Docs & To Do

Documentation and feature status for Curator.

Scan Pipeline
The full scan pipeline runs when you click Scan on a repo

Pipeline Stages

  1. Discover — Clone repo, find .md files, apply platform + repo exclusions, skip >500KB
  2. Classify — Gemini splits each .md by headings, assigns context_type per section, flags redundant content (cached by content hash)
  3. Validate Docs — Gemini checks doc sections against actual code; verdicts: confirmed, corrected, or uncertain
  4. Code Generators — Run all code-based generators (stack, gotchas, content-model, custom-modules, entities, etc.)
  5. Validate Exceptions — Check for classification conflicts and duplicates
  6. Assemble — Merge doc sections into code-generated content under "From Project Documentation" markers
  7. QA Agent — Check coverage gaps, token budgets, staleness, gitignore hygiene
  8. Write-back — Commit .curator/*.md + .claude/skills/curator/SKILL.md to GitHub
Context Files
Context Hub generates different context sets based on platform type
Drupal 8+ repos
STACK / PROJECT / GOTCHASDrupal/PHP versions, modules, hosting, known issues
CONTENT_MODEL / PARAGRAPHS / ENTITIES / VIEWS / FORMSContent types, paragraph components, custom entities, views, forms
CUSTOM_MODULES / INTEGRATIONSCustom hooks, services, routes, plugins, third-party APIs
THEME_*_overview / templates / scss / js / components / visualPer-theme analysis (6 files per theme)
JIRA_SOFTWARE_* / JIRA_SERVICEDESK_*Historic + recent learnings from Jira
CONFLUENCE / VISUAL_CONTEXTConfluence docs, computed CSS from live site
Drupal 7 repos
STACK / GOTCHAS / CUSTOM_MODULES / ENTITIESD7-specific: .info INI parsing, sites/all/modules/, Features module, ctools plugins
CONTENT_MODELFrom Features exports and .install files (limited — D7 stores config in DB)

No Paragraphs for D7. Same external sources (Jira, Confluence) as D8+.

Adaptrix / Other repos
OVERVIEW / ARCHITECTURE / DATA_MODEL / APIRuntime, structure, schemas, endpoints
DEPENDENCIES / CONFIG / FRONTEND / SERVICES / TESTINGTooling, env, UI, jobs, test coverage
DECISIONS / DEPLOYMENT / ONBOARDING / GENERAL_DOCSFrom discovered .md files in the repo
Combined output
AGENTS.mdAll context concatenated into a single file
.curator/*.mdIndividual context files committed to repo
.claude/skills/curator/SKILL.mdClaude Code skill file (MCP-first or files-only mode)
MCP Server
AI assistants query Curator for targeted context via MCP

Transports

HTTP — /api/mcp for Cursor and similar IDEs
stdio — curator-mcp CLI for Claude Desktop

Tools

get_context — Query with github_repo + natural language. Claude selects 4-10 relevant files.
check_context — Check freshness (stale vs up-to-date)
list_available_context — Show all context files with status
list_repos — List available repositories

Authentication

X-API-Key or Authorization: Bearer using the ACP_API_KEY env var.

Repo Write-back
Context files are committed directly to GitHub repos after each scan

After a successful scan, Curator commits generated context files to the repo:

.curator/*.md — Individual context files (stack.md, gotchas.md, etc.)
.claude/skills/curator/SKILL.md — Claude Code skill file

Files are SHA-compared to avoid unnecessary commits. Requires a GitHub token with write access.

Features

✓
Theme Context
Extract SASS variables, Twig templates, build tasks, and library definitions from themes (6 files per theme)
✓
Stack Context
Drupal/PHP versions, enabled modules by category, hosting platform detection
✓
Visual Context
Capture computed CSS styles from live site URLs for design consistency
✓
Repo Write-back
Auto-commit .curator/*.md context files and .claude/skills/curator/SKILL.md to GitHub repos
✓
Scan Pipeline
Full automated pipeline: discover .md files, classify with Gemini, validate against code, run code generators, assemble, combine, write-back, QA
✓
Doc Discovery & Classification
Discover existing .md files in repos, classify sections by context type using Gemini, validate accuracy against code, merge into generated context
✓
Drupal 7 Support
Full platform type with D7-specific generators: .info INI parsing, sites/all/modules/ paths, Features module, hook detection, ctools plugins
✓
MCP Server
HTTP + stdio transports for AI IDEs. Claude-powered query routing selects 4-10 relevant context files per query
✓
Claude Code Skills
Generate .claude/skills/curator/SKILL.md with MCP-first or files-only mode for Claude Code integration
✓
QA Agent
Post-scan quality checks: coverage gaps, token budgets, staleness detection, gitignore hygiene
11
Scheduled Regeneration
Automatic regeneration on webhook push events or scheduled intervals
12
Bulk Operations
Generate/refresh all context files for all repos in one action

10 done · 2 pending