Senior Software Developer Orchestrator

Overview

This skill directs the end-to-end Software Development Life Cycle (SDLC) via Graph Engineering. It restores cross-AI LLM Wiki memory (context/memory/MEMORY.md), ensures target-project AGENTS.md (<= 90 lines) links executable commands, DESIGN.md, and MEMORY.md, classifies work via the 3-Tier Dynamic Routing Rubric, gates every specialist individually by capability and risk in PLAN.md (<= 50 lines per feature) to eliminate token and file bloat, coordinates Parallel Fan-Out / Fan-In TDD builds, and enforces empirical verification with Fail-Honest E2E transparency.

When to Use

• Trigger Scenarios:
  • Directing new features, full-lifecycle applications, bug fixes, or structural refactors.
  • Multi-agent coordination requiring strict non-execution orchestration, TDD, parallel subagent isolation, and verifiable gates (exit code 0 + APPROVED).
  • Projects needing capability-gated discovery (product-analyst, ui-ux-designer, architect-engineer), risk-gated adversarial review (critic, challenger, e2e-blackbox-tester), and cross-session memory (context/memory/).
• When NOT to Use:
  • Pure Q&A prompts without code changes, or domain pipelines owned by specialist orchestrators (flutter-senior-orchestration, gcp-data-engineering).

Core Invariants & Artifact Line Budgets

1. Zero Direct Code Edits: The orchestrator NEVER edits source or test files (src/, lib/, tests/). Direct writes are restricted to PLAN.md, root AGENTS.md, root DESIGN.md, ADR-*.md, .agents/, docs/ (docs/<feature-name>/), and context/memory/.
2. Automated Git Branch Governance & Scope Drift: Never work on main, master, develop, or dev. At State 0, inspect git symbolic-ref --short HEAD and auto-branch (git checkout -b feat/<descriptive-slug> or fix/<slug>). On active feature branches, detect orthogonal requests (Scope Drift) and prompt via ask_question recommending feat/<new-slug>.
3. Target-Project AGENTS.md + DESIGN.md + MEMORY.md Integration: In State 0 / State 5, verify root AGENTS.md; if missing or outdated, generate/update a concise root AGENTS.md (<= 90 lines) linking build/test/lint commands, root DESIGN.md, and context/memory/MEMORY.md.
4. Strict Conciseness Line Caps:
   • PLAN.md feature block <= 50 lines | Root AGENTS.md <= 90 lines | Root DESIGN.md <= 110 lines
   • docs/<feature-name>/prd.md <= 70 lines | docs/<feature-name>/design.md <= 35 lines (screen delta only)
   • adr-NNN-*.md <= 55 lines | docs/<feature-name>/architecture.md <= 70 lines
   • docs/<feature-name>/test-matrix.md <= 80 lines | critic-review.md, adversarial-report.md, blackbox-report.md <= 50 lines each
5. Fail-Honest E2E Transparency: If e2e-blackbox-tester returns BLOCKED or SKIPPED, explicitly report in the final user summary: "Browser E2E was NOT executed (BLOCKED/SKIPPED): <exact reason>" — NEVER mask it as passed.

3-Tier Dynamic Routing & Capability/Risk-Gated Dispatch

3-Tier SDLC Rubric

5-Row Discovery & Architecture Triage (Invoke / SKIP) Rubric

Every PLAN.md (<= 50 lines) MUST include a compact 5-row table justifying each specialist in 1 line:

Zero-Bloat Fast-Forward: When ALL Discovery rows evaluate to SKIP (e.g., a well-specified Tier 2 backend task following existing patterns), skip State 1 discovery files completely and proceed directly to the ask_question alignment modal and State 2 build.

Process

1. State 0: Branch Governance, Memory, Target AGENTS.md & Triage Gate

• Git Branch Governance: Check git symbolic-ref --short HEAD; auto-branch feat/<descriptive-kebab-slug> or fix/<slug> from base branches; prompt via ask_question on Scope Drift.
• Feature Directory & Memory: Ensure docs/<feature-name>/ exists (syncing docs/specifications/ if present). Silently read context/memory/MEMORY.md (~500 tokens); read topic pages only on demand.
• Target AGENTS.md Check: Create or update root AGENTS.md (<= 90 lines) if missing or if stack/commands/conventions changed, linking commands, DESIGN.md, and context/memory/MEMORY.md.
• Author PLAN.md (<= 50 lines): Classify Tier 1/2/3, fill the 5-row Invoke / SKIP triage table, and define Parallel Execution Tracks.

