Files
scarf/CLAUDE.md
T
Alan Wizemann bb750e237e docs: CLAUDE.md — add Template Configuration section
Documents the v2.3 configuration feature for future agent sessions:
manifest schemaVersion 2 shape, supported field types, Keychain storage
conventions (service/account naming with project-path hash suffix), the
uninstaller's config-items cleanup path, exporter behaviour (schema
forwarded, values stripped), and the catalog site's schema display.

Includes the "Schema is Swift-primary" note so future edits to
TemplateConfigField.FieldType go through the right order of updates —
Swift first, then Python mirror, then widgets.js, then UI controls,
then tests on both sides. Schema drift between Swift + Python
validator would accept bundles the app later refuses at install
time, which is a catastrophic UX failure for the catalog.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-23 17:14:29 +02:00

15 KiB

Scarf — macOS GUI for the Hermes AI Agent

Project Structure

scarf/scarf/           Xcode project root (PBXFileSystemSynchronizedRootGroup — auto-discovers files)
  scarf/               Main app target source
    Core/Services/     HermesDataService, HermesFileService, HermesLogService, ACPClient, HermesFileWatcher
    Core/Models/       Plain structs: HermesSession, HermesMessage, HermesConfig, etc.
    Features/          MVVM-F feature modules (Dashboard, Sessions, Activity, Chat, Memory, Skills, Cron, Logs, Settings)
    Navigation/        AppCoordinator, SidebarView
  docs/                PRD, Architecture, Discovery notes
  standards/           Copied development standards (read-only reference)

Architecture Rules

  • MVVM-F: Features never import sibling features. Cross-feature goes through services.
  • AppCoordinator: Single @Observable coordinator for all navigation state, injected via .environment().
  • No external dependencies: System SQLite3, Foundation JSON, AttributedString markdown.
  • Read-only DB access: Never write to ~/.hermes/state.db. Only write to memory files and cron jobs.
  • Sandbox disabled: App reads ~/.hermes/ directly.
  • Swift 6 concurrency: @MainActor default. Services use nonisolated + async/await.

Key Paths

  • Hermes home: ~/.hermes/
  • SQLite DB: ~/.hermes/state.db (WAL mode, read-only)
  • Config: ~/.hermes/config.yaml
  • Memory: ~/.hermes/memories/MEMORY.md, ~/.hermes/memories/USER.md
  • Sessions: ~/.hermes/sessions/session_*.json
  • Cron: ~/.hermes/cron/jobs.json
  • Logs: ~/.hermes/logs/errors.log, ~/.hermes/logs/gateway.log
  • ACP: hermes acp subprocess (stdio JSON-RPC)

Build

xcodebuild -project scarf/scarf.xcodeproj -scheme scarf -configuration Debug build

Releases

Shipped via a single local script. Never run manual xcodebuild archive / notarytool / gh release create steps — use the script so nothing is skipped or misordered.

./scripts/release.sh <version>           # full release: notarize → appcast → gh-pages → tag
./scripts/release.sh <version> --draft   # draft: everything builds + notarizes, but appcast/tag are skipped

The script bumps version, archives Universal (arm64 + x86_64) + ARM64-only variants, signs with Developer ID, notarizes via xcrun notarytool (keychain profile scarf-notary), staples, EdDSA-signs the appcast entry with Sparkle's key, pushes the appcast to gh-pages, and creates a GitHub release with both zips attached. Draft mode stops after the release is uploaded so the current version stays "latest" until explicitly promoted.

Release notes convention: write them to releases/v<version>/RELEASE_NOTES.md BEFORE running the script — it's auto-included in the version-bump commit and used as the GitHub release body. If absent, a placeholder is used.

Canonical prompts (any of these trigger the flow):

  • "Release v1.6.2" — full release
  • "Release v1.6.2 as draft" — draft mode
  • "Prepare v1.6.2 release notes from recent commits, then release" — generate notes first, then run

Prerequisites (one-time, already set up on Alan's machine): Developer ID Application cert in login Keychain (team 3Q6X2L86C4), notarytool keychain profile scarf-notary, Sparkle EdDSA private key in Keychain item https://sparkle-project.org, gh-pages branch + GitHub Pages enabled. See the header of scripts/release.sh and the Releases section in README.md for details.

Wiki

Public documentation lives in the GitHub wiki at https://github.com/awizemann/scarf/wiki. The wiki is a separate git repo cloned to .wiki-worktree/ in the repo root (gitignored, sibling to .gh-pages-worktree/). Internal dev notes stay in scarf/docs/; the wiki is for public-facing reference.

