Home / Projects / Orchestrata Ultimate / Decisions / ADR index / ADR-0125

ADR-0125: The host Claude writes no orchestrata code; it perfects the briefs, checks and workflow and measures the worker models

← ADR-0124 · ADR-0126 →

The record

Status: Accepted (Mohamed's ruling, 2026-10-22) · Date: 2026-10-22

Context

• On 2026-10-22 the work was already split. The host Claude (flagship model) wrote briefs and dispatched Sonnet agents to build. It then checked their results and logged each brief defect in plans/briefs/IMPROVEMENT-LOG.md (D1 to D21, findings F1 to F5), with a template (plans/briefs/TEMPLATE-sonnet-agent.md) and a queue of mechanisms for repeated defects.
• Qwen and DeepSeek are available as workers through Hermes, which reports answered_by, fell_back and cost_usd per run, and through pi-task (/opt/pi-worker).
• ADR-0124 puts orchestrata's runtime in a test container and a prod container. This ADR says where the host Claude stands relative to them and to the code.

Decision (Mohamed, 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"

1. The host Claude stays production Claude, on the host and on the flagship model.
2. 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. Worker models write it: Sonnet (Claude subagents), Qwen and DeepSeek (Hermes, pi-task).
3. Its work is the briefs, the checks and the workflow.
4. It measures each worker model's performance and each workflow change. A change is an improvement only if the measurements improve.

Alternatives rejected

• The flagship model also writes code when that is quicker. That blurs which model produced what and leaves no measurement of the workers. Mohamed ruled it out.
• Workers write code without a measured record. Then "Sonnet is good enough for X" or "DeepSeek is cheaper for Y" is an impression, not data.

Consequences

• Every orchestrata code change is dispatched with a brief in plans/briefs/, built by a worker and checked by execution before commit.
• Each dispatch gets a measurement row: date, brief, task class, model id, wall time, cost, first-pass check result, defects by class (the improvement log's classes), rework rounds and template version. Where the rows live, and their exact fields, are settled in the first measurement brief.
• Decision records, rulings, briefs, plans, docs and wiki pages are workflow artifacts, not code, and stay with the host Claude.
• Mechanism owed (RUL-041): a user-level PreToolUse guard refusing main-session writes to code files in orchestrata checkouts and worktrees while allowing worker subagents. It first requires measuring whether the hook input tells the main session from a subagent. A worker builds it.

Provenance

• Mohamed, 2026-10-22, the verbatim ruling above, given while planning ADR-0124's containers.
• Wiki ruling: rulings/RUL-041-host-claude-writes-no-code-perfects-briefs-checks-workflow.md.

Related ADRs

• Also mentioned: ADR-0124

Rulings cited

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

Specs cited

None (or no spec page exists for the ones named).