![\"\u6a2a\u6728\uff1a\u57fa\u4e8e\u8bc1\u636e\u7684\u67b6\u6784\u8bc4\u5ba1\u4e0e\u8bbe\u8ba1\u6280\u80fd\u96c6\u5408\"](\"https:\/\/ai.jxgzhc.cn\/wp-content\/uploads\/2026\/09\/hengmu.jpg\")
<\/p>\n<\/p>\n
\u83b7\u5f97\u57fa\u4e8e\u9879\u76ee\u5b9e\u636e\u7684\u67b6\u6784\u98ce\u9669\u6e05\u5355\u3001\u53ef\u843d\u5730\u4fee\u590d\u65b9\u6848\u3001\u5408\u89c4\u76ee\u6807\u67b6\u6784\u8bbe\u8ba1\u4e0e\u786e\u5b9a\u6027\u8d28\u91cf\u68c0\u67e5\u7ed3\u8bba\u3002<\/strong><\/p>\n Website \u00b7 Why Hengmu \u00b7 Hengmu is a local-first, evidence-bound architecture review and target-design Current-state assessment and target design are equal entry paths. Existing The architecture view includes performance efficiency, reliability, security If you are looking for a repository-local architecture review, AI-agent It works at two levels:<\/p>\n one repository, using a project-specific Profile, constraints, critical a portfolio of repositories, looking for duplication, stack sprawl, shared Capability Current-state assessment Target architecture design Decisions and plans Evidence governance Packages and IDE compatibility<\/p>\n Hengmu publishes two packages from the same source. The package choice changes Package Codex package Agent Plugins package The portable package contains Agent Skills and does not include mcp.json or a Use the compatibility matrix for the current evidence Install in your IDE<\/p>\n Download the two ZIP files and their .sha256 files from the same From the download directory, use shasum on macOS or sha256sum on Linux:<\/p>\n shasum -a 256 -c hengmu-<version>.zip.sha256 Keep the extracted directory at a stable absolute path and call it Codex and ChatGPT desktop<\/p>\n Extract the Codex ZIP into a personal plugin directory, for example { Restart ChatGPT desktop, open Plugins , select Hengmu Local , and install Cursor<\/p>\n Extract the Agent Plugins ZIP into ~\/.cursor\/plugins\/local\/hengmu , then mkdir -p ~\/.cursor\/plugins\/local Invoke \/hengmu audit this repository , or ask for the same outcome in natural VS Code and GitHub Copilot<\/p>\n Enable chat.plugins.enabled , run Chat: Install Plugin From Source , and { Open Chat: Open Customizations \u2192 Plugins to confirm installation. VS Code copilot plugin install qingye-lab\/hengmu Start a new Copilot CLI session after installation and invoke \/hengmu … . Kiro<\/p>\n Kiro discovers workspace Skills under .kiro\/skills\/ ; it does not use Hengmu's mkdir -p .kiro\/skills .kiro\/resources Open Agent Steering & Skills in Kiro and confirm all nine Hengmu Skills are Prepare the shared Python runtime<\/p>\n Hengmu's Skills are portable, while its deterministic helpers require Python cd "$HENGMU_ROOT" The last command should print architecture_tool.py 1.0.4 . On Windows Why Hengmu<\/p>\n Most code and architecture reviews stop too early: they produce observations. Typical review failure A model sees a large file or a singleton and declares an architecture problem. A team names a required stack but has no target architecture or explicit trade-offs. A missing capability is mentioned as criticism but never designed. Every project copies the same architecture prompt and slowly diverges. Each repository looks reasonable in isolation while the portfolio duplicates infrastructure. A prose policy says \u201cmust\u201d but automation cannot prove it. Hengmu is intentionally not a generic \u201cbest practices\u201d checklist. A rule is Quick start<\/p>\n 1. Prepare the runtime<\/p>\n Hengmu supports Python 3.11\u20133.13. The runtime is local: it requires no hosted git clone https:\/\/github.com\/qingye-lab\/hengmu.git python3 -m venv .venv On Windows PowerShell, activate the environment with:<\/p>\n .venv\\Scripts\\Activate.ps1<\/p>\n 2. Prepare the repository<\/p>\n HENGMU_ROOT=\/path\/to\/hengmu<\/p>\n python3 "$HENGMU_ROOT\/resources\/scripts\/architecture_tool.py" \\ The command creates a facts-derived repository-local control plane when it is .architecture\/ 3. Choose the outcome in Codex<\/p>\n The examples below use Codex's $hengmu invocation syntax. In another Agent For an existing