Kural dry-run plugin for ChatGPT and Codex

This folder is a new OpenAI Plugins package using portable Agent Plugins 1.0.0. It is not custom GPT Actions, the retired ai-plugin.json beta, a UI, or a deployed robot integration.

Status and contents

plugin.json: portable identity at the package root.
mcp.json: one local stdio server, kural-skills-mcp, with no arguments.
skills/kural-dry-run/SKILL.md: request-planning and dry-run instructions.
package.py: optional deterministic ZIP builder; standard library only.
validation/: official HTTPS schema snapshots and local validation evidence.

Local adapter verified; platform installation remains unverified. The shared adapter belongs in the repository root Python package, not this plugin folder. Its argument-free entrypoint, kural-skills-mcp, passed local stdio MCP discovery and call tests with 20 forced-dry-run tools. The root native suite passed all 46 tests. Stdio is the only transport; no HTTP transport is provided. That evidence does not verify a ChatGPT/Codex client, account, tunnel, hosted endpoint, or robot connection.

Verified local adapter interface

The shared adapter forces DryRunBackend; there is no live mode, gateway selection, robot connection, or environment switch that enables robot control. Its locally verified MCP tools are:

kural_capabilities: dry-run mode, honest capability limits and SDK catalog.
kural_dry_run_<canonical_skill> for each of these 18 skills: move_forward, move_backward, move_left, move_right, steer_left, steer_right, steer_back_left, steer_back_right, turn_left, turn_right, base_velocity, lift_up, lift_down, home, wave, point, inspect, stow.
kural_dry_run_stop: immediate SDK dry-run stop result. It sends no robot STOP, confirms no physical stop, and is not an emergency-stop control.

Primitive tool inputs must come from the current SDK catalog. Discover schemas rather than guessing parameters. kural_capabilities returns a standalone object with protocol, version, mode, no_live_backend, tools, skills and limitations. Successful primitive and stop calls return structuredContent with mode: "dry_run", execution: "validated intent only; no robot motion" and result containing the SDK's SkillResult.as_dict(). execution is a string, not an object. Tool errors are not successful result envelopes. The nested SDK result identifies the operation, skill, state, reason, normalized request and available evidence. Physical verification flags remain false. Say “validated dry run; no robot motion” only when the actual result supports that statement. A validation error is not an accepted dry run. No asynchronous start/status/cancel/resume, network retry idempotency, autonomous velocity loop, or live lifecycle contract is offered.

kural-skills serve is the native JSON-lines session. It is not MCP, even though MCP stdio also uses JSON messages. Do not put it in this MCP configuration.

Local installation: future manual steps

Use Python 3.11 or newer. From the kural-skills repository root, after the adapter has passed its local tests:

uv venv .venv --python 3.12
uv pip install --python .venv/bin/python -e '.[mcp]'
.venv/bin/python -m unittest discover -s tests -v
.venv/bin/kural-skills-mcp --help

This installs the SDK and its optional official MCP dependency. No OpenAI API key is required for local dry-run execution. These commands have not been run as a plugin installation. Do not reinstall over an active environment without reviewing its current dependencies.

In a local compatible plugin client with Agent Plugins and stdio support, install/import this folder using that client's local-package or local-marketplace flow. Point its plugin source at chatgpt/, not the repository root. Put the repository's .venv/bin on that client's process PATH before starting it; mcp.json resolves the executable token kural-skills-mcp through that PATH. A terminal PATH change does not change an already running desktop client's PATH. If the client cannot resolve the token, configure its process launch environment explicitly. No machine-specific absolute path is embedded in the portable package.

Client support and account/workspace policy must be checked separately. Merely copying a package folder does not install it or prove tool execution. Test MCP initialization, discovery, capabilities, one bounded dry-run request, invalid parameters and the dry-run stop tool. Confirm there is no robot process involved.

ChatGPT web: separate private access approval required

ChatGPT web cannot be assumed to launch this local executable. The stdio manifest is for local compatible clients, not a hosted endpoint. For a private ChatGPT test, separately approve and configure Secure MCP Tunnel to the installed, tested MCP adapter. Check current account/workspace eligibility and the tunnel client's documented stdio support. No tunnel command, credentials, auth server, hosted service, endpoint or ChatGPT plugin has been configured here.

The documented personal flow is Settings → Security and login → Developer mode, then ChatGPT Plugins → plus → Tunnel (or an actual approved MCP URL). Current UI labels and availability need an account check. After a successful personal connection, use ChatGPT Work and @ to invoke the plugin. These are future steps, not a claim of account access or end-to-end compatibility. Do not submit the local executable token as an MCP URL. There is no placeholder remote URL in mcp.json and no deployable remote configuration in this package.

Private tunnel access is not public hosting, public directory acceptance, authorization for robot movement, or a safety boundary. Any remote/auth setup requires a separate review; the package supplies no credentials or auth policy.

Deterministic packaging

From the repository root:

.venv/bin/python chatgpt/package.py --output /tmp/kural-dry-run-plugin.zip

The ZIP places plugin.json, mcp.json, README.md and the skill at its root (no enclosing chatgpt/ folder). The script writes only to the chosen path, which must be outside this folder. Its fixed allowlist excludes schema reports, source code, scripts, credentials, virtual environments, robot assets and meshes. Packaging does not install the Python adapter or launch a client. Repeated builds of unchanged files produce identical bytes.

A local ZIP is not a public-directory-ready deployment. Official public submission requires a stable public HTTPS MCP endpoint and review; a private tunnel or local stdio package does not meet that requirement. This package has no hooks, UI, .app.json, registered connector, or public submission. Obtain separate approval before changing transport or attempting publication.

Validation and limits

See validation/report.json for fetched schema URLs, SHA-256 values and checks. Manifest validation proves document shape, not MCP interoperability or installation. SDK validation limits are not calibrated robot safety limits. This package does not test obstacles, geometry, clearance, measured movement, mechanical stopping, simulation or hardware. It does not add controllers, depth or perception code. Only Kural's authoritative robot package may define robot assets and controls.

Official references, checked over HTTPS:

The repository's research is in docs/AI-AGENT-INTEGRATIONS.md; that relative link applies to the source checkout, not a standalone distributed ZIP.