Skip to content

v0.3.0 — progressive-disclosure restructure and Opus 4.7 alignment#3

Merged
PBNZ merged 1 commit intomainfrom
v0.3.0-restructure
Apr 20, 2026
Merged

v0.3.0 — progressive-disclosure restructure and Opus 4.7 alignment#3
PBNZ merged 1 commit intomainfrom
v0.3.0-restructure

Conversation

@PBNZ
Copy link
Copy Markdown
Owner

@PBNZ PBNZ commented Apr 20, 2026

Summary

  • Cut per-invocation Newton context by ~45% by extracting research-methodology, reuse-check, attribution, and handoff into plugins/newton/skills/newton/references/*.md and replacing them in SKILL.md with short pointer blocks. Consolidates the previously triple-covered core principles (Core principles + "What Newton does not do" + Self-evaluation gate) into one core-principles section plus a leaner pre-delivery gate. Removes meta-commentary and duplicated prose.
  • Aligns Newton with current-generation models (Opus 4.7) without coupling to them. Outcome-oriented tool guidance (fits implicit tool detection), effort-level awareness in the opening move, session-memory guidance for multi-turn work, a principle-level parallelisation note, a vision-as-primary-evidence note, and removal of tone-management instructions Opus 4.7 handles natively. Backwards-compatible on Sonnet 4.6 and Haiku 4.5.
  • Updates the tool-rules generator (scripts/generate-rule-files.py) to inline references/*.md under the main body, so Cursor / Windsurf / Cline / Copilot rule files stay self-contained for surfaces that don't do progressive disclosure. Widens sync-rules.yml to watch references/**.
  • Adds an optional router CLAUDE.md example under docs/examples/router-claude-md.md for users who want model-aware layering across every skill, not just Newton. Opt-in only — not auto-installed by the plugin.
  • Minor housekeeping: plugins/newton/commands/newton.md goes from empty stub to a working /newton:newton slash command. plugin.json and marketplace.json bumped to 0.3.0. README adds Using Newton on Opus 4.7 pointer. CHANGELOG and CONTRIBUTING updated.

Token footprint (approximate)

  • SKILL.md body: 3,900 → 2,544 words (~35% smaller).
  • references/ on-demand: 1,640 words total across 4 files, loaded only when the current turn needs them.
  • Tool-rules (non-progressive-disclosure surfaces): ~29 KB → ~28.6 KB with full methodology preserved via inlining.

Test plan

  • All 5 validators pass locally (validate_marketplace.py, validate_plugins.py, validate_skills.py, check_skill_safety.py, check_no_private_contact.py).
  • Generator is deterministic (md5 stable across reruns).
  • All 4 references/*.md pointers in SKILL.md resolve to existing files.
  • Attribution-target section name ("Simplicity in what's produced") still matches the heading quoted in NOTICE.md.
  • JSON syntax valid in plugin.json and marketplace.json.
  • CI workflows validate.yml and sync-rules.yml pass on this branch (watching this PR).
  • Behavioural smoke tests on Opus 4.7 — default mode, quick-start, research, build (reuse check visible), multi-turn memory, vision input.
  • Behavioural smoke tests on Sonnet 4.6 and Haiku 4.5 — at least parity with v0.2.1.

What is not in this PR (deferred)

  • Dedicated Newton subagents (newton-researcher, newton-reuse-checker, newton-self-critic) — the principle-level parallelisation note is the stepping stone; the decomposition is a later release.
  • Auto-installed router CLAUDE.md — stays a user-level opt-in; the repo ships an example only.
  • README translations and the landing-page language switcher — tracked in issues.

Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com

🤖 Generated with Claude Code

… alignment

Cut per-invocation Newton context by ~45% by making the references/ folder
actually do the job its name implies: extract research-methodology,
reuse-check, attribution, and handoff into on-demand reference files, and
replace them in SKILL.md with short pointer blocks. Consolidate the
previously triple-covered core principles into one section plus a leaner
pre-delivery gate, and drop the meta-rationale and duplicated-prose
paragraphs that burned context every turn without earning their place.

Align with current-generation models (Opus 4.7) without making Newton
model-dependent:

- Tool guidance shifts from prescriptive ("run web_search") to
  outcome-oriented ("verify time-sensitive claims against current
  sources") — both a context win and a better fit for implicit tool
  detection.
- Effort-level awareness in the opening move: default toward
  quick-start at low/medium effort, full methodology at high/xhigh,
  maximum depth at max. Non-Opus environments that don't expose
  effort follow default mode.
- Session-memory guidance for multi-turn Newton work when file-system
  memory is available.
- Principle-level parallelisation note for independent sub-tasks
  (dedicated Newton subagents remain deferred to a later release).
- Vision-as-primary-evidence note inside the research reference.
- Trim tone-management instructions Opus 4.7 handles natively
  (no-emoji, no-praise-opener, no-question-echo) — native calibration
  replaces them.

