Skip to content

native-vs-dotforge-boundary.md

Latest commit

 

History

History
56 lines (46 loc) · 3.07 KB

native-vs-dotforge-boundary.md

File metadata and controls

56 lines (46 loc) · 3.07 KB
Raw
globs docs/v4/*.md, behaviors/*, stacks/*, skills/*, .claude/rules/domain/*.md
description Native-first boundary — what dotforge keeps vs cedes to native Claude Code, and the method for deciding
domain dotforge-meta
last_verified 2026-06-03

Native vs dotforge Boundary

Governing principle

If Claude Code resolves it natively, ADOPT the native solution. dotforge only owns what natively has no equivalent. Scope shrinks as Claude Code grows — that is correct.

Method (mandatory before any scope decision)

Verify the CURRENT native state against official docs (code.claude.com/docs, anthropics/claude-code) BEFORE classifying. A scope call on stale assumptions is wrong by default — this is why /forge watch (keeping domain rules current) is the INPUT that makes boundary decisions correct, not meta-work. Never cede a capability without confirming the native feature actually covers the real case.

Classification (verified 2026-06-03)

KEEP — no native equivalent:

  • Domain rules — curated encyclopedia of CC internals + business domain. Core asset.
  • Cross-project propagation with merge — forge:section markers + /forge sync. Native symlinks/global CLAUDE.md only COPY, never merge per-project customization.
  • Behaviors v3 — SPLIT verdict (validated 2026-06-04). Portfolio scan: 0 overrides across all 12 projects over ~7 weeks (behaviors live since v3.0, 2026-04-13), including production (TRADINGBOT, cotiza). Only 4/12 adopted behaviors at all.
    • KEEP the graduated escalation engine (silent→nudge→warning→soft_block). It is exercised and works — the happy path is "verify", not "override" (observed live: verify-before-done soft_blocked a push, resolved by running tests). hookify is only binary warn/block; the graduated escalation is a real, used delta.
    • RETIRE/SIMPLIFY the auditable override trail — the v3-vs-hookify differential is empirically refuted (0 uses). The v4 override-capture loop (process-override-log.sh, audit item B3, SessionStart wiring) processes a log that is always empty: machinery for an event that does not occur. Dead weight, not delta.
  • Registry + audit cross-project — REORIENT: audit must measure good use of NATIVE features (auto-memory active, sandbox set, deny rules present, /init run), not presence of dotforge machinery.

CEDE — native covers it:

  • Individual learning capture → native auto-memory (per-project, ~/.claude/projects/<p>/memory/)
  • Identical shared rules (no merge needed) → native symlinks in .claude/rules/ + global CLAUDE.md
  • Workflows / orchestration → /workflows, Agent Teams, /deep-research (already ceded v4)
  • Base CLAUDE.md generation → /init
  • One-shot code review → /code-review --comment/--fix
  • Model routing as a system → /effort (keep only as documentation)

Anti-pattern

Building atop native internals that may change (compiled hooks depend on the hook API shape). Every breaking upstream change forces a re-tune. Keep the native-dependent surface minimal and validate delta demand before expanding it.