v0.4.0

Added

• nd plugin: onboarding skill (frontmatter name: onboard; the command is
  /nd:onboarding, from the folder name), moved in from personal
  ~/.claude/skills — a workflow for producing a code-grounded onboarding
  documentation set (C4 context/containers/components, engineering, AI design,
  infrastructure, deployment, ownership map, fork-style ONBOARDING) for any repo,
  with a three-layer common/ + per-track layout when the repo holds several
  discrete products. Ships checker.py (citation/dangling-path/mermaid/link/
  track-leak acceptance gate with --layout, --track NAME:REGEX, --ai-doc,
  --agents-doc), briefs.md (subagent and reviewer briefs) and
  objective-template.md (goal-mode objective). What was verified, and how, is in
  the next entry and in nd/evals/README.md.
• nd plugin: onboarding now runs outside the original harness (subagent and
  goal-tool fallbacks, explicit phase order, an inventory step before briefing) and
  checker.py enforces the code-only rule: it fails inadmissible citations (README,
  CLAUDE.md, docs/, reference/) outside an "Existing in-repo prose" section, line
  and symbol anchors that don't exist, TODO/TBD placeholders, and ONBOARDING entries
  that mention a sibling without linking it; paths match case-exactly, and MIME
  types, ~/@ strings and path:func() anchors are handled. Agents now report in
  fixed plain-text headings instead of JSON. The minimum-citations rule now counts
  unique citations (path plus anchor) rather than unique files, and the rules say a
  file the docs call absent is named in plain text. SKILL.md gained a workflow
  checklist, a Python preflight (an unrun gate is reported NOT RUN, never PASS), a
  "running short" rule, and a verification-depth choice (full or targeted); track
  detection and gate details moved to tracks.md and gate.md. Reviewers are
  capped by group size, work alone (no helper agents) and must return a report with a
  Covered line, after a real run showed one oversized reviewer stall, spawn helpers
  that applied nothing, and leave the orchestrator to reconcile about 35 findings.
  The checker now also keeps the README exemption open across sub-headings, counts
  route-group paths like src/app/(auth)/login/page.tsx and bare *.config.* files,
  ignores symbols that occur only in comment lines, and gate.md has a failure-to-action
  table, a definition of "fails twice", and an always-pass---root rule. The brief
  tells writers to backtick a file only when citing its real path and to write "to-do
  comments" instead of the literal placeholder words; in a mixed repo agents.md
  covers agents and workloads, and the reviewer count for tracked repos is stated
  (and marked unmeasured).
  84 tests in tests/onboarding/ (the checker, plus consistency tests that tie the
  brief, gate.md and SKILL.md to the checker).
  Evidence, with its limits: 4 eval cases. The trigger case fires on natural phrasing
  (5/5 with the plugin); its Δ +1.00 is structural, since a baseline cannot load a skill
  it lacks. The near-miss does not over-fire. The applied case passes but does not
  discriminate from a baseline (Δ 0.00). The full-workflow case is one run each on two
  real repos, both single-layout Next.js apps: the second finished (score 0.82, 10 agents,
  824 claims checked and 68 corrected by the agents' own count) and its output passes the
  checker when run afterwards, but its final-report grader was skipped, not passed, and
  the agent had no shell to run the checker itself. That run used a repo about eight times
  smaller than the one where a reviewer stalled, so it does not show the review-group cap
  fixes that stall. The tracked layout, round-2 fixes, targeted depth and Haiku/Opus
  have not been exercised by an agent.