harness-bootstrap, end to end

purple = the model, billed green = script, free blue = you red = guardrail

.claude/ empty

agents?rules?guardrails?tasks?

Mode first - and it is not about size

greenfield empty or near-empty repo

brownfield source exists, agents will modify it

audit agents analyze, a human applies every fix

PURPLE - a model is required

Codebase analysis - MANDATORY stack, modules, data layer, integrations, conventions, risky ops, existing .claude/, git reality

→

Inventory Report

modules → dev agents conventions → rules risky ops → deny + hooks

↓

You confirm or correct it no file is generated before it is confirmed

PURPLE - a model is required

Intake questionnaire AskUserQuestion, max 4 per call. Brownfield pre-fills from evidence and asks only what the code cannot decide

→

Detect + confirm tools scan .cursor/ .codex/ AGENTS.md, then ask: Claude Code / Cursor / Codex

→

One-screen setup plan created / kept / modified, roster with model + effort. You confirm

PURPLE - a model is required

Roster Tier 0 unconditional, preset S / M / L, explicit model: AND effort:

→

Skills: match, review, wire matched to the stack found, reviewed before suggested, you choose - then it's wired in

→

Detect the dev OS Windows to .ps1, macOS or Linux to .sh

→

vars.json vars + flags: ui, db, ai, audit, tdd, ddd, deploy_ask, windows / posix

GREEN - a script, deterministic and free

scaffold.py --dry-run →

scaffold.py a deterministic copy of assets/

ADDEDfile did not exist

KEPTexists, byte-identical

CONFLICTexists and differs - not written

↓

Reconcile by hand: keep / adapt / add / flag exits 0

GREEN - a script, deterministic and free

scaffold.py --dry-run → scaffold.py

unresolved {{ VAR }} → exit 1

a placeholder shipped into a live rule file matches nothing and fails silently

PURPLE again - what no template can know

Model-authored gap-fill orchestrator routing table, dev-agent scopes, tech-stack.md, coding-standards.md, git-workflow.md, settings.json allow / deny for the real stack, .env.example

→

Orchestration wiring master-plan index table, Phase 1 seeded from the analysis gap list, AGENTS.md entry point, CLAUDE.md as a thin @AGENTS.md import

Quality gate structure / cost / safety / grounding / handoff

→ all green

Smoke-test the loop →

Build both graphs code-graph.py + docs-graph.py, then graph-html.py: harness-graph.html + specs-graph.html

→ Port to your tools

↓ Harness runs under orchestration

Read it as three bands

analyse + decidea model is required

→

scaffolda script, for free

→

gap-fill + wirewhat no template knows

AGENT HARNESS BOOTSTRAP