Skip to content

CHANGELOG.md

Latest commit

 

History

History
63 lines (43 loc) · 5.79 KB

CHANGELOG.md

File metadata and controls

63 lines (43 loc) · 5.79 KB
Raw

Changelog

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[4.0.0] - 2026-07-30

A major alignment release with Anthropic's "The new rules of context engineering for Claude 5 generation models" and "Effective context engineering for AI agents", plus a repositioning to tool-agnostic / multi-model (Claude 5, GPT-Sol, GLM-5.2, and others).

This release is the outcome of a full architectural audit (docs/enhancements/ce-claude5-audit.md) and a 6-phase roadmap. It removes ~1,600 lines of duplicated/bloated content while preserving the public skill/agent contract. The MAJOR bump is driven by the removal of the applyTo frontmatter field, which Copilot consumers may have depended on for deterministic context-scoping.

⚠️ Breaking changes

  • applyTo removed from all instructions and authoring docs. This field is GitHub Copilot / VS Code-specific and is not recognized by other harnesses (Claude Code, Cursor, Windsurf, OpenCode). Instructions now activate by description matching — the same portable mechanism skills use.
    • Migration: If you relied on applyTo for context-scoping in Copilot, the instruction description now carries the target-file-type signal (e.g., "Applied to .spec.ts files"). Per-tool adapters can re-introduce scoping where a harness supports it.
  • docs/references/section-details-guide.md reduced to a redirect stub. Its content was consolidated into docs/skill-anatomy.md as the single source of truth. Update any deep links to point at the corresponding section of skill-anatomy.md.
  • Four docs/enhancements/implementation-plan-* files moved to docs/archive/enhancements/. They were completed/stale planning logs marketed as "future improvements". Paths are preserved under archive/.

Context engineering (Claude 5 alignment)

  • Progressive disclosure consolidated. The Playwright locator-priority table is now single-source (locator-strategies-priority.md); verbatim copies in playwright-e2e-testing/SKILL.md and regression-best-practices.md replaced with pointers.
  • Memory/workspace files slimmed to Anthropic's "lightweight, gotchas-only" guidance: CLAUDE.md 105 → 18 lines, AGENTS.md 222 → 41 lines. Domain locator tables and orchestration patterns removed (they live in skills/docs).
  • Authoring docs de-fragmented. Six overlapping authoring files consolidated around docs/skill-anatomy.md as the canonical source (frontmatter rules, progressive loading, resource types, naming each now appear exactly once).
  • Instructions made lean. selenium-webdriver-java.instructions.md 607 → 32 lines, cicd-testing.instructions.md 222 → 29 lines. Full templates/workflows moved to the matching skills' references/; playwright-typescript (29 lines) was already the model.
  • Agents de-duplicated (conservative, multi-model). The Test Orchestration Pattern (TOP) Constitution was repeated across 6 agents with copy-paste leakage (e.g., api-tester-specialist carried an XPath rule that "is irrelevant for API"). Duplicate Must/Must-Not pairs, repeated prohibitions (the Thread.sleep rule appeared ~5× in one file), boilerplate sections, and full tutorial code removed. Each rule now appears once per agent.
    • Design note: Constitutions are kept slim but retained (not converted to open-ended heuristics) because the repo is genuinely multi-model; models with more variable judgment benefit from explicit constraints. The anti-pattern that was removed — duplication and inapplicable-rule leakage — is universal.
  • infer convention documented. Only qa-orchestrator sets infer: false (dispatcher, never auto-activates); specialists omit it (default true, auto-selectable). Now documented in authoring-agents.md and with an inline comment.

Tool-agnostic / multi-model repositioning

  • The repo is now explicitly declared tool-agnostic and multi-model (Claude 5, GPT-Sol, GLM-5.2, …) across AGENTS.md, CLAUDE.md, README.md, and the authoring guides. Prior "optimized for GitHub Copilot" framing removed.
  • Agent frontmatter fields (tools, mcp-servers, handoffs, infer, target) documented as optional adapter fields specific to Copilot/VS Code in authoring-agents.md — the agent body (role, Constitution, workflow) is the portable part.

Linter alignment

  • lint-skills.mjs S4: description limit raised 600 → 1024 chars (Anthropic's official Agent Skills limit).
  • lint-skills.mjs S7: back-link-header check demoted ERROR → WARNING (cosmetic; does not change agent behavior).

Documentation

  • Added docs/enhancements/ce-claude5-audit.md — the full architectural review (gap analysis, adversarial review, roadmap) this release implements.
  • Refreshed README.md, docs/getting-started.md, docs/copilot-setup.md: fixed a dead link to the archived migration plan, corrected the Instructions count (7 → 3), replaced the retired "Flaky Test Hunter" persona with "Playwright Test Healer", and aligned framing with the tool-agnostic positioning.

Summary of size reductions

Area Before After
CLAUDE.md 105 18
AGENTS.md 222 41
selenium-webdriver-java.instructions.md 607 32
cicd-testing.instructions.md 222 29
test-refactor-specialist.agent.md 606 272
selenium-test-specialist.agent.md 333 87
section-details-guide.md 247 5 (stub)
Net across the release — −~1,600 lines

[3.2.0] - 2026-07-30 (prior release)

Instructions cleanup v3.2 — context engineering optimization (constitution, tool cleanup, skill independence, size optimization). See git history for details.