system, start with current-state assessment:<\/p>\n Use $hengmu to audit this repository. $hengmu is the only Skill name you need to remember. Invoke it by itself to $hengmu For a new system or major redesign, add and approve a Design Brief, then ask for if [ ! -e \/path\/to\/your-project\/.architecture\/architecture-design-brief.yaml ]; then python3 "$HENGMU_ROOT\/resources\/scripts\/architecture_tool.py" \\ The copied template is deliberately draft . Before changing it to approved , $hengmu design an open target architecture from the approved Design Brief The current Brief 1.1 path produces a proposed Decision 1.4 with the complete The audit path can be run directly: the Skill invokes the preparation command The project Profile decides which qualities and specialist reviews matter. project: 4. Validate the result<\/p>\n python3 "$HENGMU_ROOT\/resources\/scripts\/architecture_tool.py" \\ python3 "$HENGMU_ROOT\/resources\/scripts\/architecture_tool.py" \\ python3 "$HENGMU_ROOT\/resources\/scripts\/architecture_tool.py" \\ The gate returns 0 for pass, 1 for policy failure, and 2 for invalid How it works<\/p>\n Hengmu separates model judgment from deterministic trust. It accepts two source The diagram is maintained as Establish facts and intent. Inspect the repository and bind the Profile; Load context. Bind constraints, critical flows, selected Rule Packs, and Assess or design. Existing systems produce candidate findings for Decide. Record the selected target, rejected alternatives, trade-offs, Authorize. A named decision maker accepts, rejects, or supersedes the Plan. Turn an accepted remediation or Greenfield target into ordered Gate. Apply deterministic contract, finding, change, or release policy One method, many projects<\/p>\n A repository should not carry a private copy of the architecture method. profile.yaml \u2014 project type, critical qualities, and required reviews;<\/p>\n constraints.md \u2014 real technical, product, regulatory, and team limits;<\/p>\n critical-flows.md \u2014 business and runtime paths that must not regress;<\/p>\n architecture-design-brief.yaml \u2014 target intent, quality scenarios, reviews\/ \u2014 candidate and verified Reviews, Decisions, Plans, and evidence Portfolio review adds the missing system-of-systems view: which capabilities Workflows<\/p>\n The installable plugin exposes one stable entry point and eight focused What you type $hengmu $hengmu audit\/ai\/mobile\/portfolio \u2026 $hengmu verify \u2026 $hengmu decide \u2026 $hengmu design\/specify\/constrain \u2026 $hengmu plan \u2026 $hengmu gate \u2026 Commands are optional. Natural language such as Focused workflow contracts<\/p>\n Phase Audit Audit Audit Audit Verify Decide Plan Enforce $hengmu may also explain the read-only lifecycle state and the next valid Project-owned quality evidence<\/p>\n Hengmu treats language linters and quality analyzers as optional Evidence Provider discovery distinguishes an applicable marker, project configuration, A Provider pass proves only the captured command, executable, declared project Knowledge curation is deliberately maintainer-only. Its source workflow lives Trust model<\/p>\n Hengmu's trust boundary is simple:<\/p>\n A model may propose. Evidence, authority, provenance, and policy decide what A trusted Review binds the reviewed repository identity and Git state, exact The deterministic runtime provides:<\/p>\n JSON Schemas for project, review, decision, plan, policy, baseline, risk machine-readable core and domain Rule Packs with complete-coverage checks;<\/p>\n sourced Knowledge Packs selected under explicit context budgets;<\/p>\n opt-in Evidence Providers with no-shell execution, safe environment Git evidence resolution, exact hashes, signature verification, SARIF, review Gate stages are cumulative:<\/p>\n Stage contract finding change release Read the assurance model for threats, controls, and Trusted review and evidence commands python3 resources\/scripts\/architecture_tool.py validate-review \\ python3 resources\/scripts\/architecture_tool.py verify-evidence \\ python3 resources\/scripts\/architecture_tool.py verify-review-signature \\ Task-scoped Knowledge selection python3 resources\/scripts\/architecture_tool.py select-knowledge \\ python3 resources\/scripts\/architecture_tool.py validate-knowledge-context \\ Governance modes<\/p>\n Not every project needs the same ceremony.