watch per-hop SNR"
    category = "Explore"
    order = 30  # see a node above, walk to it here

    async def prompt_params(self, ctx: AppContext) -> dict[str, Any] | None:
        """Nothing to gather here — the target picker lives inside :meth:`run`.

        The picker has to *stay pushed* while the live screen it opens runs: that is what
        makes Esc out of a trace one pop back onto the row it was launched from. A prompt
        gathered here would have to resolve — and pop — before the tool runs, so the picker
        moved into :meth:`_run_live` along with the loop that owns it.

        Args:
            ctx: Shared application context.

        Returns:
            ``{"live": True}`` — the menu's marker for the interactive path.
        """
        return {"live": True}

    async def execute(self, ctx: AppContext, params: dict[str, Any]) -> ToolResult:
        """Run — without the outer run row on the live path.

        The live screen opens one ``runs`` row *per trace* (matching what a scripted
        invocation records), so wrapping the whole screen session in another row would
        double-log it. Scripted runs keep the base class's logging.

        Args:
            ctx: Shared application context.
            params: Parameters for this invocation.

        Returns:
            The :class:`ToolResult` from :meth:`run`.
        """
        if params.get("live"):
            return await self.run(ctx, params)
        return await super().execute(ctx, params)

    async def run(self, ctx: AppContext, params: dict[str, Any]) -> ToolResult:
        """Open the live screen (menu) or run one persisted trace (CLI).

        Args:
            ctx: Shared application context.
            params: ``live`` from the menu (the target is picked on the screen); or
                ``target`` (str), optional ``path``, and the injected ``_run_id``
                from the CLI.

        Returns:
            A :class:`ToolResult` with the trace's outcome.
        """
        if params.get("live"):
            return await self._run_live(ctx)
        return await self._run_cli(ctx, params)

    # -- interactive (menu) -------------------------------------------------------

    async def _run_live(self, ctx: AppContext) -> ToolResult:
        """Loop the target picker and the live screen until the picker itself is left.

        The picker **stays pushed** for the whole visit (the app's hub idiom), so the live
        trace screen nests *above* it: Esc out of a trace is one pop, landing back on the
        row it was launched from with the cursor, the sort and any typed filter intact, and
        Esc from the picker is the second pop, out to the menu. The picker used to be a
        one-shot prompt that resolved and popped before the screen opened, which left the
        trace screen sitting directly on the menu — so a single Esc skipped the picker
        entirely and there was no way back to the list except reopening the tool.

        Every returning trace ages the row it walked, and the list is *sorted* by that
        column, so the lanes are re-read and swapped in place afterwards: the node just
        traced rises to the top of the ``TRACED`` sort with the highlight riding it. The
        ages come from stored history alone, so coming back costs no device round-trip.

        Args:
            ctx: Shared application context.

        Returns:
            A :class:`ToolResult` counting the screens opened and the traces run.
        """
        from ..ui.trace_screen import open_trace
        from ..ui.tui.screen import CANCEL

        built = await self._build_picker(ctx)
        if built is None:
            # Nothing to list (or no full-screen session): the free-text prompt is a
            # one-shot, with no list underneath to come back to.
            target = await self._prompt_target(ctx)
            if target is None:
                return ToolResult(summary={})
            return ToolResult(summary={"sessions": 1, "traces": await open_trace(ctx, target)})

        picker, contacts = built
        sessions = 0
        traces = 0
        async with ctx.ui.session.stay(picker) as visit:
            while True:
                chosen = await visit.result()
                if chosen is CANCEL or chosen is None:  # Esc — the pop out to the menu
                    return ToolResult(summary={"sessions": sessions, "traces": traces})
                traces += await open_trace(ctx, chosen.name)
                sessions += 1
                picker.update_rows(self._picker_rows(ctx, contacts))

    async def _prompt_target(self, ctx: AppContext) -> str | None:
        """Ask for a target by hand — the way in when there is no list to pick from.

        Reached with no known contacts at all (an empty list would be nothing to pick from)
        and on any surface without a full-screen session. It is also the only way to trace a
        bare key prefix for a node the device doesn't carry as a contact.

