- Add Program.print_override and lazy stdlib injection: only programs that
call 打印/print load std:io and route to io.写值 (ADR-0030).
- Record resource.open/read/write/close as trace effect events so the stdlib
print path is auditable; HostOp::Print no longer runs on that path.
- Permit only stdio/stdout (existing console) in strict; file/network
resource opens remain permission.denied.
- Exclude std: modules from the human-readable checked-module count; update
the Python analysis consumer to accept std: instance ids.
- Add print-routing and strict-console tests, ADR-0030 and evidence.
Gate: fmt, clippy -D warnings, 418 tests passed.
WHY
- macOS/Windows failed to compile because `cfg!` left `linux::probe()` in the
type-checked branch (E0433).
- GitHub-hosted Linux can pass the strict probe and then fail the real launch
with EAGAIN; the CLI test treated the inspect snapshot as a permanent promise.
- The security matrix shared process-id-based scratch paths, so parallel
`cargo test --workspace` made its own tests clobber each other.
- Condition paragraphs silently ignored executable `Strong`/`Emph`, and
condition parse spans included the code-span backticks.
- The corpus had no native power/control-flow/inert fixtures and no independent
generic Tool consumers.
WHAT
- Dispatch strict detection through `#[cfg]` functions; add a non-Linux
compile/fail-closed regression test.
- Split `linux::run` into an injectable probe/spawn seam; test probe-OK ->
launch-EAGAIN returns LaunchFailed and never executes unrestricted.
- Make the Linux CLI contract accept only (a) successful isolated run or
(b) non-zero, diagnosed, fail-closed refusal with no project stdout.
- Give each security test a unique project/outside path and compile the helper
once.
- Reject executable roles in a condition paragraph; map condition expressions
from the literal bytes inside the backticks.
- Add 10 AST, 8 semantic, 4 compile-negative and 2 runtime-negative fixtures
for power, branch/loop conditions, source spans, inert prose/code fences and
error boundaries.
- Add `sample-view`/`sample-knowledge`/`sample-proof` independent Tool v1
packages and an end-to-end Analysis/Trace consumer test.
- Add a locked path-dependency content-tamper test.
TESTS
- cargo fmt --all -- --check
- cargo clippy --workspace --all-targets -- -D warnings
- cargo test --workspace --locked
- G-C VM lane (separate evidence commit/run)
CI_URL
- none from this host: the checkout has a Gitee remote and no GitHub
credentials. Windows/macOS CI remains UNVERIFIED.
BREAKING_CHANGE
- An executable node in a list-item condition paragraph is now a hard
`markdown.invalid_construct` error instead of being silently ignored.
SECURITY_SCOPE
- strict remains fail-closed; no downgrade, no new capability, L2 still
rejected. G-C still requires the dedicated self-hosted lane for CI-signed
evidence.
Strict mode now re-executes the CLI inside a Bubblewrap supervisor with
user/mount/PID/net/UTS/IPC namespaces, no-new-privs, all capabilities dropped,
a read-only /project, private /proc and /dev, rlimits, output capture and
wall-clock process-group cleanup. The child verifies PID 1, NoNewPrivs, empty
CapEff, an unshared uid map and loopback-only networking before project code.
- PlatformCapabilities::detect runs a real bwrap probe; no probe, no strict
- strict L1 supports pure Marqdo + Math; Foreign/ABI/worker/kernel/grants/trace
are rejected with exit 2 before sandboxing
- forged MARQDO_STRICT_ACTIVE is rejected; missing bwrap fails closed
- add adversarial helper, 7-test escape/resource matrix, seeded security lane
script and self-hosted security workflow
- executed G-C lane as uid 1000 in a disposable Ubuntu 26.04 VM: 7/7 PASS
plus positive CLI and refusal matrix; evidence in docs/evidence
- ADR-0021, spec/security.md, STATUS/COMPATIBILITY/README/CHANGELOG updated;
ADR-0019 remains for non-Linux/fail-closed rules
A document with an executable foreign.call is statically checked without a
marker side effect, then run to show the same provider does execute. This
guards the 'check/inspect cannot execute untrusted code' contract.
