fix(backends,storage,core): setuid in exec wrapper, missing-column migration, runner config dir for mirror, tagged prompt sources

src/beaver_gateway/agents/claude.py

@@ -1,8 +1,9 @@
 """Claude agent definition, backed by the Claude Agent SDK.
 
 The system prompt is either ``system_prompt`` verbatim or, when
-``prompt_sources`` is set, the concatenation of those files assembled at
-every session spawn (see ``core/prompt.py``). ``skill_sets`` are
+``prompt_sources`` is set, the concatenation of those files (or
+``(tag, file)`` pairs) assembled at every session spawn (see
+``core/prompt.py``). ``skill_sets`` are
 directories of ``<skill>/SKILL.md`` folders; each becomes a local SDK
 plugin. Nothing from disk is loaded otherwise: the adapter runs with
 ``setting_sources=[]``.
@@ -16,6 +17,7 @@ from pathlib import Path  # noqa: TC003 - pydantic runtime
 from pydantic import BaseModel, ConfigDict, Field
 
 from beaver_gateway.agents.base import BaseAgent
+from beaver_gateway.core.prompt import PromptSource  # noqa: TC001 - pydantic runtime
 
 __all__ = ["ClaudeAgent", "ClaudeOptions"]
 
@@ -49,6 +51,6 @@ class ClaudeOptions(BaseModel):
 class ClaudeAgent(BaseAgent):
     cwd: Path
     system_prompt: str = ""
-    prompt_sources: tuple[Path, ...] = ()
+    prompt_sources: tuple[PromptSource, ...] = ()
     skill_sets: tuple[Path, ...] = ()
     options: ClaudeOptions = Field(default_factory=ClaudeOptions)

src/beaver_gateway/backends/claude_sdk.py

@@ -13,8 +13,9 @@ Events on the wire are the Anthropic ``MessageStreamEvent`` family: one
 rebased across the API calls claude makes inside the turn.
 
 Process isolation: claude is spawned through a small exec wrapper that
-drops every inherited environment variable outside a whitelist, and,
-when ``RunnerConfig.user`` is set, under that uid.
+drops every inherited environment variable outside a whitelist and, when
+``RunnerConfig.user`` is set, switches to that uid before exec (done in the
+wrapper rather than via ``subprocess(user=...)``, which uvloop rejects).
 """
 
 from __future__ import annotations
@@ -204,7 +205,7 @@ class ClaudeSdkBackend:
         self._mcp_disallowed = _mcp_disallowed(agent, mcp_tool_names or {})
         self._sessions: dict[str, _Live] = {}
         self._reaper: asyncio.Task[None] | None = None
-        self._uid = _resolve_uid(self._runner.user)
+        self._uid, self._gid = _resolve_ids(self._runner.user)
         self._wrapper: Path | None = None
 
     @property
@@ -411,6 +412,7 @@ class ClaudeSdkBackend:
         env = dict(opt.env)
         if self._runner.home is not None:
             env["HOME"] = str(self._runner.home)
+            env.setdefault("CLAUDE_CONFIG_DIR", str(self._runner.home / ".claude"))
         plugins = self._plugins()
         system_prompt = (
             prompt_assembly.assemble(agent.prompt_sources)
@@ -430,7 +432,6 @@ class ClaudeSdkBackend:
             cwd=str(agent.cwd),
             add_dirs=list(opt.add_dirs),
             env=env,
-            user=self._runner.user,
             cli_path=str(self._exec_wrapper(extra_keep=tuple(env))),
             include_partial_messages=opt.include_partial_messages,
             session_store=self._store,
@@ -472,6 +473,8 @@ class ClaudeSdkBackend:
             target=json.dumps(target),
             keep=json.dumps(keep),
             prefixes=json.dumps(list(ENV_KEEP_PREFIXES)),
+            uid=json.dumps(self._uid),
+            gid=json.dumps(self._gid),
         )
         digest = hashlib.sha256(script.encode("utf-8")).hexdigest()[:12]
         path = self._work_dir / f"claude-exec-{digest}.py"
@@ -525,9 +528,15 @@ import sys
 TARGET = {target}
 KEEP = set({keep})
 PREFIXES = tuple({prefixes})
+UID = {uid}
+GID = {gid}
 env = {{
     k: v for k, v in os.environ.items() if k in KEEP or k.startswith(PREFIXES)
 }}
+if UID is not None and os.getuid() != UID:
+    os.setgroups([])
+    os.setgid(GID)
+    os.setuid(UID)
 os.execve(TARGET, [TARGET, *sys.argv[1:]], env)
 """
 
