Changelog

All notable changes to the Claude Commands Library.

[5.2.0] - May 2026

Real Claude Code subagents migration. Multi-agent research is now backed by actual Anthropic subagent infrastructure (isolated context per specialist, per-agent model routing, Task tool invocation), not prose simulation.

Added

5 real subagents in .claude/agents/ — research-explore (haiku, blue), research-pattern (haiku, green), research-security (sonnet, red), research-performance (sonnet, orange), research-citation (haiku, purple). Each carries valid Anthropic YAML frontmatter (name, description, tools, model, color) and a focused system-prompt body. See Anthropic frontmatter spec.
tests/validate-subagent-invocation.ps1 — verifies every @research-X mention in prompt-research/SKILL.md resolves to a real agent file, and every agent file is mentioned in the orchestrator skill.
tests/validate-no-aspirational-config.ps1 — verifies every top-level JSON config key has a live consumer (advisory).
docs/V5.2-PREFLIGHT.md — full migration plan that gated Phase B.

Changed

prompt-research/SKILL.md — Specialist Selection Rules section, Orchestration Flow rewritten as explicit Task-tool invocation pattern per Anthropic parallel research guidance, Performance Expectations inline, Aggregation step appends citation index entries from main thread (subagents are read-only).
research-adapter.md debloated 2 411 → 627 words. Kept only Phase 0 research-specific add-ons (complexity scoring, strategy selection, Step 4/5 overrides, performance expectations). Removed Integration Wiring (lives in SKILL), Configuration References (skill reads configs directly), Error Handling (Anthropic default + auto-compaction), Version History (CHANGELOG-skills.md), Related Components.
caching-strategy.md — Subagents-and-caching note added (separate cache lifecycle; do not list subagent files in main session cache_control blocks).
Installer v5.1.0 → v5.2.0 — deploys .claude/agents/; verifies all 5 expected research subagents present; v5.2 announcement block.
validate-library-references.ps1 v1.2 → v1.3 — adds Suite 6 (subagent existence) and Suite 7 (frontmatter validation).

Removed

library/research-agent-{explore,pattern,security,performance, citation}.md (5 files, ~6 000 words; content migrated to .claude/agents/).
config/agent-roles.json (408 lines): per-agent metadata distributed to subagent frontmatter, body content, and skill orchestration; all aspirational fields (timeout_seconds, performance., error_handling., parallel_execution., agent_integration., metadata.*) dropped.

Tools narrowing

Bash removed from research-explore and research-security allowlists. Read-only research is fully served by Read + Grep + Glob; Bash widened attack surface unnecessarily.
Glob added to research-security and research-performance for pattern-based file enumeration without resorting to Bash.

Anthropic alignment

"Subagents cannot spawn other subagents. If your workflow requires nested delegation, use Skills or chain subagents from the main conversation."
— Anthropic subagent docs

This is why orchestration (iteration loop, gap analysis, citation index persistence, aggregation) lives in prompt-research/SKILL.md (main thread), not in any subagent.

[5.0.0] - May 2026

Added

Session Memory — PreCompact hook writes a diary entry to .claude/memory/diary/YYYY-MM-DD-session-N.md before every compaction. Captures task summary, design decisions (with WHY), challenges, solutions, and observed user preferences. Per-project; never committed.
/reflect-diary skill — reads accumulated diary entries, identifies patterns (2+ occurrences), and proposes updates to project-profile.md. Nothing is applied without explicit user approval.
Global hook installer — install-claude-commands.ps1 now deploys pre-compact.sh to ~/.claude/hooks/ and registers the PreCompact event in ~/.claude/settings.json on any machine.