        Args:
            ctx: Shared application context.

        Returns:
            The typed target, or ``None`` if it was left blank or cancelled.
        """
        entered = await ctx.ui.text("Target node (name or key prefix):")
        return entered.strip() if entered else None

    async def _build_picker(
        self, ctx: AppContext
    ) -> tuple[ContactListScreen, list[Contact]] | None:
        """Build the trace-target picker, or ``None`` when there is no list to draw.

        The same ``NAME · TRACED · HEARD · PKTS · KEY`` lanes, Ctrl+arrow sort ring, and
        type-to-filter the Contacts screen and the courier recipient draw — but with every
        node type listed (a trace answers *can I reach this node?* for a repeater or room
        just as for a companion) and an extra ``TRACED`` lane: how long ago each node was
        last traced. The list opens sorted by ``TRACED`` descending, so the most recently
        traced node leads and never-traced ones gather at the bottom. Enter commits the
        highlighted node as the target.

        Args:
            ctx: Shared application context.

        Returns:
            The screen paired with the contacts it lists (kept so the lanes can be
            re-read after a trace), or ``None`` when there is nothing to list — see
            :meth:`_prompt_target`.
        """
        from ..ui.contactlist import (
            TRACE_LANES,
            TRACE_SORT_COLUMNS,
            TRACE_SORT_OPENS_ASCENDING,
            ContactListScreen,
        )
        from ..ui.surface import TuiUi
        from ..ui.timemachine_screen import _routing_prefix_bytes
        from ..ui.widgets import ContactsSort

        # Through the session cache: this picker runs on every Trace open, and the contacts
        # table is a slow round-trip on a busy node — re-reading it here (in front of the
        # already-cached trace screen) is what kept opening Trace feeling like a stall. See
        # :class:`~meshterm.services.device_state.DeviceState`.
        contacts = await ctx.devstate.contacts()
        if not contacts or not isinstance(ctx.ui, TuiUi):
            return None

        picker = ContactListScreen(
            "Trace target — pick a target",
            rows=self._picker_rows(ctx, contacts),
            prefix_bytes=await _routing_prefix_bytes(ctx),
            sort=ContactsSort.from_name("traced", TRACE_SORT_COLUMNS, TRACE_SORT_OPENS_ASCENDING),
            footer_hint="↑↓ move · ^←→↑↓ sort · type to filter · Enter select · Esc back",
            lanes=TRACE_LANES,
        )
        return picker, contacts

    @staticmethod
    def _picker_rows(ctx: AppContext, contacts: list[Contact]) -> list[ContactRow]:
        """The picker's lane data for ``contacts``, read fresh from stored history.

        Both the ``TRACED`` ages and the packet counts come from the repository, so this is
        cheap enough to redo every time a trace hands the picker back — which is what keeps
        the lane the list is *sorted* by honest about the walk that just happened.

        Args:
            ctx: Shared application context.
            contacts: The device's contacts, as listed.

        Returns:
            One :class:`~meshterm.ui.contactlist.ContactRow` per contact, in the order
            given (the screen sorts them).
        """
        from ..ui.contactlist import ContactRow
        from ..ui.widgets import _contact_pkts

        traced = _last_traced_by_name(contacts, ctx.repo.target_last_traced())
        counts = {n.node: n.count for n in ctx.repo.heard_nodes() if n.node}
        return [
            ContactRow(
                value=c,
                name=c.name,
                key=c.public_key or c.key_prefix or "",
                node_type=c.node_type,
                last_seen=c.last_seen,
                count=_contact_pkts(c, counts),
                last_traced=traced.get(c.name),
            )
            for c in contacts
        ]

    # -- scripted (CLI) -------------------------------------------------------------

    async def _run_cli(self, ctx: AppContext, params: dict[str, Any]) -> ToolResult:
        """Run one trace, persist it, and print the route and per-hop summary.

        One transmission per invocation is a hard rule on the CLI — repeaters
        penalize (and can blacklist) nodes that burst traffic — so scripted sampling
        means invoking again (with your own pacing between runs).

