"""Agent loop: the grounded-turn document tools (phase 37, task 03; the
harness-aligned ``ls``/``read``/``grep`` surface, phase 70).

Probe verdict (task 01 — ``uv run python -m scripts.llm_probe --tools``
run live against aipi): **``probe: turbo tool_calls=supported 2026-08-26``**
— ``turbo`` answers OpenAI ``tools`` requests with
``finish_reason="tool_calls"`` and streams the calls as indexed
``delta.tool_calls`` partials (id + name on the first partial, arguments
in fragments). This module therefore uses the **native tool-calling
path**: tool calls arrive as :class:`app.rag.llm.ToolCallPiece` values
from ``chat_stream(messages, tools=AGENT_TOOLS)``. The prompt-based
JSON-block fallback (documented in the task file) is *not* implemented —
it exists only for a "not supported"/"intermittent" verdict, and the
probe came back "supported".

Loop contract (one grounded chat turn; the API layer wires this in,
task 04):

1. The model is offered the three OpenAI functions in :data:`AGENT_TOOLS`
   for the whole turn — phase 45 removed the phase-37 per-tool budgets
   (owner permission 2026-08-27, ``TODO.md`` L8: "allow the LLM to make
   as many tool calls as it wants"): ``ls``, ``read`` and ``grep`` can
   each be called as many times as the model needs, re-lists and
   re-greps included. With ``settings.agent_max_rounds``
   (``BOR_AGENT_MAX_ROUNDS``, default 10) at 0 the loop makes exactly
   one request with ``tools=None`` — byte-identical to the
   pre-phase-37 chat path (the kill switch). Phase 70 (owner permission
   2026-09-03: "match existing harnesses as much as possible") renamed
   and reshaped the tools to the harness-trained surface —
   ``ls(path?)`` / ``read(path)`` / ``grep(pattern, path?)``, the
   pi.dev tool shapes the model was trained on: the combined
   ``source/path`` string is the canonical document identity in every
   tool argument, refusal, and result header, and the old two-argument
   split (with its self-correction and "teach the split" refusals) is
   gone — the model's combined form is now simply correct. The
   phase-68 A5 match/output contract rides along under the new name.
2. Each tool call the model emits is executed server-side against
   Postgres only (no LLM, no network): ``ls`` returns the indexed
   catalog — one ``source: X | path: Y | title: Z`` line per document
   (phase 63: labeled fields — unambiguous for LLM parsing),
   ``GET /api/docs`` order (uncapped in v1; the UI never shows it, only
   the model does) — optionally scoped to one source name (a ``path``
   argument matching no source name is a refusal; a registered source
   with no indexed documents lists as ``0 documents:`` and counts) —
   ``read`` takes the combined ``source/path`` string, splits it at the
   FIRST ``'/'`` (source names are directory basenames — they can never
   contain ``'/'``), and returns the document's **full** content
   (A7-revised contract: never truncated) — and ``grep`` greps the
   indexed documents (or the one document a combined ``source/path``
   names) for a case-insensitive fixed substring and returns up to 20
   ``source/path:line: text`` match lines (owner-locked A5, phase 68),
   each line truncated to 200 chars. A grep is a **locator**, not a
   context-adder: it never appends to the answer context (only
   ``read`` does — ``holder.read_docs`` is untouched by a grep).
3. Rejected calls get a one-line refusal and count in nothing
   (``holder.tool_calls`` tracks executed calls only): unknown tool name
   → ``"Unknown tool."``; a ``read`` without a usable ``path`` (missing,
   blank or non-string) → ``"read requires a string argument
   'path'."``; a ``grep`` without a usable ``pattern`` (missing, blank
   or non-string) → ``"grep requires a string argument
   'pattern'."``; a scoped ``ls`` whose ``path`` matches no source name
   → ``"No source named '…' — check the ls output."``; a document
   already in context (seed or previously read) → ``"Already in your
   context."``; an unknown document (a ``read`` or scoped ``grep`` whose
   combined ``source/path`` matches nothing — a bare source name, which
   can never be a document, included) → ``"No document at '…' — check
   the ls output."`` with the argument echoed as passed (the model sees
   its own form). A grep that ran but found nothing is NOT a rejection
   — its ``"No matches for …"`` line is a (counted) result. A rejected
   call still consumes a *round* in the loop, so a pathological stream
   that keeps emitting rejected calls is bounded by the cap (point 4).
4. Every call the model emits is appended back to the message history as
   the assistant tool-call message + the tool result (refusals included),
   consumes one round, and the model is called again. At the round cap —
   ``max_rounds = settings.agent_max_rounds`` (``BOR_AGENT_MAX_ROUNDS``,
   default 10) — the loop forces one final retried no-tools request
   (``chat_stream_retried`` with ``tools=None``) and returns: the cap is
   the **only** forced exit (besides "the stream carried no calls"), and
   it bounds pathological rejected-call streams.
5. A rare stream that carries both content and a tool call keeps the
   content (it was already emitted) **and** still runs the tool.
6. *holder* (an :class:`AgentHolder`) records the read documents and the
   number of executed tool calls (re-lists included); the API layer
   (task 04) reads it after the stream to extend ``done.sources`` /
   ``query_log.sources`` and the per-turn log line (``tool_calls=N``).
7. Retries (phase 67, owner-locked A2): every model request — each tool
   round and the forced final ``tools=None`` call — goes through
   ``chat_stream_retried``: a round that dies before its first piece is
   restarted with the SAME messages (up to ``settings.llm_retries``
   restarts, a flat ``settings.llm_retry_delay`` between attempts, each
   preceded by a :class:`app.rag.llm.RetryPiece` the API layer turns into
   an SSE ``retry`` frame); a round that already streamed a piece fails
   the turn as before (no partial answer is ever redone). Retries are
   invisible to the round cap: a round that needed a retry still consumes
   exactly one round. With ``settings.llm_retries=0`` every request is a
   single plain attempt (the pre-phase-67 path).

