# Golden Path

Caveman-full. Internal doc.

## What golden path is

Golden path = opinionated, supported, paved way from idea to production. Term from Spotify
(Netflix calls it "paved road"). Not the *only* way, the *recommended* way. Five properties:

- **Opinionated**: one pipeline, repo conventions. No bikeshedding the flow.
- **Supported**: sensors + evals gate every stage. Harness catches drift, not you.
- **Optional**: step off any time, run stages solo. No enforcement, no punishment.
- **Self-serviceable**: one command. No tickets, no waiting on a platform team.
- **Transparent / extensible**: every stage names what ran (sensors, guides, evals).
  Override per repo via `conventions/`. Abstraction you can see through and bend.

## harness-kit golden path

One command. Idea → merged PR. Six gated stages:

```
/golden-path
```

Runs `/product-manager:run` (prd → prp) then `/sse:run` (plan → dev → test → pr → monitor).
Same six-stage `full-run` pipeline the status bar already tracks.

**Start from a brief.** The idea-brief builder collects the inputs the PRD needs (squad,
problem, hypothesis, customers, metric) in-standard and emits a paste-ready `/golden-path`
kick: [`docs/brief/`](./brief/) (GitHub Pages: `https://pierry.github.io/harness-kit/brief/`).

```mermaid
flowchart LR
    idea([idea]) --> prd[prd]
    prd --> prp[prp]
    prp --> plan[plan]
    plan --> dev[dev]
    dev --> test[test]
    test --> pr[pr]
    pr --> merged([merged PR])

    subgraph PM["/product-manager:run"]
        prd
        prp
    end
    subgraph SSE["/sse:run"]
        plan
        dev
        test
        pr
    end
```

Each stage: writes artifact → deterministic sensors fire → LLM-judge eval (retry up to 3) →
approval marker → token accounting. See `AGENTS.md` § Pipeline stages.

Flag passthrough to the SSE half:

- `--local`, stop after test, no PR. Dev + test locally.
- `--sdd`, spec-driven loop variant (plan once, dev↔test↔eval until PRP spec met). Local only.
- `--no-monitor`, PR opens, skip auto merge-watch.

## Stepping off the path

Optional property in practice. Golden path is the paved lane, not a fence. Run any stage solo:

| Detour | Command |
|---|---|
| Just the PRD | `/product-manager:prd` |
| Just the PRP | `/product-manager:prp` |
| Just the plan | `/sse:plan` |
| Implement only | `/sse:dev` |
| Tests only | `/sse:test` |
| Open PR only | `/sse:pr` |
| Dev + test, no PR | `/sse:run --local` |
| Spec-driven loop | `/sse:sdd` |
| Deploy a static site | `/sse:firebase-publish` |
| Resume where left off | `/pipeline:continue` |
| Abandon active run | `/pipeline:reset` |

Same sensors, same evals, same artifacts. You lose the one-command convenience, not the support.

## Per-discipline paving

Golden path canon: one path per engineering discipline. harness-kit does this in the dev stage,
auto-selects the repo's discipline conventions:

- `.claude/conventions/backend.md`
- `.claude/conventions/web.md`
- `.claude/conventions/mobile.md`
- `.claude/conventions/devops.md`

When a file exists, project wins over SSE defaults. These overrides *are* the per-discipline
paving. See `.claude/agents/staff-software-engineer/guides/conventions-override.md`.

## What runs behind the curtain

Transparency property. Golden path abstracts, never hides. Every stage summary names the
sensors, evals, guides, and refs that ran, not generic "done", actual names. Example from a
plan stage:

```
sensors: plan-structure ok, plan-feasibility ok
eval:    plan-quality 8/10 (attempts: 1)
guides:  pipeline.md, coding-style.md, skills/{area}/SKILL.md
refs:    prp/{feature_id}.md, conventions/{area}.md
```

Want to see deeper? Read the `sensors/`, `evals/`, `guides/` under
`.claude/agents/<agent>/`. Plain markdown. Nothing sealed.

## See also

- `AGENTS.md`, registry, routing, pipeline-stages spec
- `.claude/commands/golden-path.md`, the command
- `.claude/agents/staff-software-engineer/guides/sdd-loop.md`, spec-driven variant