Removed

8 commands/skills deleted: /prompt-hybrid, /prompt-technical, /prompt-article, /prompt-dotnet, /prompt-react, /deploy, /new-stack, /reflect
5 config files deleted: cache-config.json, verification-config.json, learning-config.json, predictive-intelligence.json, ai-fluency.json
Library subdirectories flattened: adapters/, intelligence/, agents/, orchestration/ — all files moved to library/ root

Changed

Skills-only format — all three commands live in .claude/skills/ with YAML frontmatter; .claude/commands/ directory removed
Library flat structure — 14 files at library/ root: prompt-perfection-core.md, readme-adapter.md, research-adapter.md, caching-strategy.md, model-router.md, execution-plan-template.md, orchestration files, research agent files
install-claude-commands.ps1 — v5.0.0: removes commands/ on upgrade, checks 3 skills, deploys hooks/ directory, adds Install-GlobalHook function
README.md rewritten — 70 lines, honest 3-command description + Session memory section
CLAUDE.md updated — v5.0.0, 3-command table, flat library architecture
package.json version 5.0.0
.claude/skills/ now includes reflect-diary/SKILL.md
.claude/hooks/pre-compact.sh added as reference copy in repo

Kept

/prompt, /prompt-article-readme, /prompt-research — all maintained
Phase 0 shared foundation (prompt-perfection-core.md)
Interaction Protocol from v4.8 (global, CLAUDE.md)
Memory system (project-profile.md, sessions.md, prompt-patterns.md)
All VitePress documentation (updated to reflect 3-command scope)

Why

Repo had grown to 11 commands, most untested and unused in practice. Reduced to the three that are actually run regularly.

[4.9.0] - April 2026

Added

Opus 4.7 Optimisation — token + context savings targeted at claude-opus-4-7. Seven additive changes:
- Prompt caching strategy (.claude/library/caching-strategy.md) with cache_control: ephemeral breakpoints. ~90% input cost reduction on warm cache hits (5m default, 1h beta).
- Fast Path in Phase 0 — trivial prompts (score < 5) short-circuit to Steps 0.1 → 0.5 → 0.6. Saves ~40% Phase 0 tokens.
- Model tier split — opus → opus-fast (4.6, 200K ctx, 4K thinking) vs opus-smart (4.7, 1M ctx beta, 8K thinking, interleaved thinking beta). Router picks on depth-vs-latency.
- Per-tier thinking budget — haiku 0 / sonnet 2K / opus-fast 4K / opus-smart 8K in model-tiers.json.
- Context-editing adapter — clear_tool_uses_20250919 for /prompt-research multi-iteration loops.
- Memory-tool adapter — bridge to native memory_20250818 with Phase A/B/C/D progressive migration.
- Consolidated CHANGELOG-skills.md — Version History deduped across seven skills (~5K tokens saved per skill load).
- Batch API hint in /reflect (50% cost on SDK-driven runs).

Changed

prompt-perfection-core.md v2.1 → v2.2 (Fast Path; Mermaid → text)
model-router.md v1.0 → v2.0 (smart vs fast decision)
model-tiers.json v1.0 → v2.0 (opus split; caching; thinking modes)
All seven skills: Version History → CHANGELOG-skills.md pointer
install-claude-commands.ps1 v4.8.0 → v4.9.0

Notes

Backward compatible: legacy "opus" hints resolve to opus-smart
Adapters are opt-in; they document strategy for SDK integration

[4.8.0] - April 2026

Added

Interaction Protocol in CLAUDE.md — global session-level directive that applies to ALL interactions, not just /prompt* slash commands
Four rules enforced globally:
- Language — user píše po slovensky; Claude odpovedá po slovensky; tool calls / commits / code / docs v angličtine; technické termíny verbatim
- Plan-First Execution — pred akoukoľvek zmenou: summary + 2-3 options (pre non-trivial tasks) + execution plan + explicit approval
- Proactive Option-Finding — Claude actively navrhuje lepšie cesty PRED exekúciou; pomenuje tradeoff, odporúča, rozhodnutie user-a
- Never Auto-Execute — žiadne git commit, git push, npm install, installer runs bez explicitného súhlasu
New documentation page /guide/interaction-protocol covering the rules, rationale, exceptions, and worked examples
v4.8 What's New card on the docs home feature grid
Sidebar entry "Protocols > Interaction Protocol (v4.8)" under /guide/

Changed

CLAUDE.md: project version v4.7.0 → v4.8.0; Skills promoted to PRIMARY in the architecture section (was grouped with legacy commands)
install-claude-commands.ps1: v4.7.0 → v4.8.0 with new feature announcement block listing the four protocol rules
.claude/memory/MEMORY.md: updated to reflect v4.8.0

Rationale

Previously, plan-first discipline (Curiosity Gate, Options-First, Execution Plan, Approval Gate) lived only inside /prompt* skills. Normal conversation bypassed it — leading to unwanted auto-execution and token waste when assumptions were wrong. v4.8 fixes the root cause: CLAUDE.md is loaded into every session automatically, so the same discipline applies everywhere.