A verified environment_dir was only prepended to PATH, so built-in providers
still ran the host interpreter and external entrypoints could live outside the
verified tree.
- external entrypoints must resolve inside environment_dir
- built-in providers cannot be bound to an environment (fail closed)
- bound children get exactly the environment bin on PATH, no host fallback
- declared interpreter_version is verified against the resolved program
- built-in descriptors now also run environment validation
- ADR-0020 + spec/foreign.md document the binding and limitations
- positive/negative tests for binding, outside-entrypoint, builtin binding and
version drift
The macOS job failed with no_matching_artifact because the integration test
copied the Linux-only example manifest and built a Linux-named .so. Tests now
generate a manifest for the actual host triple, use the platform dynamic
library name and compiler flags, and never rewrite the tracked example
manifest.
- add a cross-platform no_matching_artifact test that runs without a C compiler
- add real tamper and ABI-major-mismatch rejection tests
- make worker fixtures platform-native; Windows remains explicitly unsupported
(no cc) rather than counted as a pass
- example build scripts emit host-specific build/ fixtures (gitignored)
- CI smoke points at the generated fixtures
The Windows CI failure was a checkout byte-identity problem, not a Markdown
semantics change: CRLF goldens compared with LF parser output produced
'0 line(s) differ'.
- add .gitattributes: text files are LF, native products are binary, and the
intentional CRLF fixture stays a byte fixture
- report the first differing byte and whether only line endings differ
- fail early when a golden contains CR instead of the opaque mismatch
- assert exactly one CRLF fixture and that its golden is canonical LF
- add a markdown test proving CRLF/LF have equal structure while spans remain
byte offsets into the original source
Add ADR-0019 freezing the strict availability contract: strict remains
unavailable with an actionable reason on every platform until runtime probes
and filesystem/network/process/resource escape tests pass in a dedicated
security lane. Add a CLI test asserting inspect --platform reports unavailable
and worker processes are not presented as isolation. No G-C claim is made.
- add environment_dir to KernelDescriptor and verify its deterministic
file-set digest at registration and immediately before every spawn
- reject missing digests, relative/missing dirs, symlinks, special files and
oversized environments; tampering fails with EnvironmentIntegrity
- inject MARQDO_FOREIGN_ENV and prepend <env>/bin to PATH for the child
- add a real provider test covering success and post-registration tampering
- document that uv/npm environment generation and lock-v3 integration remain
open
- enforce at most 64 capability requests and 64 KiB per serialized request
in both the worker handler and the host bridge
- return structured capability.budget_exceeded / request_too_large errors and
kill the worker on a host-side violation
- add a real C loop fixture and CLI test proving a runaway plugin is stopped
without producing file contents
- update ADR-0018, plugin-abi and STATUS
Replace the /bin/sh tool fixtures with a std-only fixture compiled by rustc,
and use platform EXE_SUFFIX for entrypoints. Windows/macOS workspace tests now
run the whole Tool protocol suite instead of skipping it. Keep C ABI/worker
and package/Git tests Unix-only and document that remaining gap.
