Exercise composed cloud and local role variants with real questions #652

New issue

Closed

opened 2026-07-23 13:57:50 +00:00 by coilyco-ops · 0 comments

coilyco-ops commented

2026-07-23 13:57:50 +00:00

Member

Context

The structural designer smoke proves projection but never asks a model to act from the composed role. The first authenticated designer Codex question succeeded and identified the designer role, while also exposing a full Docker writable layer during Codex skill initialization.

Scope

Keep Codex auth read-only and stage it into the ephemeral composed HOME.
Give AOS runtime HOME and temporary composition output bounded ephemeral storage so harness initialization does not depend on the shared Docker writable-layer budget.
Add a reusable one-role-at-a-time question probe for cloud Codex and the configured open-source harness/model path.
Ask a role-specific real question and require each response to identify the role inferred from loaded instructions.
Exercise designer through both paths first, then every canonical role through the open-source path one by one.
Keep credentials, backend addresses, and model-provider details out of tracked transcripts and public output.

Acceptance

Authenticated designer Codex answers its design question as designer without writable-layer warnings.
The designer open-source variant answers its design question as designer.
Engineer, director, QA, advisor, ops, PM, designer, social, sales, and customer-success each answer one role-specific open-source question and identify the correct composed role.
The probe fails on a missing role marker or harness failure, records no secrets, and has a documented Ward entry point.
Unit tests and the full repository hook suite pass.

## Context The structural designer smoke proves projection but never asks a model to act from the composed role. The first authenticated designer Codex question succeeded and identified the designer role, while also exposing a full Docker writable layer during Codex skill initialization. ## Scope * Keep Codex auth read-only and stage it into the ephemeral composed HOME. * Give AOS runtime HOME and temporary composition output bounded ephemeral storage so harness initialization does not depend on the shared Docker writable-layer budget. * Add a reusable one-role-at-a-time question probe for cloud Codex and the configured open-source harness/model path. * Ask a role-specific real question and require each response to identify the role inferred from loaded instructions. * Exercise designer through both paths first, then every canonical role through the open-source path one by one. * Keep credentials, backend addresses, and model-provider details out of tracked transcripts and public output. ## Acceptance * Authenticated designer Codex answers its design question as designer without writable-layer warnings. * The designer open-source variant answers its design question as designer. * Engineer, director, QA, advisor, ops, PM, designer, social, sales, and customer-success each answer one role-specific open-source question and identify the correct composed role. * The probe fails on a missing role marker or harness failure, records no secrets, and has a documented Ward entry point. * Unit tests and the full repository hook suite pass.

coilyco-ops referenced this issue

2026-07-23 14:55:36 +00:00

Measure fine-grained local composed-role instruction adherence #653

coilyco-ops referenced this issue from a commit

2026-07-23 15:04:16 +00:00

test(aos): exercise composed roles through real models (closes #652)

coilyco-ops closed this issue

2026-07-23 15:04:16 +00:00

coilyco-ops referenced this issue

2026-08-05 18:32:39 +00:00

Stage usable Codex auth for standalone acompose inference #900