Scaffolding guardrail (phase 71, deterministic only — owner permission
2026-09-03: "deterministic guardrails only right now, forget using a
model for that"): every model request (each round, the forced final,
and any recovery) runs its ``delta.content`` through a fresh caller-
owned :class:`app.rag.scaffolding.ScaffoldingFilter`, so raw
``<|tool_call_start|>…<|tool_call_end|>`` tokens can never reach the
user as answer text. A round that ends with NO visible content AND a
non-empty strip (the scaffolding was the whole "answer") gets exactly
ONE bounded recovery: one extra request with ``tools=None``, the same
messages with :data:`CORRECTION_INSTRUCTION` folded into the original
single system message, a fresh filter, and the same phase-67 retry
budget. A recovery that also comes back empty — or a round with no
strip and no content (today's empty/thinking-only answer) — settles as
before; a second empty reply raises :class:`MalformedReplyError` (the
API layer turns it into the dedicated error frame). A round with real
visible content plus scaffolding needs no recovery (the clean content
stands), and a scaffolding-only round that also carried tool calls
needs none either (the tool ran) — the policy keys on the no-calls
exit only. No model participates in detection or repair: the
registry + the fixed retry policy are the whole guardrail.

The DB accessors (:func:`list_catalog`, :func:`list_source_names`,
:func:`find_document`, :func:`all_documents`) and the
:func:`grep_document` line matcher are module-level functions so unit
tests can monkeypatch them without a database.
"""
from __future__ import annotations

import json
import logging
from collections.abc import AsyncIterator, Sequence
from dataclasses import dataclass, field
from typing import Any, cast

from sqlalchemy import select
from sqlalchemy.orm import Session

from app.config import Settings
from app.models import Document
from app.rag.git_sources import effective_sources
from app.rag.llm import (
    LLMClient,
    LLMError,
    RetryPiece,
    StreamPiece,
    ToolCallPiece,
    chat_stream_retried,
)
from app.rag.scaffolding import ScaffoldingFilter
from app.rag.source_removal import resolve_source_name

logger = logging.getLogger("app.agent")

