Write

drwn write materializes the effective harness into downstream local agent tool files. It is the command that turns the resolved model into Claude, Codex, Cursor, and MCP config state.

Preview first:

drwn write --dry-run
drwn write --json

Write all effective state:

drwn write

Limit the operation:

drwn write --target=claude
drwn write --skills-only
drwn write --mcp-only
drwn mcp write --dry-run

drwn write --force repairs drift only inside paths or MCP fields already recorded as drwn-owned. It never claims foreign files or fields.

Write at explicit machine scope rather than project scope:

drwn write --scope machine --dry-run
drwn write --scope machine

--scope machine ignores project config and projects strict machine intent to user-scope tool configs (~/.claude/, ~/.codex/, ~/.cursor/). The write record lives at ~/.agents/drwn/global-write-record.json.

Project writes use <project>/.agents/drwn/write-record.json and project intent only. User-home capabilities may remain ambient in the downstream client but are never copied into project declarations.

A first machine write refuses an existing unrecorded destination or same-ID MCP field with MACHINE_PROJECTION_CONFLICT, even when bytes are identical. Removal deletes only unchanged prior-owned state; drifted and foreign state is preserved.

Enforcement flags:

drwn write --strict-hooks   # fail if any hook policy file cannot be materialized
drwn write --strict         # treat all warnings as errors

--strict-hooks is relevant when you expect hook policies to be present; without it, missing policies are warnings rather than failures.

When locked cards declare optional MCP servers, drwn write includes an optional MCP report. Human output lists each card server as active, skipped, or shadowed. JSON output includes the same data under optionalMcpReport. Skipped card MCPs can be enabled for the current project with:

drwn add mcp <server-name>
drwn write --dry-run