Merge branch 'codex/simp-prune-bash-surface' into codex/simp-prune-sandbox-surface

# Conflicts:
#	docs/config-catalog.md
#	packages/bash/bash-sandbox/src/index.ts
#	packages/sandbox/sandbox-local/README.md
#	packages/sandbox/sandbox-local/src/index.ts
This commit is contained in:
Tianyi Cui
2026-07-14 19:09:35 +08:00
562 changed files with 4349 additions and 12273 deletions

View File

@@ -1,10 +1,12 @@
# @deepseek-ai/dsh-sandbox-local
Local implementation of the [`@deepseek-ai/dsh-sandbox`](../sandbox/) seam: wraps a caller's argv in a platform confinement runner. Selection is BY PLATFORM, resolved once and cached: each platform names its runner chain, a chain of one is selected directly (probing arbitrates between candidates — a sole candidate leaves nothing to arbitrate), and a chain of several is probed functionally in preference order. Linux: [`bwrap`](https://github.com/containers/bubblewrap) when its probe passes, else the [`landlock-run`](https://www.npmjs.com/package/node-addon-landlock-run) Landlock launcher (kernel confinement that needs no userns/mount privileges — see the [sandbox RFC](../../../docs/rfc/implemented/feature/2026-07-06-sandbox.md) for the prebuilt-binary decision and profile-parity notes); darwin: `sandbox-exec` speaking a Seatbelt (SBPL) profile, unprobed. A platform with no chain means `confine()` FAILS CLOSED with the seam's structured `SANDBOX_UNAVAILABLE` error (win32 today: a reserved, deliberately empty chain awaiting an AppContainer-family runner); an unprobed runner that turns out unusable fails closed at EXECUTION instead — it refuses to run the command, and every wrap's `runnerFailureSignatures` let the consumer classify that as a sandbox failure rather than a task failure. Never a silent unconfined passthrough on any path.
Local implementation of the [`dsh-sandbox`](../sandbox/) seam. It selects and caches one platform runner: Linux prefers a working `bwrap` then Landlock; macOS uses Seatbelt. Multiple candidates are probed in order, while a sole candidate is selected directly.
The package root exports the default and named `LocalSandboxProvider` plugin, `Config`, and its public test-injection seam; platform profile builders stay internal.
Policy is per call (`SandboxPolicy`: mode + workspace root); the provider holds only the mechanism and the cached ladder verdict. Every wrap reports the selected runner's `enforcement` (`full`, or `partial` on an older Landlock ABI that governs only a subset of accesses — read from the launcher's `--probe` report line) and its `denialSignatures` — the stderr dialect that rung's kernel speaks on a denied file effect (EROFS text under bwrap, EACCES under Landlock, EPERM under Seatbelt), which stderr-inferring consumers match instead of a cross-runner union. A non-empty `runnerCommand` config is the operator's assertion of a runner that fully enforces the bwrap-shaped profile: the ladder and probes are skipped (the wrap carries both Linux denial dialects, the mechanism being unknown) — also the deterministic fake-runner seam for keyless test tiers. Its runner-failure dialect is the OUTER shell's argv0-scoped failure shapes (`exec: <argv0>: not found`, `<argv0>: No such file or directory`, `<argv0>: Permission denied`) — the consumer re-joins the wrap through `bash -c 'exec …'`, so a missing or unexecutable configured runner classifies as a sandbox failure (fail closed at execution), never as a failing command or a denial. `probeTimeoutMs` (default 5000) bounds each functional probe, the escape hatch for hosts slow enough that a timed-out probe would otherwise misread as `SANDBOX_UNAVAILABLE`.
Unsupported platforms and unusable runners fail closed with `SANDBOX_UNAVAILABLE`; execution never silently falls through unconfined. Each wrap carries runner-failure signatures so consumers can distinguish a broken sandbox from a command failure. The [sandbox RFC](../../../docs/rfc/implemented/feature/2026-07-06-sandbox.md) owns selection rationale and profile differences.
Policy is per call; the provider stores only the mechanism and cached runner verdict. Each wrap reports enforcement completeness plus backend-specific denial and runner-failure signatures. `runnerCommand` is an operator assertion of a bwrap-shaped runner and skips probes, but missing or unexecutable commands still fail closed at execution. Because its mechanism is unknown, it carries both Linux denial dialects. `probeTimeoutMs` bounds functional probes. The [sandbox RFC](../../../docs/rfc/implemented/feature/2026-07-06-sandbox.md) owns selection and failure semantics.
The Seatbelt profile is allow-default with `(deny file-write*)` plus write allow-lists, so exactly the mode's promised file effects are governed: `read-only` grants the `/dev/null` literal alone; `workspace-write` adds the workspace root, `/tmp`, and the per-user darwin temp dir (`os.tmpdir()` — the platform's real temp area for mkstemp-family tools), every root canonicalized because Seatbelt matches resolved paths (`/tmp` IS `/private/tmp`). Apple marks the `sandbox-exec` CLI deprecated but ships it on every macOS; the functional probe is what fails closed if that ever changes.

