[prellr]

Summary

Rail-2 of the Drafts germination rail. Gives the lead agent Aye the ability to converse in the draft-level conversation rail shipped in Rail-1, routes her turn's general talk there (reserving annotations for anchored structure), and updates her identity to the finalized "soul" plus the rail capability. Backend/daemon only — no TypeScript touched.

Built on Rail-1's draft_message table + GET/POST /api/drafts/{id}/messages (already agent-attribution-ready via resolveDraftAuthor).

What's in it

1. CLI — Aye reads + posts to the rail (server/cmd/multica/cmd_draft.go)

• multica draft say <id> --body "…" → POST /api/drafts/{id}/messages (mirrors runDraftReply; rejects empty --body; agent attribution automatic via X-Agent-ID/X-Task-ID).
• multica draft messages <id> → GET /api/drafts/{id}/messages (mirrors runDraftAnnotations; table + json).
• Both registered in the draft group + help.

2. Turn prompt routing (server/internal/daemon/prompt.go, buildDraftTurnPrompt)

• Reads the rail at turn start; routes her general response to the rail via say; reserves reply/annotate for anchored points.
• Principle: "Talk lands in the rail; structure lands in the doc (annotations/suggestions)." Existing no-body-edit / no-issue-creation guards unchanged.

3. Identity (server/internal/handler/aye.go)

• ayeInstructions (Layer 1) → the approved soul (Disposition incl. system-steward, Inquisitiveness, memory loop incl. working memory, Orchestration).
• ayeSkillContent (Layer 2) → conversation-rail commands + the "talk in the rail, structure in the doc" rule; rest of the skill unchanged.

4. Backfill updates existing Aye (server/internal/handler/aye_backfill.go)

• BackfillAyeAgents now also refreshes an already-seeded Aye's instructions + attached skill content when they drift from the constants (so the new soul + rail-aware skill reach already-seeded prod agents), in addition to seeding where missing.
• New sqlc queries UpdateAgentInstructions (by agent id + workspace id) and UpdateAgentSkillContent (by agent id + skill name via the agent_skill join), both :execrows and gated on a real diff so a matching Aye is a no-op.
• Keyed strictly on AyeAgentID → a user-created agent is never touched. Best-effort + logged; never blocks startup.

Verification

• make sqlc — clean, no drift.
• go build ./..., go vet ./... — clean.
• go test ./internal/handler -run 'Aye|Draft' ./internal/daemon — green (incl. the new backfill-update test and the updated draft-turn prompt test).
• Built multica draft --help lists say + messages; both --help render; empty say --body is rejected.

Deploy note

CLI (say/messages) + the turn prompt live in the daemon binary → daemon rebuild. The identity constants + backfill run in the backend → backend rebuild.

Does NOT touch feat/drafts-3b-spike. Do not merge — for review.

🤖 Generated with Claude Code