CHANGELOG.md

Changelog

v1.10.1 (2026-04-18)

Marketplace preparation + CodeBuddy removal

• Removed CodeBuddy support: .codebuddy-plugin/ directory removed, CODEBUDDY.md marked deprecated, all CodeBuddy references cleaned from Skills, Commands, Agents, templates, and docs. Plugin now targets Claude Code exclusively.
• English-first documentation: README.md, CONTRIBUTING.md, CHANGELOG.md, all 8 SKILL.md files, 9 Command files, 4 Agent files, architecture.md, and marketplace-submission.md translated to English. Chinese README preserved as README.zh-CN.md.
• Plugin metadata in English: plugin.json description, userConfig fields, and keywords; marketplace.json description and version bump.
• New: harness:archive Skill (P0-1): 4-step workflow — archive completed specs to docs/archive/ (preserving git history via git mv), check doc-code consistency, run architecture health scan, generate structured archive report.
• Enhanced: harness:plan (P0-2): now enforces <action>/<verify>/<done> triple structure for every task. Reads rigid vs flexible constraints from features.json — rigid items must map to tasks, flexible items are advisory. 100% rigid constraint coverage check before execution.
• Enhanced: session-start.sh (P0-3): added features.json summary output (in-progress feature, next feature, stats) and 5-step ceremony chain checklist for Agent.
• Expanded safety hooks: pre-protect-env upgraded with comprehensive secret detection patterns (SSH keys, service accounts, .netrc, AWS credentials), dangerous command blocking (force-push, hard-reset, chmod 777, curl-pipe-to-shell), and inline secret detection.
• Marketplace submitted: v1.10.1 submitted to Anthropic Plugin Directory for review (Claude Code + Claude Cowork).

v1.10.0 (2026-04-12)

Unified harness: namespace

• Simplified plugin name: harness-engineering → harness, so users see harness:init instead of harness-engineering:harness-init
• Skills renamed: harness-init → init, harness-audit → audit, harness-evolve → evolve, using-harness → router, writing-plans → plan, verification → verify (tdd unchanged)
• Commands renamed: standardized to verb-noun format: arch-scan → scan-arch, entropy-scan → scan-entropy, assign-features → assign, context-dump → dump, doc-sync → sync-docs, trim-claudemd → trim
• Full reference update: all SKILL.md cross-references, eval JSON, architecture.md, marketplace docs, README, CONTRIBUTING, and 36+ other files updated in sync
• Breaking: old slash commands (/harness-init, etc.) no longer work; use the new names (/harness:init, etc.)

v1.9.3 (2026-04-12)

Cross-platform hooks + upstream version tracking

• Cross-platform polyglot hook wrappers (inspired by obra/superpowers polyglot-hooks):
  • Added .cmd polyglot entry for each hook script (valid as both CMD and bash); auto-detected by Windows Git Bash
  • hooks.json registration entries changed from .sh to .cmd
  • Kept .sh files for backward compatibility; added extensionless bash logic files (.cmd delegates to them)
  • On Windows, silently succeeds (exit 0) when bash is not found, avoiding workflow blocking
• Upstream version tracking: three workflow Skills (writing-plans / tdd / verification) now include upstream + harness-delta metadata in their headers, pinned to obra/superpowers @ 917e5f5, documenting each Skill's delta from upstream
• Updated README: scripts directory description, cross-platform compatibility table

v1.9.2 (2026-04-11)

Superpowers workflow integration

• Added skills/writing-plans/: pre-implementation planning Skill, triggered for tasks >30 min or touching 3+ files; outputs tasks.md with a human confirmation gate
• Added skills/tdd/: TDD workflow Skill (RED → GREEN → REFACTOR), bound to the 1% rule — automatically triggered on any code writing
• Added skills/verification/: pre-completion verification Skill with four-layer checks (Functional / Quality / Architecture / Integration)
• Added scripts/session-start.sh: SessionStart hook that reads claude-progress.json on session open, displaying in-progress tasks, to-do count, and blockers; triggers an archive reminder when completed items >= 10
• Updated hooks/hooks.json: registered SessionStart event
• Updated skills/using-harness/SKILL.md: added Steps 4-6 documenting trigger conditions for writing-plans / tdd / verification
• Updated skills/harness-init/SKILL.md: init artifact table and file structure diagram now include the three new Skills and session-start.sh
• Updated docs/templates/generic/CLAUDE.md.template: added "Workflow Skill auto-trigger" section
• Updated references/HarnessEngineering.md: added Section L (Superpowers Integration) with comparison table, integration points, full execution chain diagram, 14-Skill mapping; reference table now includes three new obra/superpowers entries

