feat(rag): pass chat history with prior thinking to the LLM

Phase 74 (TODO.md L4): a follow-up question now reaches the model WITH
the conversation so far — every prior user/brain turn and the prior
thinking blocks on brain turns (preserve-thinking) — while
POST /api/chat stays stateless (A10): the client provides the history
in the request body and the server stores nothing new.

Server (task 01):
- ChatRequest.history: optional list[HistoryTurn] (who: user|brain,
  text, optional thinking) — absent/empty keeps the request
  byte-identical to pre-phase-74 (the two-message [system, user]
  request; the kill-switch semantics are pinned in the integration
  suite).
- app.rag.prompts.history_to_messages: pure mapper — walks the turns
  newest-first against the settings budgets (history_max_turns=40 /
  history_max_chars=24000, BOR_HISTORY_MAX_TURNS /
  BOR_HISTORY_MAX_CHARS); a capped turn is dropped WHOLE (never cut
  mid-answer); the kept window is returned oldest-first; brain turns
  carry their thinking as reasoning_content (A4) only when
  non-empty.
- Both branches feed it: the deflected path splices it between the
  system prompt and the current user message (the phase-71 recovery
  still rebuilds from messages[1:]), the grounded agent receives
  run_agent(..., history=hist); llm.py's message params widen to
  list[dict[str, Any]] (string-only messages stay byte-identical on
  the wire — the SDK passes message dicts through verbatim).
- The per-turn log line (PLAN §9) gains history_msgs=N after
  kb_chars=N.
- Pins: tests/unit/test_history.py (mapper: mapping, reasoning
  gating, both budgets, drop-whole, ordering, empty default),
  tests/unit/test_config.py (the two settings + env overrides),
  tests/unit/test_agent.py (the history splice + the default),
  tests/integration/test_chat_api.py (deflected AND grounded forward
  the history incl. reasoning_content, no-history byte-identity, 422
  pins, the log field).

Client (task 02):
- runTurn — the single funnel for fresh send / phase-49 retry /
  phase-53 stale-regen — sends history = the conversation record
  minus the current question, with thinking only on brain records
  that streamed one (undefined drops the key from the JSON, the
  record's convention); the question is never duplicated into the
  history.

Wire proof (task 03):
- The mock's echo my history marker (HISTORY_TRIGGER) answers with
  the deterministic history echo — history: N prior messages; last
  answer tail: <last 24 chars>; thinking: yes|no — checked BEFORE
  the DEFLECT_MODE branch (like TABLE_TRIGGER), so it fires on both
  turn branches whatever the gate says; the module docstring records
  the user/assistant-only history invariant that keeps every
  existing (tool-result-classified) marker flow unaffected.
- tests/e2e/test_llm_history.py (isolated): a grounded follow-up and
  a deflected follow-up both receive history: 2 prior messages +
  thinking: yes + the byte-exact tail of turn 1's answer (derived
  from the persisted bor.chat.v1 record — the same array the client
  maps into the body); a cold start receives history: 0 prior
  messages / last answer tail: none / thinking: no.
- Regressions green in isolation: chat_rag, chat_history (phase 50),
  agent_document_tools, harness_aligned_tools, stop_generation,
  retry_answer, response_to_docs.

app/rag/agent.py

@@ -176,7 +176,7 @@ import logging
 import re
 from collections.abc import AsyncIterator, Sequence
 from dataclasses import dataclass, field
-from typing import Any, cast
+from typing import Any
 
 from sqlalchemy import select
 from sqlalchemy.orm import Session
@@ -821,6 +821,7 @@ async def run_agent(
     seed_docs: Sequence[Document],
     settings: Settings,
     holder: AgentHolder,
+    history: Sequence[dict[str, Any]] = (),
 ) -> AsyncIterator[StreamPiece | ToolCallPiece | RetryPiece]:
     """Run the grounded-turn tool loop, yielding every stream piece.
 
@@ -830,6 +831,18 @@ async def run_agent(
     SSE ``retry`` events. After the loop finishes, *holder* carries the
     read documents and the executed tool-call count (re-lists included).
 
+    History (phase 74, TODO L4): *history* is the client's prior turns
+    already mapped to model messages by
+    :func:`app.rag.prompts.history_to_messages` (trimmed newest-first
+    against the settings budgets; assistant turns carry their prior
+    thinking as ``reasoning_content``). It is spliced between the system
+    prompt and the current user message —
+    ``[system, *history, user]`` — and everything downstream (the tool
+    rounds, the phase-71 recovery rebuilding from ``messages[1:]``, the
+    retry restarts) already operates on that one ``messages`` list,
+    unchanged. ``()`` (the default) keeps the pre-phase-74 two-message
+    request byte-identical.
+
     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
@@ -862,6 +875,7 @@ async def run_agent(
     """
     messages: list[dict[str, Any]] = [
         {"role": "system", "content": system_prompt},
+        *history,  # phase 74: the client's prior turns (empty by default)
         {"role": "user", "content": user_message},
     ]
     # Phase 45: no per-tool budgets — the tools stay offered for the
@@ -892,7 +906,7 @@ async def run_agent(
         # a quiet no-op.
         stream = chat_stream_retried(
             llm,
-            cast("list[dict[str, str]]", messages),
+            messages,
             tools=tools,
             retries=settings.llm_retries,
             delay=settings.llm_retry_delay,
@@ -948,7 +962,7 @@ async def run_agent(
             recovery_filter = ScaffoldingFilter()
             recovered = chat_stream_retried(
                 llm,
-                cast("list[dict[str, str]]", messages_recovered),
+                messages_recovered,
                 tools=None,
                 retries=settings.llm_retries,
                 delay=settings.llm_retry_delay,
@@ -1030,7 +1044,7 @@ async def run_agent(
             final_filter = ScaffoldingFilter()
             final = chat_stream_retried(
                 llm,
-                cast("list[dict[str, str]]", messages),
+                messages,
                 tools=None,
                 retries=settings.llm_retries,
                 delay=settings.llm_retry_delay,

