Summary

• Adds cedar-configuration/using-and-configuring-extracted-fields.md , org/user field scoping, the live/liveReason freshness signal (design pending a separate list-fields code change), other write gates independent of scope, and who can actually change what. Previously this content didn't exist and the org/user field model had no dedicated resource at all.
• Links the new resource from cedar-configuration/SKILL.md's catalog, fields-and-values.md, building-an-agent.md, and playbook-authoring/SKILL.md , confirmed via grep it was reachable from nowhere before this.
• scope-and-sharing.md: orphan cleanup on promote-to-org is now mandatory unless the user asks to keep a copy (was "optionally").
• fields-and-values.md: corrected a false claim that agentRequiresApproval fields notify via Slack automatically , verified against the real write path that creating the approval task is a silent DB insert with no notification of any kind.
• playbook-authoring/SKILL.md: adds a "know who you're talking to" role-awareness rule (admin → team-oriented suggestions, member → own config only), a "scope-blind rewrite" guardrail (an additive change should default to a patch, not a full-body rewrite), a document-only-output confirmation rule, the exact-field-key rule, and the approval-required grouped-instruction pattern for crm-updater.

All found and fixed while auditing two newly-added Aspire org agents and a user's custom agent this session , see session notes for the underlying incidents (Slack channel resolved to a per-rep-scoped lookup that silently failed for private channels; a field referenced by description instead of its real key; a subagent silently downgraded to a generic output because the authoring session had no visibility into a prior chat thread's proven-working recipe).

Test plan

• Grepped the full apps/server/.claude/skills/ tree for stale references to the renamed file , none found
• Confirmed every resource named in cedar-configuration/SKILL.md's catalog exists on disk
• Static discoverability trace against 3 realistic authoring prompts (documented in session notes); one routing gap found and fixed (the exact-field-key rule was only reachable from §1a, invisible to §1b's custom-agent flow , moved to §0, which both pass through)
• This is a documentation-only change (.claude/skills/**/*.md) , no code paths affected, no tests applicable

🤖 Generated with Claude Code

https://claude.ai/code/session_01F4mjBtjYZyxnVvLQpehDSb

Greptile Summary

This PR expands agent-authoring guidance and adds supporting configuration, field-liveness, Slack-channel discovery, authorization, and test-run functionality.

• Adds extracted-field scope, freshness, ownership, approval, and aggregation guidance.
• Exposes live and liveReason through list-fields.
• Adds role-aware bot-channel discovery and rejects known multi-workspace ambiguity.
• Adds a chat-accessible subagent test-run action and documents that suppressed delivery does not suppress durable writes.
• Improves stale-owner handling, although organization owners can still be lost through filename shadowing.

Confidence Score: 3/5

The PR is not yet safe to merge because organization-owned fields can be incorrectly treated as unowned, and delivery-suppressed test runs still perform durable mutations.

The new owner-validation path derives its live-agent set from a filename-deduplicated listing, so a same-named user agent can hide a still-runnable organization owner and allow the CRM updater to write its field. The previous preview-run finding also remains outstanding: the updated implementation and documentation now acknowledge that deliverForReal: false suppresses delivery only, while database writes, documents, tasks, and other durable effects still occur.

Files Needing Attention: apps/server/src/services/aop/field-ownership.ts, apps/server/s