- add request-scoped capability_request/capability_reply/capability_error
frames to the worker transport (ADR-0018); no C ABI change
- worker routes its capability handler through stdin/stdout, the host serves
requests inside the existing invoke deadline and binds identity from the
spawned worker, never from frame text
- both native and worker plugins now execute a granted fs.read through
ProjectFs with audit; unsupported capabilities return capability.unavailable
- add a real C capability plugin and CLI test for granted read and denial
- document the bridge and that strict OS isolation (H3-B) is still refused
- CapabilityBroker::execute separates authorization from execution; native
plugins requesting fs.read get real project file bytes through ProjectFs
- unsupported capabilities (fs.write/net/process/console/vendor without a
handler) return capability.unavailable instead of empty success
- keep audit decisions for granted/denied/unavailable without logging contents
- document that worker capability bridging, live ABI session identity and OS
isolation (H3-B) remain open; no G-C claim is made
- add environment_id/environment_digest/interpreter_version to KernelDescriptor
- require external CLI descriptors to pin an absolute executable and a full
provider digest
- verify the digest at registration and immediately before every spawn;
tampering after registration fails with ForeignError::Integrity before spawn
- validate digest formats and reject relative external entrypoints
- pin the example provider digest and add an integrity test
- document that uv/npm environment materialization and lock integration remain
open; strict isolation is unchanged and still refused
- replace the hand-rolled version matcher with semver; require valid tool
versions and marqdo ranges
- declare SupportedProtocols (tool 1, analysis 2, trace 2) and negotiate tool
Analysis/Trace declarations at install and dispatch; disjoint/empty sets fail
closed with tool.protocol_incompatible
- record a v2 whole-bundle manifest (path/size/sha256 + deterministic bundle
digest), copy exactly that bundle into staging, re-verify it, and recompute it
before every dispatch
- reject auxiliary-file, extra-file, missing-file and entrypoint tampering,
record/command mismatches, traversal and a source root record.json
- remove tools by detaching with rename before deletion
- expose negotiated protocol environment variables to declared tools
- update tool-v1 and STATUS; add negotiation/bundle tests
- bind traces to the logical project identity and the exact emitted Analysis
digest; every span-bearing event carries module_id
- add host TraceState budgets (100k events / 32 MiB defaults), CLI
--trace-max-events/--trace-max-bytes, terminal trace.truncated events, and a
final serialized byte cap; budgets never change run.status
- disable semantic recording when no trace is requested
- cap worker frames at 8 MiB, align request/result/capability wire caps, make
clean EOF distinct from protocol errors, and mark/kill a worker after a
transport failure instead of reusing it
- add trace-v2 spec/schema, independent trace consumer, CI check and tests
- record the S2/C state in STATUS
- add ProjectIdentity (manifest root instance or workspace snapshot digest)
and remove absolute-path root ids from Analysis bytes
- replace the resources count with an addressable array over the two real
marqdo-ir resource kinds (formula, code), including unit/span/payload digest
and code language
- project diagnostics by module/relative path/span, removing machine paths
- define the digest as SHA-256 of the exact emitted bytes so consumers can
verify it; Trace binds to the same digest
- add analysis-v2 spec and JSON schema, supersede v1, update the independent
Python consumer, schemas and docs
- tests: location-independent byte equality and resource addressing
RuntimeError and CallFrame now carry a logical ModuleId and an origin
(source/callsite). Errors are attributed at function/body boundaries so the
innermost callee module wins; argument-binding errors and host/plugin/foreign
boundaries are attributed to the caller callsite without fabricating a foreign
stack.
The CLI resolves (module_id, span) through LoadedProject::locate instead of
indexing every span against the entry file. Unknown modules and out-of-range
spans keep the module id and byte span but emit null line/file. Human output and
frames now point at the correct file.
Record the decision in ADR-0017 and supersede the ADR-0006 limitation. Add an
end-to-end imported-module division-by-zero test.
- lock the full transitive closure with per-instance alias edges, root
manifest digest, full content digests and structured path/git sources
- reject v1/v2 locks; validate ids, digests, edges, reachability and sources
- resolve every locked edge by (parent instance, alias) and fail closed on
missing locks/edges/records, digest or source drift and root drift
- require a lock for run/check/inspect when dependencies exist and map lock
errors to exit 1
- make sync strict in both modes, never rewrite the lock, and materialize the
whole closure with env.json schema 2
- verify Git cache markers, require full commit hashes and confirm HEAD
- stop add/remove from patching the lock; regenerate example locks
- update package-format.md and STATUS for S1
Add content_profile_version 1 with one selection shared by hash_tree and
materialization: exclude .git/.marqdo, include target, reject symlinks and
special files, propagate walk errors, enforce file/size/depth budgets, and use
a domain-separated length-prefixed encoding (fixes the old NUL-delimiter
ambiguity). Materialization hashes exactly the bytes it writes and verifies the
expected digest before publishing.
Record the decision in ADR-0016 and update the package-format hash contract.
The property test builds all fixtures in code but asserted that
tests/properties/ existed. Git does not track empty directories, so a fresh
clone failed before exercising the invariance checks. Drop the assertion and
the now-unused repo_root helper; keep every prose-invariance assertion intact.
Record the S0/H0 local gate and the pending CI URL in STATUS.