Migration

No breaking changes. Existing skills and commands continue to work. Users will notice Claude now asks for approval more consistently in free-form chat tasks (not only inside /prompt).

[4.7.0] - April 2026

Added

Dynamic model routing — new config model-tiers.json maps prompt types, complexity scores, and file counts to the best Claude tier (haiku / sonnet / opus)
Model Router (library/model-router.md) — decision algorithm used by every skill: hard de-escalation, hard escalation, by complexity score, by prompt type, by file count. Each skill emits a MODEL HINT line before the approval gate
Execution Plan Template (library/execution-plan-template.md) — mandatory pre-approval plan: Goal, Files, Steps, Tools, MODEL HINT, Risks, Verification, Estimated effort, Assumptions
Step 0.25 Curiosity Gate in Phase 0 — confidence score 0-100 + mandatory assumption ledger when confidence < 100%. Forces explicit uncertainty accounting
Step 0.35 Options-First — 2-3 alternatives shown BY DEFAULT for Task/Feature/Bug Fix/Refactor/Config (was "only when multiple exist"). Each option carries model tier + cost estimate
Step 0.55 Execution Plan + Model Selection — mandatory block before approval for all non-Question prompts
Approval gate: switch [tier] response — request a model change before approving, then re-approve

Changed

prompt-perfection-core.md: v2.0 → v2.1 (new steps, updated flowchart)
skills/prompt/SKILL.md: v4.0 → v4.1
skills/prompt-hybrid/SKILL.md: v4.0 → v4.1
skills/prompt-dotnet, prompt-react, reflect, deploy, new-stack: v2.0 → v2.1
Legacy commands: Phase 0 v2.1 Add-ons block added
Installer: v4.6.0 → v4.7.0, verifies all new library files + 7 skills

Performance

Estimated token savings: 30-45% via Haiku default for simple tasks
Clearer execution plans reduce re-prompting iterations

Upgrade Guide (4.6 → 4.7)

No breaking changes. Skills are backward compatible. Run the installer:

powershell

.\install-claude-commands.ps1

Or update an existing installation:

powershell

.\install-claude-commands.ps1 -Force

[4.6.0] - April 2026

Added

Anti-Hallucination Contract in prompt-perfection-core.md v2.0
- NEVER rules: prohibit inventing file paths, versions, config keys
- ALWAYS rules: require grounding every fact to a source file
- Grounding Protocol: map each stated fact to the file it was read from
HARD-GATE blocks in all 7 skills (prompt, prompt-hybrid, prompt-dotnet, prompt-react, reflect, deploy, new-stack)
- Each skill now has a pre-flight checklist that must pass before output
- Skill-specific checks (e.g. .csproj must be read before .NET facts)
NEVER sections in every skill — explicit per-domain hallucination prohibitions
Mermaid flowchart for Phase 0 flow in core library
Chain-of-Thought REASONING block required in Step 0.1 output
Few-shot examples in /prompt and /prompt-hybrid — correct vs incorrect output
Version history sections added to all skills that were missing them

Changed

prompt-perfection-core.md: v1.6 → v2.0
skills/prompt/SKILL.md: v3.0 → v4.0
skills/prompt-hybrid/SKILL.md: v3.0 → v4.0
All other skills: v1.0 → v2.0
Installer: v4.5.0 → v4.6.0, skills/ directory now verified, skills count shown in summary

Fixed

/prompt STARTUP template: removed hardcoded version v4.5.0 — now reads from project-profile.md

Upgrade Guide (4.5 → 4.6)

No breaking changes. Skills are backward compatible. Run the installer to get the latest version:

powershell

.\install-claude-commands.ps1

[4.5.0] - March 2026

Removed

/session-start and /session-end — deleted (both skills and old commands/ format)
- Replaced by Claude Code's built-in auto-memory system (~/.claude/projects/.../memory/)
- Auto-memory loads context into every conversation automatically without a command
- session-adapter.md library file also removed (orphaned after deletion)