<\/p>\n Mode Advisory Governed Enforced See governance modes for adoption guidance. Documentation<\/p>\n Read this Target architecture Assurance model Governance modes Evaluation guide Knowledge authoring Compatibility Host compatibility Support and feedback 1.0 migration Release verification Roadmap Implementation matrix Dogfood review history Visual assets Accepted architecture decisions live in docs\/decisions . Development<\/p>\n python3 -m venv .venv python3 scripts\/validate_repository.py Build and verify both deterministic plugin archives:<\/p>\n python3 scripts\/package_plugin.py –format codex –output-dir dist The Codex archive is hengmu-<version>.zip . The portable Agent Plugins CI runs the supported Python boundary on Linux, macOS, and Windows. Tagged Feedback and compatibility reports<\/p>\n Real user feedback is welcome, including reports that a Skill works differently For an IDE or host report, include the package format, Hengmu version or commit, Non-goals<\/p>\n Hengmu does not:<\/p>\n autonomously approve architecture decisions, risk, or releases;<\/p>\n turn every detected technology, pattern, or large file into a Finding;<\/p>\n discover unrelated repositories without an explicit portfolio registry;<\/p>\n implement the audited product's remediation by itself;<\/p>\n replace dedicated security, privacy, performance, legal, or compliance prove that a system is secure or correct.<\/p>\n Contributing<\/p>\n Focused issues and pull requests are welcome. Start with Changes to public schemas, CLI behavior, policy, trust boundaries, or persisted When a Review or Selector Runtime binds source commits, preserve those commits Credits and license<\/p>\n Hengmu is a \u9752\u91ce open-source project: The README's editorial illustration system was created with PAAD-derived concepts retain attribution in NOTICE and The software is licensed under the MIT License . The \u9752\u91ce wordmark \u4f5c\u8005\uff1aqingye-lab \u5f00\u53d1\u8005 \/ \u804c\u573a\u4eba \uff5c GitHub Stars 1 \uff5c \u6807\u7b7e\uff1a\u7f16\u7a0b\u5f00\u53d1 \u9879\u76ee\u7ba1\u7406 \u5df2\u8ba4\u8bc1 \u5f00\u6e90\u8bb8\u53ef: MIT<\/p>\n\u8be6\u7ec6\u4ecb\u7ecd<\/h2>\n
English \u00b7 \u7b80\u4f53\u4e2d\u6587<\/p>\n
Install \u00b7
Quick start \u00b7
How it works \u00b7
Workflows \u00b7
Trust model \u00b7
Compatibility \u00b7
Feedback \u00b7
Documentation<\/p>\n
tool for Codex and compatible Agent Plugins. It turns repository facts,
approved design intent, and explicit constraints into evidence-bound
current-state assessments, open or constrained target architectures, traceable
decisions, executable plans, and deterministic policy results.<\/p>\n
systems can move from candidate findings through independent verification to a
remediation decision. New systems and major redesigns can start from an approved
Design Brief and produce a complete target architecture without inventing a
Review or Findings. Required, preferred, and prohibited constraints are
challenged inputs, never proof of feasibility or fitness.<\/p>\n
and privacy boundaries, data and API contracts, observability, testing,
deployment, technical debt, proportionality, technology selection, and
operating reality. Dedicated lenses cover AI-agent context and economics,
Memory, tool authority, privacy, behavior evidence, technology evolution,
mobile systems, and multi-project portfolios.<\/p>\n
architecture audit, target architecture design, architecture decision
governance, remediation planning, or a deterministic quality gate, Hengmu is
built for that workflow. It is not a hosted architecture service or a generic
code linter.<\/p>\n
flows, rules, and review history;<\/p>\n
capability, ownership conflicts, data flows, and hidden coupling.<\/p>\n
What Hengmu does<\/p>\n
Reviews boundaries and engineering qualities, then keeps findings candidate-only until independent verification resolves their evidence.<\/p>\n