v1.9.1 (2026-04-10)

harness-init Phase 5: archive mechanism + features.json tiering strategy

• Changed features.json positioning: from "optional for long-cycle projects" to "optional for solo / required for multi-person or multi-Agent"; init now auto-decides whether to generate it based on team size
• Added a two-file responsibility comparison table (who writes / what it records / token growth trend / multi-Agent conflict risk)
• Added extension field documentation for multi-person/multi-Agent scenarios (owner, depends_on, files_owned, worktree, acceptance)
• Added archive mechanism: triggers archiving when completed_features exceeds 10 entries, preventing token bloat (~8000 tokens after 6 months); init now also generates docs/archive/ directory skeleton and archive strategy README
• Three archive rules written into AGENTS.md (archive threshold / done-entry compression strategy / Agent read-scope restriction)
• Added docs/templates/generic/archive-readme.md.template

v1.9.0 (2026-04-10)

Claude Marketplace support (Plan B)

• marketplace.json: added .claude-plugin/marketplace.json for community marketplace subscription distribution. Users can subscribe with auto-updates via:

  /plugin marketplace add https://raw.githubusercontent.com/huangbaixun/harness-engineering/main/.claude-plugin/marketplace.json
  

• plugin.json improvements: added homepage and repository fields (required for official marketplace submission); added userConfig (team_name, default_tech_stack) — Claude Code prompts the user to fill these on enable, no manual configuration needed
• ${CLAUDE_PLUGIN_ROOT} path fix: all internal path references in harness-init (template directory, init.sh.template) now use ${CLAUDE_PLUGIN_ROOT} prefix, ensuring correct path resolution in marketplace cache mode
• Init artifact table fix: .claude/hooks/ paths changed to $TOOL_DIR/hooks/, consistent with the tool-agnostic architecture
• Keyword expansion: plugin.json keywords now include codebuddy, team, sprint for better marketplace discoverability

v1.8.0 (2026-04-08)

Tool-agnostic architecture — full CodeBuddy compatibility

• AGENTS.md as cross-tool universal memory file (ADR 0005):
  • Added AGENTS.md: unified memory file serving as the single source of truth for all project rules, readable by both Claude Code and CodeBuddy
  • CLAUDE.md reduced to a 2-line wrapper that directs Claude Code users to AGENTS.md
  • Added CODEBUDDY.md: 2-line wrapper that directs CodeBuddy users to AGENTS.md
• Removed hardcoded tool paths:
  • Added .codebuddy-plugin/plugin.json: CodeBuddy plugin manifest (v1.8.0)
  • harness-init Phase 2 six-layer table now uses $TOOL_DIR variable instead of hardcoded .claude/
  • harness-init Phase 3 file structure diagram updated to show dual-tool compatible layout with AGENTS.md hierarchy
  • harness-audit SKILL.md tool detection logic and memory file diagnostics adapted for multi-tool support (merged before v1.8.0)
• Template updates:
  • Added docs/templates/generic/AGENTS.md.template: includes $TOOL_DIR no-hardcode convention
  • docs/templates/generic/init.sh.template added 10-line tool detection block that auto-detects CodeBuddy / Claude Code and exports $TOOL_DIR and $TOOL_NAME
• ADR 0005: documents the tool-agnostic architecture decision, with three-option comparison and Agent constraint rules
• docs/architecture.md updated: reflects dual-tool compatible structure, directory diagram includes .codebuddy-plugin/, CODEBUDDY.md, AGENTS.md, and references/team-parallel-development.md

v1.7.0 (2026-04-06)

Multi-person collaboration + features.json lifecycle improvements

• Added /assign-features command: sprint feature assignment planner inspired by superpowers writing-plans, five-phase workflow —
  • Phase 1: analyze features.json dependency graph, compute startable (all dependencies done?) and criticality (transitive closure block count)
  • Phase 2: auto-read CLAUDE.md ## Team Members section; prompt and write back if missing
  • Phase 3: four-rule assignment algorithm (file conflict detection / load cap protection / critical path priority / layer affinity)
  • Phase 4: generate sprint-kickoff.sh with per-person sections containing git claim + worktree + Agent launch commands
  • Phase 5: append assignment records to claude-progress.json sprint_history for traceability
• Added commands/evals/evals.json: 4 command test cases (standard 3-person sprint, file conflict detection, dependency-unlock scheduling, overload protection)
• Added references/team-parallel-development.md: parallel development guide for multi-person full-stack teams, synthesizing Anthropic Agent Teams official docs, OpenAI Codex team practices, and a 16-Agent parallel stress test case study on a C compiler; covers features.json parallel field upgrades, Git Worktree isolation configuration, three division-of-labor models, and design principles for reducing human-Agent dependencies
• harness-init Phase 5 added features.json usage rules: "Agent read-only principle" and "cancel-don't-delete principle" (status=cancelled + cancelled_reason, never delete entries)

v1.6.0 (2026-04-06)