Update the wiki when:

  • A new feature module is added under scarf/scarf/scarf/Features/ → extend the relevant User Guide page.
  • A new core service is added under Core/Services/ → extend Core-Services.md.
  • Architecture changes (AppCoordinator, transport, MVVM-F rule, sandbox) → Architecture-Overview.md + the specific sub-page.
  • Hermes version bumps in this file → Hermes-Version-Compatibility.md.
  • scripts/release.sh completes a full (non-draft) release → bump latest-version on Home.md + append to Release-Notes-Index.md.
  • Keyboard shortcut or sidebar section changes → Keyboard-Shortcuts.md / Sidebar-and-Navigation.md.

Skip for: bug fixes with no user-observable change, pure refactors, typos, test-only changes, internal cleanups.

./scripts/wiki.sh pull                         # always first
# edit .wiki-worktree/*.md with normal tools
./scripts/wiki.sh commit "docs: describe X"   # runs secret-scan
./scripts/wiki.sh push                         # runs secret-scan again, then push

Never commit API keys, tokens, .env files, private keys, or real hostnames/IPs to the wiki. The script's two-pass secret-scan blocks common token patterns and a user-maintained blocklist at scripts/wiki-blocklist.txt (gitignored). Do not bypass without explicit approval. Full workflow on the wiki itself at .wiki-worktree/Wiki-Maintenance.md.

Hermes Version

Targets Hermes v0.9.0 (v2026.4.13). Log lines may carry an optional [session_id] tag between the level and logger name — HermesLogService.parseLine treats the session tag as an optional capture group, so older untagged lines still parse.

Project Templates

Scarf ships a .scarftemplate format (v1 as of 2.2.0) for sharing pre-packaged projects across users and machines. A bundle is a zip containing:

  • template.json — manifest (id, name, version, contents claim)
  • README.md — shown in the install preview sheet
  • AGENTS.md — required; the Linux Foundation cross-agent instructions standard — every template is agent-portable out of the box
  • dashboard.json — copied to <project>/.scarf/dashboard.json
  • instructions/… — optional per-agent shims (CLAUDE.md, GEMINI.md, .cursorrules, .github/copilot-instructions.md)
  • skills/<name>/… — optional; installed to ~/.hermes/skills/templates/<slug>/ (namespaced so uninstall is rm -rf on one folder)
  • cron/jobs.json — optional; registered via hermes cron create with a [tmpl:<id>] … name prefix and immediately paused
  • memory/append.md — optional; appended to ~/.hermes/memories/MEMORY.md between <!-- scarf-template:<id>:begin/end --> markers

Key services: ProjectTemplateService.swift (inspect + validate + plan), ProjectTemplateInstaller.swift (execute a plan), ProjectTemplateExporter.swift (build a bundle from a project), ProjectTemplateUninstaller.swift (reverse an install using the lock file). UI in Features/Templates/. The scarf://install?url=<https URL> deep link + file:// URLs for .scarftemplate files are handled by TemplateURLRouter.swift and onOpenURL in scarfApp.swift. A <project>/.scarf/template.lock.json uninstall manifest is written after every install and drives the uninstall flow.

Uninstall semantics: driven by the lock file. Only files listed in lock.projectFiles are removed from the project dir; user-added files (e.g. a sites.txt created on first run) are preserved. If every file in the dir was installed by the template, the dir is removed too; otherwise the dir stays with just the user's files. Skills namespace is always removed wholesale (it's isolated). Cron jobs are removed via hermes cron remove <id> after resolving each lock-recorded name. Memory block is stripped between the begin/end markers, leaving the rest of MEMORY.md intact. No "undo" — uninstall is destructive; to re-install, run the install flow again. Uninstall UI lives on the project-list context menu and the dashboard header (only shown when the selected project has a lock file).

Never let a template write to config.yaml, auth.json, sessions, or any credential path — the v1 installer refuses. If you extend the format, treat the preview sheet as load-bearing: the user's only trust boundary is that the sheet is honest about everything that's about to be written.

Template configuration (v2.3, schemaVersion 2)

Templates can declare a typed configuration schema in template.json's new config block. The installer renders a Configure step between the parent-directory pick and the preview sheet; values land at <project>/.scarf/config.json (non-secret) and in the login Keychain (secret). A post-install Configuration button on the dashboard header (shown when <project>/.scarf/manifest.json exists) opens the same form pre-filled for editing.

Manifest shape:

{
  "schemaVersion": 2,
  "contents": { "dashboard": true, "agentsMd": true, "config": 2 },
  "config": {
    "schema": [
      {"key": "site_url", "type": "string", "label": "Site URL", "required": true},
      {"key": "api_token", "type": "secret", "label": "API Token", "required": true}
    ],
    "modelRecommendation": {
      "preferred": "claude-sonnet-4.5",
      "rationale": "Tool-heavy workload — reasoning helps."
    }
  }
}

Supported field types: string, text, number, bool, enum (with options: [{value, label}]), list (itemType "string" only in v1), secret. Type-specific constraints (pattern, min/max, minLength/maxLength, minItems/maxItems) are optional. secret fields must not declare a default — the validator refuses.

