{
  "markdown": "# WORKS Public Verifier\n\n<!-- mcp-name: io.github.DYZCODE/works-public -->\n\nWORKS independently checks an immutable public GitHub commit against a pinned\nstatic contract and returns a signed receipt. The pilot never executes\nrepository commands and supports only exact lowercase 40-character commit SHAs.\n\n## One-command install\n\nInstall in every detected supported agent without supplying a Vercel credential:\n\n```console\nnpx github:DYZCODE/works-plugins\n```\n\nUse `--client codex` or `--client claude` to select only one host. Restart the\nhost after the installer completes.\n\nAfter success, the installer sends one bodyless aggregate install signal. It\ncontains no installation ID, account, repository, payload, or cookie and can be\ndisabled with `--no-signal`.\n\n## Install in Codex\n\nAdd this repository as a marketplace:\n\n```console\ncodex plugin marketplace add DYZCODE/works-plugins --ref v0.1.13\n```\n\nRestart the ChatGPT desktop app, open Plugins, choose **WORKS Plugins**, and\ninstall **WORKS Public Verifier**.\n\n## Install in Claude Code\n\nAdd the same repository as a marketplace and install the plugin:\n\n```console\nclaude plugin marketplace add DYZCODE/works-plugins@v0.1.13\nclaude plugin install works-public@works-plugins\n```\n\n## MCP Registry\n\nThe official MCP Registry metadata points to a dedicated three-tool endpoint.\nIts frozen Codex gate passed with 20 of 20 eligible verifications, zero\nverification in 20 ineligible controls, 100% eligible response fidelity, and no\nfinal host errors. Ineligible controls had a 34.8955-second median and a\n67.353-second maximum final wall time. One control recovered on its single\nallowed retry.\n\nThe Registry is currently in preview. The Codex and Claude plugin path remains\navailable so supported hosts can also install the strict selection and\nabstention skill.\n\nThe `v0.1.6` Registry attempt was rejected before publication because its\nGitHub namespace casing did not match the OIDC grant. `v0.1.7` published the\ncorrect Registry metadata but retained the previous Codex adapter version.\n`v0.1.8` aligned the Codex, Claude, and Registry release metadata. `v0.1.9`\nadds the one-command installer and its bodyless aggregate install signal without\nchanging the validated verification tool or trust anchor. `v0.1.10` makes the\ninstaller replace older tag-pinned marketplace snapshots before reinstalling.\n`v0.1.11` publishes the fresh paired multistack result and adds a locked,\ntestable installer package.\n`v0.1.12` publishes the rejected v3 scale result and its successful targeted\nv4 corrective canary without changing the public static verifier or trust anchor.\n`v0.1.13` publishes the passing reserved v4b holdout and advances only the\ntechnical evidence state; billing, the verifier, and the trust anchor are\nunchanged.\n\n## Pilot evidence\n\nThe frozen Codex gate passed with 20 of 20 eligible selections, zero activation\nin 20 ineligible controls, complete receipt fidelity, and a 0.5905-second median\neligible-minus-control wall-time delta. One ineligible control recovered on its\nsingle allowed retry.\n\nClaude marketplace and plugin manifests validate, but autonomous Claude\nselection was not executed for this release because the local test account\nrequired reauthentication. No cross-host claim is made.\n\nThe fresh repair-agent pilot then compared 20 randomized pairs across Node,\nPython, monolithic, and layered fixtures. WORKS passed 20/20 while the control\npassed 6/20, a 70-point lift. It stayed within the frozen wall-time and verifier\ncall budgets; median input tokens increased by 53.24%. Raw results, score, hashes,\nmethod, and limitations are in [`evidence/`](evidence/README.md).\n\nThe next frozen scale run stopped at 34 tasks with verdict `kill_or_pivot`.\nWORKS retained a 23.33-point success lift and neutral median wall time, but\nmissed the 25% token limit at 26.44% and exhausted its final error budget. A\ntargeted four-case corrective canary then passed 4/4 with 68.32% lower median\ninput tokens and 67.08% lower median wall time than the corresponding v3 arms.\n\nThe reserved fresh v4b holdout then completed 23 decisive pairs. WORKS passed\n21/23 while the same model without WORKS passed 7/23, a 60.87-point lift.\nMedian wall time fell 45.01% and median input tokens fell 20.44%. All frozen\nruntime, fixture, verifier-call, selection, and error gates passed. One WORKS\narm timed out, and two Python repairs failed in both arms. This clears the\ntechnical pilot gate; it does not prove organic routing, real-repository\nperformance, retention, willingness to pay, or a durable moat.\n\n## Pilot scope\n\n- Plugin MCP endpoint: `https://works-runner.vercel.app/mcp`\n- Registry MCP endpoint: `https://works-runner.vercel.app/mcp-registry`\n- Trust anchor: `https://works-runner.vercel.app/trust`\n- Contracts: `node-package` and `env-safety`\n- Source: public GitHub repositories at immutable commits\n- Execution: static checks only\n\nThe automatic verification tool is read-only and stores no verification\ntelemetry, repository URL, commit SHA, source, tool body, credentials, or\ninstallation identifier. Its signed receipt is returned to the caller only.\n\nA passing static receipt proves only the outcomes declared by its contract. It\ndoes not prove builds, tests, deployments, payments, runtime behavior, or market\nvalidation.\n",
  "bytes": 5292,
  "sha": "9a11c1759e7e25f723330b9e2e42b2d1ea1190bea09ad48309f135289b4eafce",
  "repo_slug": "dyzcode/works-plugins",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/mcp_io_github_dyzcode_works_public_87cf4f41/readme"
}