# Marqdo 3.0.0-beta.1 — known limitations This file ships in the release archive as `KNOWN_LIMITATIONS.md`. It is part of the release record: the beta support claim is exactly what is written here. Items marked **unsupported** must fail closed, not be silently emulated. ## Platform support | Platform | Beta status | Notes | |---|---|---| | Linux x86_64 (glibc) | **Supported release platform** | The archived `marqdo` and `marqdo-provider-worker` binaries target `x86_64-unknown-linux-gnu` and are the only binaries this release claims. | | Linux aarch64 | Not built | No release artifact; source may build. | | macOS (arm64/x86_64) | Experimental, not claimed | Core CLI may work, but no macOS provider worker, no native plugin ABI claim and no macOS release artifact. | | Windows x86_64 | Experimental, not claimed | Core CLI may compile/run; the plugin/provider C ABI and provider worker are not claimed. | `pure` and `project` are **language execution modes**, not OS sandboxes. `project` may access resources only through explicit grants, but it runs as the user and is not a security boundary against hostile code. ## `strict` isolation `strict` is Linux-only and probe-gated. The beta release claims it only when the same-SHA G3 evidence recorded in `docs/evidence/` is present; otherwise this file and the release notes must say **strict: unsupported** and the runtime must fail closed (no project code executes). Even when available: * strict L1 covers the pure language plus native Math; * strict L2 (Worker/Foreign/Provider inside strict) is **unsupported** and every such flag is refused with exit 2; * rlimits are used for resource limits, not cgroup-v2 accounting; * the network is denied with a network namespace, not a seccomp allowlist; * a PASS is evidence for isolation correctness, never for performance. ## Extensions * **Callable plugin ABI v1**: native and worker transports exist. Native loads into the host process and requires the explicit unsafe `--trust-native-plugin`; a C ABI is not a sandbox. Worker is crash isolation with handshake/call timeouts, not an OS sandbox. Wasm, hot unload and a full worker capability broker are unsupported. Windows C ABI is not claimed. * **Resource Provider v2**: Linux x86_64 worker transport only. The `marqdo-provider-worker` helper must be next to the `marqdo` executable; missing/renamed helpers fail closed (`ext.worker_unavailable`) and PATH is never consulted. A provider is not a sandbox. * **Foreign kernels**: any language can be registered with `kernel.json`, but the kernel process is a normal child process; it is not sandboxed, and the host must refuse unauthorized, digest-mismatched, timed-out or wire-incompatible kernels. * **CLI tools**: a tool is an ordinary user-authorized executable and does **not** inherit `pure`/`project`/`strict` guarantees from the documents it reads. `MARQDO_BIN`, `MARQDO_TOOL_PROTOCOL`, `MARQDO_PROTOCOL_ANALYSIS` and `MARQDO_PROTOCOL_TRACE` are the only helpers the CLI injects. * **Extension selection** (`marqdo.yaml.extensions`, lock v4) is never authorization. `--allow-plugin`, `--allow-provider` and `--allow-foreign` are still required. ## Compatibility and versioning * `3.0.0-beta.1` is a **prerelease**, not stable and not LTS. Breaking changes inside the 3.0 beta series are allowed and announced in the changelog. * The product version is independent of protocol versions: Analysis v2, Trace v2, Execution Events v1, Tool v1, Plugin ABI v1, Foreign wire v1, Provider manifest v2 and lock v4 keep their own major numbers. Do not infer them from `3.0.0-beta.1`. Execution Events v1 is a post-beta addition and is not in the tagged beta archive. * The official bundled stdlib keeps its independent package version `0.1.0`. A plugin `.mq.md` package version such as `0.1.0` is also independent. * `project.marqdo` in `marqdo.yaml` is enforced in this release. SemVer prerelease rules apply: `>=3.0.0,<4.0.0` does **not** accept a beta core; beta users must declare a prerelease range such as `>=3.0.0-beta.1,<4.0.0`. * Old `cflmy/marqdo` source syntax, package format and plugin ABI v2 are not compatible. There is no `--legacy` or downgrade path; migration is a separate documented workflow. ## Artifact integrity Archives are accompanied by SHA256 checksums and `RELEASE_INFO.json` (`source_sha`, target triple, toolchain, binary hashes). Checksums prove content integrity only; they are **not** a publisher signature. A digest obtained from an untrusted Git URL does not make the code trustworthy. ## Execution Events v1 (post-beta addition) * Implemented: live stdout/stderr chunks, lifecycle/terminal events, call instance ids, branch/loop/effect events, explicit bounded value observation. * Not implemented: memory region/budget metadata (OBS-06/P2), raw memory content, breakpoints/stepping/time travel, and a bidirectional debug protocol. Do not infer a debugger from this protocol. * `--mode strict` currently rejects `--event-stream` explicitly; there is no silent fallback to `project`. * The live lane is tested on Linux x86_64; macOS/Windows are unverified for this option like the rest of the beta CLI. * In-process native plugins are rejected in event mode (`ext.event_stream_unsupported_native_plugin`) because they share the process and can write the protocol fd directly; worker plugins remain usable. * Observation lowers confidentiality when values are explicitly enabled. Text needs a second opt-in; View/agents must not enable these flags by default. ## Deliberately deferred strict L2, Wasm plugins, macOS/Windows provider worker, automatic uv/npm environment creation, a distributed package registry, JIT execution, built-in Knowledge/Proof/View commands, a mature async/concurrency model, and platform installers. The planned direction is external Tool repositories consuming the public protocols, not new core commands.