#: The three agent tools (phase 70: the harness-aligned surface —
#: ``ls`` / ``read`` / ``grep``, the pi.dev tool shapes the model was
#: trained on, replacing the phase-37 list/read and phase-68 search
#: names): OpenAI function
#: definitions passed as ``tools=AGENT_TOOLS`` to ``chat_stream`` for
#: the whole grounded turn — phase 45 removed the per-tool budgets; the
#: round cap (``BOR_AGENT_MAX_ROUNDS``) is the only bound. The combined
#: ``source/path`` string is the canonical document identity in every
#: argument (phase 70, owner permission 2026-09-03).
AGENT_TOOLS: list[dict[str, Any]] = [
    {
        "type": "function",
        "function": {
            "name": "ls",
            "description": (
                "List the indexed documents as `source: X | path: Y | "
                "title: Z` lines."
            ),
            "parameters": {
                "type": "object",
                "properties": {
                    "path": {
                        "type": "string",
                        "description": (
                            "Source name to list one source's documents "
                            "(e.g. 'homelab'); omit to list every "
                            "document."
                        ),
                    }
                },
                "required": [],
            },
        },
    },
    {
        "type": "function",
        "function": {
            "name": "read",
            "description": (
                "Add the full content of one indexed document to your "
                "context."
            ),
            "parameters": {
                "type": "object",
                "properties": {
                    "path": {
                        "type": "string",
                        "description": (
                            "The document to add to your context, as the "
                            "combined `source/path` string exactly as "
                            "shown in the `ls` output (e.g. "
                            "'homelab/active/container_caddy/caddy.md')."
                        ),
                    }
                },
                "required": ["path"],
            },
        },
    },
    {
        "type": "function",
        "function": {
            "name": "grep",
            "description": (
                "Search the indexed documents for an exact string "
                "(case-insensitive) and return up to 20 matching lines "
                "as `source/path:line: text` — a locator, not a "
                "context-adder: read the winner with `read`."
            ),
            "parameters": {
                "type": "object",
                "properties": {
                    "pattern": {
                        "type": "string",
                        "description": (
                            "The exact text to search for (a plain "
                            "substring, not a regex)"
                        ),
                    },
                    "path": {
                        "type": "string",
                        "description": (
                            "Limit the search to one document, as a "
                            "combined `source/path` string from the "
                            "`ls` output (omit to search every "
                            "document)."
                        ),
                    },
                },
                "required": ["pattern"],
            },
        },
    },
]

#: Tool refusal texts (phase 37): rejected calls count in nothing
#: (``holder.tool_calls`` tracks executed calls); the round cap bounds
#: their pathological repetition (phase 45).
ALREADY_IN_CONTEXT = "Already in your context."
UNKNOWN_TOOL = "Unknown tool."
MISSING_READ_ARGS = "read requires a string argument 'path'."
MISSING_SEARCH_ARGS = "grep requires a string argument 'pattern'."

#: The harness-owned recovery line (phase 71, task 03) — folded into the
#: ORIGINAL single system message of the one bounded recovery request
#: (``system_prompt + "\n" + CORRECTION_INSTRUCTION``; provider-safe,
#: the user message stays last). Verbatim constant: the E2E mock
#: (task 05) keys on a stable substring of it, so it must not drift.
CORRECTION_INSTRUCTION: str = (
    "Your previous reply contained raw tool-call markup, which is not "
    "interpreted here. Answer the user's question directly in plain "
    "text — no tool syntax."
)


class MalformedReplyError(LLMError):
    """The model kept replying in raw tool-scaffolding (phase 71).