@@ -550,12 +559,11 @@ def _claude_binary() -> str:
     return found
 
 
-def _resolve_uid(user: str | None) -> int | None:
+def _resolve_ids(user: str | None) -> tuple[int | None, int | None]:
     if user is None:
-        return None
-    if user.isdigit():
-        return int(user)
-    return pwd.getpwnam(user).pw_uid
+        return None, None
+    record = pwd.getpwuid(int(user)) if user.isdigit() else pwd.getpwnam(user)
+    return record.pw_uid, record.pw_gid
 
 
 def _chown_tree(root: Path, uid: int) -> None:

src/beaver_gateway/core/prompt.py

@@ -3,7 +3,9 @@
 The gateway holds no prompt text: an agent names its granules (paths from
 ``config.py``) and :func:`assemble` concatenates them in that order, so the
 result is byte-for-byte identical for every session of the same agent as
-long as the files are. Each granule's hash is logged at assembly so a
+long as the files are. A source is a path, or a ``(tag, path)`` pair whose
+content is wrapped in ``<tag>...</tag>`` - the markup lives here, the vault
+keeps plain markdown. Each granule's hash is logged at assembly so a
 drifted prompt can be traced to the file that changed.
 """
 
@@ -17,18 +19,24 @@ from typing import TYPE_CHECKING
 if TYPE_CHECKING:
     from collections.abc import Iterable
 
-__all__ = ["assemble"]
+__all__ = ["PromptSource", "assemble"]
 
 _log = logging.getLogger("beaver_gateway.core.prompt")
 
+PromptSource = str | Path | tuple[str, str | Path]
 
-def assemble(sources: Iterable[str | Path]) -> str:
+
+def assemble(sources: Iterable[PromptSource]) -> str:
     parts: list[str] = []
     for source in sources:
-        path = Path(source)
+        tag, raw = source if isinstance(source, tuple) else (None, source)
+        path = Path(raw)
         text = path.read_text(encoding="utf-8").strip()
         digest = hashlib.sha256(text.encode("utf-8")).hexdigest()[:12]
-        _log.info("prompt granule %s sha=%s bytes=%d", path, digest, len(text))
-        if text:
-            parts.append(text)
+        _log.info(
+            "prompt granule %s tag=%s sha=%s bytes=%d", path, tag, digest, len(text)
+        )
+        if not text:
+            continue
+        parts.append(f"<{tag}>\n{text}\n</{tag}>" if tag else text)
     return "\n\n".join(parts) + "\n"

src/beaver_gateway/storage/db.py

@@ -18,6 +18,7 @@ import json
 from datetime import UTC, datetime
 from typing import TYPE_CHECKING, Any
 
+from sqlalchemy import inspect, text
 from sqlalchemy.ext.asyncio import create_async_engine
 from sqlmodel import SQLModel, select
 from sqlmodel.ext.asyncio.session import AsyncSession
@@ -27,6 +28,7 @@ from beaver_gateway.storage.models import AuditLog, Token, Usage
 if TYPE_CHECKING:
     from collections.abc import Sequence
 
+    from sqlalchemy import Connection
     from sqlalchemy.ext.asyncio import AsyncEngine
 
 
@@ -72,9 +74,14 @@ class Database:
         self._engine: AsyncEngine = create_async_engine(_to_async_url(url), echo=False)
 
     async def create_all(self) -> None:
-        """Idempotent ``CREATE TABLE IF NOT EXISTS`` for every model."""
+        """Create missing tables, then add columns models gained since.
+
+        Nullable columns only - that is the whole migration story until
+        Alembic is worth it.
+        """
         async with self._engine.begin() as conn:
             await conn.run_sync(SQLModel.metadata.create_all)
+            await conn.run_sync(_add_missing_columns)
 
     def session(self) -> AsyncSession:
         """Open a fresh :class:`AsyncSession` (use as ``async with``).
@@ -90,6 +97,19 @@ class Database:
         await self._engine.dispose()
 
