Summary

Phase 1 of the metadata-field-conventions unification (apps/server/docs/design/subagent-metadata-field-conventions.md , "The system fix, finalized").

Cedar ships nine system agents, but "is this a system agent" was decided in two places that could disagree:

• SYSTEM_SUBAGENT_FILENAMES named only five (crm-updater, strategist, next-steps, meeting-prep, daily-agenda).
• agent-read.ts compensated by OR-ing it with a document's own system: true frontmatter claim to catch the other four (drafter, pipeline-review, coach-meeting, coach-weekly).
• aop-agents.ts's own SubagentDocAgent.isSystemDefault used only the first signal , no compensation , so it under-reported those four.

This PR replaces the Set with one SYSTEM_SUBAGENTS registry naming all nine, each tagged with its dispatch kind (custom / scheduled / generic). Being in the registry is now what "system agent" means, not a document's claim about itself.

What's explicitly NOT in this PR (deferred by design)

• Tool-allowlist curation for pipeline-review/coach-meeting/coach-weekly , the registry's generic kind reserves a toolAllowlist field, left undefined on purpose. Curating real allowlists needs production tool-call data (the same diligence STRATEGIST_TOOLS got), which is its own follow-up (Phase 3).
• The shared generic dispatcher itself , runSubagentTool.ts's existing fallback (readSubagentFull + runAgent) already serves every non-custom agent today, registered or not, and is unchanged here. Nothing about dispatch behavior changes in this PR.
• runSubagentTool.ts's synthetic AgentDescriptor.isSystemDefault (a separate computation, hardcoded false in the generic-fallback branch) , this one feeds markExecutionAsFailedTool's failure-alert suppression, not UI classification. Conflating it with the registry would silently suppress production monitoring alerts for pipeline-review/coach-meeting/coach-weekly. Verified and deliberately left untouched.
• Path B (chat) wiring onto the registry , Phase 2 of the same doc.

Verified

• New system-subagents-registry.test.ts , asserts the full nine-agent shape and dispatch classification (all passing).
• agent-read.ts's isSystemDefault change verified as behavior-preserving: its only consumers (AgentActionMenu.tsx's delete warning, agent-groups.ts's System section) already saw all nine agents correctly via the old OR , this just removes the second, disagreeable signal.
• aop-agents.ts's isSystemDefault change verified as safe: its only consumers (aop.ts/aop-agents.ts tRPC routes, Files-tab agent listing, backfill scripts) are UI-read-only, no execution-critical path.
• pnpm --filter @zero/server run types , clean.
• pnpm deps:check , no violations (1770 modules, 9251 dependencies cruised).
• oxlint --deny-warnings on all touched files , clean.
• Full agent-workspace suite (330 tests) + the three router tests + strategist-seed/reference-resolver , all green.

🤖 Generated with Claude Code

https://claude.ai/code/session_01E3AhwqADAFeqe9wEhz6Mk7

Greptile Summary

The PR centralizes system-subagent identity and dispatch metadata in SYSTEM_SUBAGENTS, then updates both agent-listing paths and scheduled-trigger inference to use it.

• Covers all nine canonical system-agent slugs.
• Preserves the established custom, scheduled, and generic dispatch classifications.
• Removes frontmatter as an authority for system-default classification.
• Needs an own-property membership check and a test tied to independent canonical definitions.

Confidence Score: 4/5

The PR appears safe to merge after addressing two non-blocking correctness and test-maintainability concerns.

The registry matches the current canonical agents and dispatch paths, but prototype-chain membership can misclassify a specially named user agent, and the new registr