Changed

Universal skills — all project-specific hardcoded values removed from all skills
- /deploy and /new-stack: removed lab463-web, linkvault-internal, postgresql-db hardcodes
- All server config now read dynamically from ~/.claude/memory/personal-profile.md
- personal-profile.md now requires: POSTGRES_CONTAINER, FRONTEND_NGINX_CONTAINER
- /new-stack Docker network references use [POSTGRES_STACK]-internal (dynamic)
- /new-stack Dockerfile .NET version now detected from .csproj TargetFramework
- /prompt-hybrid context display: reads from project-profile.md, no hardcoded project facts
- /prompt-hybrid threshold corrected: Very High is 20+ (was 15+, mismatched CLAUDE.md)
- /prompt-dotnet best practices header: version-agnostic (was ".NET 10")
- /prompt-dotnet Quick Reference: version-agnostic (was ".NET 10 Quick Reference")
/deploy: safer defaults
- Default output is now copy (show script) — was ambiguous between copy and execute
- Explicit guard: commands are NOT auto-executed without run confirmation
- DB migrations detection added to Phase 1 pre-deploy checks (was a footnote)
/reflect: scope and quality limits
- Steps 2+3 merged (removed duplication)
- Max 5 observations per session
- Priority field added to each observation
- Grouped approval (HIGH/MEDIUM) replaces per-item yes/no loop

Fixed

/prompt-dotnet: scan fallback when .csproj not found (3 options presented to user)
/prompt-react: scan fallback when package.json not found; monorepo support (searches frontend/, client/, web/, app/ subdirectories before failing)
/session-start: removed hardcoded version v4.2.0 and hardcoded stack info
/session-end: context limit warning added; sessions.md trimmed to max 10 entries; pattern detection scoped to current session only (no cross-session inference)
Installer: removed "private repository" error message (repo is public)
CLAUDE.md: commands table updated — removed session commands, added /deploy and /new-stack

Upgrade Guide (4.4 → 4.5)

/session-start and /session-end no longer exist. If you relied on them:

Context is now auto-loaded via Claude Code's built-in memory system
Historical sessions.md data is preserved (not deleted by installer)

powershell

.\install-claude-commands.ps1

[4.4.0] - March 2026

Added

/prompt-dotnet - .NET project-aware prompt perfection skill
- Scans .csproj, Program.cs, appsettings.json, Dockerfile, docker-compose.yml
- Auto-detects: TargetFramework, architecture (Minimal API vs Controllers), auth, ORM, DB, Docker
- Applies .NET 10 best practices automatically based on detected stack (EF Core, PostgreSQL, JWT, CORS)
- Presents recommendations consistent with existing project patterns
- Related commands: /prompt-react, /deploy
/prompt-react - React project-aware prompt perfection skill
- Scans package.json, vite.config.ts, tsconfig.json, src/ structure, .env
- Auto-detects: React version, TypeScript strict mode, router, state management, data fetching, base path
- Applies React best practices automatically (hooks, TS types, TanStack Query, Vite env vars, SPA subpath)
- Presents recommendations consistent with existing project patterns
- Related commands: /prompt-dotnet, /deploy

Fixed

Installer critical bug - skills/ directory was listed in $obsoleteDirs and was deleted on every update
- skills/ removed from $obsoleteDirs, added to $directoriesToDeploy
- All skills (prompt-dotnet, prompt-react, deploy, new-stack, etc.) now survive updates

Upgrade Guide (4.3 → 4.4)

No breaking changes. Two new skills added.