Key services: TemplateConfig.swift (schema + value models + Keychain ref helpers), ProjectConfigKeychain.swift (thin SecItemAdd/Copy/Delete wrapper; the only Keychain user in Scarf today), ProjectConfigService.swift (load/save config.json, resolve secrets, cache manifest, validate schema + values). UI in Features/Templates/ViewModels/TemplateConfigViewModel.swift + Features/Templates/Views/TemplateConfigSheet.swift.

Secret storage. Keychain service name is com.scarf.template.<slug>, account is <fieldKey>:<project-path-hash-short>. The path-hash suffix means two installs of the same template in different dirs don't collide on Keychain entries. Values in config.json are "keychain://service/account" URIs — never plaintext. The bytes hit the Keychain only on form commit, so cancelling never leaves orphan entries.

Uninstall. TemplateLock v2 gains config_keychain_items and config_fields arrays. The uninstaller iterates each URI through SecItemDelete before removing the lock file. Absent items (user hand-cleaned) are no-ops.

Exporter. Carries the schema from <project>/.scarf/manifest.json through into exported bundles, never values. Exporting never leaks anyone's secrets. schemaVersion bumps to 2 only when a schema is forwarded; schema-less exports stay at 1.

Catalog site. tools/build-catalog.py mirrors the Swift schema validator. Each v2 template's template.json is copied into .gh-pages-worktree/templates/<slug>/manifest.json and the site's widgets.js calls ScarfWidgets.renderConfigSchema to display the schema on the detail page (display-only — the form lives in-app).

Schema is Swift-primary. If TemplateConfigField.FieldType gains a new case, update in order: TemplateConfig.swift (model + validation), tools/build-catalog.py (SUPPORTED_CONFIG_FIELD_TYPES + type-specific rules), widgets.js (summariseConstraint), TemplateConfigSheet.swift (new control subview), tests on both sides. Schema drift between validator + installer is the kind of bug users only notice after shipping.

Template Catalog

Shipped community templates live at templates/<author>/<name>/ (one level down — templates/CONTRIBUTING.md explains the submission flow for authors). The catalog site is generated from this directory and served at awizemann.github.io/scarf/templates/ alongside the Sparkle appcast — the two coexist on the gh-pages branch but touch completely disjoint paths.

Pipeline:

  • Validator + regenerator: tools/build-catalog.py is stdlib-only Python (3.9+). It walks templates/*/*/, validates every .scarftemplate against its manifest claim (mirrors the Swift ProjectTemplateService.verifyClaims invariants), enforces a 5 MB bundle-size cap, scans for high-confidence secret patterns, checks staging/ matches the built bundle byte-for-byte, and emits templates/catalog.json. Tested by tools/test_build_catalog.py — 16 tests covering every validation path.
  • Wrapper: scripts/catalog.sh mirrors the scripts/wiki.sh shape with check / build / preview / serve / publish subcommands. publish runs a second-pass secret-scan against the rendered site before committing + pushing gh-pages.
  • Site source: site/index.html.tmpl + site/template.html.tmpl are {{TOKEN}}-substitution templates. site/widgets.js (~300 lines of vanilla JS) is the dogfood — renders a ProjectDashboard JSON into HTML using the same widget vocabulary the Swift app uses, so each template's detail page shows a live preview of its post-install dashboard.
  • Install-URL hosting: raw-served from main at https://raw.githubusercontent.com/awizemann/scarf/main/templates/<author>/<name>/<name>.scarftemplate. No per-template Releases ceremony.
  • CI gate: .github/workflows/validate-template-pr.yml runs the Python validator + its own test suite on every PR that touches templates/, the validator, or its tests. Failures post a comment on the PR with the last 3 KB of the validator log.

Maintainer workflow on merge to main:

./scripts/catalog.sh build       # regenerate templates/catalog.json + .gh-pages-worktree/templates/
./scripts/catalog.sh publish     # secret-scan rendered output + commit + push gh-pages

Same cadence as scripts/release.sh (manual, auditable, no auto-deploy). Runs stay isolated: release.sh only touches appcast.xml on gh-pages; catalog.sh only touches templates/ on gh-pages. Never push catalog output on a release cadence or vice versa.

Schema is Swift-primary. When ProjectDashboardWidget.type gains a new case or ProjectTemplateManifest adds a field, update Swift first, then mirror into tools/build-catalog.py (SUPPORTED_WIDGET_TYPES, _validate_manifest, _validate_contents_claim) so the web validator stays honest. The Python test suite's real-bundle test catches drift on the example template but not on the full widget vocabulary — add a synthetic fixture to test_build_catalog.py for any new widget type.