Tailored Build - nguyenhx2/agent-harness-bootstrap GitHub Wiki

Tailored Build

English · 日本語

spec-builder ends with a contract. harness-bootstrap begins with a codebase. The step between them is where the team gets decided, and most agent kits do not have that step at all.

The problem it answers

The usual agent kit ships everything - every agent, a hundred skills, every rule and hook - installed before anyone has read the codebase. That costs more than it looks:

You pay context for all of it, in every session. A rule that matches no file in your repo is a tax on every request.
Seats end up with no owner. An agent nobody routes work to is a name that makes the routing table harder to read.
The advice goes generic. Guidance true in any repository is rarely specific enough to act on in yours.
Nothing is visibly load-bearing. When everything is installed, nobody can safely remove anything.

Completeness gets mistaken for fit. Only fit is worth paying for.

What decides what

The roster comes from evidence and nothing else: what the contract says must be built, what the codebase analysis found, and the answers only you can give.

flowchart LR
    C[/"Codebase analysis"/]
    S[/"The contract"/]
    A["Your answers"]
    B["Roster and flags"]
    SE[/"Seats"/]
    RU[/"Path-scoped rules"/]
    SK[/"Skill shortlist"/]

    C --> B
    S --> B
    A --> B
    B --> SE
    B --> RU
    B --> SK

    classDef mod fill:#5A189A,stroke:#240046,stroke-width:1px,color:#E0AAFF
    classDef hum fill:#1D3557,stroke:#0B1B2B,stroke-width:1px,color:#A8DADC
    classDef art fill:#495057,stroke:#212529,stroke-width:1px,color:#F8F9FA
    class B mod
    class A hum
    class C,S,SE,RU,SK art

Decided from	What it decides
The modules that actually exist	One dev agent each, scoped to real paths. No module, no seat.
The repo's purpose	Code, research, strategy or docs, in any combination. A research repo gets `researcher` and `checker`, a docs repo `writer` and `checker`, and none of them gets a test suite or a code review gate.
The contract and your answers	Which of the 18 seats are filled. A run installs 3 to 17 of the 18 seats, never all of them by default.
The manifests in your repo	Which skills are proposed. You choose from that shortlist.
The paths that exist	Which rules are path-scoped, which keeps most rule content out of the default session.

The numbers are measured, not claimed

The leanest answers give 3 seats (a documentation or research repo). A code repo with databases, tests and a long-lived project gives 17.

scripts/check_numbers.py computes that range with the scaffolder's own selection function, over every combination of the flags that gate an agent, so it cannot drift from a real run. Change which agents are gated and every place that quotes the range, this page included, fails until it is updated.

Discovery and wiring are two problems

flowchart TD
    M[/"Your manifests"/]
    P["Propose a shortlist"]
    U{"You pick"}
    N["/new-skill"]
    W["/skill-wire to a seat"]
    G[/"Graph node"/]

    M --> P --> U
    U -->|"a fit"| W
    U -->|"nothing fits"| N --> W
    W --> G

    classDef det fill:#2D6A4F,stroke:#081C15,stroke-width:1px,color:#D8F3DC
    classDef mod fill:#5A189A,stroke:#240046,stroke-width:1px,color:#E0AAFF
    classDef hum fill:#1D3557,stroke:#0B1B2B,stroke-width:1px,color:#A8DADC
    classDef art fill:#495057,stroke:#212529,stroke-width:1px,color:#F8F9FA
    class P,N,W mod
    class U hum
    class M,G art

Skill Discovery - you cannot choose from a catalogue you have never read. It reads this project's manifests (Python, Ruby, .NET, Java and Kotlin, JavaScript, monorepo markers), works out the stack, and only then proposes. Marketplaces are a source, not an authority. Nothing installs unless you pick it.
/new-skill - when nothing fits a recurring workflow, it first decides whether the workflow earns a project skill at all (refusing is normal), then scaffolds one and lints it with skill-check.py.
/skill-wire - installed is not reachable. It connects a skill to the seat or command that will use it and records that as a node in the graph, not a convention in a document.

Because the wire is a node, harness-view shows which skills are wired and which merely sit on disk, and assess reports the difference as a finding.

It is checkable

A seat nobody owns, a skill nobody is told to use and a rule that matches no file are all costs. harness-view assess names all three:

harness-view assess .

"Tailored" is a word anyone can put in a README. Three to seventeen of eighteen, and a finding when a seat has no module is something you can check.