        Args:
            ctx: Shared application context.
            params: ``target`` (str), optional ``path``, and the injected ``_run_id``.

        Returns:
            A :class:`ToolResult` with the trace's outcome.
        """
        return await _trace_once_cli(
            ctx, params["_run_id"], target=params["target"], path_spec=params.get("path")
        )

    def register_cli(self, app: typer.Typer) -> None:
        """Register the ``trace`` subcommand.

        Args:
            app: The Typer application.
        """
        from ..cli import run_tool_command

        @app.command(
            name=self.name, help="Run a single path trace to a target and show per-hop SNR"
        )
        def _trace(
            target: str = typer.Option(..., "--target", "-t", help="Target node name/prefix"),
            path: str | None = typer.Option(
                None,
                "--path",
                "-p",
                help="Force a route: comma-separated contact names/hex prefixes (e.g. 3d,f2,3d)",
            ),
        ) -> None:
            tool_params: dict[str, Any] = {"target": target}
            if path:
                tool_params["path"] = path
            run_tool_command(self, tool_params)


@register
class TracePathTool(Tool):
    """Walk a hand-composed route — no target — and see how far it carries."""

    name = "trace-path"
    title = "Trace path"
    icon = "👣"
    help = "Walk a route you compose, hop by hop"
    category = "Explore"
    order = 32

    async def prompt_params(self, ctx: AppContext) -> dict[str, Any] | None:
        """No parameters to gather — a path walk has no target to pick.

        Args:
            ctx: Shared application context.

        Returns:
            ``{"live": True}`` — the route itself is composed on the screen.
        """
        return {"live": True}

    async def execute(self, ctx: AppContext, params: dict[str, Any]) -> ToolResult:
        """Run — without the outer run row on the live path (see :class:`TraceTool`).

        Args:
            ctx: Shared application context.
            params: Parameters for this invocation.

        Returns:
            The :class:`ToolResult` from :meth:`run`.
        """
        if params.get("live"):
            return await self.run(ctx, params)
        return await super().execute(ctx, params)

    async def run(self, ctx: AppContext, params: dict[str, Any]) -> ToolResult:
        """Open the live path-walk screen (menu) or run one persisted trace (CLI).

        Args:
            ctx: Shared application context.
            params: ``live`` from the menu; or ``path`` (str, required) and the
                injected ``_run_id`` from the CLI.

        Returns:
            A :class:`ToolResult` with the walk's outcome.
        """
        if params.get("live"):
            from ..ui.trace_screen import open_trace_path

            traces = await open_trace_path(ctx)
            return ToolResult(summary={"traces": traces})
        return await _trace_once_cli(
            ctx, params["_run_id"], target=PATH_TRACE_TARGET, path_spec=params["path"]
        )

    def register_cli(self, app: typer.Typer) -> None:
        """Register the ``trace-path`` subcommand.

        Args:
            app: The Typer application.
        """
        from ..cli import run_tool_command

        @app.command(
            name=self.name,
            help="Walk a composed route once (no target) and show per-hop SNR",
        )
        def _trace_path(
            path: str = typer.Option(
                ...,
                "--path",
                "-p",
                help="The whole walk: comma-separated contact names/hex prefixes "
                "(must end within earshot of this node)",
            ),
        ) -> None:
            run_tool_command(self, {"path": path})


async def _trace_once_cli(
    ctx: AppContext, run_id: int, *, target: str, path_spec: str | None
) -> ToolResult:
    """Run one persisted trace and print the route and per-hop summary (CLI body).

    Shared by both scripted trace commands: ``trace`` passes its target (path
    optional — the device routes without one), ``trace-path`` passes the
    :data:`~meshterm.core.models.PATH_TRACE_TARGET` sentinel with a required path.

    Args:
        ctx: Shared application context.
        run_id: The already-open run row to record under.
        target: The label the trace persists under.
        path_spec: The forced route as typed (names/hex, comma-separated), if any.