powershell

.\install-claude-commands.ps1

[4.3.0] - March 2026

Added

Skills Format - Commands migrated to new .claude/skills/ format
- YAML frontmatter: name, description, disable-model-invocation, argument-hint
- description field enables Claude to auto-invoke skills when contextually relevant
- disable-model-invocation: true on workflow commands (session-start, session-end, reflect)
- New skills: prompt/SKILL.md, prompt-hybrid/SKILL.md, session-start/SKILL.md, session-end/SKILL.md, reflect/SKILL.md
- Old .claude/commands/ files still work unchanged (backward compatible)
STARTUP Section in all prompt skills - Project context loaded before any analysis
- Reads project-profile.md, sessions.md, prompt-patterns.md on every invocation
- Shows "CONTEXT LOADED" summary so user sees what is pre-filled
- Eliminates repeated context questions across sessions
project-profile.md v2.0 - Populated with real project data
- Previously contained only empty section headers
- Now includes: project identity, full tech stack, infrastructure, project structure, command list, architecture overview, user preferences, workflows, known gotchas
- All Phase 0 commands immediately benefit via Step 0.2a Memory Recall
Auto Memory (~/.claude/projects/.../memory/MEMORY.md)
- Created MEMORY.md with project key facts for new session bootstrapping
- Follows Claude Code native auto-memory format (first 200 lines loaded per session)

Changed

prompt-perfection-core.md → v1.6
- Step 0.2a renamed to "ALWAYS LOAD FIRST" with stronger emphasis
- Now reads three memory sources: project-profile.md + sessions.md + prompt-patterns.md
- Shows "CONTEXT LOADED" brief summary on every invocation
- Backward compatible: still works without any memory files
CLAUDE.md - Memory Recall section updated
- References now-populated project-profile.md v2.0
- Lists all three memory files to check before asking questions
- Documents .claude/skills/ as new preferred format
session-end skill - Now includes check for project-profile.md updates
- After each session, checks if new permanent facts should be added to profile
- Ensures project-profile.md stays current with project evolution

Improved

Commands no longer ask for project context the user has already provided
Skills with description field are suggested by Claude when contextually relevant
Session continuity improved through richer memory loading at startup

Upgrade Guide (4.2 → 4.3)

No breaking changes. New .claude/skills/ directory works alongside existing .claude/commands/.

powershell

# Optional: re-run installer to get latest files
.\install-claude-commands.ps1

[4.2.0] - February 2026

Added

Memory Recall System
- New .claude/memory/project-profile.md structured fact store
- Phase 0 Step 0.2a reads profile before completeness check
- Pre-fills known facts with "(from project profile)" attribution
- Step 0.3 skips questions already answered by profile
- First-use opt-in prompt for profile creation
Profile Extraction in /session-end (v3.0)
- New Step 1.5 extracts structured facts from session
- Categories: Infrastructure, Tech Stack, Deployment, Project Structure, User Preferences, Workflows
- Merges without duplicating existing facts
- Shows extraction summary
Memory Recall in CLAUDE.md
- Root CLAUDE.md instructs Claude to check profile for all interactions
- Works outside slash commands too

Changed

project-knowledge.md slimmed from 440 to ~40 lines (headers only, grows on use)
architectural-context.md slimmed from 653 to ~45 lines (headers only, grows on use)
Installer (v4.2.0) now deploys new memory files on update without overwriting existing ones
Installer deploys all memory templates on fresh install

Improved

All Phase 0 commands gain memory recall automatically (no individual command changes needed)
Fewer repeat questions across sessions

[4.1.0] - January 2026

Added

AI Fluency Framework Integration (Anthropic's 4Ds Model)
- Delegation Assessment - Explicit human vs AI task distribution
  - Problem Awareness (goal clarity, scope definition, success criteria)
  - Platform Awareness (AI capabilities matching)
  - Task Delegation (autonomous, review, collaborative, human-only modes)
- Interaction Mode Detection - Automation, Augmentation, or Agency modes
- Performance Description - AI behavior preferences (concise/detailed, formal/casual)
- Discernment Hints - Product, Process, and Performance evaluation criteria
- Diligence Summary - Track AI-generated content requiring verification
- New configuration: .claude/config/ai-fluency.json
Skill Reflection System (/reflect)
- Analyze conversations and propose skill improvements
- Signal detection: corrections, successes, edge cases, preferences
- Priority-coded change proposals (HIGH/MED/LOW)
- Direct skill file modifications with user approval
- Observation persistence for later review
- Integration with learning system

Improved

Phase 0 completeness check expanded from 6 to 9 criteria
Learning system now tracks reflection signals
Better integration between commands
/session-end now includes Diligence Summary section

