v1.5.0

Added

• Voice Directive: writing style guidelines + banned AI-slop words (delve, crucial, robust, etc.) to enforce professional agent output quality
• Completion Status Protocol: DONE / DONE_WITH_CONCERNS / BLOCKED / NEEDS_CONTEXT status framework with hard evidence gates — prevents vague "it's done" claims
• Confusion Protocol: STOP-and-ask guardrail for high-stakes ambiguity (architecture decisions, data model changes, destructive scope) — agents no longer guess on critical decisions
• Proactive Skill Routing: 28 intent-to-skill auto-dispatch mappings — users no longer need to memorize all 27 skill names; say "bug" → bug-hunter, "deploy" → qa-engineer + code-reviewer
• AskUserQuestion Decision Brief: structured user question format with D headers, ELI10 explanations, completeness scores, and forced recommendation — replaces ad-hoc agent questions
• AI Effort Compression Reference: human-vs-AI time comparison table (100x boilerplate, 30x features, 5x architecture) to reframe build-vs-skip scope decisions
• Learning Log (scripts/log-learning.sh + CLAUDE.md Operational Self-Improvement section): cross-session knowledge compounding — skills log durable project quirks, fixes, and patterns to ~/.virtual-company/projects/<slug>/learnings.jsonl; loaded at session start
• Preamble Tier System (T1-T4): preamble-tier field in all 27 SKILL.md frontmatters — controls which guardrail sections are active (T1=minimal documentation, T4=full security gates) to save tokens on simple skills while applying full discipline to critical ones
• Skill Template Pipeline (scripts/gen-skills.sh + skills/_templates/ directory): SKILL.md.tmpl master template with {{PLACEHOLDER}} tokens resolved by 6 reusable resolver components (voice, completion, confusion, ask-user-format, effort-table, learning-log) — edit one resolver file to propagate changes to ALL skills; bug-hunter has a working .spec proof of concept

Changed

• CLAUDE.md expanded from 346 to ~530 lines with 8 new operational sections inspired by gstack v1.58 architecture patterns
• All 27 SKILL.md files: preamble-tier field added to frontmatter (4x T1, 16x T2, 5x T3, 2x T4)
• All internal references changed from ~/.gstack/ to ~/.virtual-company/ (learning log storage, script paths)
• validate-skill.sh: now checks for preamble-tier field in SKILL.md files
• scripts/README.md: documented log-learning.sh and gen-skills.sh scripts
• hooks.json: welcome banner bumped to v1.5.0
• .claude-plugin/plugin.json: v1.4.2 → v1.5.0, updated description with new features
• New directories: skills/_templates/ (master template + 6 resolvers)
• New files: skills/_templates/SKILL.md.tmpl, 6 resolver .md files, scripts/gen-skills.sh, scripts/log-learning.sh
• All 27 skills now have SKILL.md.spec source files generated from clean SKILL.md content + resolver placeholders
• All 27 SKILL.md files regenerated via gen-skills.sh --all with Voice, Completion, Confusion, and Learning Log injected from resolvers (+68 lines per skill)

v1.4.2

Fixed

• disallowedTools corrected from bare YAML string to proper list [Edit, Bash] in all 5 advisory agents — previously the string literal was not parsed as a list, leaving agents potentially unrestricted
• tech-lead agent now has explicit disallowedTools: [] — was the only agent missing this field entirely
• qa-engineer agent now allows Bash (required to run test suites); only Edit remains blocked
• validate-skill.sh calls in commands/skills.md fixed — script expects a single file path, not a directory
• --category architecture example corrected to valid category (logic)
• PostToolUse lint hook now tries npm, bun, yarn, and pnpm (was bun-only)

Added

• SessionEnd hook — teardown message on session end
• PostCompact hook — context-resume prompt after compaction
• PreToolUse and PreCompact hooks now documented in CLAUDE.md (were present in hooks.json but invisible to agents)
• commands/ directory added to CLAUDE.md Technical Architecture section

