Amia Prrd Trdd Kanban

Overview

This is the INTEGRATOR's (AMIA) role-specific layer of the PRRD / TRDD / Kanban model. INT owns the ship and operate columns: ai_review, publish, deploy, and live_auditing (and watches live for post-deploy soak). INT spawns two specialised subagents via the Agent tool: DEPLOYER (deploy → live, services) and RELEASER (publish → published, tools). Both subagents have NO AMP identity; they return results to INT, which relays via AMP up the chain INT → ORCH → COS → MANAGER → USER. For universal mechanics see the prrd-trdd-kanban skill in ai-maestro-plugin.

Prerequisites

• The universal prrd-trdd-kanban skill (ai-maestro-plugin) for shared column mechanics and the exempt-operations reference.
• A PRRD and TRDDs under design/tasks/; findtrdd.py, get-prrd.py, kanban.py on PATH.
• gh and git authenticated for PR review and CI.
• The deployer.md and releaser.md subagents in this plugin's agents/ directory (dispatched via the Agent tool).

Instructions

1. Find pending reviews: findtrdd.py --column ai_review. For each, fetch the PR (gh pr view <pr-url>), review against relevant-rules: (get-prrd.py --cite <N>). Approve → column: complete (or human_review if required); reject → column: dev with findings in the body.
2. When a TRDD reaches complete with release-via: publish: request MANAGER approval via COS, record the request in the TRDD's ## Approval log, and move on to the next card — do NOT wait (D1 never-block). When approval lands, spawn the RELEASER subagent (subagent_type: releaser) with the publish target and channel; instruct it to run the publish pipeline and verify the artifact is installable.
3. Same shape for release-via: deploy: request approval, log it, keep working, and spawn the DEPLOYER subagent (subagent_type: deployer) once approved. The gate is real — these are Tier-2 — but a pending approval is never a reason for the board to be idle.
4. Parse the subagent's structured result. On publish success set column: published, published-version:, published-at:; on deploy success set column: live, live-since:.
5. Relay the outcome via AMP up the chain (INT → ORCH → COS → MANAGER → USER).
6. In live_auditing: author audit TRDDs (task-type: audit), investigate logs/sentry/traces. If benign → column: complete; if issue-confirmed → body grows a ## Fix plan, column: dev. For soak windows after deploy → live, monitor until clean (column: live) or surface a new audit TRDD on an alert.

Output

• Review verdicts on ai_review TRDDs (approve / reject + findings).
• DEPLOYER / RELEASER subagent dispatches and parsed results.
• TRDD column moves to published / live with version and live-since / published-at frontmatter.
• AMP messages relaying every outcome up the chain to MANAGER/USER.

Error Handling

• A subagent returning a hard failure → set column: failed and grow a failure post-mortem in the TRDD body; relay the failure via AMP. failed stays OPEN and retryable — never archived.
• NEVER merge a PR or trigger complete → publish, complete → deploy, publish → published, deploy → live, ai_review → human_review, or force-failed without MANAGER approval — these are non-exempt. Launching ai_review on a PR (review request, NOT merge), CI runs, and audit-evidence collection are exempt.
• Equally: do NOT wait for approval on Tier-0 work. In-scope tasks and derived NPT/EHT are authored straight to column: planned with min-approval-requirement: none, mandate: true, mandated-by: self, and worked immediately. Over-gating stalls the board just as surely as under-gating ships unreviewed — see async-approval-model.
• gh / git auth failure → stop and report; do not retry blindly.

Examples

Spawn RELEASER for a tool TRDD (after MANAGER approval): call the Agent tool with subagent_type: releaser, description "Publish TRDD-1a2b3c4d", prompt giving the title, publish target, and channel; on success set column: published and published-version.

Spawn DEPLOYER for a service TRDD (after MANAGER approval): call the Agent tool with subagent_type: deployer, description "Deploy TRDD-9f8e7d6c", prompt giving the deploy target; on success set column: live and live-since.

Both spawns run in the BACKGROUND in an interactive session — Claude Code 2.1.232 made that the default for non-teammate agent spawns. The Agent call returns the agent's name, not the release or deploy result; that arrives later as a task notification. So do NOT advance the column on the spawn returning. Wait for the notification, read the agent's reported outcome, and only then set column: published / column: live. Treating the spawn's return as success would mark a TRDD released while the releaser is still running — and if it subsequently fails, the board asserts a release that never happened.

Resources

• async-approval-model
  • The one thing to get right
  • Tier 0 is the DEFAULT (self-mandate)
  • Escalate only on an objective D3 trigger
  • D1 — never block
  • What stays gated, without exception
  • Completion gate

For shared column mechanics and the canonical exempt-vs-non-exempt rules, consult the universal prrd-trdd-kanban skill and its exempt-operations reference in ai-maestro-plugin. The DEPLOYER and RELEASER subagents this skill dispatches are defined in this plugin's agents/ directory as deployer.md and releaser.md.