Skip to main content
Glama
ValentinoWang

Control-PromotionMCP

by ValentinoWang
OverviewSchemaRelated ServersScoreDiscussions
Python
Local

control-promotion-mcp

Portable governance control-promotion CLI and MCP server.

This repository turns recurring engineering experience into a structured control lifecycle:

raw evidence
  -> reusable observation
  -> docs / Skill / scoped AGENTS
  -> static guard / QA harness
  -> type, schema, or contract prevention
  -> retired guard

The package has three layers:

control_promotion        # deterministic core and CLI
control_promotion_mcp    # read-only MCP server
.control-promotion.yaml  # project adapter

Why MCP

MCP lets a server expose callable tools, readable resources, and reusable prompts over JSON-RPC. The 2025-06-18 specification defines stdio and Streamable HTTP transports; stdio messages are newline-delimited JSON-RPC, and Streamable HTTP uses POST requests to a single MCP endpoint. This server follows that model for a read-only governance control plane.

References:

Related MCP server: scan-your-ai-toolkit

Install

pip install control-promotion-mcp

For local development:

python -m venv .venv
. .venv/bin/activate
pip install -e .

CLI

control-promotion inspect --project-root .
control-promotion validate-adapter .control-promotion.yaml
control-promotion validate-catalog references/smell-catalog.yaml
control-promotion validate-guard-spec examples/guard-specs/good-creation-table-contract.yaml
control-promotion run-guard-fixtures \
  examples/guard-specs/good-creation-table-contract.yaml \
  --command "python examples/mock-guards/creation_table_guard.py" \
  > /tmp/creation-table-proof.yaml
control-promotion validate-fixture-execution \
  /tmp/creation-table-proof.yaml \
  --guard-spec examples/guard-specs/good-creation-table-contract.yaml
control-promotion review-guard \
  examples/guard-specs/good-creation-table-contract.yaml \
  --execute "python examples/mock-guards/creation_table_guard.py"
control-promotion route \
  --failure-class frontend_semantic_metric_without_source \
  --detectability static \
  --recurrence repeated \
  --harm high
control-promotion review --candidate candidate.yaml --format markdown

Candidate file:

candidate_text: |
  frontend-metric-source-guard prevents hard-coded semantic KPI literals.
evidence:
  paths:
    - scripts/quality/check_frontend_metric_source_guard.py
  commands:
    - bash scripts/quality/run_frontend_metric_source_guard.sh --mode ci
context:
  recurrence: repeated
  harm: high

MCP stdio

{
  "mcpServers": {
    "control-promotion": {
      "command": "control-promotion-mcp",
      "args": [
        "--project-root",
        ".",
        "--adapter",
        ".control-promotion.yaml",
        "--mode",
        "stdio"
      ]
    }
  }
}

MCP HTTP

control-promotion-mcp \
  --project-root . \
  --adapter .control-promotion.yaml \
  --mode http \
  --host 127.0.0.1 \
  --port 8765

The V1 HTTP server exposes POST /mcp and returns one JSON response. It binds to localhost by default and rejects non-local Origin headers. It intentionally does not expose write tools.

Exposed MCP Tools

  • inspect_project

  • evaluate_control_candidate

  • route_control_destination

  • validate_smell_catalog

  • validate_project_adapter

  • validate_guard_spec

  • review_guard_quality

  • validate_fixture_execution

  • render_smell_gate_report

  • check_ssot_links

evaluate_control_candidate also returns an abstraction_review block. This block flags guard-quality issues that routing alone cannot catch, including incident-string denylist overfit, fixed current-file allowlists, missing positive/negative fixtures, missing canonical contracts, missing scoped surface discovery, and missing exception policies.

Example high-risk result:

decision: refactor_before_promote
control_level: L5_static_quality_guard
abstraction_review:
  specificity_risk: high
  overfit_signals:
    - literal_incident_phrase_denylist
    - fixed_current_file_allowlist
    - missing_targeted_fixtures
    - missing_canonical_contract
  missing_abstraction:
    - canonical_contract
    - deprecated_alias_set
    - scoped_surface_discovery
    - exception_policy
  recommendation: refactor_before_promote

GuardSpec Promotion Gate

Starting in v0.3.0, static quality guards should provide a GuardSpec before promotion to L5. GuardSpec is the evidence layer that proves the guard protects a reusable invariant instead of one observed incident.

