Nw DeliverSAFE

AI agents that guide you from idea to working code, with you in control at every step.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: nw-deliver
description: "Orchestrates the full DELIVER wave end-to-end (roadmap > execute-all > finalize). Use when all prior waves are complete and the feature is ready for implementation."
user-invocable: true
argument-hint: '[feature-description] - Example: "Implement user authentication with JWT"'
---

# NW-DELIVER: Complete DELIVER Wave Orchestrator

**Wave**: DELIVER (wave 6 of 6)|**Agent**: Main Instance (orchestrator)|**Command**: `/nw-deliver "{feature-description}"`

## LANGUAGE CONVENTION FRAME (read FIRST — overrides all examples below)

**Code examples in this skill use Python syntax for illustration only.** They are NOT prescriptive about target language. nWave is language-agnostic per the "genericity and agnosticism" mandate (2026-05-24).

**Before crafting**, detect the target project's language from manifest files: `package.json` → TypeScript/JS; `Cargo.toml` → Rust; `go.mod` → Go; `pyproject.toml`/`setup.py`/`Pipfile` → Python; `pom.xml`/`build.gradle` → Java/Kotlin; `*.csproj`/`*.fsproj` → C#/F#; `Gemfile` → Ruby; `Package.swift` → Swift.

**When the target language is NOT Python**: adapt every code example to target conventions (imports, type system, test-framework idioms, file extensions, directory layout). Project conventions ALWAYS WIN over examples below.

**Empirical anchor**: skill examples being Python-only caused LLM to emit Python code in greenfield TS project. Fix per F-SKILL-EXAMPLES-LANGUAGE-LEAK. Connects [[feedback_language_adapter_plugin_architecture_2026_05_24]].

## Overview

Orchestrates complete DELIVER wave: feature description → production-ready code with mandatory quality gates. You (main Claude instance) coordinate by delegating to specialized agents via Task tool. Final wave (DISCOVER > DISCUSS > SPIKE > DESIGN > DEVOPS > DISTILL > DELIVER).

Sub-agents cannot use Skill tool or `/nw-*` commands. You MUST:
- Read the relevant command file and embed instructions in the Task prompt
- Remind the crafter to load its skills as needed for the task (skill files are at `~/.claude/skills/nw-{skill-name}/SKILL.md`)

## Output Tiers (per D2)

Provenance: feature `lean-wave-documentation` — D2 (schema-typed sections), D10 (one-line expansion descriptions). Tier-1 [REF] sections (always emitted) + Tier-2 EXPANSION CATALOG items (lazy, on-demand) are the two output bands. Implementation details live in code; the wave-delta sections are pointers + structured summaries. Full contract: `nWave/skills/nw-density-resolution-contract/SKILL.md`.

### Tier-1 [REF] — always emitted

Under `## Wave: DELIVER / [REF] <Section>` headings:

- Implementation summary — one-paragraph description of what shipped (no design rationale)
- Files modified — categorized list (production, tests, docs) with one-line per file
- Scenarios green count — `<N> of <M>` from the `.feature` file with timestamp
- DoD check — itemized pass/fail against the DISCUSS Definition of Done items
- Demo evidence — captured stdout/exit-code per Elevator Pitch demo command (Phase 3.5 gate)
- Quality gates — per-phase outcomes (refactor, review, mutation, integrity)
- Pre-requisites — DISTILL scenarios + DESIGN component manifest the implementation depended on

### Tier-2 EXPANSION CATALOG — lazy, on-demand (per D10)

Rendered under `## Wave: DELIVER / [WHY|HOW] <Section>` only when requested via `--expand <id>` (DDD-2), the broad wave-end menu (`expansion_prompt = "ask"`), a declared trigger under `expansion_prompt = "ask-intelligent"`, `mode = "full"` auto-expansion, or an ad-hoc user request mid-session.

| Expansion ID | Tier label | One-line description |
|---|---|---|
| `refactoring-journal` | [HOW] | L1-L6 refactoring log with rationale per micro-transformation |
| `retrospective-notes` | [WHY] | 5 Whys analysis on issues encountered, lessons learned, what to repeat/avoid |
| `performance-measurements` | [WHY] | Benchmarks, profiling output, latency/memory deltas vs baseline |
| `alternative-implementations-rejected` | [WHY] | Implementation approaches tried and rejected with one-paragraph reason each |
| `mutation-testing-report` | [HOW] | Mutmut/Pitest output: kill rate, surviving mutants, mitigation actions |
| `architecture-decision-deviations` | [WHY] | Where DELIVER deviated from DESIGN and the back-propagation logged in upstream-issues.md |
| `coverage-deltas` | [HOW] | Per-module coverage delta with rationale for any drops |
| `expansion-catalog-rationale` | [WHY] | Why this set of expansions, why these defaults, why D10 enforces one-line descriptions |

## Density resolution (per D12)

Call `resolve_density(global_config)` from `scripts/shared/density_config.py` after reading `~/.nwave/global-config.json` (missing/malformed = empty dict). Returns `mode` (`"lean"` | `"full"`) + `expansion_prompt` (`"ask"` | `"ask-intelligent"` | `"always-skip"` | `"always-expand"` | `"smart"`) per the D12 cascade (resolver-internal, DDD-5 — do NOT replicate locally). Branch on `density.mode` for what to emit; branch on `density.expansion_prompt` at wave end for menu behaviour. DELIVER declares no `ask-intelligent` triggers: emit no menu and the shared-contract no-trigger skip event; do not invent triggers. Full cascade detail, branch semantics, ad-hoc override workflow: `nWave/skills/nw-density-resolution-contract/SKILL.md`.

## Telemetry (per D4 + DDD-6)

Every expansion choice emits a `DocumentationDensityEvent` (dataclass at `src/des/domain/telemetry/documentation_density_event.py`) via `event.to_audit_event()` → `JsonlAuditLogWriter().log_event(...)`. Schema fields per D4: `feature_id`, `wave`, `expansion_id`, `choice`, `timestamp`. For this wave the schema declares `"wave": "DELIVER"`. Use helper `scripts/shared/telemetry.py:write_density_event(...)` — do NOT write JSONL directly.

Wave-specific signal: a DELIVER wave recording `choice = "expand"` for `retrospective-notes` indicates the team needed deeper learning capture; over time the data drives whether retrospective notes should be promote

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.