View File

@@ -72,17 +72,7 @@ export interface Config {
probeTimeoutMs?: number
}
/**
* Functional `bwrap` probe: can it actually build the read-only profile on
* this host? (`--version` alone would miss a disabled unprivileged user
* namespace.) Synchronous by design — it runs once, lazily, before the first
* confined wrap, and the chain's verdict is cached for the provider's
* lifetime. `timeoutMs` bounds the probe (the `probeTimeoutMs` config).
* The Landlock rung needs no such helper: resolution (`launcherPath`) and
* the functional probe (`probe`) come from `node-addon-landlock-run`, the
* package family that ships the launcher binary itself, so the probe-report
* parsing can never drift against the binary.
*/
/** Probe whether `bwrap` can create the profile; the provider caches the bounded result. */
function defaultProbeBwrap(timeoutMs: number): boolean {
const probe = spawnSync('bwrap', ['--ro-bind', '/', '/', '--dev', '/dev', '--proc', '/proc', '--die-with-parent', '--', 'true'], {
timeout: timeoutMs,

View File

@@ -10,22 +10,11 @@ import { LocalSandboxProvider } from '@deepseek-ai/dsh-sandbox-local'
import { bwrapProfileArgs } from '../src/profiles.ts'
/**
* KEYLESS bwrap integration proof for the BACKEND: the REAL `bwrap` confining
* REAL processes through `confine()` + a direct spawn of the returned argv.
* Nothing is forced off: bwrap is the ladder's FIRST rung, so a passing probe
* selects it naturally — the wrap shape assertion pins that. Verifies the
* WORLD (files exist or don't) and that the kernel's denial text matches the
* dialect the wrap advertises; the through-`ctx.bash` consumer proof lives
* with `@deepseek-ai/dsh-bash-sandbox`.
*
* Self-skips wherever the functional probe fails — no `bwrap` on PATH, or a
* host that denies unprivileged user namespaces (the probe is the same
* profile the provider enforces, so skip conditions match runtime exactly).
*
* Workspaces for the workspace-write tests live under the HOME directory on
* purpose: bwrap's `/tmp` is an EPHEMERAL mount (the documented
* bwrap-profile difference — pinned by its own test below), so only a
* workspace OUTSIDE `/tmp` proves the workspace-root rebind itself.
* Keyless backend integration through `confine()` and a real bwrap process. With no rung forced,
* a passing probe must select the first rung. Tests assert world effects, wrap shape, and that the
* kernel denial matches the advertised dialect; consumer coverage lives in dsh-bash-sandbox.
* Skips when bwrap or user namespaces are unavailable. HOME-based workspaces avoid bwrap's
* ephemeral `/tmp`, so workspace-write actually proves the workspace-root rebind.
*/
const probe = spawnSync('bwrap', [...bwrapProfileArgs({ mode: 'read-only', workspaceRoot: '/' }), '--', 'true'], { timeout: 5_000, stdio: 'ignore' })

