Files
ironclaw/crates/Dockerfile.sandbox
Illia Polosukhin 0af0267125 feat(engine-v2): per-project sandbox (Phases 1–7) (#2211)
* feat(engine-v2): mount-backend abstraction for per-project sandbox (Phase 1)

Adds the engine-side `MountBackend` trait + minimal `WorkspaceMounts` registry
and a host-side bridge interceptor that routes sandbox-eligible tool calls
(`file_read`, `file_write`, `list_dir`, `apply_patch`, `shell`) through a
backend when their path argument starts with `/project/`. Default behavior is
unchanged: until `EffectBridgeAdapter::set_workspace_mounts(Some(...))` is
called (Phase 6), the interception path is dormant.

This is the first phase of the per-project sandbox plan
(`docs/plans/2026-04-10-engine-v2-sandbox.md`) and a deliberately small subset
of the unified Workspace VFS proposed in nearai/ironclaw#1894 — just enough
abstraction so the sandbox can be a `MountBackend` rather than a special case
in the bridge. When #1894's full mount table lands, the sandbox backend slots
in unchanged.

Engine crate (`crates/ironclaw_engine/src/workspace/`):
- `mount.rs` — `MountBackend` trait, `MountError` (NotFound / InvalidPath /
  PermissionDenied / Io / Tool / Backend / Unsupported), `DirEntry`,
  `EntryKind`, `ShellOutput`
- `filesystem.rs` — `FilesystemBackend`: passthrough host-fs implementation
  with two-layer path validation (lexical reject of absolute / `..`, then
  symlink-escape canonicalization). `read`/`write`/`list` fully implemented;
  `patch`/`shell` return `Unsupported` so the bridge falls through to the
  host tool until Phase 5
- `registry.rs` — `WorkspaceMounts` per-project registry with lazy
  `ProjectMountFactory`, longest-prefix-match resolution, cached and
  invalidatable

Bridge (`src/bridge/sandbox/`):
- `intercept.rs` — `maybe_intercept` and `SANDBOX_TOOL_NAMES`. Returns
  `Handled(json)` on a successful backend dispatch, `FellThrough` for
  non-sandbox tools, host paths, missing path params, or `Unsupported`
  backend ops
- `effect_adapter.rs` — `workspace_mounts` field + `set_workspace_mounts`
  setter; interception block in `execute_action_internal` right before
  `execute_tool_with_safety`, gated on the optional mount table

Tests (31 new):
- 17 engine workspace unit tests covering trait error mapping, path safety
  (lexical + symlink), longest-prefix routing, and lazy factory caching
- 9 bridge sandbox unit tests including `intercept_actually_dispatches_into_backend`
  (counting backend) which proves the interceptor reaches the backend
- 5 integration tests in `tests/engine_v2_sandbox_integration.rs` driving
  `EffectBridgeAdapter::execute_action()` end-to-end per the
  "Test Through the Caller" rule (`.claude/rules/testing.md`), including
  a host-path-falls-through test that asserts the sandbox tempdir was
  not touched, and a `..`-escape test that verifies no `/etc/passwd`
  content leaks even after safety-layer redaction

Drive-by: feature-gate two pre-existing dead-code helpers in
`crates/ironclaw_skills/src/parser.rs` on `#[cfg(feature = "registry")]` to
match their only call site, fixing a pre-existing clippy warning that blocked
the workspace's `-D warnings` policy when `ironclaw_skills` is built with
`default-features = false` (as the engine crate does).

Verification:
- `cargo fmt --check` clean
- `cargo clippy --all --benches --tests --examples --all-features` zero warnings
- 31 / 31 new tests passing; no existing tests broken

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(engine-v2): per-project sandbox — Phases 2–7 + live Docker e2e test

Completes the per-project sandbox plan (docs/plans/2026-04-10-engine-v2-sandbox.md
Phases 2–7), building on Phase 1's mount-backend abstraction (#2211).

Phase 2 — Project workspace folder:
- `Project.workspace_path: Option<PathBuf>` field + `with_workspace_path()`
- Host-side `project_workspace_path()`, `ensure_project_workspace_dir()` (creates
  `~/.ironclaw/projects/<id>/` mode 0700, idempotent)
- `FilesystemMountFactory` taking a `ProjectPathResolver` closure (decoupled from
  `Store`); wired into `EffectBridgeAdapter` via `set_workspace_mounts()`

Phase 3 — Standalone daemon binary:
- `src/bin/sandbox_daemon.rs` — NDJSON over stdin/stdout, health/shutdown/execute_tool
- Constructs ReadFileTool/WriteFileTool/ListDirTool/ApplyPatchTool/ShellTool with
  `base_dir=/project` (override via `IRONCLAW_SANDBOX_BASE_DIR`)

