RUL-041 — The host Claude writes no orchestrata code; it perfects the briefs, checks and workflow and measures the worker models

Ruling

Mohamed, 2026-10-22, verbatim:

"this claude here remains production claude but it writes no code when working on my project just perfects the briefs checks and workflow with the flagship model while measuring sonnect and qwen and deepseek performance and workflow improvements"

• The host Claude stays production Claude, on the host and on the flagship model. The test and prod containers of ADR-0124 are not where it runs.
• On orchestrata it writes no code. Code is anything executed, imported or loaded: source, tests, scripts, hooks, systemd units, pyproject.toml and uv.lock, terraform, CI. The worker models write it: Sonnet (Claude subagents), Qwen and DeepSeek (Hermes, pi-task).
• Its work is the briefs, the checks and the workflow: writing and correcting each brief, checking each worker's result by execution, and improving the template, the improvement log and the mechanisms that catch repeated defects.
• It measures. For each dispatch: the model, the time, the cost, the first-pass result of the checks, the defects the check found, and the rework rounds. A workflow change counts as an improvement only if those measurements improve.

Decision record: ADR-0125 (docs/adr/0125-the-host-claude-writes-no-orchestrata-code-it-perfects-briefs-checks-and-workflow.md).

Evidence

• Practice on 2026-10-22 already took this shape: briefs in plans/briefs/, the Sonnet brief template, and an improvement log recording defects D1 to D21 and findings F1 to F5.
• Hermes reports answered_by, fell_back and cost_usd for every Qwen and DeepSeek run, so those runs can be measured without new tooling.

Enforcement

• none (judgment)

Mechanism (RUL-014)

Judgment-only at recording time. Owed (ADR-0125): a user-level PreToolUse guard that refuses Write, Edit and Bash writes to code files in orchestrata checkouts and worktrees from the main session while allowing worker subagents. Building it first requires measuring whether the hook input tells the main session from a subagent. A worker builds the guard. When it lands, this section names its commit and the Enforcement line moves to hook.

Verification

• Each orchestrata code commit after 2026-10-22 traces to a worker dispatch with a brief in plans/briefs/.
• Each dispatch has a measurement row (model, time, cost, first-pass check result, defects, rework).
• Once the guard lands: a main-session Edit of a .py file in an orchestrata worktree is refused, and the same edit from a worker subagent goes through.