sanad-agent: Skill for Codex

Ground Sanad Work in External Evidence

Use one completion-oriented workflow for both evidence authoring and implementation. Publicly acknowledge that Sanad learns from external reference projects while keeping source-specific navigation details in the ignored local evidence store.

Choose the mode

• Author or refresh: Use when planning a new reference-informed task, when no packet exists, or when a source revision changed.
• Consume: Use when implementing or reviewing a task with a current packet.
• Audit: Use after implementation to compare the result with adopted and adapted obligations.

Read references/evidence-packet-authoring.md completely before authoring, refreshing, discovering, or downloading sources.

Resolve before implementation

1. Run scripts/resolve_packet.sh <task-id> from the primary checkout or a linked worktree.
2. Follow the returned status:
   • ready: read the packet completely and continue in Consume mode.
   • authoring_required: enter Author mode and create the missing store data or packet.
   • refresh_required: enter Refresh mode and reconcile the changed source.
   • source_unavailable: exhaust catalog, public discovery, and authorized access options before blocking.
3. Read the governing AGENTS.md and license for each selected source tree.
4. Read any navigation_aid paths returned by the resolver when they accelerate discovery. Treat them as optional maps, not evidence authority.
5. Inspect every mandatory source file, symbol, and test in the packet. Reports are navigation aids only; source and tests are authoritative.
6. If a source working tree is dirty, determine whether the changed files overlap mandatory evidence. Do not treat unrelated dirt as a blocker and do not overwrite user changes.

Do not stop merely because a packet, catalog entry, or local clone is absent. Build or repair the recoverable evidence first, rerun the resolver, then continue implementation.