Phase 4 — Dockerfile.sandbox:
- Multi-stage build: rust-slim builder (+ python3 for pyo3) compiles sandbox_daemon;
  debian-slim runtime with tini PID 1, common build tools, `/project` mount target

Phase 5 — ProjectSandboxManager + ContainerizedFilesystemBackend:
- protocol.rs: Request/Response/RpcError matching daemon wire format
- transport.rs: `SandboxTransport` trait (seam for testing without Docker)
- containerized_backend.rs: `ContainerizedFilesystemBackend` impls `MountBackend`,
  translates relative→`/project/<rel>`, maps tool-error→MountError
- docker_transport.rs: real bollard exec session, serialized Mutex, lazy reconnect
- lifecycle.rs: deterministic `ironclaw-sandbox-<pid>` naming, ensure_running/stop/remove
- manager.rs: `ProjectSandboxManager` per-project transport cache

Phase 6 — Router gating on ENGINE_V2_SANDBOX:
- `engine_v2_sandbox_enabled()` helper (truthy: 1/true/yes/on)
- Router selects `ContainerizedMountFactory` when enabled + Docker reachable;
  falls back to `FilesystemMountFactory` with warning otherwise

Live e2e bugs caught and fixed:
- Shell without explicit `workdir` defaulted to host (not sandbox); fixed by
  defaulting to `/project/` in `extract_path_param`
- `ContainerizedFilesystemBackend::shell` parsed `stdout`/`stderr` but host
  ShellTool returns merged `output` field; fixed with fallback key lookup
- SANDBOX_TOOL_NAMES only had v2 names (`file_read`/`file_write`) but host
  registry uses v1 names (`read_file`/`write_file`); added both aliases

Tests (62 sandbox-related, all green):
- 27 bridge sandbox unit tests (intercept, workspace_path, factory, protocol,
  lifecycle, containerized_backend with ScriptedTransport mock)
- 7 containerized-backend tests (including 2 regression tests for the shell bugs)
- 5 engine v2 sandbox integration tests (EffectBridgeAdapter end-to-end)
- 5 daemon binary smoke tests (real subprocess + NDJSON I/O)
- 17 engine workspace unit tests
- 1 live Docker e2e test: agent clones nearai/ironclaw into sandbox, renames
  to megaclaw via sed, verifies with grep — 70s, $0.09, recorded trace committed

Verification:
- `cargo fmt --check` clean
- `cargo clippy --all --benches --tests --examples --all-features` zero warnings
- All 62 sandbox tests passing; no existing tests broken

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: replace .expect() with Result in DockerTransport::ensure_session

CI's no-panics checker flagged the .expect("just inserted") in production
code. Replace with .ok_or_else() returning MountError::Backend.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: multi-tenant project paths + unify sandbox env var with v1

Two issues addressed:

1. Project workspace paths now namespace by user_id:
   `~/.ironclaw/projects/<user_id>/<project_id>/` instead of
   `~/.ironclaw/projects/<project_id>/`. Prevents filesystem collisions
   in multi-tenant deployments where two users could theoretically have
   the same project UUID.

2. Sandbox enablement now reads `SANDBOX_ENABLED` (same env var as v1
   sandbox) in addition to `ENGINE_V2_SANDBOX`. Either being truthy
   enables the per-project sandbox. This means a single flag governs
   sandbox behavior regardless of engine version, while the v2-specific
   override remains available for transitional setups.

Tests: 30 bridge sandbox unit tests passing (added multi-tenant path
tests + env var combination tests).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR review — TOCTOU race, shell env passthrough, canonicalize guard

Three issues flagged by the code review bot on #2211:

1. TOCTOU race in WorkspaceMounts::resolve (HIGH): Added double-checked
   locking — re-check the cache after acquiring the write lock so two
   threads racing on the same project's first access don't both call
   factory.build(). The second thread finds the insert from the first.

2. Shell intercept ignores env parameter (MEDIUM): The shell arm in
   maybe_intercept was passing HashMap::new() instead of forwarding
   the tool call's env map. Fixed to parse parameters["env"] and pass
   it through to backend.shell().

