Problem
The install-framework.sh script uses cp -r .architecture/.architecture/.architecture/* .architecture/ which blindly copies the entire .architecture/ directory from the cloned repo — including 40+ framework-internal files (11 ADRs, 10 reviews, 8 comparisons, research docs, a 50KB deferrals.md, methodology documents). These are the framework's own work product and should never appear in a target project.
The MCP server (mcp/index.js) takes a completely different approach — generating everything programmatically with hardcoded templates. Neither approach is correct: blind copy installs cruft, hardcoded generation creates duplication.
Additionally, the triple-nesting clone path (.architecture/.architecture/.architecture/) is confusing and error-prone.
Approach
Manifest-based selective installation:
- Clone repo to
/tmp (not nested inside .architecture/) with --depth 1
- Copy only files listed in a manifest (
.install-manifest in repo root)
- Create empty directory structure for project-specific content
- Initialise config from template
- Let the skill's interpretive steps handle content generation (members.yml, principles.md, initial analysis)
What gets copied from repo (templates + agent docs):
templates/ directory (9 files: ADR template, review template, config template, etc.)
agent_docs/ directory (3 files: progressive disclosure reference docs)
What gets created as empty directories:
decisions/adrs/, reviews/, recalibration/, comparisons/
What the skill generates (not the script):
members.yml — from scratch based on project analysis
principles.md — from scratch based on detected tech stack
config.yml — initialised from templates/config.yml
reviews/initial-system-analysis.md — from skill's analysis step
Skill instruction changes:
- Steps 3 & 4: change from "customize" to "create from scratch"
- Remove assumption that members.yml/principles.md exist before the skill runs
Acceptance Criteria
Context
Problem
The
install-framework.shscript usescp -r .architecture/.architecture/.architecture/* .architecture/which blindly copies the entire.architecture/directory from the cloned repo — including 40+ framework-internal files (11 ADRs, 10 reviews, 8 comparisons, research docs, a 50KB deferrals.md, methodology documents). These are the framework's own work product and should never appear in a target project.The MCP server (
mcp/index.js) takes a completely different approach — generating everything programmatically with hardcoded templates. Neither approach is correct: blind copy installs cruft, hardcoded generation creates duplication.Additionally, the triple-nesting clone path (
.architecture/.architecture/.architecture/) is confusing and error-prone.Approach
Manifest-based selective installation:
/tmp(not nested inside.architecture/) with--depth 1.install-manifestin repo root)What gets copied from repo (templates + agent docs):
templates/directory (9 files: ADR template, review template, config template, etc.)agent_docs/directory (3 files: progressive disclosure reference docs)What gets created as empty directories:
decisions/adrs/,reviews/,recalibration/,comparisons/What the skill generates (not the script):
members.yml— from scratch based on project analysisprinciples.md— from scratch based on detected tech stackconfig.yml— initialised fromtemplates/config.ymlreviews/initial-system-analysis.md— from skill's analysis stepSkill instruction changes:
Acceptance Criteria
/tmp, not nested inside.architecture/.install-manifestchecked into repo listing exactly what to copy/createContext
mcp/index.jssetupArchitecture()) proves the concept: new projects don't need framework files.claude/plans/20260210-direction.md