WHY - The code fixes existed locally without same-SHA release material; the G0..G12 gate matrix, G-C VM evidence, RC artifact and explicit non-claims must be reviewable together. - The release-candidate workflow packaged no consumers and did not verify its own checksum in the clean-install directory. WHAT - Record the current Linux/VM evidence atf075c96and keep remote CI/G-C as UNVERIFIED; overall decision DO NOT RELEASE. - Add G-C evidence (7/7 + positive CLI + 47-binary workspace run), RC report with gate matrix, clean-install and checksum result. - Update STATUS, COMPATIBILITY, CHANGELOG, README, tests/README, mapping spec and releasing procedure for the new condition/span/power/Tool behavior. - Harden release-candidate packaging: include tests/consumers, SOURCE_SHA, rustc -vV, verify the checksum before extraction, and run independent Analysis/Trace validation from the extracted archive. - Record the local convergence note in NEXT_RELEASE_CI_CONVERGENCE.md. TESTS - Local fmt/clippy/workspace PASS; G-C VM PASS atf075c96. - Linux RC: sha256 -c OK and clean-install smoke PASS. CI_URL - none: Gitee remote, no GitHub credentials. Windows/macOS and same-SHA G-C CI remain the blocking gates. BREAKING_CHANGE - none in this docs/workflow commit. SECURITY_SCOPE - strict remains fail-closed; no new capability or platform claim. G-C is local-VM evidence, not a CI-signed security URL.
5.8 KiB
5.8 KiB
Changelog
All notable changes to the Marqdo CLI workspace. The workspace version is
currently 0.1.0; no release tag has been cut. Dates use ISO 8601.
Unreleased (stabilization candidate)
Added
.gitattributesbyte-identity policy: canonical LF for text files and one explicit CRLF byte fixture for Markdown line-ending semantics.- Byte-level AST golden diagnostics: first differing byte, byte windows and an explicit "only line endings differ" classification.
- Cross-platform Native ABI test fixtures generated for the host target, plus
no_matching_artifact, artifact-tamper and ABI-major-mismatch negative tests. - Host-specific example plugin builds under
examples/plugin-*/build/; the tracked manifests are never rewritten by tests or CI. - Foreign pinned-environment binding (ADR-0020): an external provider's
entrypointmust resolve insideenvironment_dir, built-in providers cannot be bound to one, the child gets only the environment executable directory onPATH, andinterpreter_versionis verified before execution. - Linux strict isolation (ADR-0021): a Bubblewrap supervisor with user/mount/
PID/net/UTS/IPC namespaces, no-new-privs, capability dropping, a read-only
/project, rlimits, output/wall-clock supervision and child self-verification. Probe-gated; L1 supports pure Marqdo + Math only. - Adversarial G-C lane (
tests/security/escape_helper.c,crates/marqdo-sandbox/tests/security_linux.rs,tests/security/run-lane.sh) covering filesystem,/proc, network, process, privilege and resource escapes. 7/7 PASS in a disposable Ubuntu 26.04 VM. - Release documentation (
COMPATIBILITY.md,docs/releasing.md) and the Linux security-gate evidence report (docs/security-gate.md). check_never_spawns_a_foreign_kernel: proves staticcheckdoes not spawn a declared Foreign provider, whilerunon the same document does.- Power/control-flow corpus: 10 new AST goldens, 8 semantic goldens, 4
compile-negative fixtures and 2 runtime-negative fixtures covering
right-associative
**, unary-minus precedence, parentheses, float negative exponents, Strong+CodeSpan payloads, branch/loop conditions, inert prose and code fences, and source spans inside one/two-backtick code spans. - Independent Tool Protocol v1 consumers
sample-view,sample-knowledgeandsample-proof(analysis/trace readers) with install, dispatch and remove tests; none is a hard-coded CLI command. - Locked path-dependency content-tamper test: a changed dependency is rejected
without rewriting
marqdo.lockor the materialized.marqdo/env. - G-C lane honours
CARGO_TARGET_DIR, so a run from a cleangit archivewith an external target directory uses the binary it just built.
Changed
docs/STABLE_RELEASE_IMPLEMENTATION.mdis the single next-phase implementation plan. Older plan documents are historical input, not current status.- AST conformance tests fail early when a golden contains CR bytes instead of reporting an opaque mismatch.
- Worker plugin fixtures use the platform dynamic-library name and compiler flags.
- List-item condition paragraphs now accept only inert description; an
executable
Strong/Emphthere is a hardmarkdown.invalid_constructerror instead of being silently swallowed. Condition expression spans are mapped from the literal bytes inside the code-span backticks. - Linux strict
inspect --platformremains a snapshot, not a promise. A run that loses namespace/resources between probe and launch must fail closed.
Fixed
- Windows CI AST golden failure caused by CRLF checkout of canonical JSON
goldens (reproduced with a fresh
core.autocrlf=trueclone). - macOS CI plugin failure caused by assuming a Linux
.soand Linux-only manifest target. - macOS/Windows
E0433: platform detection now uses#[cfg]functions instead ofcfg!aroundlinux::probe(), so the Linux module is not referenced on other targets. - Linux strict probe→launch race: the CLI contract and the sandbox seam now require a non-zero, diagnosed, fail-closed refusal when the real supervisor launch fails after a successful probe.
- Parallel security tests shared process-id scratch paths and deleted each other's project; every case now has a unique path and the helper is compiled once.
- Condition parse diagnostics pointed at the backtick instead of the missing operand; the payload map now excludes delimiters.
Security
- Linux
strictis implemented and probe-gated. It isAvailableonly when a real Bubblewrap smoke test succeeds; otherwise it staysUnavailableand refuses. There is no silent downgrade toproject. - Strict L1 covers pure Marqdo and native Math; Foreign, native/worker plugins, kernels, external grants and trace files are rejected before the sandbox starts. Strict L2 inheritance is future work.
- Capability brokers and worker processes remain cooperative authorization/crash isolation, not an OS sandbox.
- The G-C lane was re-run at
f075c96from a cleangit archivein a disposable Ubuntu 26.04 VM as ordinary userdev: 7/7 adversarial tests, positive CLI contract and forged-marker rejection PASS. This is still a local VM result, not a self-hosted GitHub security-lane URL.
Baseline 0.1.0 (before this stabilization slice)
- Markdown profile
marqdo.md.v1with 150 AST goldens. - Native Rust interpreter, analysis and execution conformance corpus.
- lock v3 package-instance graph with full transitive closure, verified Git
cache and atomic
.marqdo/envmaterialization (ADR-0016). - Analysis v2 and Trace v2 with independent Python consumers and JSON Schemas.
- Tool Protocol v1 with SemVer/protocol negotiation and whole-bundle digests.
- Plugin ABI v1 native and worker transports with a brokered
fs.readcapability path (ADR-0018). - Pinned Foreign provider/environment digests and one-shot Python/Node kernels.