    Raised ONLY by the recovery policy — :func:`run_agent` (grounded
    path) and ``app.api.chat`` (deflected path) — when the one bounded
    ``tools=None`` recovery still comes back with no visible content.
    It is never raised from inside a stream, so
    :func:`app.rag.llm.chat_stream_retried`'s retry-before-first-piece
    rule never sees it. The API layer catches it BEFORE the generic
    :class:`LLMError` handler and settles the turn with the dedicated
    "malformed reply" error frame (no ``done``, no ``query_log`` row).
    Deterministic only (owner permission 2026-09-03): no model
    participates in detection or repair.
    """

#: Search caps (owner-locked A5, phase 68): a global per-call match cap
#: (across documents, in catalog order) and a per-match-line char limit.
SEARCH_MAX_MATCHES = 20
SEARCH_LINE_LIMIT = 200

#: No-match result lines (templates — the pattern is truncated to 100
#: chars before formatting, to keep a long pattern from bloating the
#: tool result). A no-match line is a *result* of an executed grep,
#: not a refusal (see the module docstring, point 3).
NO_MATCHES = "No matches for '{pattern}' in the knowledge base."
NO_MATCHES_SCOPED = "No matches for '{pattern}' in {source}/{path}."


def list_catalog(db: Session) -> list[tuple[str, str, str]]:
    """Every indexed document as ``(source, path, title)``.

    Ordered by ``(source, path)`` — the same order as ``GET /api/docs``.
    Module-level (not a method) so unit tests can monkeypatch it.
    """
    rows = db.execute(
        select(Document.source, Document.path, Document.title).order_by(
            Document.source, Document.path
        )
    ).all()
    return [(source, path, title) for source, path, title in rows]


def list_source_names(db: Session) -> list[str]:
    """Every registered source name, deduped, in registry order.

    The source registry (the ``git_sources`` rows — the
    ``BOR_GIT_SOURCES`` env fallback while the table is empty) is the
    source of truth for *source* names independent of document count:
    a registered source with no indexed documents still lists (as
    ``0 documents:`` — the scoped ``ls`` must not refuse it as unknown).
    Names resolve exactly as the import pipeline indexes them
    (:func:`app.rag.source_removal.resolve_source_name` — reuse, or a
    scoped ``ls`` would judge the wrong names unknown, the phase-69
    "RAG consistent with the registry" invariant); two rows resolving
    to the same name (the phase-69 sibling case) share documents, so
    the name is listed once. Module-level (not a method) so unit tests
    can monkeypatch it.
    """
    rows, _origin = effective_sources(db)
    names: list[str] = []
    for row in rows:
        name = resolve_source_name(row)
        if name not in names:
            names.append(name)
    return names


def find_document(db: Session, source: str, path: str) -> Document | None:
    """The indexed document at ``(source, path)``, or ``None``.

    Module-level (not a method) so unit tests can monkeypatch it.
    """
    return db.scalar(
        select(Document).where(Document.source == source, Document.path == path)
    )


def _resolve_path(db: Session, combined: str) -> tuple[Document | None, str, str]:
    """The combined ``source/path`` identity → document (phase 70).

    The canonical document identity in every tool argument, refusal and
    result header is the combined string exactly as printed in the
    ``ls`` output, the ``Document …`` result headers, and the grep
    result lines. Source names are directory basenames (``app.rag.importer``:
    ``source = root.name``) and can never contain a ``'/'``, so the
    split at the FIRST slash is exact: the part before is the source
    name, the part after is the path. Returns ``(doc, source, path)``
    with the split pair (so callers can echo the canonical form, e.g.
    the scoped no-match line); no ``'/'`` in the argument →
    ``(None, combined, "")`` — a bare source name is never a document
    (no DB lookup; the refusal echoes the argument as passed).
    """
    if "/" not in combined:
        return None, combined, ""
    source, _, path = combined.partition("/")
    return find_document(db, source, path), source, path


def all_documents(db: Session) -> list[Document]:
    """Every indexed document (full rows), ordered by ``(source, path)``
    — catalog order.

