feat(sandbox): vendor Anthropic document + frontend skills into image

Bakes docx, pdf, pptx, xlsx, and frontend-design skills from
github.com/anthropics/skills into /opt/claude-skills/ at image build time
(sparse-checkout, shallow clone). On each chat turn the pipe symlinks them
into $CLAUDE_CONFIG_DIR/skills/ via the new ensure_skills() method.

Symlinks (not copies) mean:
  - zero per-chat disk cost regardless of user count;
  - image rebuilds that add or update a skill propagate automatically to
    existing chats on their next turn (ln -sfn replaces stale targets);
  - skills are read-only from the user's perspective — any attempt to
    mutate them hits /opt/claude-skills, which is root-owned.

The skill list is surfaced as a valve (SKILLS) so operators can disable
individual skills without a rebuild; it must remain a subset of what the
Dockerfile CLAUDE_SKILLS arg pulled in.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Thomas Friedel 2026-04-18 23:44:20 +02:00
parent 6af5180513
commit 5c41f57f0b
2 changed files with 64 additions and 2 deletions

View File

@ -269,6 +269,30 @@ class OpenTerminalClient:
async for _ in self.stream_output(handle):
pass
async def ensure_skills(
self, user_id: str, workdir: str, skills: List[str]
) -> None:
"""Symlink system-vendored skills from /opt/claude-skills/ into the
chat's CLAUDE_CONFIG_DIR so Claude picks them up at startup. Uses
`ln -sfn` so image rebuilds that change the skill set propagate to
existing chats on their next turn (the old symlink is replaced, not
skipped). No-op if the skills list is empty."""
if not skills:
return
for s in skills:
# Skill names are baked into the image path; reject anything
# that could break out of the /opt/claude-skills/ prefix.
if not re.match(r"^[a-zA-Z0-9_-]+$", s):
raise OpenTerminalError(f"unsafe skill name: {s!r}")
parts = [f"mkdir -p {workdir}/.claude/skills"]
parts += [
f"ln -sfn /opt/claude-skills/{s} {workdir}/.claude/skills/{s}"
for s in skills
]
handle = await self.start(user_id, " && ".join(parts))
async for _ in self.stream_output(handle):
pass
async def run_capture(self, user_id: str, command: str) -> str:
"""Run a short command and return its merged stdout/stderr as a single
string. For one-shot utility commands (ls, cat of tiny files) where
@ -715,6 +739,16 @@ class Pipe:
"{chat_id} is substituted. Artifacts persist across turns."
),
)
SKILLS: str = Field(
default="docx,pdf,pptx,xlsx,frontend-design",
description=(
"Comma-separated list of skill names to symlink into each "
"chat's .claude/skills/ from /opt/claude-skills/ in the "
"sandbox image. Must match directories baked into the image "
"at build time (see sandbox/Dockerfile CLAUDE_SKILLS arg). "
"Empty string disables."
),
)
def __init__(self) -> None:
self.valves = self.Valves()
@ -810,6 +844,13 @@ class Pipe:
) as client:
await client.ensure_dir(sandbox_user, workdir)
skills = [
s.strip()
for s in (self.valves.SKILLS or "").split(",")
if s.strip()
]
await client.ensure_skills(sandbox_user, workdir, skills)
# Cache miss = either a brand-new chat OR an OWUI restart
# dropped the in-process map (#5). Claude's own on-disk
# session JSONL tells us which — look there before giving

View File

@ -17,11 +17,32 @@ RUN npm install -g "@anthropic-ai/claude-code@${CLAUDE_CODE_VERSION}" \
# bubblewrap + socat are the Linux backends for Claude Code's built-in
# `/sandbox` feature. Without them, Claude falls back to unsandboxed bash —
# our managed-settings.json sets failIfUnavailable=true so that silent
# degradation becomes a hard error instead.
# degradation becomes a hard error instead. git is needed for the skill
# vendoring step below.
RUN apt-get update \
&& apt-get install -y --no-install-recommends bubblewrap socat \
&& apt-get install -y --no-install-recommends bubblewrap socat git ca-certificates \
&& rm -rf /var/lib/apt/lists/*
# Vendor a curated set of Anthropic skills into /opt/claude-skills. Each
# chat's CLAUDE_CONFIG_DIR/.claude/skills/ symlinks here at runtime, so all
# users share a read-only copy without per-chat duplication. To update the
# set, change the sparse-checkout list and rebuild — running chats pick up
# the new set on their next turn (symlinks re-point).
ARG CLAUDE_SKILLS="docx pdf pptx xlsx frontend-design"
RUN set -eux; \
mkdir -p /opt/claude-skills; \
tmp="$(mktemp -d)"; \
git clone --depth 1 --filter=blob:none --sparse \
https://github.com/anthropics/skills.git "$tmp"; \
cd "$tmp"; \
paths=""; for s in $CLAUDE_SKILLS; do paths="$paths skills/$s"; done; \
git sparse-checkout set $paths; \
for s in $CLAUDE_SKILLS; do \
cp -r "skills/$s" "/opt/claude-skills/$s"; \
done; \
cd /; rm -rf "$tmp"; \
ls /opt/claude-skills/
# Managed (enterprise) settings live at the system-wide path and take
# precedence over per-user and per-project settings. This is how we pin
# Bash egress + WebFetch allowlists: users can extend via their own