View File

@@ -10,22 +10,10 @@ import { launcherPath } from 'node-addon-landlock-run'
import { LocalSandboxProvider } from '@deepseek-ai/dsh-sandbox-local'
/**
* KEYLESS Landlock integration proof for the BACKEND: the REAL npm-distributed
* `landlock-run` launcher (`node-addon-landlock-run`) confining REAL processes through `confine()` + a direct
* spawn of the returned argv, with the bwrap rung forced off so the ladder
* lands on the launcher. Verifies the WORLD (files exist or don't), not the
* wrapper argv alone; the through-`ctx.bash` consumer proof lives with
* `@deepseek-ai/dsh-bash-sandbox`.
*
* Self-skips when the running kernel does not enforce Landlock (or this
* platform has no launcher package — the probe cannot pass then). The
* binary itself arrives with `pnpm install`, so absence is not a checkout
* state.
*
* Workspaces live under the HOME directory on purpose: `workspace-write`
* grants the host `/tmp` wholesale (the documented Landlock-profile
* difference), so only a workspace OUTSIDE `/tmp` proves the workspace-root
* grant itself.
* Keyless backend integration through `confine()` and the registry `landlock-run` launcher, with
* bwrap forced off. Tests assert real world effects; consumer coverage lives in dsh-bash-sandbox.
* Skips when the platform package or enforcing kernel is unavailable. HOME-based workspaces avoid
* Landlock's wholesale `/tmp` grant, so workspace-write proves the workspace-root grant itself.
*/
const probe = spawnSync(launcherPath(), ['--probe'], { timeout: 5_000, encoding: 'utf8' })

View File

@@ -250,19 +250,15 @@ describe('the platform chains', () => {
})
it('a rogue chain entry throws via the probe walk\'s exhaustiveness guard (closed union)', async () => {
// Same convention as the wrap switch below: the union is closed, so a
// runner added later fails to compile at the probe switch instead of
// silently selecting without a probe. Only a cast can reach the guard.
// Same convention as the wrap switch below: the union is closed, so a runner added later
// fails to compile at the probe switch instead of silently selecting without a probe.
const { sandbox } = await setup({}, { chain: ['chroot', 'bwrap'] as unknown as readonly ['bwrap'] })
expect(() => sandbox.confine(['true'], RO)).toThrow('unreachable variant')
})
it('a rogue cached runner tag throws via the exhaustiveness guard (closed union)', async () => {
// The wrap switches on the chain verdict's runner tag and ends with
// assertNever: a rogue tag (only reachable by a cast — the union is
// closed and chainVerdict writes only its own literals) must throw, so a
// runner added later fails to compile at the switch instead of silently
// wrapping with another runner's dialect.
// Only a cast can create this rogue closed-union tag. It must hit `assertNever`, ensuring a new
// runner cannot silently use another runner's wrap or denial dialect.
const { sandbox } = await setup()
;(sandbox as unknown as { selectedRunner: unknown }).selectedRunner = { runner: 'chroot', enforcement: 'full' }
expect(() => sandbox.confine(['true'], RO)).toThrow('unreachable variant')

View File

@@ -7,30 +7,14 @@ import { fileURLToPath } from 'node:url'
import { afterAll, beforeAll, describe, expect, it } from 'vitest'
/**
* KEYLESS publish-path rehearsal for this package's own distribution: the
* provider must work from its PACKED tarball plus its REGISTRY launcher
* dependency, not the git checkout. `pnpm pack` produces the EXACT bytes
* `pnpm publish` would upload; this suite packs the workspace closure
* (`dsh-sandbox-local` + its `@deepseek-ai` peers), installs the tarballs
* into a throwaway consumer OUTSIDE the repo — npm resolving the
* `node-addon-landlock-run` dependency (and its os/cpu-selected platform
* package) from the public registry, the real consumer path — and drives
* the INSTALLED packages under plain `node`: no tsx, no tsconfig paths, no
* workspace resolution, so a `files`-list omission, a broken launcher
* dependency, or a mode-stripped binary fails here instead of at the first
* real install.
* Keyless publish-path rehearsal. It packs the package and workspace peers, installs those exact
* tarballs in an external plain-Node consumer, and lets npm resolve the registry Landlock launcher
* plus its platform package. No tsx, path mapping, or workspace resolution can hide missing files,
* dependency errors, or lost executable modes.
*
* World-proofs: the registry-installed launcher carries this host's ELF
* architecture and IS executable (a tarball that loses the mode bit would
* otherwise masquerade as a non-enforcing kernel — the fail-closed branch
* below must never absorb that), and the installed provider confines a real
* process THROUGH it (bwrap forced off) — or fails closed when the running
* kernel does not enforce Landlock, which is itself the installed
* fail-closed contract. Byte provenance of the launcher is the
* `node-addon-landlock-run` repository's own release-pipeline concern.
*
* Self-skips off Linux or when the built `lib/` is absent (run
* `pnpm run build` first — CI's landlock legs do).
* The installed launcher must match the host architecture, remain executable, and either confine a
* real process with bwrap disabled or fail closed on a non-enforcing kernel. Skips off Linux or
* before `pnpm run build`; launcher byte provenance belongs to its upstream release pipeline.
*/
const packageDir = fileURLToPath(new URL('..', import.meta.url))
@@ -84,11 +68,8 @@ describe.skipIf(!packable)('sandbox-local: packed-tarball distribution (publish-
tarballs.push(lines[lines.length - 1] as string)
}
// A real consumer: plain ESM project, tarballs installed by npm — the
// peer ranges (^0.0.1) resolve to the tarball versions, cordis pins to
// the peer range's rc, and `node-addon-landlock-run` (with its
// os/cpu-selected platform package, an OPTIONAL dependency of the entry
// — so no `--omit=optional` here) comes from the public registry.
// Peer ranges resolve to the tarballs; Cordis is pinned to their peer range. Do not omit optional
// dependencies because the launcher selects its OS/CPU package through one.
writeFileSync(join(consumerDir, 'package.json'), JSON.stringify({ name: 'dsh-packed-consumer', private: true, type: 'module' }))
const install = spawnSync('npm', ['install', '--no-audit', '--no-fund', ...tarballs, 'cordis@4.0.0-rc.6'], {
cwd: consumerDir,