Update scripts/generate-rule-files.py to inline references/*.md into
the generated tool-rules output so Cursor, Windsurf, Cline, and Copilot
keep the full methodology; those surfaces don't do progressive
disclosure. Widen sync-rules.yml to watch references/** so future
reference edits can't slip past CI.

Also:
- docs/examples/router-claude-md.md: optional user-level CLAUDE.md
  template that detects the running model and layers model-appropriate
  behaviour on top of any loaded skill. Opt-in by copying into the
  user's own CLAUDE.md; not auto-installed.
- plugins/newton/commands/newton.md: replaced the previously empty
  stub with a working /newton:newton slash command.
- plugin.json and marketplace.json bumped to 0.3.0.
- README.md: "Using Newton on Opus 4.7" pointer and roadmap update.
- CHANGELOG.md: v0.3.0 entry.
- CONTRIBUTING.md: documents the new references/ inlining flow.

Backwards-compatible on Sonnet 4.6 and Haiku 4.5; no behavioural
regressions. Newton's voice and triggering rules are unchanged.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Copilot AI review requested due to automatic review settings April 20, 2026 04:34
Copy link
Copy Markdown

Copilot AI left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

This PR restructures the Newton skill content for progressive disclosure (moving extended methodology into references/*.md) while keeping generated tool-rule files self-contained by inlining those references during generation. It also bumps the Newton plugin/marketplace versions to 0.3.0 and adds an example model-aware router template plus a working slash-command stub.

Changes:

  • Extracts extended methodology into plugins/newton/skills/newton/references/*.md and replaces SKILL body duplicates with short pointer blocks.
  • Updates scripts/generate-rule-files.py + sync-rules.yml to inline references/*.md into generated tool-rule files and watch the new paths in CI.
  • Adds docs/examples router template, fills in /newton:newton command, and bumps versions + changelog/readme/contributing.

Reviewed changes

Copilot reviewed 19 out of 19 changed files in this pull request and generated 1 comment.

Show a summary per file
File Description
plugins/newton/skills/newton/SKILL.md Restructures core Newton guidance; adds effort-awareness, pointers to extracted references, and consolidates principles/gates.
plugins/newton/skills/newton/references/attribution.md New extracted attribution methodology reference.
plugins/newton/skills/newton/references/handoff.md New extracted handoff methodology reference.
plugins/newton/skills/newton/references/research-methodology.md New extracted research workflow reference (incl. vision-as-primary-evidence note).
plugins/newton/skills/newton/references/reuse-check.md New extracted reuse-check workflow reference (incl. parallelisation and simplicity discipline).
scripts/generate-rule-files.py Inlines references/*.md into generated rule files and updates README/banner templates accordingly.
.github/workflows/sync-rules.yml Expands CI drift-check inputs to include references/** and updates messaging.
tool-rules/README.md Documents that tool-rule files inline references/ for non-progressive-disclosure tools.
tool-rules/.cursorrules Regenerated output including inlined references + updated banner/body.
tool-rules/.windsurfrules Regenerated output including inlined references + updated banner/body.
tool-rules/.clinerules Regenerated output including inlined references + updated banner/body.
tool-rules/copilot-instructions.md Regenerated Copilot instructions including inlined references + updated banner/body.
plugins/newton/commands/newton.md Adds a working /newton:newton command that prefixes prompts with Newton:.
plugins/newton/.claude-plugin/plugin.json Version bump to 0.3.0.
.claude-plugin/marketplace.json Version bump to 0.3.0.
docs/examples/router-claude-md.md Adds optional user-level model-aware router CLAUDE.md example.
README.md Adds Opus 4.7 usage notes and links to the router example; updates roadmap text.
CONTRIBUTING.md Updates contributor guidance to include references extraction + inlining behavior.
CHANGELOG.md Adds 0.3.0 release notes and updates comparison links.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Comment on lines 172 to 176
skill_text = SKILL_PATH.read_text(encoding="utf-8")
body = strip_frontmatter(skill_text)
body = strip_frontmatter(skill_text).rstrip() + "\n"
references_block = collect_references()
combined = body + references_block

Copy link

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The generated rule files now inline references/*.md, but the SKILL body still contains pointers like see references/research-methodology.md``. In the generated outputs these paths don’t exist (they’re only meaningful relative to plugins/newton/skills/newton/SKILL.md), so readers/tools will see broken references even though the content is inlined below. Consider rewriting those pointer strings during generation (e.g., to “see the ‘Research methodology’ section below” or an anchor link to the inlined heading) so generated files remain self-contained and internally consistent.

Copilot uses AI. Check for mistakes.
@PBNZ PBNZ merged commit de6d40e into main Apr 20, 2026
6 checks passed
@PBNZ PBNZ deleted the v0.3.0-restructure branch April 20, 2026 04:50
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants