Skip to content

init-agent-os ​

Bucket: workflow · Invocation: manual · /init-agent-os (direct Claude), /agent-os:init-agent-os (Claude plugin), or $init-agent-os (Codex)

Sets up managed global policy or seeds one repository's project policy.

global verifies plugin visibility, checks drift, and applies the managed policy blocks through the deterministic script. The explicit invocation authorizes those setup writes.

Repository mode is facts-first: it inspects instruction and policy surfaces, project docs, build and test configuration, CI, and evidence of conditional concerns before asking anything. Before writing, every live host-consumed policy surface has unambiguous ownership and precedence; shared identical content or a clear canonical cross-reference is sufficient, while leaving conflicting applicable rules elsewhere is forbidden. If target ownership cannot be resolved, it asks one ownership question and blocks the write. A transient completeness ledger ensures every applicable material field reaches evidenced or confirmed, N/A, explicitly deferred/open, or conflicting. It asks only unresolved material questions, one at a time, with a recommendation and consequence. The mandatory core is policy ownership, delivery and external-effect boundaries, verification and ground truth, coherent planning identities/claims/handoffs/readiness/dependencies, epic ownership, and durable lesson maintenance. Frontend/rendered QA, package/release/docs, CI/deploy, auth/security/billing, database/migration/concurrency, batch/monorepo, and native/desktop sections activate only when repository evidence supports them.

Each core field reaches one allowed terminal state before writing. Explicit deferrals are recorded concisely under Open setup questions without guessing; only an unresolved policy ownership/write-target conflict blocks the write. The written policy contains only durable behavior-changing defaults and relevant sections; it omits inactive modules, N/A fields, the ledger, and interview history, preserves unrelated instructions, and shows all changed policy surfaces in the resulting diff. The invocation itself authorizes this setup write.

Managed block edits remain script-owned and stop on duplicate or malformed markers.

A personal framework, published in the open.