View File

@@ -10,20 +10,11 @@ import { LocalSandboxProvider } from '@deepseek-ai/dsh-sandbox-local'
import { seatbeltProfileArgs } from '../src/profiles.ts'
/**
* KEYLESS Seatbelt integration proof for the BACKEND: the REAL macOS
* `sandbox-exec` confining REAL processes through `confine()` + a direct
* spawn of the returned argv, with the Linux rungs forced off so the ladder
* lands on Seatbelt. Verifies the WORLD (files exist or don't) and that the
* kernel's denial text matches the dialect the wrap advertises; the
* through-`ctx.bash` consumer proof lives with `@deepseek-ai/dsh-bash-sandbox`.
*
* Self-skips wherever the functional probe fails — every non-macOS host, or
* a macOS whose `sandbox-exec` refuses the profile.
*
* Workspaces for the workspace-write tests live under the HOME directory on
* purpose: `workspace-write` grants `/tmp` and the per-user temp dir
* wholesale (the documented Seatbelt-profile temp areas), so only a
* workspace OUTSIDE both proves the workspace-root grant itself.
* Keyless backend integration through `confine()` and a real macOS Seatbelt process, with Linux
* rungs forced off. Tests assert world effects and that the kernel denial matches the advertised
* dialect; consumer coverage lives in dsh-bash-sandbox. Skips off macOS or when the profile probe
* fails. HOME-based workspaces avoid Seatbelt's wholesale temp-directory grants, so
* workspace-write proves the workspace-root grant itself.
*/
const probe = spawnSync('sandbox-exec', [...seatbeltProfileArgs({ mode: 'read-only', workspaceRoot: '/' }), '--', 'true'], { timeout: 5_000, stdio: 'ignore' })

View File

@@ -12,7 +12,17 @@ Implementations: [`@deepseek-ai/dsh-sandbox-local`](../sandbox-local/) (Linux: `
## Model Experience
Indirectly, through [`dsh-bash-sandbox`](../../bash/bash-sandbox/README.md) and `dsh-tool-bash`, which render this seam's enforcement facts as the exact denial or `SandboxUnavailableError` text documented by the consumer, with retained tokens added only for a denial or failed confinement.
### Confinement error, indirectly
**What the model sees**: Through [`dsh-bash-sandbox`](../../bash/bash-sandbox/README.md) and [`dsh-tool-bash`](../../bash/tool-bash/README.md), failure to enforce a requested mode produces code `SANDBOX_UNAVAILABLE` and the exact error below. An execution-time runner failure adds ` Runner failure: <detail>`.
**Token effect**: Conditional error text is visible for that call and retained in history until compaction.
#### Exact error
```markdown
sandbox mode "<mode>" is requested but no sandbox backend is usable on this host; refusing to run the command unconfined. Install bubblewrap or run a Landlock-enforcing kernel (Linux), ensure sandbox-exec is usable (macOS) — Windows has no confinement backend yet — or switch the consumer to danger-full-access.
```
## Known Limitations and Deferred Work