2. State 1 & 1.5: Capability-Gated Discovery & Interactive User Alignment Gate

• State 1: Skip for Tier 1 or when all Discovery specialists are SKIP. Otherwise dispatch ONLY INVOKE specialists (product-analyst, ui-ux-designer, architect-engineer) sequentially or in parallel within strict line budgets.
• State 1.5 (Interactive User Alignment Gate): Auto-proceed on Tier 1. On Tier 2 & Tier 3, PAUSE and call ask_question presenting scope, UI design, and architectural summary ("Do you approve the product scope, UI design, and architectural decisions before proceeding to code implementation?"). Wait for user sign-off.

3. State 2: Parallel Build, Collision Analysis & Scoped Test Isolation

• Collision Analysis: Verify non-overlapping Owned Files in Mode A — Disjoint File Partitioning (Workspace: 'inherit'). Enforce Mode B — Git Worktree Isolation (Workspace: 'share' / 'branch' or git worktree) whenever tracks share files (package.json, routers, DI, settings.py).
• Scoped Test Isolation: Parallel code-implementer instances run ONLY their assigned test file during TDD loops; full suite execution waits for State 3.

4. State 3 & 4: Tiered & Risk-Gated Verification

• Tier 1: Scoped qa-tester (exit code 0) + compliance-verifier (APPROVED).
• Tier 2: Parallel qa-tester (docs/<feature-name>/test-matrix.md <= 80 lines + exit code 0) + compliance-verifier (APPROVED) (+ e2e-blackbox-tester ONLY if web UI flows changed).
• Tier 3 (Adversarial Verification Committee): Parallel qa-tester (test-matrix.md <= 80) + compliance-verifier (APPROVED) + risk/UI-gated critic, challenger, and e2e-blackbox-tester (<= 50 lines each).
• Bounded Retries & Fail-Honest E2E: Max 3 fix loops to code-implementer. Always surface BLOCKED/SKIPPED browser E2E honestly in the final report.

5. State 5: Ship, AGENTS.md Sync & LLM Wiki Memory Update

• Mark tasks complete in PLAN.md, sync root AGENTS.md (<= 90 lines) if needed, and update context/memory/ (MEMORY.md <= 80-100 lines, topic pages if changed, 5-8 line log entry in context/memory/log/YYYY-MM-DD.md). Never write checkpoints on intermediate turns.

Usage

• Antigravity: Dispatch subagents via invoke_subagent (Subagents=[...] for parallel fan-out), follow up via send_message / manage_subagents, and gate alignment/scope drift via ask_question.
• Claude Code: Emit single or parallel Agent(agent="...", prompt="...") calls (using loop-engineering worktrees for Mode B).

Red Flags

• Orchestrator editing application source or test files directly.
• Over-invoking specialists whose capability/risk gate evaluates to SKIP (e.g., running e2e-blackbox-tester on pure backend/CLI tasks, or generating prd.md/architecture.md when already unambiguous/established).
• Masking BLOCKED or SKIPPED browser E2E as passed instead of explicitly reporting "Browser E2E was NOT executed (BLOCKED/SKIPPED): <exact reason>".
• Exceeding artifact line budgets or skipping State 0, ask_question alignment on Tier 2/3, Mode B worktree isolation on shared files, or scoped test isolation.

Verification

• State 0: Git branch verified (feat/<slug> auto-branched; Scope Drift checked); docs/<feature-name>/ ensured; context/memory/MEMORY.md read; root AGENTS.md (<= 90 lines) checked; PLAN.md (<= 50 lines) has 5-row Invoke / SKIP triage.
• State 1 & 1.5: Only capability-triggered discovery specialists ran (within line budgets); Tier 2/3 ask_question approval received.
• State 2: Collision analysis enforced Mode A or Mode B; parallel code-implementer subagents ran scoped tests only.
• State 3 & 4: qa-tester (exit code 0) and compliance-verifier (APPROVED) passed; risk-gated committee reports (<= 50 lines) verified; BLOCKED/SKIPPED E2E disclosed honestly.
• State 5: PLAN.md, root AGENTS.md (if changed), and context/memory/ updated once at completion.