 
+def _add_missing_columns(conn: Connection) -> None:
+    inspector = inspect(conn)
+    for table in SQLModel.metadata.sorted_tables:
+        existing = {c["name"] for c in inspector.get_columns(table.name)}
+        for column in table.columns:
+            if column.name in existing:
+                continue
+            kind = column.type.compile(conn.dialect)
+            conn.execute(
+                text(f"ALTER TABLE {table.name} ADD COLUMN {column.name} {kind}")
+            )
+
+
 # ---- Token CRUD ---------------------------------------------------------
 
 

tests/test_claude_sdk_backend.py

@@ -25,7 +25,12 @@ from claude_agent_sdk import (
 
 from beaver_gateway.agents.base import ExposedMcp
 from beaver_gateway.agents.claude import ClaudeAgent, ClaudeOptions
-from beaver_gateway.backends.claude_sdk import ClaudeSdkBackend, UsageEvent, fingerprint
+from beaver_gateway.backends.claude_sdk import (
+    ClaudeSdkBackend,
+    RunnerConfig,
+    UsageEvent,
+    fingerprint,
+)
 from beaver_gateway.core.transcript import messages_from_entries
 from beaver_gateway.core.turn_capture import TurnCapture
 
@@ -359,7 +364,9 @@ async def test_prompt_sources_are_assembled(cwd: Path) -> None:
     (cwd / "a.md").write_text("alpha\n")
     (cwd / "b.md").write_text("\nbeta\n\n")
     backend = _backend(
-        cwd, InMemorySessionStore(), prompt_sources=(cwd / "a.md", cwd / "b.md")
+        cwd,
+        InMemorySessionStore(),
+        prompt_sources=(("role", cwd / "a.md"), cwd / "b.md"),
     )
     await _drain(
         backend.complete(
@@ -368,7 +375,33 @@ async def test_prompt_sources_are_assembled(cwd: Path) -> None:
             conversation_id="c",
         )
     )
-    assert FakeClient.instances[0].options.system_prompt == "alpha\n\nbeta\n"
+    assert (
+        FakeClient.instances[0].options.system_prompt
+        == "<role>\nalpha\n</role>\n\nbeta\n"
+    )
+
+
+async def test_runner_user_lands_in_wrapper(cwd: Path) -> None:
+    import os
+    import pwd
+
+    me = pwd.getpwuid(os.getuid())
+    backend = _backend(cwd, InMemorySessionStore())
+    backend._runner = RunnerConfig(user=me.pw_name, home=cwd)
+    backend._uid, backend._gid = me.pw_uid, me.pw_gid
+    await _drain(
+        backend.complete(
+            agent=backend.agent,
+            messages=[{"role": "user", "content": "x"}],
+            conversation_id="c",
+        )
+    )
+    opts = FakeClient.instances[0].options
+    assert opts.env["HOME"] == str(cwd)
+    assert opts.env["CLAUDE_CONFIG_DIR"] == str(cwd / ".claude")
+    wrapper = Path(opts.cli_path).read_text()
+    assert f"UID = {me.pw_uid}" in wrapper
+    assert "os.setuid(UID)" in wrapper
 
 
 async def test_close_disconnects(cwd: Path) -> None:

tests/test_db_migrate.py

@@ -0,0 +1,33 @@
+import sqlite3
+import tempfile
+from pathlib import Path
+
+from sqlmodel import select
+
+from beaver_gateway.storage import Database
+from beaver_gateway.storage.models import Conversation
+
+
+async def test_create_all_adds_missing_columns() -> None:
+    path = Path(tempfile.mkdtemp(prefix="beaver-migrate-")) / "old.db"
+    raw = sqlite3.connect(path)
+    raw.execute(
+        "CREATE TABLE conversations (id INTEGER PRIMARY KEY, frontend VARCHAR NOT NULL, "
+        "external_id VARCHAR NOT NULL, agent_name VARCHAR NOT NULL, "
+        "created_at DATETIME NOT NULL, updated_at DATETIME NOT NULL)"
+    )
+    raw.execute(
+        "INSERT INTO conversations VALUES (1, 'markdown', 'x', 'a', '2026-01-01', '2026-01-01')"
+    )
+    raw.commit()
+    raw.close()
+
+    db = Database(f"sqlite:///{path}")
+    await db.create_all()
+    async with db.session() as session:
+        conv = (await session.exec(select(Conversation))).one()
+        assert conv.session_id is None
+        conv.session_id = "sid"
+        session.add(conv)
+        await session.commit()
+    await db.dispose()