横木：基于证据的架构评审与设计技能集合

Website · English · 简体中文

Why Hengmu · Install · Quick start · How it works · Workflows · Trust model · Compatibility · Feedback · Documentation

Hengmu is a local-first, evidence-bound architecture review and target-design 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.

Current-state assessment and target design are equal entry paths. Existing 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.

The architecture view includes performance efficiency, reliability, security 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.

If you are looking for a repository-local architecture review, AI-agent 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.

It works at two levels:

• one repository, using a project-specific Profile, constraints, critical flows, rules, and review history;
• a portfolio of repositories, looking for duplication, stack sprawl, shared capability, ownership conflicts, data flows, and hidden coupling.

Packages and IDE compatibility

Hengmu publishes two packages from the same source. The package choice changes the discovery manifest and active host UI projection; it does not fork the underlying Skills or local architecture runtime.

The portable package contains Agent Skills and does not include mcp.json or a 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.

Use the compatibility matrix for the current evidence 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.

Install in your IDE

Download the two ZIP files and their .sha256 files from the same 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:

From the download directory, use shasum on macOS or sha256sum on Linux:

shasum -a 256 -c hengmu-<version>.zip.sha256
shasum -a 256 -c hengmu-<version>-agent-plugins.zip.sha256

Keep the extracted directory at a stable absolute path and call it 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/.

Codex and ChatGPT desktop

Extract the Codex ZIP into a personal plugin directory, for example ~/.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):

{
  "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"
    }
  ]
}

Restart ChatGPT desktop, open Plugins, select Hengmu Local, and install 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.

Cursor

Extract the Agent Plugins ZIP into ~/.cursor/plugins/local/hengmu, then 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:

mkdir -p ~/.cursor/plugins/local
ln -s /absolute/path/to/hengmu ~/.cursor/plugins/local/hengmu

Invoke /hengmu audit this repository, or ask for the same outcome in natural 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.

VS Code and GitHub Copilot

Enable chat.plugins.enabled, run Chat: Install Plugin From Source, and 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:

{
  "chat.pluginLocations": {
    "/absolute/path/to/hengmu": true
  }
}

Open Chat: Open Customizations → Plugins to confirm installation. VS Code 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:

copilot plugin install qingye-lab/hengmu
copilot plugin list
copilot

Start a new Copilot CLI session after installation and invoke /hengmu .... See the official VS Code Agent Plugins and GitHub Copilot CLI plugin documentation.

Kiro

Kiro discovers workspace Skills under .kiro/skills/; it does not use Hengmu's 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:

mkdir -p .kiro/skills .kiro/resources
cp -R "$HENGMU_ROOT/skills/." .kiro/skills/
cp -R "$HENGMU_ROOT/resources/." .kiro/resources/

Open Agent Steering & Skills in Kiro and confirm all nine Hengmu Skills are 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.

Prepare the shared Python runtime

Hengmu's Skills are portable, while its deterministic helpers require Python 3.11–3.13, PyYAML, and jsonschema. Install the locked dependencies into the python3 environment exposed to the IDE agent, or launch the IDE from this activated environment:

cd "$HENGMU_ROOT"
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

The last command should print architecture_tool.py 1.0.4. On Windows 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.

Why Hengmu

Most code and architecture reviews stop too early: they produce observations. Hengmu is designed around a longer, evidence-bound engineering decision chain.

Hengmu is intentionally not a generic “best practices” checklist. A rule is 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.

Quick start

1. Prepare the runtime

Hengmu supports Python 3.11–3.13. The runtime is local: it requires no hosted service, telemetry, credentials, network access, or MCP server.

git clone https://github.com/qingye-lab/hengmu.git
cd hengmu

python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install --require-hashes -r requirements-runtime.lock
python3 scripts/validate_repository.py

On Windows PowerShell, activate the environment with:

.venv\Scripts\Activate.ps1

2. Prepare the repository

HENGMU_ROOT=/path/to/hengmu

python3 "$HENGMU_ROOT/resources/scripts/architecture_tool.py" \
  prepare-project-audit --repo /path/to/your-project

The command creates a facts-derived repository-local control plane when it is missing, or validates and reuses the existing one without overwriting it:

.architecture/
├── profile.yaml
├── repository-facts.yaml
├── constraints.md
├── critical-flows.md
├── gate-policy.yaml
├── baseline.yaml
├── risk-acceptances.yaml
├── evidence-providers.yaml
├── evidence/
├── rules/
├── runs/
└── reviews/

3. Choose the outcome in Codex

The examples below use Codex's $hengmu invocation syntax. In another Agent 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.