    The whole-KB ``grep`` path loads all contents in this one bulk query
    (catalog order is the locked match order, owner-locked A5).
    Module-level (not a method) so unit tests can monkeypatch it.
    """
    return list(
        db.execute(
            select(Document).order_by(Document.source, Document.path)
        ).scalars()
    )


def grep_document(content: str, pattern: str) -> list[tuple[int, str]]:
    """Every line of *content* that contains *pattern*, in file order.

    Case-insensitive **fixed substring** (owner-locked A5: no regex — no
    ReDoS surface, a simple contract for the model). Returns
    ``(1-based line number, line.rstrip())`` pairs; an empty *content*
    never matches a non-empty pattern.
    """
    needle = pattern.lower()
    return [
        (number, line.rstrip())
        for number, line in enumerate(content.split("\n"), start=1)
        if needle in line.lower()
    ]


@dataclass
class AgentHolder:
    """Per-turn agent state the API layer reads after the stream (task 04).

    ``read_docs``: the documents ``read`` added to the context, in read
    order (deduped — re-reading a document appends nothing).
    ``tool_calls``: how many tool calls executed (re-lists included);
    rejected calls (unknown tool, unknown/missing arguments or document,
    already-in-context) do not count. Drives the per-turn log line's
    ``tool_calls=N`` field (task 04).
    ``scaffold_stripped``: how many chars of tool-scaffolding the
    turn's filters removed across the turn's requests (rounds + the
    forced final + any recovery, phase 71) — drives the per-turn log
    line's ``scaffold_stripped=N`` field on grounded turns (the
    deflected path computes its own total in ``app.api.chat``).
    """

    read_docs: list[Document] = field(default_factory=list)
    tool_calls: int = 0
    scaffold_stripped: int = 0


def _execute_tool(
    db: Session,
    call: ToolCallPiece,
    seed_docs: Sequence[Document],
    holder: AgentHolder,
) -> str:
    """Execute one tool call server-side (DB only).

    Returns the tool result text. A successful call bumps
    ``holder.tool_calls`` (a successful ``read`` also appends the
    :class:`Document` to ``holder.read_docs``; a ``grep`` never does —
    it is a locator, locked A5); rejected calls return their refusal
    line and count in nothing. A grep that ran but found nothing is
    still a successful (counted) call — its no-match line is a result,
    not a refusal. Document targets are combined ``source/path``
    strings, resolved by :func:`_resolve_path` (the canonical identity,
    phase 70).
    """
    if call.name == "ls":
        raw_path = call.arguments.get("path")
        scope = raw_path.strip() if isinstance(raw_path, str) else ""
        rows = list_catalog(db)
        if scope:
            if scope not in list_source_names(db):
                return f"No source named '{scope}' — check the ls output."
            rows = [row for row in rows if row[0] == scope]
        listing = f"{len(rows)} documents:\n" + "\n".join(
            f"source: {source} | path: {path} | title: {title}"
            for source, path, title in rows
        )
        holder.tool_calls += 1
        return listing
    if call.name == "read":
        raw_path = call.arguments.get("path")
        arg = raw_path.strip() if isinstance(raw_path, str) else ""
        if not arg:
            return MISSING_READ_ARGS
        known = {(doc.source, doc.path) for doc in (*seed_docs, *holder.read_docs)}
        # The dedupe check needs no DB: the split pair of a combined
        # identity that is in context is in `known` as-is (the resolve
        # below would find the same document).
        if "/" in arg:
            src, _, p = arg.partition("/")
            if (src, p) in known:
                return ALREADY_IN_CONTEXT
        doc, _source, _path = _resolve_path(db, arg)
        if doc is None:
            # Echo the argument as passed — the model sees its own form
            # (a bare source name can never be a document, no DB lookup).
            return f"No document at '{arg}' — check the ls output."
        holder.read_docs.append(doc)
        holder.tool_calls += 1
        return f"Document {doc.source}/{doc.path}:\n{doc.content}"
    if call.name == "grep":
        raw_pattern = call.arguments.get("pattern")
        pattern = raw_pattern.strip() if isinstance(raw_pattern, str) else ""
        if not pattern:
            return MISSING_SEARCH_ARGS
        raw_path = call.arguments.get("path")
        scope = raw_path.strip() if isinstance(raw_path, str) else ""
        scoped_to: tuple[str, str] | None = None
        if scope:
            target, src, p = _resolve_path(db, scope)
            if target is None:
                return f"No document at '{scope}' — check the ls output."
            docs: list[Document] = [target]
            scoped_to = (src, p)  # the resolved (canonical) identity
        else:
            docs = all_documents(db)
        matches: list[str] = []
        for doc in docs:
            for lineno, line in grep_document(doc.content, pattern):
                matches.append(
                    f"{doc.source}/{doc.path}:{lineno}: {line[:SEARCH_LINE_LIMIT]}"
                )
                if len(matches) >= SEARCH_MAX_MATCHES:
                    break
            if len(matches) >= SEARCH_MAX_MATCHES:
                break  # the global cap is hit — stop scanning
        holder.tool_calls += 1  # the grep executed (no-match counts too)
        # Locked A5: a grep never adds context — read_docs untouched.
        if not matches:
            shown = pattern[:100]  # keep a long pattern short in the line
            if scoped_to is not None:
                # The scoped no-match line is keyed on the resolved
                # source/path (== the argument, stripped).
                return NO_MATCHES_SCOPED.format(
                    pattern=shown, source=scoped_to[0], path=scoped_to[1]
                )
            return NO_MATCHES.format(pattern=shown)
        return "\n".join(matches)
    return UNKNOWN_TOOL


