agents-delivery-workflow

name	agents-delivery-workflow
description	Non-trivial delivery start→finish, validation routing, meta/subagent pointers, and completion gate before done/commit.

Delivery workflow

Canon: AGENTS.md, docs/agent_knowledge_base.md, docs/agents_quick_reference.md. Default: Plan -> Execute -> Verify -> Report. Triggers: docs/agents_quick_reference.md#automatic-workflow-triggers.

• Plan: non-trivial -> tasks/cursor/todo.md; existing code uses context ladder; delegation -> agents-meta-behavior.
• Plan: do not edit until 95% confident; if vague/risky, define boundaries, data flow, failure handling, smallest proof.
• Plan: one loop (plan -> tool -> observe -> revise); branch only when risk pays.
• Execute: reuse seams; keep Presentation -> Domain <- Data; update DI/routes/l10n/codegen when touched.
• Execute: UI/design/Mix -> DESIGN.md + docs/design_system.md; prefer AppTheme, buildAppMixScope, AppStyles, UI.
• Execute: File verified reusable conclusions into source doc, docs/changes/, docs/plans/, or tasks/lessons.md.
• Verify: docs/ai_code_review_protocol.md, smallest matching validation; empty/truncated tool output is not proof.
• Verify: Self-verify final response vs request, changed files, proof, blockers, residual risk.
• Report: Report only after Verify. Surgical diff: each line traces to request or required validation/doc update.

Docs/agent changes: targeted docs/link checks, ./tool/check_agent_knowledge_base.sh, host-template drift path; escalate to ./bin/checklist when guidance changed materially.

Multi-agent

Coordinator-as-hub. Gate before fan-out.

• Team when >=2: blast radius, cross-layer read, high-risk logic, separate implement/review bars, or user asked plan+implement+verify.
• Single otherwise; tie-break single. Tracker: Benefit: team - <reason> or Benefit: single - <reason>.
• single: normal loop; no tasks/cursor/team/<run-id>/.
• team: create tasks/cursor/team/<run-id>/ (goal.md, findings.md, plan.md, diff-summary.md/diff.md, review.md). Coordinator owns artifacts + validation.
• Spawn specialists with inline context, never path-only. Max two Implementer fix loops unless user extends.
• Specialist returns summary + final result + verified artifacts only, not full transcript/reasoning dump.

Doctrine + matrix: docs/agent_knowledge_base.md#multi-agent-hub; roles/redaction: agents-meta-behavior.

Completion gate

• AI review gate + non-trivial tracker proof.
• UI/design: ./tool/check_design_md.sh if DESIGN.md edited; ./tool/run_mix_lint.sh for Mix token/style edits.
• UI/app: prefer app-visible proof.
• Finish gate: edge cases, failure paths, readability, operational clarity, breakage impact.
• Resolve selected validation failures; add focused tests when behavior warrants.

Commit rules

Inspect git status + git diff; stage intended scope only; check git diff --staged. Commit message: imperative, <=72 char subject, why when not obvious, tests run/not run, no assistant mention. Full bar: docs/testing_overview.md, docs/validation_scripts.md.