← ClaudeAtlas

specify-an-isolated-harness-startup-probelisted

Design a source-discovery probe for a fresh harness that can detect selected material and unexpected inherited instructions before a task run.
alisonjieli-png/loop-engine · ★ 2 · AI & Automation · score 68
Install: claude install-skill alisonjieli-png/loop-engine
# Specify an isolated harness startup probe ## Use case Use when a new harness process is expected to see only selected step material, while its client may also read ancestor, user, bundled, or compatibility locations. Produce a probe specification, not a launched process. ## Required inputs - Client identity and version with documented or observed discovery locations. - The selected item manifest, proposed work, home, and configuration roots, and the startup observation interface. - A declared policy for built-in material and for any unavoidable client source. ## Procedure 1. Enumerate every plausible instruction, skill, configuration, and extension source for the exact client version, including ancestors and isolated versus real home. Mark unverified discovery rules as unknown. 2. Name one harmless selected canary and one harmless unselected canary that a future authorized fixture can place at distinct sources. Use synthetic text, never a credential or customer material. 3. Specify the expected observed source set before startup. A selected skill listed in discovery is evidence of discovery only; plan separate evidence for loading and use. 4. Identify where a parent process, inherited environment, shared cache, symlink, or client built-in could carry prior task material into this step. Give each a negative observation that would reveal it. 5. Define a refusal when the probe sees an extra source, misses a required source, cannot inspect a source, or resolves a path out