A preflight, with a clear boundary
Castles evaluates an explicit binding and reports findings. The existing lifecycle owner remains the authority for the sandbox.
1. Parse the exact binding
The binding contains the supported profile, exact sandbox ID, canonical owner-agent record ID and expected generation. Missing or malformed values fail closed. Castles does not list sandboxes, choose one by name or adopt a substitute.
These values identify the requested record and resource. The owning service must authenticate the caller and derive its own tenant; caller-supplied IDs do not establish ownership.
2. Inspect through one seam
SandboxInspectionClient exposes one operation: inspectBoundSandbox. It is the entire provider-facing surface available to Castles. A future implementation must be backed by the existing Coven Agents lifecycle owner and perform an exact-ID read.
inspectBoundSandbox(
{ agentId, expectedGeneration, sandboxId },
{ signal }
)The current implementation supplies synthetic fixture clients only. The fixture inspection contract is a proposal carried by fixtures; it is not a shipped live-provider contract. The upstream inspection issue tracks the missing capability.
3. Validate before reporting
The evaluation checks the returned binding and resource identity, generation, lifecycle, attempt fence, credential boundary, private exposure, restricted egress, image, platform, SDK generation and resource policy. Provider responses are untrusted input.
An ambiguous lookup, different returned sandbox or unresolved prior attempt requires reconciliation. An ownership refusal is unauthorized. Unsafe or missing policy metadata is incompatible. No failure triggers a retry against another target or an automatic repair.
4. Separate policy, observation and evidence
The result keeps source-pinned policy in sourcePolicy and sanitized reported metadata in observed. Target IDs appear as redacted aliases. Raw provider text and raw errors are not displayed.
The client determines the evidence class. A fixture cannot promote itself to live-provider evidence. The isLiveReady() consumer guard requires both live evidence and a ready status; see the result reference.
The trust boundary
| Component | Responsibility |
|---|---|
| Castles | Parse the binding, evaluate the profile, and emit a bounded result. No lifecycle mutation or credential custody. |
| Canonical lifecycle owner | Authenticate the caller, derive the tenant, resolve the canonical record and own provider I/O and lifecycle operations. |
| Operator / maintainer | Supply the explicit target and authorize any named live window or separate lifecycle operation. |
| Fixture client | Replay synthetic data for offline checks. It proves no real ownership, containment or readiness. |
Castles has no server, database, second provider adapter or assignment bot. It holds no provider, database or harness credential. The documentation site is static and does not invoke the CLI or inspect sandboxes.
Stop/start, cancellation containment and safe workspace reuse remain separately gated work. A finished session or disconnected client is not proof that an earlier attempt is contained.