app/rag/llm.py

@@ -298,7 +298,7 @@ class LLMClient:
         return vec
 
     async def chat(
-        self, messages: list[dict[str, str]], model: str | None = None
+        self, messages: list[dict[str, Any]], model: str | None = None
     ) -> str:
         """One-shot (non-streaming) completion (A5 extended, phase 30).
 
@@ -341,12 +341,20 @@ class LLMClient:
 
     async def chat_stream(
         self,
-        messages: list[dict[str, str]],
+        messages: list[dict[str, Any]],
         tools: list[dict[str, Any]] | None = None,
         scaffolding: ScaffoldingFilter | None = None,
     ) -> AsyncGenerator[StreamPiece | ToolCallPiece, None]:
         """Stream assistant pieces from the chat model (PLAN A5/A15, phase 17).
 
+        Messages are passed to the request body VERBATIM: string-only
+        ``{role, content}`` dicts are byte-identical on the wire to the
+        pre-phase-74 requests, and an assistant message may additionally
+        carry ``reasoning_content`` (the client's prior thinking, phase
+        74 — the same wire field the model uses for its OWN reasoning on
+        the response side; the ``openai`` SDK passes message dicts
+        through untouched, so no transport change).
+
         ``stream=True`` against the OpenAI-compatible endpoint, yielding
         typed :class:`StreamPiece` values. Wire convention (verified live
         against aipi's ``turbo`` on 2026-08-23): the model's reasoning
@@ -488,7 +496,7 @@ class LLMClient:
 
 async def chat_stream_retried(
     llm: LLMClient,
-    messages: list[dict[str, str]],
+    messages: list[dict[str, Any]],
     *,
     tools: list[dict[str, Any]] | None = None,
     retries: int = 0,

app/rag/prompts.py

@@ -54,10 +54,12 @@ not the wording, so that contract is unchanged.
 from __future__ import annotations
 
 from collections.abc import Sequence
+from typing import Any
 
-from app.config import get_settings
+from app.config import Settings, get_settings
 from app.models import Document
 from app.rag.retriever import TRUNCATION_MARKER
+from app.schemas import HistoryTurn
 
 #: PLAN §6 verbatim (line wrapping included); ``{relevance}`` is filled by
 #: :func:`_base`.
@@ -166,6 +168,62 @@ TOOLS_SECTION: str = (
 )
 
 
+def history_to_messages(
+    history: Sequence[HistoryTurn],
+    settings: Settings,
+) -> list[dict[str, Any]]:
+    """Client-provided chat history → model messages (phase 74, TODO L4).
+
+    The ``POST /api/chat`` ``history`` (the client's prior turns, oldest
+    first) becomes the message block that sits between the system prompt
+    and the current user message — so a follow-up question reaches the
+    model together with the exchange so far, on BOTH turn branches
+    (the deflected path and the grounded agent).
+
+    Trimming (owner-locked A3, 2026-09-08): the turns are walked
+    **newest-first** and kept while BOTH budgets hold — the turn count
+    stays ≤ ``settings.history_max_turns`` and the cumulative chars
+    (``len(text) + len(thinking or "")`` per turn) stay ≤
+    ``settings.history_max_chars``. A turn that would overflow either
+    remaining budget is DROPPED WHOLE — never cut mid-answer — and the
+    walk stops there, so the kept history is always the contiguous
+    newest window (the oldest turns are the ones dropped; ``0`` on
+    either budget yields ``[]`` — the pre-phase-74 behavior). The kept
+    turns are returned in chronological (oldest → newest) order.
+
+    Mapping (owner-locked A4, 2026-09-08): ``who="user"`` →
+    ``{"role": "user", "content": text}``; ``who="brain"`` →
+    ``{"role": "assistant", "content": text}`` plus
+    ``"reasoning_content": thinking`` ONLY when *thinking* is
+    non-empty — the preserve-thinking wire convention
+    :mod:`app.rag.llm` already reads on the response side
+    (``delta.reasoning_content``), which is what keeps the owner's
+    preserve-thinking models carrying the reasoning chain forward.
+
+    Pure and side-effect free (no I/O) — unit-testable in isolation.
+    """
+    kept: list[HistoryTurn] = []
+    chars = 0
+    for turn in reversed(history):
+        if len(kept) >= settings.history_max_turns:
+            break
+        size = len(turn.text) + len(turn.thinking or "")
+        if chars + size > settings.history_max_chars:
+            break
+        kept.append(turn)
+        chars += size
+    messages: list[dict[str, Any]] = []
+    for turn in reversed(kept):
+        if turn.who == "user":
+            messages.append({"role": "user", "content": turn.text})
+            continue
+        message: dict[str, Any] = {"role": "assistant", "content": turn.text}
+        if turn.thinking:
+            message["reasoning_content"] = turn.thinking
+        messages.append(message)
+    return messages
+
+
 def _base(relevance: str) -> str:
     if relevance not in ("HIGH", "LOW"):
         raise ValueError(f"relevance must be HIGH or LOW, got {relevance!r}")