amoa-prrd-trdd-kanban

Overview

This is the ORCHESTRATOR's role-specific layer of the PRRD / TRDD / Kanban model. ORCH is the team's traffic controller and owns the RED (blocked) column — TRDDs whose blocked-by: list is non-empty are the fundamental source of project delays, and ORCH's central job is to minimise the time TRDDs spend there. ORCH owns three columns: todo (claims promoted TRDDs from AMAMA), dispatch (assigns designed TRDDs to agents and moves to dev), and blocked 🔴. It reads but does not own dev, testing, ai_review, human_review for upward status. For the universal mechanics, see the prrd-trdd-kanban skill in ai-maestro-plugin.

Prerequisites

• The universal prrd-trdd-kanban skill (ai-maestro-plugin) for shared mechanics, transition numbers, and the exempt-operations matrix.
• A PRRD plus a populated design/tasks/ tree of TRDD .md files — the TRDD files are the single source of truth.
• AI Maestro Plugin (AMP) installed for inter-agent messaging; ORCH routes every cross-team message through its CHIEF-OF-STAFF (COS).

Instructions

1. Claim a todo: findtrdd.py --column todo, pick highest-priority (oldest breaks ties), read body + frontmatter for intent.
2. Delegate design to ARCHITECT: AMP-send via COS "delegate TRDD- to ARCHITECT", then edit column: design and bump updated:.
3. On design→dispatch, ARCHITECT signals via AMP (a 1→N split leaves N child TRDDs in dispatch, parent superseded); verify each child's task-type:, test-requirements:, review-requirements:, eht:.
4. Assign each dispatch TRDD: set assignee: (skill matches task-type:, capacity via kanban.py --group-by assignee), column: dev, bump updated:, AMP-send the assignee via COS.
5. Run kanban.py --red every session. For each TRDD in the 🔓 BLOCK-CLEARING PRIORITY ranking, raise priority: toward 1 in proportion to unblocks_count; assign unassigned blockers now; recurse into chained (transitively blocked) blockers first.
6. Escalate non-exempt actions via COS — cross-team reassignment, human_review, force-failed. Exempt (no approval): dispatch→dev assignment, red-column priority bumps, within-team reassignment.

Output

• TRDD frontmatter edits: column:, assignee:, priority:, pre-block-column:, updated: — written to the .md files.
• AMP status messages to COS (and through COS to MANAGER for any non-exempt request or escalation).
• Red-column priority actions: bumped priority: on blockers plus priority-ping AMP messages to their assignees.

Error Handling

• Blocker cannot clear within the team → escalate to the team's COS with the chain of blocked-by: refs.
• Cross-team dependency or reassignment → route to MANAGER via COS; never reassign across teams unilaterally.
• A stuck TRDD that needs force-failed or human_review → request MANAGER approval via COS (non-exempt).
• Unsure whether an action is exempt → treat as non-exempt and request approval. Conservative default.

Examples

# Start every session by inspecting the RED column priority ranking
kanban.py --red

# See what is waiting to be assigned, then assign by capacity
findtrdd.py --column dispatch
kanban.py --group-by assignee

Single-writer-per-domain and NPT/EHT collision avoidance

Every mutable surface in design/ has exactly one owning instance at a time (PRRD S5.1). A TRDD .md file is write-locked by the instance named in its current-owner: frontmatter field; a non-owner may mutate only the coordination fields (column:, assignee:) and must delegate any body or requirement edit to the owner or take an explicit claim (set current-owner: to itself) before writing. Never blind-write a file another instance owns — concurrent writes corrupt the single source of truth.

When authoring derived NPT/EHT child TRDDs, prevent cross-instance collisions (PRRD S6.1): each child declares in its body the files/domains it will touch, and before creating it, scan the open TRDDs (findtrdd.py --column dev / dispatch / testing) for a domain overlap. On collision, serialise the children — make one blocked-by: the other so only one writes the shared domain at a time — or merge them into a single TRDD, rather than letting two instances edit the same files in parallel. The task-comprehension handshake (the amoa-implementer-interview-protocol skill) surfaces these domains up front by asking each MEMBER which files/domains it will touch.

Resources

For shared mechanics, transition numbers, and the approval matrix, read the universal prrd-trdd-kanban skill in ai-maestro-plugin — in particular its exempt-operations.md (which transitions need MANAGER approval) and column-transitions.md (the full numbered transition list). The existing amoa-kanban-management skill manages an AI Maestro server-backed board and coexists with this file-based flow: the server board is a UI mirror, while the TRDD .md files remain the source of truth and kanban.py drives in-session red-column decisions. The ORCHESTRATOR persona lives in the agent's main-agent definition.