async def run_agent(
    llm: LLMClient,
    db: Session,
    *,
    system_prompt: str,
    user_message: str,
    seed_docs: Sequence[Document],
    settings: Settings,
    holder: AgentHolder,
) -> AsyncIterator[StreamPiece | ToolCallPiece | RetryPiece]:
    """Run the grounded-turn tool loop, yielding every stream piece.

    Every piece (``thinking`` / ``content`` / tool calls /
    :class:`RetryPiece`) is yielded as it arrives; the API layer (task 04)
    turns tool-call pieces into SSE ``tool`` events and retry pieces into
    SSE ``retry`` events. After the loop finishes, *holder* carries the
    read documents and the executed tool-call count (re-lists included).

    Retries (phase 67, owner-locked A2): every model request goes through
    :func:`chat_stream_retried` — a failed round is retried **before** its
    first piece (same messages, ``settings.llm_retries`` restarts, a flat
    ``settings.llm_retry_delay``); a round that already streamed pieces
    fails the turn as before.

    Scaffolding recovery (phase 71, deterministic only): every request —
    each round, the forced final, and any recovery — runs its content
    through a fresh :class:`app.rag.scaffolding.ScaffoldingFilter`. A
    round that ends with NO visible content but a non-empty strip (the
    scaffolding was the whole "answer") gets exactly ONE recovery:
    ``tools=None``, :data:`CORRECTION_INSTRUCTION` folded into the
    original single system message (the rest of the history — user
    message and tool results — unchanged), a fresh filter, the same
    retry budget. A clean recovery ends the turn; a second empty reply
    raises :class:`MalformedReplyError` (terminal — the API layer turns
    it into the dedicated error frame). A round with visible content
    plus scaffolding needs no recovery (the clean content stands), and
    a scaffolding-only round that also carried tool calls needs none
    (the tool ran) — the policy keys on the no-calls exit only. The
    per-span strip warning log (each span truncated to 200 chars) is
    the capture mechanism for new registry entries; *holder* accumulates
    the turn's ``scaffold_stripped`` total for the API layer's log line.

