Add Agents skill and BYO provider guidance

[calvin-archastro]

Review on ArchCode

Problem and author intent

External coding agents need an installable guide to ArchDev's Agents command family and BYO model access. Session execution, Factory automation, provider authentication, and local-work recovery have different prerequisites and ownership. Without that distinction, an agent can install Jobs unnecessarily, choose the wrong model route, or mistake stored OAuth credentials for successful subscription access.

This adds a companion Agents skill covering the public execution surface and settings provider, including ChatGPT/Grok subscription OAuth and explicitly chosen API keys.

What changed

• Add independent Agents skill packaging and pinned-installer Unix/PowerShell bootstraps, with explicit Agents/provider help-signature checks.
• Cover agent-only setup, interactive/headless execution, resume/stop, Factory, workflows, images, effective definitions, worktrees, and retained local work.
• Document separate ArchDev/provider authentication, browser/device-code login, account ordering, credential removal versus email-list edits, cooldown reset, aliases, and explicit model selection.
• Explain that BYO catalogs are bundled discovery rather than account-access proof; existing-account login may skip reauthentication, while public logout affects the entire provider.
• Keep ordinary-session delegation/docs commands distinct from Factory's control tools and read-only definitions distinct from coding workers.
• Add README discovery and a PR CI packaging/bootstrap job.
• Include the requested Jobs clarification: Tasks are optional; jobs repo submit replaces archdev push, and --task only associates existing Tasks.

Scope: CLI distribution, agent documentation, and CI; no backend/frontend runtime changes. Users can independently install the skill and have their coding agent connect model access and operate the existing CLI.

Risk

Low to medium: executable additions mirror existing skill bootstraps; instructions govern account and automation operations. They preserve provider/publication intent, session ownership, and explicit destructive-reset boundaries. The change is primarily instructions split into focused references; bootstrap and discovery ship together so the skill installs independently.

Testing

Canonical packaging/bootstrap proof: tests/agents-skill.sh, the complete script target. It crosses the real skills-manager process boundary for global and repository installs, checks the packaged references, then executes the installed bootstrap against a fixture installer/CLI. It asserts cold installation, exact executable identity, reuse, upgrade of an older CLI returning generic help, and failure without misleading stdout. This is not a live OAuth/provider end-to-end proof.

Passed:

• bash tests/agents-skill.sh
• Skill frontmatter validation, Bash syntax checks, and git diff --check
• Custom-definition example through parseSubagentConfig; alias structure through validateArchdevModelAliases
• Actual built CLI help checks for Agents execution and provider login
• Independent source walkthrough of SSH subscription login, account removal/reset/logout, headless continuation, Factory one-offs, session control, definitions, and cleanup; identified caveats were corrected

Automatic coverage: .github/workflows/installer-smoke-test.yml → agents-skill → bash tests/agents-skill.sh, on every PR, main push, and manual dispatch. This distribution repository has no Aster target; the workflow runs the script directly without test filters or skips.

Follow-ups and limitations

Live browser/device OAuth, subscription/API requests, and Windows bootstrap were not exercised. These require the user's authorized account and an appropriate native environment; follow-up verification must cross the real login/provider boundary before claiming subscription access works. The new CI job covers installation/bootstrap, not live model access. PowerShell bootstrap has no dedicated Agents CI execution yet. No accounts, credentials, daemon registrations, or live automation were changed for validation.