Methodology deepening + Skill TDD improvements

• Added references/HarnessEngineering.md: comprehensive methodology handbook distilled from Anthropic, OpenAI, InfoQ, and Hacker News best practices, serving as the primary source for plugin design
• harness-init added Phase 0 existing-project detection: scans for existing CLAUDE.md before init, branching into new project / existing project / corrupted file paths; existing-project mode offers three options (incremental supplement / optimize and consolidate / full rebuild) with mandatory .bak backup before execution
• harness-init added init artifact manifest: lists all output files at the top of the Skill so users know what will be generated before triggering
• Added docs/templates/generic/init.sh.template: session startup script template that runs before each new Claude Code session, displaying progress/feature list/architecture doc entry points (inspired by walkinglabs/learn-harness-engineering harness-creator pattern)
• Evals restructured: split from a single file into per-skill directories, converted to skill-creator compatible format (with assertions[] array for objective grading by a grader subagent)
  • skills/harness-init/evals/evals.json (3 evals, including new existing-project detection case)
  • skills/harness-audit/evals/evals.json (1 eval)
  • skills/harness-evolve/evals/evals.json (1 eval)
  • skills/using-harness/evals/evals.json (2 evals)
  • evals/agents/coding-agent.json (3 evals)
  • evals/evals.json converted to an index file
• README.md added "Methodology Reference" section with artifact preview table after the quick-start second step
• CLAUDE.md added methodology handbook reference link

v1.5.1 (2026-04-05)

• Fixed: release.yml version validation and tag matching (initial release correction)

v1.5.0 (2026-04-05)

Open-source preparation

• Added LICENSE (MIT)
• Added CONTRIBUTING.md: Skill TDD contribution workflow, language template contribution guidelines, hook script conventions, PR format
• Rewrote README.md: targeted at first-time users, 3-step quick start, badges, collapsible file inventory
• Added .github/workflows/release.yml: auto-packages .skill and creates GitHub Release on semver tag, with manifest version validation
• Added .github/workflows/validate.yml: auto-validates plugin structure, manifest, hooks.json, and eval format on PR
• Added .github/ISSUE_TEMPLATE/: bug_report and feature_request templates
• Added .github/PULL_REQUEST_TEMPLATE.md: includes Skill TDD checklist

v1.4.1 (2026-04-05)

• Fixed: removed path declaration fields from plugin.json, resolving "agents: Invalid input" loading error
• Relies on Claude Code auto-discovery of default directories; only metadata fields retained

v1.4.0 (2026-04-05)

• Changed: restructured to comply with official Claude Code plugin spec
  • Manifest moved to .claude-plugin/plugin.json
  • commands/, agents/, skills/ moved to plugin root
  • Hooks converted to hooks/hooks.json JSON registration format
  • Hook scripts migrated to scripts/, using ${CLAUDE_PLUGIN_ROOT} path variable
  • author field changed to object format

v1.3.0 (2026-04-05)

• Added using-harness meta-Skill (forced intent-recognition trigger, based on obra/superpowers 1% rule)
• Evals expanded to 9 test cases with 6 new stress tests
• coding-agent now embeds a two-stage mandatory review (Spec Compliance → Code Quality)

v1.2.0 (2026-04-05)

• Added coding-agent (Sonnet model, long-cycle multi-session coding, per handbook Section F.4)
• Added Java language stack template (JUnit 5 + Mockito + Checkstyle + SpotBugs)
• Language templates expanded to five

v1.1.0 (2026-04-05)

• Added explore-agent (Haiku model, context-efficient exploration subagent)
• Added code-review-agent (Sonnet model, code quality inferential sensor)
• Added /entropy-scan command (fourth-category garbage collection: code entropy detection)
• Added plugin.json version manifest

v1.0.0 (2026-04-04)

• Initial release: three core Skills + seven Commands + five Hooks + multi-language templates + security-reviewer