- pass workers an explicit temporary manifest so plugin.yaml naming is not a
hidden constraint
- remove temporary manifests after handshake
- reconcile STATUS with implemented multi-export, broker, worker, handle,
Git/env and CI behavior; document strict isolation as the remaining blocker
- add a hidden __plugin-worker mode that loads one native library in a child
process and serves length-prefixed JSON invoke frames
- add host WorkerPlugin with ready handshake, export cross-check, per-call
timeout and crash/EOF propagation
- route kind: worker without in-process trust; native still requires explicit
trust; wasm is rejected explicitly
- report worker_processes=true while keeping strict OS isolation unavailable
- test successful worker calls and worker crash isolation
- update plugin ABI spec and R2 status
- add runnable examples for an unknown-name Tool, multi-export plugin,
external kernel provider and same-name multi-version packages
- add JSON Schemas for Analysis v1, Trace v1 and Tool v1 and validate them in CI
- extend CI with protocol/consumer smoke tests and a Windows/macOS cargo check
- add --kernel provider descriptors and R6 example integration tests
- document the examples and update M0.6 status
- add content-addressed Git checkout cache for exact revisions; no hooks,
submodules or floating branches
- resolve transitive @alias imports from each package's own manifest
- materialize locked path/Git packages into .marqdo/env atomically and prefer
a matching env copy in the resolver
- make sync a real environment sync while keeping lock immutable
- kill Unix foreign-kernel process groups so grandchildren do not survive
- add Git/env/transitive integration tests and update package/security specs
- add check --emit with a deterministic, read-only semantic projection:
package-aware modules, signatures, normalized bodies, refs, calls,
diagnostics and source digests
- remove stale analysis artifacts on compile errors; atomic writes
- add run/call --trace-file with interpreter call/branch/loop/return events
and host effect events under one sequence; redact values and record error
status
- add Analysis/Trace v1 specs, ADR-0015 and an independent Python consumer
- add contract tests for determinism, stale artifacts, trace ordering/status
and secret redaction
- add tool install/list/remove and unknown-command dispatch without hardcoded
tool names or clap variants
- validate and copy local tool packages atomically, reject entrypoint escapes
and symlinks, record and re-verify SHA-256
- forward OsString args, stdio, cwd and exit status without a shell; never
execute PATH programs implicitly
- add black-box tests for arbitrary names, Unicode args, built-in conflicts,
traversal, version mismatch, tampering and PATH hijacking
- add ADR-0013 and mark R3 complete
- add PluginHandleTable with owner, run id and generation scoping
- reject forged, cross-owner, cross-run and post-unload handles
- reject any plugin_handle returned by plugin or foreign-kernel output,
including nested values
- add ADR-0014 and update value-wire/security specs
- register every plugin export by its complete dotted name; reject conflicting
loads atomically instead of overwriting the first namespace
- route plugin calls by full export name and expose loaded descriptors to the
runtime
- add CapabilityBroker for core grants and vendor policy, with audit records
- accept syntactically valid vendor capability requests in manifests while
still denying them until a trusted policy exists
- add check --plugin static signature validation without dlopen
- test multi-export C plugin routing, conflict rollback, broker decisions and
static arity rejection
- add KernelDescriptor/KernelRegistry with built-in Python/Node providers
- route run/call through the registry; external providers need an explicit
executable and wire version
- expose provider registration on StdHost and use it in the foreign adapter
- add mocklang registration/call, duplicate-label, wire-version and timeout tests
- update spec/foreign.md and M0.6 progress
- add CallableId/Descriptor/Origin and CallableRegistry in marqdo-ir
- include builtins, Marqdo functions and declared ABI exports in one static
lookup model with deterministic conflict rules
- validate declared plugin export signatures in the frontend without loading
any shared library
- add registry and frontend static-plugin-signature tests
- replace path-only ModuleId with {instance, relative_path} and canonical
<instance>::<path> encoding
- add PackageResolver/ResolvedPackage in marqdo-package for path sources;
hash package trees deterministically for content-addressed instance ids
- move import resolution out of the CLI loader; reject package escapes
- emit lock_version 2 with instance ids and migrate v1 lockfiles
- update references/package-format specs and the function-ref golden
- add package resolver tests for stable identity across path moves
- host: `打印` checks the console capability at the execution point (pure and
project modes grant it, but the boundary is now real instead of decorative).