    ``seed_docs`` are the documents the retrieval already put in context
    (they shape the *system_prompt* the caller built); re-reading one of
    them is rejected as "Already in your context." — the rejection counts
    in nothing, but it still consumes a round.
    """
    messages: list[dict[str, Any]] = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_message},
    ]
    # Phase 45: no per-tool budgets — the tools stay offered for the
    # whole turn, bounded by the round cap. ``0`` is the no-tools kill
    # switch: exactly one request with ``tools=None`` (the pre-phase-37
    # path).
    max_rounds = settings.agent_max_rounds
    tools: list[dict[str, Any]] | None = AGENT_TOOLS if max_rounds > 0 else None
    rounds = 0
    while True:
        calls: list[ToolCallPiece] = []
        # Phase 71: one fresh filter per round (one per model request —
        # the retry attempts of this logical request share it: a restart
        # only happens while the filter was never fed). Content-only:
        # thinking pieces pass through raw.
        round_filter = ScaffoldingFilter()
        round_content = 0  # visible (clean) content chars this round
        # Phase 48: bind the round's stream so a consumer abandon
        # (GeneratorExit into the yield below) tears down the in-flight
        # model stream deterministically — not GC-dependent. Phase 67:
        # the round goes through the retry primitive — a failure before
        # the first piece restarts the request (locked A2) after a
        # RetryPiece; closing the OUTER generator propagates GeneratorExit
        # into ``chat_stream_retried``, whose own ``finally`` closes the
        # in-flight inner ``chat_stream``, so teardown stays deterministic
        # on consumer abandon. Awaiting ``aclose()`` in the ``finally`` is
        # safe because it does not yield; on a fully consumed round it is
        # a quiet no-op.
        stream = chat_stream_retried(
            llm,
            cast("list[dict[str, str]]", messages),
            tools=tools,
            retries=settings.llm_retries,
            delay=settings.llm_retry_delay,
            scaffolding=round_filter,
        )
        try:
            async for piece in stream:
                if isinstance(piece, ToolCallPiece):
                    calls.append(piece)
                elif isinstance(piece, StreamPiece) and piece.kind == "content":
                    round_content += len(piece.text)
                yield piece
        finally:
            await stream.aclose()
        # Phase 71: the per-strip-event capture log — one warning per
        # stripped span, truncated to 200 chars (how a new scaffolding
        # format gets captured and added to the registry) — and the turn
        # total for the API layer's ``scaffold_stripped=N`` log field.
        holder.scaffold_stripped += round_filter.stripped_chars
        for span in round_filter.stripped_spans:
            logger.warning(
                "agent: stripped %d chars of tool-scaffolding in round %d: %r",
                len(span),
                rounds + 1,
                span[:200],
            )
        if not calls:
            if round_content > 0:
                return  # the answer was streamed (the clean content stands)
            if round_filter.stripped_chars == 0:
                # Empty/thinking-only answer — today's behavior, unchanged
                # (the UI handles it); the guardrail keys on a strip.
                return
            # Phase 71: the scaffolding was the whole "answer" — the ONE
            # bounded recovery (a fixed policy, not a conversation):
            # ``tools=None``, the correction folded into the ORIGINAL
            # single system message (provider-safe — the user message and
            # any tool history stay in place), a fresh filter, the same
            # phase-67 retry budget.
            logger.warning(
                "agent: round %d was pure tool-scaffolding (%d chars stripped) "
                "— running the one bounded recovery",
                rounds + 1,
                round_filter.stripped_chars,
            )
            messages_recovered = [
                {
                    "role": "system",
                    "content": system_prompt + "\n" + CORRECTION_INSTRUCTION,
                },
                *messages[1:],
            ]
            recovery_filter = ScaffoldingFilter()
            recovered = chat_stream_retried(
                llm,
                cast("list[dict[str, str]]", messages_recovered),
                tools=None,
                retries=settings.llm_retries,
                delay=settings.llm_retry_delay,
                scaffolding=recovery_filter,
            )
            recovery_content = 0
            try:
                async for piece in recovered:
                    if isinstance(piece, StreamPiece) and piece.kind == "content":