View File

@@ -1,24 +1,7 @@
/**
* The process-sandbox seam (`ctx.sandbox`): an abstract service defining WHAT
* platform confinement does — wrap a subprocess argv so it executes under a
* file-effect policy — without saying HOW. Implementations subclass
* {@link SandboxProvider} and register as the `sandbox` service;
* `@deepseek-ai/dsh-sandbox-local` (per-platform chains: Linux `bwrap` then the
* npm-distributed `landlock-run` launcher, macOS `sandbox-exec`/Seatbelt) is
* the first.
* Consumers hand over the exact argv they are about to spawn
* (`@deepseek-ai/dsh-bash-sandbox` wraps `['bash', '-c', command]`; a
* subagent backend wraps its child-agent argv) and spawn the returned argv
* instead.
*
* The seam confines SAME-WORLD subprocesses only: a backend shares the
* host's filesystem and kernel, and the policy's `workspaceRoot` names a
* real host path. Containers, microVMs, and remote executors are NOT
* backends of this seam — they are sibling implementations of whole
* capability seams (`ctx.bash`, `ctx.fs`), deployed as environment-coherent
* groups; the boundary is recorded in
* docs/rfc/implemented/feature/2026-07-06-sandbox.md.
*
* Same-world process-confinement seam: wrap exact subprocess argv under a
* host-path file policy. Containers, microVMs, and remote execution replace the
* surrounding capability seam instead; this service shares the host kernel and filesystem.
* @module @deepseek-ai/dsh-sandbox
*/
@@ -26,25 +9,10 @@ import { Context, Service } from 'cordis'
import { HarnessError } from '@deepseek-ai/dsh-llm'
/**
* File-effect policy a sandbox backend enforces on confined processes.
*
* - `read-only` — the process cannot write the filesystem anywhere; a
* write-shaped `/dev/null` sink stays available so `>/dev/null` redirects
* keep working (HOW is the backend's choice: bwrap mounts a fresh `/dev`,
* the Landlock launcher and Seatbelt grant the single `/dev/null` node).
* - `workspace-write` — writes are allowed only under the policy's
* workspace root and `/tmp`; everything else stays read-only. Which `/tmp`
* is backend-specific — an ephemeral mount under bwrap, the HOST `/tmp`
* under the Landlock launcher, the host `/private/tmp` plus the per-user
* darwin temp dir under Seatbelt: the seam promises the write boundary,
* not the mount's nature.
* - `danger-full-access` — no confinement; a consumer configured with it
* spawns its argv unwrapped and never calls the provider.
*
* The mode governs FILE effects only: network and process visibility are not
* restricted (a backend that cannot honestly enforce them must not pretend
* to). How completely the file effects themselves are enforced is likewise a
* reported fact, not an assumption — see {@link SandboxEnforcement}.
* File-effect policy for confined processes. `read-only` permits only required
* sinks such as `/dev/null`; `workspace-write` also permits the workspace and a
* backend-defined temp area; `danger-full-access` bypasses confinement. Network
* and process visibility are outside this vocabulary.
*/
export type SandboxMode = 'read-only' | 'workspace-write' | 'danger-full-access'
@@ -52,18 +20,9 @@ export type SandboxMode = 'read-only' | 'workspace-write' | 'danger-full-access'
export type ConfinedSandboxMode = Exclude<SandboxMode, 'danger-full-access'>
/**
* How completely the selected backend enforces a confined mode's file
* effects.
*
* - `full` — every file effect the mode promises to block is governed: the
* `bwrap` mount profile, a Landlock kernel enforcing the launcher's whole
* ruleset, or an operator-configured runner (configuring one asserts full
* enforcement along with existence).
* - `partial` — the backend is active but the kernel governs only the subset
* of accesses its ABI knows (an older Landlock ABI: path-based truncate is
* ungoverned before ABI v3), so a file effect the mode promises to block
* may still land. A caller that needs the mode's promise to be absolute
* must treat `partial` as outside that promise.
* Enforcement completeness for this host. `partial` means an active backend or
* older kernel ABI cannot govern every promised file effect; callers requiring
* an absolute boundary must not treat it as `full`.
*/