Minimum shape:

id: creation-table-name-contract
protected_invariant: User-facing creation output targets must use the canonical table name.
failure_class: deprecated_creation_table_alias
contract:
  owner: CREATION_OUTPUT_TABLE_CONTRACT
  canonical: 03_创作任务总表
  source: MEDIA_OS_CREATION_TASKS_URL
scan_scope:
  include:
    - docs/说明书/**
    - openclaw-agents/media/**
  exclude:
    - tmp/**
    - agents-results/**
    - tests/fixtures/**
fixtures:
  positive:
    - 产出位置：写入 03_创作任务总表
  negative:
    - 产出位置：写入创作灵感表
  near_miss:
    - 事实源字段 current_node_title 可以保存历史标题“创作灵感表”
  exception:
    - allow reason=truth-source owner=media-os expires=2026-12-31 scope=registry.json
exception_policy:
  required_fields: [reason, owner, expires, scope]
retirement:
  stronger_control: schema_or_registry_contract
  condition: All user-facing labels are rendered from the registry contract.
  action: downgrade heuristic guard to a contract smoke check.

Examples:

control-promotion validate-guard-spec examples/guard-specs/good-creation-table-contract.yaml
control-promotion run-guard-fixtures \
  examples/guard-specs/good-creation-table-contract.yaml \
  --command "python examples/mock-guards/creation_table_guard.py" \
  --format yaml
control-promotion review-guard \
  examples/guard-specs/good-creation-table-contract.yaml \
  --fixture-execution examples/fixture-execution/good-creation-table-contract.yaml
control-promotion review --candidate examples/candidates/good-creation-table-contract.yaml

control-promotion validate-guard-spec examples/guard-specs/bad-incident-string-guard.yaml
control-promotion review --candidate examples/candidates/bad-incident-string-guard.yaml --format markdown

The bad sample is intentionally blocked because it has a one-incident phrase denylist, fixed file scope, no canonical contract, no near-miss fixture, incomplete exception policy, and no retirement path.

Starting in v0.4.0, GuardSpec alone is not enough for promotion. L5 promotion requires fixture execution proof. The local CLI can execute fixtures through a narrow stdin/exit-code interface; the MCP server only validates supplied proof and does not run commands.

Starting in v0.6.0, execution proof has a stricter trust boundary. guard_spec_fingerprint detects GuardSpec drift, but it is not a signature and does not prove that execution happened. Reusing a proof file is treated as declared external evidence, even if it contains proof_origin: executed. Use review-guard --execute or review --execute when the gate decision must be based on tool-executed fixtures. This is a breaking change from v0.5.0: externally supplied executed proof no longer promotes an L5 guard.

Starting in v0.5.0, abstraction risk is based on structured GuardSpec fields and the cataloged abstraction rubric, not candidate prose. Candidate/evidence text can still produce warnings, but text warnings do not lower or raise the promotion risk score.

Exposed MCP Resources

  • control://ladder

  • control://routing-matrix

  • control://abstraction-rubric

  • control://smell-rubric

  • control://proof-obligations

  • control://retirement-policy

  • control://fingerprints

  • catalog://base

  • catalog://project

  • adapter://project

  • schema://guard-spec

  • schema://fixture-execution

  • sample://guard-spec/good-creation-table-contract

  • sample://guard-spec/bad-incident-string-guard

  • template://smell-gate-report

Exposed MCP Prompts

  • review-control-candidate

  • promote-experience

  • retire-guard

Project Adapter

Every consuming repository should keep project-specific paths and policies in .control-promotion.yaml instead of forking this server. The adapter expresses:

  • project type

  • AGENTS, Skill, docs, quality, QA, and generated paths

  • SSOT links

  • baseline quality commands

  • routing overrides

  • generated artifact and write-tool policies

Safety Model

V1 is read-only. It can inspect files, validate catalogs/adapters, classify candidates, and render reports. It does not write repository files, run arbitrary project commands, or mutate governance rules through MCP.

Future write tools should remain disabled by default, require explicit path scopes, forbid generated and secret paths, and return diffs plus verification commands before applying changes.

Development

python -m unittest discover -s tests
PYTHONPATH=src python -m control_promotion.cli validate-adapter .control-promotion.yaml
PYTHONPATH=src python -m control_promotion.cli validate-catalog references/smell-catalog.yaml
PYTHONPATH=src python -m control_promotion.cli validate-guard-spec examples/guard-specs/good-creation-table-contract.yaml
PYTHONPATH=src python -m control_promotion.cli review-guard examples/guard-specs/good-creation-table-contract.yaml --execute "python examples/mock-guards/creation_table_guard.py"

This repository intentionally does not use GitHub Actions. Run the local validation commands above before publishing a commit or tag.

Install Server
A
license - permissive license
B
quality
A
maintenance

How are these scores calculated?

Maintenance

–Maintainers
–Response time
–Release cycle
1Releases (12mo)
Commit activity

Resources

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Tools

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/ValentinoWang/Control-PromotionMCP'

If you have feedback or need assistance with the MCP directory API, please join our Discord server