For an existing system, start with current-state assessment:

Use $hengmu to audit this repository.
Treat missing capabilities as findings, but verify evidence before
recommending a structural change.

$hengmu is the only Skill name you need to remember. Invoke it by itself to see the complete capability menu, or describe the outcome in natural language:

$hengmu
$hengmu verify the latest candidate findings
$hengmu compare the queue and durable-workflow options

For a new system or major redesign, add and approve a Design Brief, then ask for an open or constrained target:

if [ ! -e /path/to/your-project/.architecture/architecture-design-brief.yaml ]; then
  cp "$HENGMU_ROOT/resources/templates/architecture-design-brief.yaml" \
    /path/to/your-project/.architecture/architecture-design-brief.yaml
fi

python3 "$HENGMU_ROOT/resources/scripts/architecture_tool.py" \
  validate-design-brief \
  /path/to/your-project/.architecture/architecture-design-brief.yaml \
  --project /path/to/your-project

The copied template is deliberately draft. Before changing it to approved, 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.

$hengmu design an open target architecture from the approved Design Brief
$hengmu constrain the target to FastAPI, PostgreSQL, and one production deployment; challenge each constraint and record rejected alternatives

The current Brief 1.1 path produces a proposed Decision 1.4 with the complete 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 … can produce a Greenfield Plan 1.3.

The audit path can be run directly: the Skill invokes the preparation command and initializes .architecture/ automatically. Use an explicitly read-only request only when you want a one-off Advisory assessment with no repository artifacts.

The project Profile decides which qualities and specialist reviews matter. The global Skill provides the method; the repository provides the truth.

project:
  name: example-service
  type:
    - ai-agent-platform
  critical_qualities:
    - traceability
    - recoverability
    - privacy
  required_reviews:
    - project-architecture
    - ai-agent-architecture

4. Validate the result

python3 "$HENGMU_ROOT/resources/scripts/architecture_tool.py" \
  validate-project /path/to/your-project

python3 "$HENGMU_ROOT/resources/scripts/architecture_tool.py" \
  gate --project /path/to/your-project --stage change

python3 "$HENGMU_ROOT/resources/scripts/architecture_tool.py" \
  gate --project /path/to/your-project \
  --decision .architecture/reviews/<greenfield-decision.yaml> --stage change

The gate returns 0 for pass, 1 for policy failure, and 2 for invalid input or configuration.

How it works

Hengmu separates model judgment from deterministic trust. It accepts two source paths, and neither a candidate audit nor a constraint assertion is policy or proof.

The diagram is maintained as Mermaid source and an editable Excalidraw scene.

1. Establish facts and intent. Inspect the repository and bind the Profile; for target design, add an approved Brief with measurable scenarios and boundaries.
2. Load context. Bind constraints, critical flows, selected Rule Packs, and task-scoped Knowledge without turning detected technology or owner assertions into proof.
3. Assess or design. Existing systems produce candidate findings for independent verification. Greenfield work uses the approved Brief to compare open or constraint-compliant targets without manufacturing Findings.
4. Decide. Record the selected target, rejected alternatives, trade-offs, complete architecture model, source bindings, and proposed status.
5. Authorize. A named decision maker accepts, rejects, or supersedes the Decision; Hengmu's router and Advisor cannot perform this transition.
6. Plan. Turn an accepted remediation or Greenfield target into ordered implementation slices, protections, rollback, stop conditions, and acceptance evidence.
7. Gate. Apply deterministic contract, finding, change, or release policy to provenance-bound artifacts.

One method, many projects

A repository should not carry a private copy of the architecture method. Instead, it carries only the context that makes its decisions different:

• profile.yaml — project type, critical qualities, and required reviews;
• constraints.md — real technical, product, regulatory, and team limits;
• critical-flows.md — business and runtime paths that must not regress;
• architecture-design-brief.yaml — target intent, quality scenarios, boundaries, and typed constraints;
• reviews/ — candidate and verified Reviews, Decisions, Plans, and evidence history.

Portfolio review adds the missing system-of-systems view: which capabilities should be shared, which boundaries must remain independent, where data moves, and where one repository can unexpectedly affect another.

Workflows

The installable plugin exposes one stable entry point and eight focused workflow Skills. Use $hengmu for normal work; direct focused invocation remains available for automation and compatibility.

Commands are optional. Natural language such as $hengmu help me compare these two technical approaches routes to the same focused workflow.

Focused workflow contracts

$hengmu may also explain the read-only lifecycle state and the next valid focused workflow from existing artifacts. It does not verify findings, accept a decision, mutate policy, or run a Gate on the user's behalf.