Changed

• Architect agent: replaced 00-tech-lead skill with 01-doc-writer — architect produces ADRs and delegates to tech-lead
• Skills lists standardized to inline [a, b, c] format across all agents
• Code-reviewer effort tuned from high to medium (matches focused 10-turn audit role)

v1.4.1

Added

• Examples section added to 19-observability-specialist SKILL.md (structured logging, Prometheus/Grafana, OpenTelemetry tracing, alert rules)

Fixed

• Stale references to deleted scripts (doc-health.sh, tsv-log.sh, dockerfile-lint.sh) in README.md, commands/skills.md, and skills/01-doc-writer/SKILL.md
• PostToolUse lint hook now gracefully skips when no linter is configured (instead of erroring)

Changed

• Agent frontmatter: qa-engineer added missing disallowedTools (Edit, Bash), fixed maxTurns spacing
• Agent frontmatter: tech-lead fixed maxTurns spacing
• docs/ directory now tracked in git (was gitignored); .pantheon/ added to .gitignore

v1.4.0

Added

• performance-profiler added to Domain Expert Registry in Infrastructure category (was missing from CLAUDE.md, README, and commands/skills.md — table listed 26 but 27 skills exist)
• Examples section added to 00-tech-lead SKILL.md (now 10/10 validation score)
• docs/plans/ directory created (referenced throughout documentation and agents but didn't exist)

Fixed

• Agent naming consistency: README used reviewer instead of code-reviewer, and security instead of security-reviewer — now matches actual agent files and CLAUDE.md
• PreToolUse security hook in hooks/hooks.json was non-functional (grepped /dev/stdin which never received content) — replaced with working prompt-type warning
• skill-generator was listed in two categories (Orchestration and Meta) in commands/skills.md — removed duplicate Meta category

Removed

• .agent/ directory with 59 legacy/non-advertised skills and placeholder hooks (already gitignored, now fully purged from disk)
• tsv-log.sh — autoresearch-specific logger, not core to virtual-company
• dockerfile-lint.sh — Docker-only tool, not general-purpose
• doc-health.sh — generic documentation checker, not central to multi-agent orchestration

Changed

• Verification scripts reduced from 8 to 5 (kept: validate-skill, security-sentinel, verify-gate, coverage-gate, audit-deps)
• Plugin description updated to reflect 5 verification scripts
• README project structure diagram now shows all actual directories (commands/, scripts/, .claude-plugin/)

v1.3.1

Fixed

• Fixed hooks.json schema to use PostToolUse correctly with required hooks arrays
• Fixed README install instructions — clarified GitHub shorthand usage vs plugin name
• Fixed plugin.json repository field from object to string format

Changed

• Updated initialization message to v1.3.1
• Lint command changed to bun run lint

v1.3.0

Added

• Mechanical overrides and agent directives for operational rules (Step 0 rule, phased execution, senior dev override, forced verification, sub-agent swarming, context decay awareness, file read budget, tool result blindness, edit integrity, no semantic search)
• Hard gates for: task completion, pre-commit, refactors, and file edits
• Hard gate enforcement section in CLAUDE.md

Changed

• Refactored all 27 skills with standardized mechanical directives and operational guidance
• Decision trees, failure modes, and red flags sections standardized across skills

v1.2.0

Added

• Token and cost awareness sections across skill documentation
• Concrete examples for cost-aware patterns

Changed

• Version bumped to 1.2.0

v1.1.0

Added

• Native plugin architecture (plugin.json, marketplace.json)
• Automated lifecycle hooks (hooks.json)
• 8 verification scripts (validate-skill, security-sentinel, verify-gate, tsv-log, coverage-gate, dockerfile-lint, audit-deps, doc-health)

Changed

• Upgraded from standalone skills to full Claude Code plugin format
• Skills moved to ./skills/ directory with numbered prefixes
• Agents moved to ./agents/ directory

v1.0.0

Added

• Initial release
• 27 domain expert skills with Hard Gates, Iron Laws, Decision Trees, Failure Modes
• 6 strategic role agents (planner, architect, tech-lead, code-reviewer, security-reviewer, qa-engineer)
• Skills command (/virtual-company:skills) for listing and filtering experts
• CLAUDE.md with orchestration flow, decision trees, and failure modes
• MIT License