Designs open or constrained targets across runtime and deployment units, data ownership, interfaces, trust boundaries, critical flows, operations, and technology choices.<\/p>\n
Compares viable options, records why alternatives lose, and turns an authorized decision into ordered remediation or Greenfield implementation slices.<\/p>\n
Binds facts, Knowledge, provenance, authority, acceptance, and deterministic policy into one auditable chain.<\/p>\n
the discovery manifest and active host UI projection; it does not fork the
underlying Skills or local architecture runtime.<\/p>\n
Manifest and contents
Intended host
What this repository verifies<\/p>\n
.codex-plugin\/plugin.json plus Codex agents\/openai.yaml metadata
Codex
Native manifest, Skill contracts, deterministic archive, and CI\/release packaging<\/p>\n
Host-neutral root plugin.json , standard skills\/ and resources\/ , plus an inert Codex manifest retained for selector provenance
Cursor, VS Code\/GitHub Copilot, and other Agent Plugins 1.0 clients
Standard manifest\/layout projection, deterministic archive, and exclusion of Codex-only agents\/openai.yaml files<\/p>\n
Cursor-specific .cursor-plugin extension. Cursor's plugin documentation
states that spec-conformant Agent Plugins load without changes; this repository
has not yet recorded a Hengmu-specific installed smoke test in Cursor or another
external IDE. For that reason, format compatibility is not presented as a
guarantee of identical commands, permissions, UI, or marketplace behavior.<\/p>\n
boundary. Kiro currently consumes the same Skills through its Agent Skills
locations rather than the root Agent Plugins manifest. Host-specific Hooks,
permissions, rules, steering, and automatic lifecycle behavior are not installed
by either Hengmu package.<\/p>\n
GitHub release . Use
hengmu-<version>.zip for Codex and
hengmu-<version>-agent-plugins.zip for the other hosts below. Verify the
download before extracting it:<\/p>\n
shasum -a 256 -c hengmu-<version>-agent-plugins.zip.sha256<\/p>\n
HENGMU_ROOT . Each host needs the complete directory, not only
skills\/hengmu , because the router, focused Skills, schemas, Knowledge, and
deterministic CLI use relative paths across skills\/ and resources\/ .<\/p>\n
~\/.codex\/plugins\/hengmu . Then add this entry to the plugins array in
~\/.agents\/plugins\/marketplace.json (merge it with any existing marketplace
instead of replacing the file):<\/p>\n
"name": "hengmu-local",
"interface": { "displayName": "Hengmu Local" },
"plugins": [
{
"name": "hengmu",
"source": {
"source": "local",
"path": ".\/.codex\/plugins\/hengmu"
},
"policy": {
"installation": "AVAILABLE",
"authentication": "ON_INSTALL"
},
"category": "Developer Tools"
}
]
}<\/p>\n
Hengmu. In Codex CLI, run \/plugins , install Hengmu from the same marketplace,
and start a new session. Invoke the router with $hengmu audit this repository
or describe the outcome naturally. In ChatGPT Work mode, select it with
@hengmu . The Codex IDE extension can use standalone Skills but does not
currently provide the plugin browser; install the full Hengmu plugin through
ChatGPT desktop or Codex CLI. See the
official OpenAI plugin installation documentation .<\/p>\n
restart Cursor or run Developer: Reload Window . Open Customize and
confirm that the Hengmu Skills are enabled. You can also symlink a checked-out
Hengmu repository while developing:<\/p>\n
ln -s \/absolute\/path\/to\/hengmu ~\/.cursor\/plugins\/local\/hengmu<\/p>\n
language. Cursor loads the portable Skills, but this package does not install
Cursor-specific rules, agents, commands, Hooks, or variables. See
Cursor's Agent Plugins installation guide .<\/p>\n
enter https:\/\/github.com\/qingye-lab\/hengmu . For a pinned local build, extract
the Agent Plugins ZIP and register its absolute directory in VS Code settings:<\/p>\n
"chat.pluginLocations": {
"\/absolute\/path\/to\/hengmu": true
}
}<\/p>\n
namespaces plugin-provided Skills with the plugin name, so invoke
\/hengmu:hengmu audit this repository in Copilot Chat or use natural language.