3. Canonicalization fails when root doesn't exist (MEDIUM): When
   self.root hasn't been created yet (first write to a new project),
   canonicalize_under_root would walk up to a real ancestor and the
   starts_with check against the non-existent root would always fail.
   Now skips canonicalization entirely when root doesn't exist — lexical
   safety is already guaranteed by safe_join.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR review round 2 — apply_patch schema, content validation, dir perms, docs

- Fix apply_patch schema mismatch: MountBackend::patch now takes
  (old_string, new_string, replace_all) matching ApplyPatchTool's
  actual contract. Previously sent {patch: diff} which would fail
  with invalid_params in the containerized daemon.
- Validate file_write content param: return error instead of silently
  writing empty string when content is missing.
- Log stderr frames from sandbox daemon at debug! instead of silently
  discarding them in docker_transport StreamReader.
- Tighten permissions on intermediate directories created by
  ensure_project_workspace_dir (projects/, <user_id>/) to 0o700,
  not just the leaf.
- Fix stale module doc in sandbox/mod.rs (referenced "Phase 5 will
  add" but all phases shipped).
- Fix doc path mismatch: workspace path is <user_id>/<project_id>/,
  not <project_id>/ (workspace_path.rs, CLAUDE.md, design plan).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR review round 3 — symlink safety, visibility, debug logging

- Close TOCTOU window in canonicalize_under_root: re-canonicalize and
  verify containment when the reassembled path exists on disk
- Fix list_dir_recursive: use symlink_metadata (lstat) so symlinks are
  detected instead of followed; validate directories against root before
  recursive traversal
- Tighten is_mountable_path to /project/, /memory/, /home/ prefixes
  instead of any absolute path (defense-in-depth)
- Narrow sandbox module visibility to pub(crate) and remove unused
  pub use re-exports
- Remove concrete types (FilesystemBackend, DirEntry, EntryKind,
  ShellOutput) from engine crate top-level re-exports; access via
  ironclaw_engine::workspace:: module path
- Add debug! tracing to sandbox intercept routing decisions
- Add read_file/write_file v1 aliases to daemon SUPPORTED_TOOLS health
  response
- Remove developer-local path from sandbox mod.rs doc comment
- Merge staging to fix CI (user_timezone field on ThreadExecutionContext)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR review round 4 — safety validation, network isolation, binary writes

- Add pre-intercept safety param validation so sandbox-dispatched calls
  go through the same checks as host-dispatched calls (#1)
- Set network_mode: "none" on sandbox containers to prevent outbound
  network access (#3)
- Reject binary content in containerized write instead of silently
  corrupting via from_utf8_lossy (#5)
- Cap list_dir depth to 10 to prevent unbounded traversal (#8)
- Change container creation log from info! to debug! to avoid breaking
  REPL/TUI output (#10)
- Make is_truthy case-insensitive so SANDBOX_ENABLED=True works (#11)
- Return error instead of unwrap_or_default for missing container ID (#12)
- Propagate set_permissions errors instead of silently ignoring (#13)
- Return error for missing daemon output key instead of defaulting to
  empty object (#14)
- Add env mutex guard in sandbox_live_e2e test (#15)
- Fix rustfmt formatting for let-chain in canonicalize_under_root

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address review round 5 — path traversal, error types, tests

Security fixes:
- Sanitize user_id in workspace path to prevent directory traversal via
  malicious user IDs containing `..` or `/`
- Add Component::ParentDir check in ContainerizedFilesystemBackend::container_path
  matching the defense-in-depth approach of FilesystemBackend::safe_join

Correctness:
- Use MountError::Tool instead of MountError::InvalidPath for missing
  tool parameters (content, old_string, new_string) — fixes confusing
  LLM-visible error messages
- Fix clippy sort_by_key suggestion in registry.rs

Cleanup:
- Remove spurious Notify import and dead _notify_link function

New tests:
- ContainerizedFilesystemBackend path traversal rejection (read + write)
- container_path unit tests for safe and unsafe paths
- Adversarial user_id test in workspace_path
- Daemon-side path traversal test in sandbox_daemon_smoke

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address review round 6 — param normalization, error types, edge cases

- Normalize sandbox params via prepare_tool_params() before validation,
  matching the host execution path (fixes inconsistent validation)
- Return ToolError::InvalidParameters instead of EngineError::Effect for
  sandbox param validation failures (consistent error surface)
- ensure_dir checks path.is_dir() not path.exists() (rejects files)
- Empty user_id returns "_anonymous" sentinel instead of empty hex string
  that would drop the tenant namespace via PathBuf::join("")
- Restore ENGINE_V2_SANDBOX env var after sandbox live E2E test
- Tighten is_mountable_path to /project/ only (no mounts for /memory/
  or /home/ yet)
- Add v1 tool name aliases (read_file, write_file) to SUPPORTED_TOOLS

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* refactor: unify sandbox env var — remove ENGINE_V2_SANDBOX, use SANDBOX_ENABLED only

Single env var controls sandboxing for both engine versions. The
transitional ENGINE_V2_SANDBOX override is removed from code, tests,
docs, and Dockerfile.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: double-checked locking in transport_for, explicit stdin close in smoke test

- ProjectSandboxManager::transport_for no longer holds the mutex across
  the Docker ensure_running await. Uses double-checked locking so
  concurrent projects initialize in parallel.
- sandbox_daemon_smoke: explicitly take() stdin before wait_with_output
  so EOF is sent even without a shutdown request.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address review — network mode, error types, race, protocol dedup

