9626c15c6b
The policy home's resolve() now stamps readDenyPaths, so every consumer that pins the resolved shape (bash-sandbox hand-off, tool-fs stamps) carries it, and three uncovered branches gained real tests: landlock reporting partial enforcement for a denial it cannot express, the policy's default under programmatic construction, and both ambient credential paths in llm-deepseek without a mounted seam.
dsh-sandbox-policy — the sandbox policy home (ctx.sandboxPolicy)
English | 中文
The single owner of sandbox-policy resolution: the deployment's default SandboxMode and fallback root, plus each session's durable mode override and immutable workspace root. Every enforcing capability family receives one resolved mode-and-root policy per call.
Why a shared home
Two families enforce the same mode vocabulary: the sandboxed bash executor (@deepseek-ai/dsh-bash-sandbox) and the sandboxed filesystem provider (@deepseek-ai/dsh-fs-sandbox). If each resolved its own mode + workspaceRoot, the two could drift into a split world — bash confined to one root while fs fences another, exactly what the sandbox RFC warns against. Both tool layers resolve policy through ctx.sandboxPolicy, and both enforcing backends consume that complete per-call result. The cross-family fs sandbox RFC records the shared-policy decision.
Config
mode— the deployment defaultSandboxMode(read-only/workspace-write/danger-full-access), validated at load. Defaultread-only(fail-safe).workspaceRoot— the fallback directoryworkspace-writemay write under for agentless calls or sessions without a cwd. Defaultprocess.cwd(), resolved to its absolute filesystem identity either way. A normal agent call uses its session header's immutablecwdinstead.
Read denials
readDenyPaths names absolute paths a confined execution must not read, whatever its mode otherwise permits. Omitted (or empty) denies the harness credential document $DSH_HOME/.env; a non-empty list replaces that default. Denials name exact paths rather than roots on purpose: denying the whole harness home would also take away the model's documented access to its own session log.
Enforcement is backend-shaped. Seatbelt appends a trailing deny file-read* file-write* (last matching rule wins) and bwrap maps /dev/null over each path after any workspace bind; Landlock grants are a pure allow-list, so a read grant on / cannot be subtracted from and confine() reports partial enforcement rather than pretending the boundary exists. danger-full-access confines nothing at all, so no denial applies there — the credential document is then protected only by its file mode, which does not stop a same-UID tool process.
Surface
ctx.sandboxPolicy.resolve({ session?, mode? })— resolves one complete per-call policy. An explicit approved mode outranks the session's lastsandbox/modeevent, which outranksdefaultMode; the session's immutablecwdis canonicalized with filesystem semantics before becomingworkspaceRoot, otherwise the configured fallback applies. Canonicalization precedes lexical normalization sosymlink/..agrees with process working-directory resolution.ctx.sandboxPolicy.defaultMode/ctx.sandboxPolicy.workspaceRoot— the deployment default and fallback root used byresolve().effectiveSandboxMode(events)— the pure fold of a session'ssandbox/modeevents (the last switch wins, orundefined), used insideresolve().setSandboxMode(session, mode)— THE write path for a per-session override: appends exactly onesandbox/modeevent. The switch IS its event; nothing mutates the mode out of band.SANDBOX_MODES— every mode, for option advertisement and runtime validation.
The optional ./invariant companion rejects a forged durable sandbox/mode event whose value falls outside that closed vocabulary; Session and its companion own the surrounding storage and core execution-enclosure rules.
The per-session store
A runtime switch is one log-only sandbox/mode event on the session it applies to. effective = explicit grant ?? fold(events) ?? deployment default, so an override survives restart by replay and two sessions never see each other's state. Workspace identity does not need another event: the immutable SessionHeader.cwd recorded at creation is the root for every call in that session. The event is log-only (the approval/* precedent): the model learns the mode from the enforcing tools' denial markers, never from the event.
Model Experience
Indirectly, through dsh-tool-bash and dsh-tool-fs, which render the effective mode this service holds in their [sandbox: …] denial markers and escalation prompts; the sandbox/mode event itself never reaches the model.
KV Cache effect
No direct invalidation; the named consumers own any request-prefix changes, and the mode is deliberately absent from the prompt.
Known Limitations and Deferred Work
- One primary workspace root per session — policy resolves
SessionHeader.cwd; extra writable roots are not part ofSandboxExecutionPolicy. - File-effect modes only —
SandboxModegoverns file effects; network and process policy are outside its vocabulary, so no knob here restricts them.