[4.0.0] - January 2026

Added

Predictive Intelligence (Phase 0.15)
- Journey stage detection (6 stages)
- Proactive domain warnings
- Pattern recognition
- Relationship mapping to previous work
- Next-steps prediction with scope options
- Smart scoping (Focused/Balanced/Comprehensive)
Configuration files
- predictive-intelligence.json for Phase 0.15
- Enhanced learning-config.json

Changed

/prompt now includes LITE predictive intelligence
/prompt-hybrid includes FULL predictive intelligence
Improved complexity detection accuracy

Documentation

VitePress documentation site launched
Comprehensive architecture documentation
Full configuration reference

[3.0.0] - December 2025

Added

Multi-Agent Research System (/prompt-research)
- 2-5 specialized agents working in parallel
- 2-4 iteration cycles with gap-driven refinement
- Lead agent orchestration
- Consensus analysis and conflict resolution
External Memory
- Persistent knowledge graph
- Research history tracking
- Cross-session relationship mapping
New Configuration Files
- orchestration-config.json - Research strategies
- iteration-rules.json - Convergence criteria
- agent-roles.json - Agent definitions
- citation-config.json - Citation formatting
- external-memory-config.json - Memory persistence

Changed

Agent result caching now includes iteration checkpoints
Learning system tracks agent effectiveness
Improved multi-agent verification

[2.0.0] - November 2025

Added

Hybrid Intelligence System
- Automatic complexity detection
- Intelligent agent spawning
- Configurable thresholds
Agent Types
- Explore Agent (Haiku, 30s)
- Plan Agent (Sonnet, 60s)
- Security Agent (Sonnet, 45s)
- Performance Agent (Sonnet, 45s)
- Pattern Agent (Haiku, 30s)
Caching System
- Content-based cache keys
- Automatic invalidation
- 10-20x speedup on cache hits
Learning System
- Pattern tracking
- User preference detection
- Smart defaults after 3+ occurrences
Multi-Agent Verification
- 2-3 agents parallel verification
- Consensus analysis
- Conflict detection
Configuration Files
- complexity-rules.json
- agent-templates.json
- cache-config.json
- verification-config.json
- learning-config.json

Changed

All commands now use library system
/prompt-hybrid replaces manual complexity selection
/prompt-technical auto-spawns agents for complex tasks

[1.0.0] - October 2025

Added

Core Commands
- /prompt - Basic prompt perfection
- /prompt-technical - Technical implementation focus
- /prompt-article - Article writing wizard
- /prompt-article-readme - README generator
- /session-start - Load session context
- /session-end - Save session context
Library Architecture
- prompt-perfection-core.md - Canonical Phase 0
- Adapter system for domain customization
- Centralized library for consistency
Phase 0 System
- 6-step prompt perfection flow
- Completeness check (6 criteria)
- Clarification questions
- Approval gate
Session Management
- Session persistence in .claude/memory/
- Context aggregation
- Cross-session continuity

Version History Summary

Version	Date	Highlight
4.4	Mar 2026	.NET + React Project-Aware Skills
4.3	Mar 2026	Skills Format + Project-Aware Commands
4.2	Feb 2026	Memory Recall System
4.1	Jan 2026	Skill Reflection System
4.0	Jan 2026	Predictive Intelligence
3.0	Dec 2025	Multi-Agent Research
2.0	Nov 2025	Hybrid Intelligence
1.0	Oct 2025	Initial Release

Upgrade Guide

From 4.1 to 4.2

Re-run installer:
powershell
```
.\install-claude-commands.ps1
```
1
The installer automatically adds new project-profile.md to your memory directory
All existing commands work unchanged - memory recall activates on first use

From 3.x to 4.x

Update library files:
powershell
```
.\install-claude-commands.ps1
```
1
Add new config file:
- Copy predictive-intelligence.json to .claude/config/
Existing commands work unchanged

From 2.x to 3.x

Re-run installer:
powershell
```
.\install-claude-commands.ps1
```
1
Add new config files:
- orchestration-config.json
- iteration-rules.json
- agent-roles.json
- citation-config.json
- external-memory-config.json