The same package can be installed directly in GitHub Copilot CLI:<\/p>\n
copilot plugin list
copilot<\/p>\n
See the official VS Code Agent Plugins
and GitHub Copilot CLI plugin
documentation.<\/p>\n
root plugin.json for this installation path. Extract the Agent Plugins ZIP to
a stable HENGMU_ROOT , then project both the Skills and their shared resources
into the repository. Run these commands only when .kiro\/resources is unused,
or merge the directories deliberately:<\/p>\n
cp -R "$HENGMU_ROOT\/skills\/." .kiro\/skills\/
cp -R "$HENGMU_ROOT\/resources\/." .kiro\/resources\/<\/p>\n
visible. Invoke \/hengmu audit this repository or use natural language. Do not
import only skills\/hengmu : the router delegates to eight sibling Skills and
those Skills require the shared runtime. See the official
Kiro Agent Skills guide .<\/p>\n
3.11\u20133.13, PyYAML, and jsonschema. Install the locked dependencies into the
python3 environment exposed to the IDE agent, or launch the IDE from this
activated environment:<\/p>\n
python3 -m venv .venv
source .venv\/bin\/activate
python3 -m pip install –require-hashes -r requirements-runtime.lock
python3 resources\/scripts\/architecture_tool.py –version<\/p>\n
PowerShell, activate with .venv\\Scripts\\Activate.ps1 . Installation does not
grant permissions or enable Hooks; review each host's agent permissions before
allowing repository writes or shell execution.<\/p>\n
Hengmu is designed around a longer, evidence-bound engineering decision chain.<\/p>\n
Hengmu's response<\/p>\n
Candidate findings must survive independent verification and evidence resolution before they become trusted.<\/p>\n
Hengmu challenges required, preferred, and prohibited constraints, compares compliant variants, and records a complete target architecture instead of treating technology names as proof.<\/p>\n
Confirmed gaps flow into solution comparison, remediation slices, rollback, tests, and acceptance criteria.<\/p>\n
One global method reads a repository-local Profile and real constraints.<\/p>\n
Portfolio review models shared capabilities, dependencies, data flow, ownership, and coupling.<\/p>\n
JSON Schemas, hashes, Git evidence, role policy, fingerprints, signatures, and stable exit codes make enforcement reproducible.<\/p>\n
useful only when it protects a declared quality or critical flow, and a
recommendation is useful only when the project can understand its cost,
dependencies, migration order, and stopping conditions.<\/p>\n
service, telemetry, credentials, network access, or MCP server.<\/p>\n
cd hengmu<\/p>\n
source .venv\/bin\/activate
python3 -m pip install –require-hashes -r requirements-runtime.lock
python3 scripts\/validate_repository.py<\/p>\n
prepare-project-audit –repo \/path\/to\/your-project<\/p>\n
missing, or validates and reuses the existing one without overwriting it:<\/p>\n
\u251c\u2500\u2500 profile.yaml
\u251c\u2500\u2500 repository-facts.yaml
\u251c\u2500\u2500 constraints.md
\u251c\u2500\u2500 critical-flows.md
\u251c\u2500\u2500 gate-policy.yaml
\u251c\u2500\u2500 baseline.yaml
\u251c\u2500\u2500 risk-acceptances.yaml
\u251c\u2500\u2500 evidence-providers.yaml
\u251c\u2500\u2500 evidence\/
\u251c\u2500\u2500 rules\/
\u251c\u2500\u2500 runs\/
\u2514\u2500\u2500 reviews\/<\/p>\n
Plugins host, install the portable archive and invoke the same Skill through
that host's documented UI or command syntax; $hengmu is not a portable
invocation contract.<\/p>\n
Treat missing capabilities as findings, but verify evidence before
recommending a structural change.<\/p>\n
see the complete capability menu, or describe the outcome in natural language:<\/p>\n
$hengmu verify the latest candidate findings
$hengmu compare the queue and durable-workflow options<\/p>\n
an open or constrained target:<\/p>\n
cp "$HENGMU_ROOT\/resources\/templates\/architecture-design-brief.yaml" \\
\/path\/to\/your-project\/.architecture\/architecture-design-brief.yaml
fi<\/p>\n
validate-design-brief \\
\/path\/to\/your-project\/.architecture\/architecture-design-brief.yaml \\
–project \/path\/to\/your-project<\/p>\n
add brief.approval with an authorized decision-maker identity and at least one
repository-relative approval-evidence path and SHA-256, plus one detached SSH