- cli: `marqdo call --json` reports "command":"call" instead of "run".
Non-error diagnostics (for example `markdown.inert_executable` for a would-be
execution node outside any function body) are now printed to stderr before
execution and included in `--json` run output, instead of being silently
hidden on a successful run.
Findings from the conformance corpus workstream, all fixed at the source:
- value/wire: enable serde_json's float_roundtrip so typed wire floats keep
their exact digits across foreign/plugin boundaries (Python pi*9 now stays
28.274333882308138).
- host: default interpreter stack budget lowered to 128 frames so deeply
recursive documents hit a structured resource.exhausted instead of exhausting
the native debug-build Rust stack; --stack remains available.
- frontend/map: calling an external HTTP(S) link is reported as ref.unresolved
(consistent with other unresolvable call targets) instead of math.unsupported.
- frontend/analyze: invalid front matter emits a single frontmatter.invalid
diagnostic (the YAML error is not repeated by the imports pass).
- spec/mapping: clarify that would-be execution nodes outside any function body
are inert with a markdown.inert_executable warning (no "return outside a
function" hard-error claim).
- tests: add the conformance corpus and harness (150 AST goldens, 36 semantic
goldens, 48 hard-error + 6 warning + 10 runtime-error negatives, CLI/module/
plugin integration tests, prose-invariance properties) and update the
three reported observations to the fixed behaviour.
Blueprint section 4.2: reference edges are static, never execution.
- For every loaded document, resolve same-document anchors, relative file
links and target anchors (analysing referenced files statically without
executing them); unresolved targets produce `ref.unresolved` errors with
source spans.
- External HTTP(S) links are treated as external references, not errors.
- `inspect --refs` continues to print the resolved edge graph as JSON.
Blueprint section 10 / P6: exports of an explicitly trusted plugin are callable
from a `.mq.md` file through a bound namespace.
- host: StdHost::load_plugin requires an explicit trust decision and a granted
plugin capability, loads the native artifact through the ABI v1 loader with
no capability handler (all plugin host-capability requests denied), and
dispatches HostOp::PluginInvoke through the typed wire request/response.
In-process native code is documented as sharing the host address space; no
isolated worker transport exists in this build.
- runtime: a qualified call whose base is not a local value, import alias or
builtin namespace is routed to the plugin namespace; plugin calls take named
arguments only and plugin values become Marqdo values (Unit when absent).
- cli: run/call gain `--plugin <path>`, `--trust-native-plugin` and reuse
`--allow-plugin <package>`; missing trust/grant/loaded namespace surface as
structured errors.
- frontend: member-call bases are namespaces (builtin/import/plugin), never
inferred parameters, so `solver.add(...)` does not invent a parameter.
- `marqdo add <name> --path <dir>` / `--git URL --rev SHA` and
`marqdo remove <name>` edit marqdo.yaml only after validating the dependency
source (path-relative, git pinned by exact rev, no duplicate entry).
Removing a dependency also drops its stale lock entry; locking stays an
explicit step so no version changes happen silently.
- Manifest serialization omits empty dependency/plugin maps and empty
capability requests, keeping generated manifests readable and stable.
`inspect --ast` and AST goldens now show `{"kind":"heading","level":2,...}`
instead of a nested `{"kind":{"kind":"heading",...}}` object, matching
spec/markdown-profile.md section 5 and the shape new contributors expect.
- calls: validate_program runs after a whole project is loaded and checks
qualified calls through explicit import aliases against the imported module's
signatures (missing function, missing/unknown arguments), plus qualified
builtin calls such as math.eval; local calls stay checked by
validate_module so diagnostics are not duplicated.
- infer: `math.eval(...)` / `foreign.call(...)` callee namespaces are builtin
namespaces, not free variable reads, so they no longer become inferred
parameters or unknown-variable errors. Explicit signatures behave likewise.
Blueprint section 8, minimum viable P3:
- marqdo.yaml parsing/validation: project identity, dependency sources (path or
git with mandatory exact rev), plugin requirements, capability requests;
absolute dependency paths, ambiguous sources and unpinned git are errors.