Create memory directories:

powershell

New-Item -ItemType Directory -Path .claude/memory -Force

From 1.x to 2.x

Back up memory files:

powershell

Copy-Item .claude/memory/* ./memory-backup/

Re-run installer:
powershell
```
.\install-claude-commands.ps1
```
1
Add configuration files from defaults

Restore memory:

powershell

Copy-Item ./memory-backup/* .claude/memory/

Backward Compatibility

All versions maintain backward compatibility:

v1.0 commands work in v4.2
Configuration is additive (new files, not breaking changes)
Memory files persist across upgrades
Project profile is opt-in (commands work without it)

Deprecation Notices

Planned for v5.0

None currently planned

Removed in v4.0

None (all features maintained)

Installation - Latest install instructions
Configuration - All config options
Troubleshooting - Common issues

Changelog ​

[5.2.0] - May 2026 ​

Added ​

Changed ​

Removed ​

Tools narrowing ​

Anthropic alignment ​

[5.0.0] - May 2026 ​

Added ​

Removed ​

Changed ​

Kept ​

Why ​

[4.9.0] - April 2026 ​

Added ​

Changed ​

Notes ​

[4.8.0] - April 2026 ​

Added ​

Changed ​

Rationale ​

Migration ​

[4.7.0] - April 2026 ​

Added ​

Changed ​

Performance ​

Upgrade Guide (4.6 → 4.7) ​

[4.6.0] - April 2026 ​

Added ​

Changed ​

Fixed ​

Upgrade Guide (4.5 → 4.6) ​

[4.5.0] - March 2026 ​

Removed ​

Changed ​

Fixed ​

Upgrade Guide (4.4 → 4.5) ​

[4.4.0] - March 2026 ​

Added ​

Fixed ​

Upgrade Guide (4.3 → 4.4) ​

[4.3.0] - March 2026 ​

Added ​

Changed ​

Improved ​

Upgrade Guide (4.2 → 4.3) ​

[4.2.0] - February 2026 ​

Added ​

Changed ​

Improved ​

[4.1.0] - January 2026 ​

Added ​

Improved ​

[4.0.0] - January 2026 ​

Added ​

Changed ​

Documentation ​

[3.0.0] - December 2025 ​

Added ​

Changed ​

[2.0.0] - November 2025 ​

Added ​

Changed ​

[1.0.0] - October 2025 ​

Added ​

Version History Summary ​

Upgrade Guide ​

From 4.1 to 4.2 ​

From 3.x to 4.x ​

From 2.x to 3.x ​

From 1.x to 2.x ​

Backward Compatibility ​

Deprecation Notices ​

Planned for v5.0 ​

Removed in v4.0 ​

Related ​

Changelog

[5.2.0] - May 2026

Added

Changed

Removed

Tools narrowing

Anthropic alignment

[5.0.0] - May 2026

Added

Removed

Changed

Kept

Why

[4.9.0] - April 2026

Added

Changed

Notes

[4.8.0] - April 2026

Added

Changed

Rationale

Migration

[4.7.0] - April 2026

Added

Changed

Performance

Upgrade Guide (4.6 → 4.7)

[4.6.0] - April 2026

Added

Changed

Fixed

Upgrade Guide (4.5 → 4.6)

[4.5.0] - March 2026

Removed

Changed

Fixed

Upgrade Guide (4.4 → 4.5)

[4.4.0] - March 2026

Added

Fixed

Upgrade Guide (4.3 → 4.4)

[4.3.0] - March 2026

Added

Changed

Improved

Upgrade Guide (4.2 → 4.3)

[4.2.0] - February 2026

Added

Changed

Improved

[4.1.0] - January 2026

Added

Improved

[4.0.0] - January 2026

Added

Changed

Documentation

[3.0.0] - December 2025

Added

Changed

[2.0.0] - November 2025

Added

Changed

[1.0.0] - October 2025

Added

Version History Summary

Upgrade Guide

From 4.1 to 4.2

From 3.x to 4.x

From 2.x to 3.x

From 1.x to 2.x

Backward Compatibility

Deprecation Notices

Planned for v5.0

Removed in v4.0

Related