signature per approver. The signatures must verify against the project's
artifact_signatures policy. Validation never treats the template authors or
a status string as approval.<\/p>\n
$hengmu constrain the target to FastAPI, PostgreSQL, and one production deployment; challenge each constraint and record rejected alternatives<\/p>\n
target architecture. Hengmu does not approve the Brief or Decision and does not
implement application code. After an authorized decision maker accepts the
Decision, $hengmu plan \u2026 can produce a Greenfield Plan 1.3.<\/p>\n
and initializes .architecture\/ automatically. Use an explicitly read-only
request only when you want a one-off Advisory assessment with no repository
artifacts.<\/p>\n
The global Skill provides the method; the repository provides the truth.<\/p>\n
name: example-service
type:
– ai-agent-platform
critical_qualities:
– traceability
– recoverability
– privacy
required_reviews:
– project-architecture
– ai-agent-architecture<\/p>\n
validate-project \/path\/to\/your-project<\/p>\n
gate –project \/path\/to\/your-project –stage change<\/p>\n
gate –project \/path\/to\/your-project \\
–decision .architecture\/reviews\/<greenfield-decision.yaml> –stage change<\/p>\n
input or configuration.<\/p>\n
paths, and neither a candidate audit nor a constraint assertion is policy or
proof.<\/p>\n
Mermaid source and an
editable Excalidraw scene .<\/p>\n
for target design, add an approved Brief with measurable scenarios and
boundaries.<\/p>\n
task-scoped Knowledge without turning detected technology or owner assertions
into proof.<\/p>\n
independent verification. Greenfield work uses the approved Brief to compare
open or constraint-compliant targets without manufacturing Findings.<\/p>\n
complete architecture model, source bindings, and proposed status.<\/p>\n
Decision; Hengmu's router and Advisor cannot perform this transition.<\/p>\n
implementation slices, protections, rollback, stop conditions, and
acceptance evidence.<\/p>\n
to provenance-bound artifacts.<\/p>\n
Instead, it carries only the context that makes its decisions different:<\/p>\n
boundaries, and typed constraints;<\/p>\n
history.<\/p>\n
should be shared, which boundaries must remain independent, where data moves,
and where one repository can unexpectedly affect another.<\/p>\n
workflow Skills. Use $hengmu for normal work; direct focused invocation
remains available for automation and compatibility.<\/p>\n
Required input
Output<\/p>\n
Declared repository context, when present
Menu and read-only next-step guidance<\/p>\n
Repository or portfolio facts and Profile
Candidate Review in the selected scope<\/p>\n
Candidate Review and resolvable evidence
Provenance-bound verified Review<\/p>\n
Verified Review or approved Design Brief
Proposed Architecture Decision<\/p>\n
Approved Brief 1.1, facts, constraints, and selected Knowledge
Proposed Decision 1.4 with an open or constrained target architecture<\/p>\n
Accepted remediation or Greenfield Decision
Ordered remediation Plan 1.2 or Greenfield Plan 1.3<\/p>\n
Schema-valid, provenance-bound artifacts
Deterministic policy result and stable exit code<\/p>\n
$hengmu help me compare these two technical approaches routes to the same
focused workflow.<\/p>\n
Skill
Responsibility<\/p>\n
project-architecture-audit
Boundaries, data ownership, contracts, reliability, security, operations, tests, deployment, debt, and proportionality in one repository.<\/p>\n
ai-agent-architecture-audit
Models, context necessity\/assembly\/compression\/cache ordering, Memory, retrieval, tools, injection, privacy, approval, recovery, version-bound behavior evidence, cost, latency, and evolution boundaries.<\/p>\n
mobile-architecture-audit
Local state, sync, migrations, background work, notifications, privacy, caching, and lifecycle behavior.<\/p>\n
portfolio-architecture-audit
Duplication, stack sprawl, shared capabilities, dependencies, data flow, ownership, and hidden coupling across projects.<\/p>\n
architecture-finding-verifier
Challenge candidates, resolve evidence, assign V0\u2013V5 verification, and produce a provenance-bound trusted Review.<\/p>\n
architecture-solution-advisor
Compare or specify open\/constrained targets; assess required conflicts, preferred trade-offs, prohibited eliminations, and target units, flows, boundaries, operations, and Knowledge.<\/p>\n