    Returns:
        A :class:`ToolResult` with the trace's outcome.
    """
    is_walk = target == PATH_TRACE_TARGET
    device = await ctx.device()

    # Our own device's name labels both endpoints of the path (first hop's origin,
    # last hop's destination); fall back to a neutral label if it's unavailable.
    self_info = await device.get_self_info()
    device_label = str(self_info.get("name") or LOCAL_DEVICE_LABEL)
    # Our own public key, so the route's endpoints (us) carry a hash like every
    # other hop, addressed at the same path-hash width.
    device_hash = str(self_info.get("public_key") or "") or None

    # Resolve repeater hashes in the results to contact names where we know them,
    # so the tables read as names instead of opaque hex prefixes.
    contacts = await device.get_contacts()
    resolve = trace_runner.make_node_resolver(contacts)

    # Resolve an optional forced path (names/hex) to the hex string the radio wants.
    # Leaving it blank is fully supported for a *target* trace: the device reuses the
    # route it already learned (or floods if it has none). A path walk always has one.
    path: str | None = None
    if path_spec:
        path = trace_runner.parse_trace_path(path_spec, contacts)
        ctx.ui.note(f"[muted]forcing path:[/muted] [brand]{path}[/brand]")
    else:
        ctx.ui.note("[muted]path: auto (device-routed)[/muted]")

    with ctx.ui.progress("trace") as progress:
        task = progress.add_task("walking the path" if is_walk else f"tracing {target}", total=1)
        result = await device.run_trace(target, path=path)
        ctx.repo.record_trace(run_id, result)
        progress.advance(task)

    # The stats panel renders route + per-hop readings for the single trace (its
    # medians collapse to the readings themselves).
    stats = TraceStats.from_traces(target, [result])
    ctx.ui.show(
        stats_panel(
            stats,
            device_label,
            resolve,
            route=result if result.success else None,
            device_hash=device_hash,
        )
    )

    summary: dict[str, Any] = {
        "target": target,
        "success": result.success,
        "hops": result.hop_count if result.success else None,
        "min_snr": result.min_snr,
        "rtt_ms": result.round_trip_ms,
    }
    if path:
        summary["path"] = path
    via = f" via [brand]{path}[/brand]" if path else ""
    if is_walk:
        if result.success:
            message = f"[ok]✓[/ok] the path came home{via}"
        else:
            message = f"[err]✗[/err] no reply{via}"
    elif result.success:
        message = f"[ok]✓[/ok] traced [brand]{target}[/brand]{via}"
    else:
        message = f"[err]✗[/err] no reply from [brand]{target}[/brand]{via}"
    return ToolResult(summary=summary, message=message)


def _last_traced_by_name(
    contacts: list[Contact], traced: dict[str, datetime]
) -> dict[str, datetime]:
    """Map each contact's name to when that node was last traced.

    The Trace picker's ``TRACED`` lane and default sort read this. Stored trace targets are
    filed exactly as the user addressed them — a contact name one day, a raw hex key prefix
    another — so each target is folded onto the contact it names (by name, case-insensitively,
    or as a prefix of a contact's public key) and the *latest* trace time wins, so a node
    traced under both spellings still shows one honest "last traced" age. This is the inverse
    of the picker's old recent-targets fold, kept per-contact rather than as a name list.

    Args:
        contacts: The device's current contacts.
        traced: ``target → last-traced time`` from
            :meth:`~meshterm.persistence.repository.Repository.target_last_traced`.

    Returns:
        ``contact name → last-traced time`` for every contact the history can place.
    """
    by_fold = {c.name.casefold(): c.name for c in contacts}
    out: dict[str, datetime] = {}

    def note(name: str, when: datetime) -> None:
        current = out.get(name)
        if current is None or when > current:
            out[name] = when

    for target, when in traced.items():
        named = by_fold.get(target.casefold())
        if named is not None:
            note(named, when)
            continue
        needle = target.lower().removeprefix("0x")
        # Only fold plausible key prefixes (≥2 bytes of hex) — a short hex-looking *name*
        # like "ace" must not be mistaken for an address.
        if len(needle) >= 4 and all(ch in _HEX_DIGITS for ch in needle):
            for contact in contacts:
                if (contact.public_key or "").lower().startswith(needle):
                    note(contact.name, when)
    return out