feat(harness): engineering agent pipeline — ship-feature, planner, engineer, ship-orchestrator

[siddhant3030]

Summary

• Adds a full engineering agent harness with two pipeline modes: command (synchronous, in-session) and agent (async, fresh-context via ship-orchestrator)
• Introduces three new agents: planner, engineer, ship-orchestrator
• Both pipelines chain: design gate → plan → implement → validate loop → design review → docs → PR
• Agents use landmarks.md for exact file:line lookups before any codebase exploration — eliminates redundant searching
• Context isolation: orchestrator reads only pipeline.md between stages; each stage runs in a fresh sub-agent context
• Adds docs/harness-evolution.md: 5-iteration improvement roadmap + command vs agent comparison framework

What changed

New agents

Agent	Role	Model
planner	Blast radius analysis + produces plan.md + research.md	Sonnet
engineer	Implements all plan milestones, milestone-scoped context reads	Opus
ship-orchestrator	Async pipeline orchestrator (agent mode)	Sonnet

New commands

• /engineering/ship-feature — synchronous, in-session orchestration
• /engineering/ship-feature-bg — thin launcher that spawns ship-orchestrator in a fresh context

Architecture reference

• backend-architecture/landmarks.md — exact file:line for auth, models, API modules, test fixtures, migration patterns
• frontend-architecture/landmarks.md — exact file:line for sidebar, auth hook, routing tree, API client, component library

Context isolation design

Orchestrator (reads only pipeline.md)
  ├── spawns: planner    → fresh context, returns plan.md
  ├── spawns: engineer   → fresh context, reads plan milestone by milestone
  ├── spawns: validator  → fresh context per retry
  ├── spawns: reviewer   → fresh context, reads diff + design artifacts
  └── spawns: docs agent → fresh context, reads spec + plan

Comparison experiment

Both modes write Mode: command / Mode: agent to pipeline.md. See docs/harness-evolution.md → Experiment 0 for the comparison framework and metrics to capture.

Test plan

• Restart Claude Code session to pick up new agent registrations
• Run /engineering/ship-feature-bg features/access-control/v1/... — verify ship-orchestrator spawns correctly
• Verify design gate fires for UI features without design.md
• Verify planner agent produces plan.md with all 9 sections including Decision Log
• Compare command vs agent mode outputs on equivalent features

🤖 Generated with Claude Code