architecture-remediation-planner
Convert an accepted remediation or Greenfield target into ordered implementation slices, migration controls where applicable, protections, stop conditions, rollback, and acceptance criteria without inventing Findings.<\/p>\n
architecture-quality-gate
Apply deterministic contract, finding, change, and release policy to trusted artifacts.<\/p>\n
focused workflow from existing artifacts. It does not verify findings, accept a
decision, mutate policy, or run a Gate on the user's behalf.<\/p>\n
Providers, not as architecture truth. The bundled catalog includes representative
providers for Python, JavaScript\/TypeScript, Rust, Go, Swift, and Kotlin\/JVM in
addition to architecture, contract, test, runtime, security, and supply-chain
providers.<\/p>\n
enablement, executable availability, and readiness. A missing executable remains
an explicit unassessed evidence surface. Hengmu never downloads, installs, enables,
or adds a dependency implicitly. If installation would materially improve the
review, it first names the exact tool, scope, version strategy, command, affected
files, and consequence, then asks for user authorization.<\/p>\n
dependency closure, isolated cache mode, configuration,
commit, and output bytes. It becomes architecture evidence only after it is bound
to an applicable invariant and independently reviewed.<\/p>\n
under maintainer\/skills\/architecture-knowledge-curator\/ and does not expand
the public end-user Skill surface.<\/p>\n
can become trusted or blocking.<\/p>\n
scope, Profile, repository facts, selected Knowledge, Rule Packs, candidate
review, verifier authority, semantic Finding fingerprints, critical-flow
coverage, and resolvable evidence.<\/p>\n
acceptance, Knowledge, provider, benchmark, and governance artifacts;<\/p>\n
allowlists, timeouts, structured-output validation, and tamper-evident run
records;<\/p>\n
diffing, artifact migration, benchmark scoring, and layered gates.<\/p>\n
Proves<\/p>\n
Schemas, provenance, identity, hashes, roles, and coverage are valid.<\/p>\n
Severity, confidence, verification, status, baseline, waiver, and risk acceptance satisfy policy.<\/p>\n
Review freshness, changed contracts, required decisions, migration compatibility, signatures, and evidence resolution are acceptable.<\/p>\n
Required evidence, decision authority, and complete remediation acceptance are present.<\/p>\n
residual risk. A passing gate proves policy evaluation of supplied artifacts;
it does not prove that the audited product is correct, secure, compliant, or
well designed.<\/p>\n
python3 resources\/scripts\/architecture_tool.py review-bindings \\
–project \/path\/to\/project \\
–candidate .architecture\/reviews\/example-candidates.yaml<\/p>\n
\/path\/to\/verified.yaml –project \/path\/to\/project<\/p>\n
–repo \/path\/to\/project –review \/path\/to\/verified.yaml<\/p>\n
–project \/path\/to\/project –review \/path\/to\/verified.yaml<\/p>\n
python3 resources\/scripts\/architecture_tool.py inspect-repository \\
–repo \/path\/to\/project \\
–output \/path\/to\/project\/.architecture\/repository-facts.yaml<\/p>\n
–facts \/path\/to\/project\/.architecture\/repository-facts.yaml \\
–profile \/path\/to\/project\/.architecture\/profile.yaml \\
–task "Current architecture audit" \\
–skill project-architecture-audit \\
–output \/path\/to\/project\/.architecture\/knowledge-selection.yaml \\
–context-output \/path\/to\/project\/.architecture\/knowledge-context.yaml<\/p>\n
\/path\/to\/project\/.architecture\/knowledge-context.yaml \\
–selection \/path\/to\/project\/.architecture\/knowledge-selection.yaml \\
–facts \/path\/to\/project\/.architecture\/repository-facts.yaml \\
–profile \/path\/to\/project\/.architecture\/profile.yaml<\/p>\n
Use when
Behavior<\/p>\n
The project needs structured architecture help without a blocking gate.
Skills produce evidence-backed artifacts; maintainers retain full judgment.<\/p>\n
Important changes need trusted review, explicit decisions, and change policy.
Provenance, authority, freshness, and Finding policy are enforced.<\/p>\n
Releases require deterministic architecture evidence and completed remediation.