- marqdo.lock: lock_version, sorted packages, canonical YAML with no
timestamps or machine paths so repeated locking is byte-identical.
- discover_root walks up to the nearest marqdo.yaml.
- Manifest::template for `marqdo init`.
Blueprint section 10: a first-class plugin ABI independent of the legacy
MARQDO_ABI_VERSION=2, with a C ABI that is explicitly not a sandbox.
- include/marqdo_plugin_abi_v1.h: frozen stable-layout C contract with
documented buffer ownership/lifetimes, size limits, NULL/zero-length
semantics, single-threaded invocation, no hot-unload policy, version
negotiation and target-triple requirements.
- marqdo-plugin-abi (host side): repr(C) mirrors with compile-time layout
assertions, major/minor negotiation (wrong major fails, minor clamps),
plugin.yaml parsing/validation, target selection and sha256 artifact
verification, wire_version 1 typed JSON request/response envelopes,
structured PluginError, and a libloading-based loader with host alloc/free,
export registration and runtime ABI re-negotiation.
- marqdo-plugin-sdk (plugin side): mq_plugin_entry! macro, panic containment,
safe byte-buffer wrappers, capability/logging context and JSON helpers; it
depends only on the shared ABI crate.
- examples/plugin-demo: dependency-free C plugin (solver.add), manifest,
build.sh that refreshes the artifact hash, README explaining that install,
inspect and load are three different trust decisions.
- Tests: C plugin cross-compiled with cc and loaded, plus wrong ABI major,
missing symbol/artifact, hash mismatch and malformed manifest cases.
Marqdo-math: limited, documented LaTeX-subset kernel (blueprint section 5).
- Exact BigRational arithmetic, structured Float only for irrational results;
Int result when exact; bounds on exponents, source size, AST nodes and result
bits so hostile input cannot hang or exhaust memory.
- Supports numbers, Unicode identifiers, \pi/\e, + - * / ^, \frac, \sqrt,
\cdot/\times, braces, parentheses, one '=' equation, and linear solving for a
target variable (F=ma solved for F/m/a).
- Unsupported constructs are structured errors; the kernel never claims any
mathematical proof.
Marqdo-foreign: explicit kernels (blueprint section 6).
- run/call/detect only; reading a document never spawns a process.
- Direct exec of python3/python/node with a temp program file, clear env,
private cwd, JSON request envelope on stdin, captured bounded stdout/stderr,
wall-clock timeout with kill, typed wire decoding for call, structured
ForeignError for spawn/protocol/timeout/oversized output.
- Process isolation only: crate docs state this is not a strict OS sandbox.
Implements the single side-effect boundary from blueprint sections 7.3 and 9:
- Capabilities: console/math, project-relative fs globs, net hosts, process,
foreign, plugins; pure/project defaults deny external I/O and grants are
explicit, checked at execution time (never at parse time).
- Budget: evaluation fuel, call-stack depth, program output bytes and a
wall-clock deadline, all reported as resource.exhausted.
- EffectTrace: op/module/span/duration/result entries with no secret data.
- StdHost: canonical Host implementation dispatching print, fs read/write and
the math/foreign kernels; all errors are structured RuntimeErrors, including
a permission.denied path with a grant hint.
- ProjectFs: project-rooted, rejects absolute paths, `..` traversal and symlink
escapes outside the root; bounded UTF-8 reads and safe writes.
- Math adapter: parse/eval/substitute through marqdo-math with exact rational
binding conversion and explicit unsupported-operation errors.
- Foreign adapter: capability-gated one-shot run/call through marqdo-foreign,
typed wire input, structured result map, timeout/output limits mapped to
budget errors.
Blueprint section 9: advertise a security guarantee only when it can be
enforced.
- SandboxMode pure/project/strict with stable names.
- PlatformCapabilities detection reports strict isolation as unavailable in
this build with an explicit reason; worker processes are reported false.
- enforce() refuses strict mode when OS-level isolation is not implemented,
instead of pretending a cooperative in-process check is a sandbox.
- Pure/project modes remain available.