← Corpus / lossless-monorepo / agent-skill

context-vigilance grows from six folders to eight — plus a universal tier and an experimental tier

plans/ and agent-skills/ earned canon the honest way: fourteen and five repos adopted them without anyone declaring them official. The skill now names them, blesses extra/ and sitemap/ as universal utilities (extra/ gitignored by default), and stakes out loops/, handoffs/, decisions/, and contracts/ as the experimental tier still finding its shape.

Path
agent-skills/changelog/2026-07-21_01.md
Authors
Michael Staton
Augmented with
Claude Code on Claude Fable 5
Tags
Context-Vigilance · Conventions · Agent-Skills

Why Care?

The context-vigilance skill’s own philosophy says the convention grows by assimilation — unplanned folders appear, prove themselves, and get promoted. Today that mechanism ran for real, backed by a tree-wide survey: plans/ exists in 14 context-v directories and agent-skills/ in 5, all organic, none of it named by the skill. A convention document that trails its own practice by that much stops being the source of truth. This release closes the gap and — for the first time — gives the skill an explicit experimental tier, so proposed folders have a documented on-ramp instead of living in ambiguity until someone notices them.

What’s New?

  • Six → Eight canonical directories. plans/ joins Prep mode (spec → plan → prompt, in descending altitude — the natural home for plan-mode output, roadmaps, and migration sequences). agent-skills/ joins Reflective mode as its executable tier: per-repo skills in the Anthropic SKILL.md shape, incubating next to the code they serve before graduating to this shared public repo.
  • Universal utility folders. extra/ (scratch and out-of-band material — now gitignored by default, making the privacy boundary structural rather than tooling-dependent) and sitemap/ (maps of a project’s own pages/routes/slides/surfaces, proven across the deck client-sites, applicable anywhere).
  • Experimental tier. loops/ (durable definitions of recurring operational loops — scope, cadence, exit conditions — pairing with agent loop-runners), handoffs/ (end-of-session state capture so the next session, human or agent, resumes without re-derivation), decisions/ (isolated artifacts of decisions made — what, when, why, alternatives passed over; kin to ADRs but not limited to architecture), and contracts/ (binding agreements in two flavors: constitution contracts — standing rules agents must always follow — and interface contracts — ironclad API-level documentation of how data must flow through the app or a microservices architecture). Explicitly marked as not-yet-consistent: divergence between repos is expected and should be surfaced, not normalized.
  • Decision tree extended to all twelve destinations, and the “seventh folder” section renamed to “a folder outside the set” — with the observed domain-specific folders (narratives/, profiles/, inquiry/, …) named as correct-where-they-are rather than drift.
  • References reconciled. philosophy.md (“Six is not the answer” → “The count is not the answer”, now citing plans/agent-skills as completed assimilations and loops/handoffs as the ones in flight) and doc-type-guide.md (tree updated to the eight).
  • The tree-wide folder survey that grounded the promotions ran against the context-vigilance-kit’s freshly re-curated sources.md (41 included context-v directories), same day — see the kit’s changelog/2026-07-21_01.md.