Project-owned quality evidence

Hengmu treats language linters and quality analyzers as optional Evidence 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.

Provider discovery distinguishes an applicable marker, project configuration, 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.

A Provider pass proves only the captured command, executable, declared project 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.

Knowledge curation is deliberately maintainer-only. Its source workflow lives under maintainer/skills/architecture-knowledge-curator/ and does not expand the public end-user Skill surface.

Trust model

Hengmu's trust boundary is simple:

A model may propose. Evidence, authority, provenance, and policy decide what can become trusted or blocking.

A trusted Review binds the reviewed repository identity and Git state, exact scope, Profile, repository facts, selected Knowledge, Rule Packs, candidate review, verifier authority, semantic Finding fingerprints, critical-flow coverage, and resolvable evidence.

The deterministic runtime provides:

• JSON Schemas for project, review, decision, plan, policy, baseline, risk acceptance, Knowledge, provider, benchmark, and governance artifacts;
• machine-readable core and domain Rule Packs with complete-coverage checks;
• sourced Knowledge Packs selected under explicit context budgets;
• opt-in Evidence Providers with no-shell execution, safe environment allowlists, timeouts, structured-output validation, and tamper-evident run records;
• Git evidence resolution, exact hashes, signature verification, SARIF, review diffing, artifact migration, benchmark scoring, and layered gates.

Gate stages are cumulative:

Read the assurance model for threats, controls, and 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.

python3 resources/scripts/architecture_tool.py review-bindings \
  --project /path/to/project \
  --candidate .architecture/reviews/example-candidates.yaml

python3 resources/scripts/architecture_tool.py validate-review \
  /path/to/verified.yaml --project /path/to/project

python3 resources/scripts/architecture_tool.py verify-evidence \
  --repo /path/to/project --review /path/to/verified.yaml

python3 resources/scripts/architecture_tool.py verify-review-signature \
  --project /path/to/project --review /path/to/verified.yaml

python3 resources/scripts/architecture_tool.py inspect-repository \
  --repo /path/to/project \
  --output /path/to/project/.architecture/repository-facts.yaml

python3 resources/scripts/architecture_tool.py select-knowledge \
  --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

python3 resources/scripts/architecture_tool.py validate-knowledge-context \
  /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

Governance modes

Not every project needs the same ceremony.

See governance modes for adoption guidance. product_mode is a declared operating tier, not a bypass: an explicitly invoked gate always evaluates its policy.

Documentation

Accepted architecture decisions live in docs/decisions. The repository's implemented target state is tracked in the target architecture implementation matrix.

Development

python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install --require-hashes -r requirements-dev.lock

python3 scripts/validate_repository.py
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

Build and verify both deterministic plugin archives:

python3 scripts/package_plugin.py --format codex --output-dir dist
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

The Codex archive is hengmu-<version>.zip. The portable Agent Plugins 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.

CI runs the supported Python boundary on Linux, macOS, and Windows. Tagged releases publish both deterministic ZIPs, their SHA-256 checksums and SPDX SBOMs, and GitHub provenance/SBOM attestations.

Feedback and compatibility reports

Real user feedback is welcome, including reports that a Skill works differently 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.

For an IDE or host report, include the package format, Hengmu version or commit, 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.

Non-goals

Hengmu does not:

• autonomously approve architecture decisions, risk, or releases;
• turn every detected technology, pattern, or large file into a Finding;
• discover unrelated repositories without an explicit portfolio registry;
• implement the audited product's remediation by itself;
• replace dedicated security, privacy, performance, legal, or compliance assessment;
• prove that a system is secure or correct.

Contributing

Focused issues and pull requests are welcome. Start with CONTRIBUTING.md, then read GOVERNANCE.md, SECURITY.md, and SUPPORT.md.

Changes to public schemas, CLI behavior, policy, trust boundaries, or persisted artifacts require compatibility analysis, tests, migration notes, and an updated architecture decision when authority changes.

When a Review or Selector Runtime binds source commits, preserve those commits with a Merge Commit. Squash or rebase merging can invalidate source ancestry and is rejected by validate-history-anchors.

Credits and license

Hengmu is a 青野 open-source project: 理性结构中的持续进化，在不确定中，持续构建。

The README's editorial illustration system was created with Ian Xiaohei Illustrations and recast with an original 青野 builder character derived from the public 青野 avatar and brand palette. The technical flow is available as Mermaid, Excalidraw, SVG, and PNG so documentation remains editable.

PAAD-derived concepts retain attribution in NOTICE and third_party/PAAD-MIT.txt.

The software is licensed under the MIT License. The 青野 wordmark identifies the originating project and is not a grant to imply endorsement.