MarkdownLLM: Skill for Claude Code

Configure a MarkdownLLM cloud environment

Resolve the named domain from the local estate; read its AGENTS.md and routed skills before domain writes. Read the framework's adapters/codex-cloud.md. Use the domain's declared framework runtime; on Windows use tools/mdllm.ps1.

Identify the primary GitHub repository Codex selects, plus every additional domain or ordinary code repository needed for the task. For each additional repository ask for branch or immutable commit, intended read/write access, publication policy and the names of credential/boundary variables if needed. The operator need not disclose token values to the agent. Use kind: domain only for a MarkdownLLM domain with framework_root: ../..; use kind: repository for ordinary source code. Read-only references use a pinned commit where possible. The primary host checkout uses the Codex PR path.

Run mdllm cloud configure <primary> --harness codex from a clean, published framework checkout. It produces a private bundle with a strict JSON manifest, bootstrap, settings guide and this skill. For multiple repositories, edit the manifest in a fresh review directory and rerender with --config <manifest>. Check the exact generated bytes before adopting them; the renderer refuses to overwrite different existing files. Commit reviewed artifacts to the owning primary repository under its publication policy. Its default branch must carry them before cloud setup runs. An unavailable framework pin is a release dependency, not permission to push the public framework. Never substitute a moving branch for the full pin.

Read generated .codex/cloud/SETUP.md and environment.json. Credential fields contain variable names, never values. A setup-only secret does not reach the working agent. If agent-time Git access is needed, use a persistent environment variable with a GitHub fine-grained token scoped to the exact repository and needed operations, or prove native Git access in that phase. Do not copy tokens into URLs, committed files, setup output or chat. Keep additional writable domains' publication manual unless their own AGENTS.md already declares literal autopush true and the operator wants that behavior; declared cannot grant new authority. Treat access: read as intention and use read-scoped credentials to enforce it. For declared, ensure the manifest branch is the remote default; the adapter also refuses automatic publication from another checked-out branch.

If the user's request includes configuring Codex, use the available browser tool and its instructions to find/create the matching repository environment, set exactly the generated workspace base, Manual setup and maintenance commands, then save and read back the fields. Preserve unrelated secrets, variables and policies. Changing the pin requires a cache reset; report that it affects a shared cache. If no browser is available, leave the bundle and exact settings for the operator. GitHub login and repository authorisation are user actions when the UI requires them. Do not invent APIs, reuse browser session tokens, or claim a save succeeded without observing it.

If launch is requested, start a task in that environment using the available product route. Verify setup completion, the primary's unchanged selected HEAD, the exact extra repository identities and revisions, credential access in the agent phase (mdllm cloud probe), domain-only floor installation, the clone local PR restraint and model-visible Tier-0 delivery (mdllm cloud start). Run mdllm cloud status for unpushed extra commits and unknown upstreams; check each remote separately. Distinguish local fixture tests, cloud setup logs, project-hook trust and actual agent delivery. Report any untested part explicitly. For each started domain, surface the full trigger evaluation, imports coverage and repository branch map in the working task.