export type SandboxEnforcement = 'full' | 'partial'
@@ -103,36 +62,25 @@ export interface ConfinedArgv {
*/
denialSignatures: readonly string[]
/**
* How the RUNNER ITSELF failing identifies itself: case-insensitive stderr
* substrings produced when the sandbox binary is missing, refuses its
* profile, or fails closed before exec'ing the command (`bwrap: `,
* `landlock-run: `, `sandbox-exec: ` — each covers both the runner's own
* error prefix and the shell's runner-not-found message). ORTHOGONAL to
* {@link denialSignatures}: a denial is the confined COMMAND being blocked
* (the sandbox working as designed); a runner failure means the command
* NEVER RAN and must surface as a sandbox failure, not a task failure —
* consumers check these signatures FIRST (a runner's own error text may
* contain denial words, e.g. an unopenable grant root reporting
* `Permission denied`).
* Case-insensitive signatures for runner failure before command execution.
* Consumers check these before denial signatures: runner failure means the
* command never ran, while denial means confinement worked and blocked it.
*/
runnerFailureSignatures: readonly string[]
}
/**
* Error `code` carried by the infrastructure error a provider throws when a
* confined policy is requested but no backend is available or usable on this
* host: confinement FAILS CLOSED (refuses to run) rather than silently
* executing unconfined. Thrown as a `HarnessError`, it reaches the model
* through the structured `{ name, code }` error channel on `tool/result`, so
* callers can distinguish "the sandbox is missing" from a failing command.
* Error code for a requested confined mode when no backend is usable. The
* provider fails closed, and `HarnessError` carries the code through
* `tool/result` so callers can distinguish missing confinement from command
* failure.
*/
export const SANDBOX_UNAVAILABLE = 'SANDBOX_UNAVAILABLE'
/**
* Thrown by {@link SandboxProvider.confine} when a confined policy is
* requested but no backend is usable on this host: confinement fails closed.
* Carries the {@link SANDBOX_UNAVAILABLE} code through the structured
* `{ name, code }` error channel.
* Thrown when {@link SandboxProvider.confine} cannot enforce the requested
* mode. Carries {@link SANDBOX_UNAVAILABLE} through the structured error
* channel.
*/
export class SandboxUnavailableError extends HarnessError {
constructor(mode: ConfinedSandboxMode, detail?: string) {
@@ -155,27 +103,10 @@ declare module 'cordis' {
}
/**
* Abstract process-sandbox service. Subclass, implement {@link confine}, and
* load the subclass as a plugin — it registers as `ctx.sandbox` (one
* implementation per context; loading a second throws, cordis' standard
* duplicate-service behavior).
*
* Semantics every implementation must honor:
* - {@link confine} either returns an argv whose runner ENFORCES the policy
* or fails closed — at `confine` time with {@link SandboxUnavailableError}
* (no backend for this host), or at EXECUTION time by the runner itself
* refusing to run the command (exiting without exec'ing it, identified by
* {@link ConfinedArgv.runnerFailureSignatures}). A silent unconfined
* passthrough is never a legal outcome on either path.
* - Probing exists to ARBITRATE between multiple candidate backends and may
* be skipped when a platform has exactly one: the sole candidate is
* selected directly and the runner's exec-time fail-closed refusal carries
* the safety property. When probing does run, it is functional (actually
* enforcing a profile, not a version check), at most once per provider
* lifetime; `confine` itself spawns nothing beyond that one-time probing.
* - The returned {@link ConfinedArgv.enforcement} states the backend's
* actual completeness for THIS host; `partial` is reported, never silently
* upgraded to `full`.
* Abstract process-sandbox service. {@link confine} must return enforcing argv
* or fail closed at wrap or runner-execution time; silent unconfined passthrough
* is forbidden. Functional probes arbitrate multi-runner chains and may be
* skipped for a sole candidate, whose own refusal remains the fail-closed end.
*/
export abstract class SandboxProvider extends Service {
constructor(ctx: Context) {