Change and release gates become required delivery controls.<\/p>\n
product_mode is a declared operating tier, not a bypass: an explicitly
invoked gate always evaluates its policy.<\/p>\n
When you need<\/p>\n
Facts, Knowledge, workflow, trust boundaries, and runtime components.<\/p>\n
Threats, guarantees, non-guarantees, and residual risk.<\/p>\n
Advisory, Governed, and Enforced adoption.<\/p>\n
Behavior benchmarks, ablation, scoring, and interpretation limits.<\/p>\n
Source quality, freshness, frontmatter, and curation rules.<\/p>\n
Supported Python, schemas, artifacts, and version boundaries.<\/p>\n
Cross-IDE outcome equivalence, package paths, and host-specific boundaries.<\/p>\n
Reproducible defects, documentation gaps, and host compatibility reports.<\/p>\n
Open\/constrained Brief\/Decision\/Plan artifacts, coexistence, and rollback.<\/p>\n
Deterministic ZIPs, checksums, SBOMs, and attestations.<\/p>\n
Canonical forward plan, evidence milestones, and conditional triggers.<\/p>\n
How review recommendations map to executable capability and evidence.<\/p>\n
How Hengmu governs its own repository.<\/p>\n
Bilingual icon, banner, editorial character, and diagram source conventions.<\/p>\n
The repository's implemented target state is tracked in the
target architecture implementation matrix .<\/p>\n
source .venv\/bin\/activate
python3 -m pip install –require-hashes -r requirements-dev.lock<\/p>\n
python3 resources\/scripts\/architecture_tool.py validate-project .
python3 resources\/scripts\/architecture_tool.py validate-history-anchors .
python3 resources\/scripts\/validate_knowledge.py
python3 -m pytest
python3 resources\/scripts\/architecture_tool.py gate –project . –stage change
python3 -m ruff check .
python3 -m ruff format –check .
python3 scripts\/audit_licenses.py<\/p>\n
python3 scripts\/package_plugin.py –format agent-plugins –output-dir dist
python3 scripts\/smoke_test_package.py \\
–format codex \\
–archive dist\/hengmu-<version>.zip
python3 scripts\/smoke_test_package.py \\
–format agent-plugins \\
–archive dist\/hengmu-<version>-agent-plugins.zip
python3 scripts\/verify_checksum.py dist\/*.zip.sha256
python3 scripts\/generate_sbom.py \\
–archive dist\/*.zip \\
–output-dir dist<\/p>\n
archive is hengmu-<version>-agent-plugins.zip and uses the host-neutral root
plugin.json . It also retains the complete .codex-plugin\/plugin.json ,
including its Codex-only interface fields, as inert provenance data required
by the shared Knowledge selector. Codex-specific skills\/*\/agents\/openai.yaml
files are excluded from the portable archive.<\/p>\n
releases publish both deterministic ZIPs, their SHA-256 checksums and SPDX
SBOMs, and GitHub provenance\/SBOM attestations.<\/p>\n
across Codex, Cursor, or another Agent Plugins host. Open a
bug report
for reproducible behavior or a
feature request
for a focused improvement. Read SUPPORT.md first.<\/p>\n
client name and version, operating system, installation path, Skill or prompt,
expected result, observed result, and sanitized logs. Do not include credentials,
private repository content, or personal data. We treat client support as
evidence-backed and time-bound: a report can improve the compatibility record,
but an untested client is not presented as verified support.<\/p>\n
assessment;<\/p>\n
CONTRIBUTING.md , then read
GOVERNANCE.md , SECURITY.md , and
SUPPORT.md .<\/p>\n
artifacts require compatibility analysis, tests, migration notes, and an
updated architecture decision when authority changes.<\/p>\n
with a Merge Commit. Squash or rebase merging can invalidate source ancestry
and is rejected by validate-history-anchors .<\/p>\n
\u7406\u6027\u7ed3\u6784\u4e2d\u7684\u6301\u7eed\u8fdb\u5316\uff0c\u5728\u4e0d\u786e\u5b9a\u4e2d\uff0c\u6301\u7eed\u6784\u5efa\u3002<\/p>\n
Ian Xiaohei Illustrations
and recast with an original \u9752\u91ce builder character derived from the public
\u9752\u91ce avatar and brand palette. The technical flow is
available as Mermaid, Excalidraw, SVG, and PNG so documentation remains
editable.<\/p>\n
third_party\/PAAD-MIT.txt .<\/p>\n
identifies the originating project and is not a grant to imply endorsement.<\/p>\n\u8bd5\u8bd5\u8fd9\u6837\u505a<\/h2>\n
\n
\n