- Change sandbox container network_mode from "none" to default bridge
  so git clone / cargo build / pip install work inside the container
- Fix binary content rejection to use MountError::Tool instead of
  MountError::InvalidPath (semantic mismatch)
- Fix list depth: use actual depth value instead of depth.max(1)
- Fix orphan container race in transport_for by holding lock across
  container creation instead of double-checked locking
- Deduplicate protocol types: daemon now imports from shared
  bridge::sandbox::protocol instead of defining its own copies
- Make bridge::sandbox pub (narrow exposure: only protocol and
  workspace_path sub-modules are pub)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: update plan doc — sandbox uses bridge networking, not network_mode=none

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-19 20:20:17 +09:00

96 lines
3.8 KiB
Docker

# syntax=docker/dockerfile:1.6
#
# Per-project sandbox container image for IronClaw engine v2.
#
# This image is the running environment for one project's `/project/` mount.
# It hosts the `sandbox_daemon` binary which speaks NDJSON over its stdin/
# stdout to the host (`docker exec -i`). The container itself is started
# once and kept alive across IronClaw restarts so installed dependencies
# (cargo, npm, apt) and build caches accumulate over the project's lifetime.
#
# Build:
#
# docker build -f crates/Dockerfile.sandbox -t ironclaw/sandbox:dev .
#
# Run (host):
#
# ENGINE_V2=true SANDBOX_ENABLED=true \
# IRONCLAW_SANDBOX_IMAGE=ironclaw/sandbox:dev \
# cargo run
#
# The host's `ProjectSandboxManager` (Phase 5) handles `docker create` /
# `start` / `exec` lifecycle. PID 1 is `tini` so child processes (the
# daemon, transient `cargo build`s, etc.) are reaped cleanly. The default
# CMD just blocks on a fifo so the container stays alive without burning
# CPU; the daemon is launched separately via `docker exec -i`.
# ── Stage 1: build the sandbox_daemon binary ─────────────────────────
#
# We compile against the same Rust toolchain the host uses. Pinning to a
# minor version keeps build-cache hit rates stable across reproducible
# builds. Override at build time with `--build-arg RUST_VERSION=<v>`.
ARG RUST_VERSION=1.92
FROM rust:${RUST_VERSION}-slim-bookworm AS builder
WORKDIR /build
# Cache deps independently of source. Python is required at *build time*
# because the `ironclaw` crate pulls in `monty` → `pyo3`, whose build script
# probes for a Python 3 interpreter during compilation.
RUN apt-get update \
&& apt-get install -y --no-install-recommends \
pkg-config libssl-dev ca-certificates git build-essential \
python3 python3-dev \
&& rm -rf /var/lib/apt/lists/*
# Copy the workspace. We bind-mount the build context as the workspace
# root because the sandbox_daemon binary lives at the workspace root and
# imports from `crates/`. Everything is needed for cargo workspace
# resolution; selectively copying would require shadowing every member.
COPY . .
# Build only the daemon binary. The rest of the workspace is unused at
# runtime — this keeps the image small.
RUN --mount=type=cache,target=/usr/local/cargo/registry \
--mount=type=cache,target=/build/target \
cargo build --release --bin sandbox_daemon \
&& cp target/release/sandbox_daemon /sandbox_daemon
# ── Stage 2: minimal runtime ─────────────────────────────────────────
FROM debian:bookworm-slim AS runtime
# tini reaps zombies (cargo/python often spawn helpers); the rest are the
# common build/test toolchain that the LLM is likely to want for any
# real-world project workspace.
RUN apt-get update \
&& apt-get install -y --no-install-recommends \
tini \
ca-certificates \
curl \
git \
jq \
openssh-client \
build-essential \
pkg-config \
libssl-dev \
python3 python3-pip python3-venv \
&& apt-get clean \
&& rm -rf /var/lib/apt/lists/*
COPY --from=builder /sandbox_daemon /usr/local/bin/sandbox_daemon
RUN chmod +x /usr/local/bin/sandbox_daemon
# Project bind mount target.
RUN mkdir -p /project
WORKDIR /project
# tini is PID 1; CMD blocks indefinitely so the container stays alive
# without consuming CPU. The host launches the daemon via `docker exec -i`
# whenever a new IronClaw session attaches to this project.
ENTRYPOINT ["/usr/bin/tini", "--"]
CMD ["sleep", "infinity"]
LABEL org.opencontainers.image.title="ironclaw-sandbox" \
org.opencontainers.image.description="Per-project sandbox container for IronClaw engine v2" \
org.opencontainers.image.source="https://github.com/nearai/ironclaw"