                        recovery_content += len(piece.text)
                    yield piece
            finally:
                await recovered.aclose()
            holder.scaffold_stripped += recovery_filter.stripped_chars
            for span in recovery_filter.stripped_spans:
                logger.warning(
                    "agent: stripped %d chars of tool-scaffolding in the "
                    "recovery after round %d: %r",
                    len(span),
                    rounds + 1,
                    span[:200],
                )
            if recovery_content > 0:
                return  # the recovery answered — the turn ends
            # The second empty reply is terminal (at most one recovery per
            # turn). Raised OUTSIDE the stream, so chat_stream_retried's
            # retry rule never sees it; the API layer catches it before
            # the generic LLMError handler.
            logger.warning(
                "agent: the recovery reply was still empty "
                "(scaffold_stripped=%d) — settling with a malformed-reply error",
                holder.scaffold_stripped,
            )
            raise MalformedReplyError(
                f"the model answered in raw tool-scaffolding twice in a row "
                f"(round {rounds + 1} plus one recovery) — no clean answer "
                "to stream"
            )
        call = calls[0]  # a stream can carry several calls; run the first
        result = _execute_tool(db, call, seed_docs, holder)
        rounds += 1  # every call the model emits consumes a round
        logger.info(
            "agent tool=%s args=%s round=%d/%d",
            call.name,
            json.dumps(call.arguments, ensure_ascii=False)[:200],
            rounds,
            max_rounds,
        )
        messages.append(
            {
                "role": "assistant",
                "content": None,
                "tool_calls": [
                    {
                        "id": call.id,
                        "type": "function",
                        "function": {
                            "name": call.name,
                            "arguments": json.dumps(call.arguments),
                        },
                    }
                ],
            }
        )
        messages.append({"role": "tool", "tool_call_id": call.id, "content": result})
        if rounds >= max_rounds:
            logger.warning(
                "agent round cap reached (rounds=%d) — forcing a final "
                "no-tools answer",
                rounds,
            )
            # Phase 48: the forced final answer gets the same explicit
            # teardown as the loop rounds (consumer abandon mid-final
            # answer must still close the model's stream). Phase 67: the
            # forced call retries under the same locked-A2 rule as the
            # loop rounds. Phase 71: the forced final runs through a
            # fresh filter too — raw scaffolding can never reach the
            # user from ANY grounded request.
            final_filter = ScaffoldingFilter()
            final = chat_stream_retried(
                llm,
                cast("list[dict[str, str]]", messages),
                tools=None,
                retries=settings.llm_retries,
                delay=settings.llm_retry_delay,
                scaffolding=final_filter,
            )
            final_content = 0
            try:
                async for piece in final:
                    if isinstance(piece, StreamPiece) and piece.kind == "content":
                        final_content += len(piece.text)
                    yield piece
            finally:
                await final.aclose()
            # Phase 71: the same capture log + turn total; a
            # scaffolding-only forced final (this turn used no recovery,
            # so nothing is doubled up) settles with the same terminal
            # malformed-reply error rather than a silently empty answer.
            holder.scaffold_stripped += final_filter.stripped_chars
            for span in final_filter.stripped_spans:
                logger.warning(
                    "agent: stripped %d chars of tool-scaffolding in the "
                    "forced final answer (round %d): %r",
                    len(span),
                    rounds + 1,
                    span[:200],
                )
            if final_content == 0 and final_filter.stripped_chars > 0:
                logger.warning(
                    "agent: the forced final answer was pure tool-scaffolding "
                    "— settling with a malformed-reply error"
                )
                raise MalformedReplyError(
                    "the forced final answer was raw tool-scaffolding — no "
                    "clean answer to stream"
                )
            return