v4.4.1 — Gentle-AI parity sync + compact rules rollout

Highlights

• Restore parity with the latest mirrored gentle-ai assets across skills, examples, installers, and docs.
• Add skill-resolver, go-testing, and skill-creator, and align delegated skill loading around compact rules injected as ## Project Standards (auto-resolved).
• Ship examples/opencode/AGENTS.md, update OpenCode configs to reference it, and expand setup/install flows plus tests to the full 15-skill set.

Verification

• Ran bash scripts/install_test.sh
• Result: 41/41 tests passed

v4.4.0 — Context-inflation delegation + skill resolution alignment

Context-Inflation Delegation + Skill Resolution Alignment

Parity with gentleman-ai v1.8.12.

New Delegation Model

Replaced 5 fragmented sections (Delegation Rules, Hard Stop, Delegate-First, Anti-Patterns, Task Escalation) with one evidence-based decision table:

Action	Inline	Delegate
Read to decide/verify (1-3 files)	✅	—
Read to explore (4+ files)	—	✅
Read as prep for writing	—	✅ with the write
Write atomic (one file, known change)	✅	—
Write with analysis (multi-file, new logic)	—	✅
Bash for state (git, gh)	✅	—
Bash for execution (test, build)	—	✅

Core principle: does this inflate my context without need? If yes → delegate.

Skill Resolution

Orchestrator passes skill paths when available (optimal). Sub-agents fall back to self-discovery if none provided (resilient). Aligned with `sdd-phase-common.md` Section A — no contradictions.

Full Changelog: v4.3.1...v4.4.0

v4.3.1 — Compact prompts + judgment-day skill

Compact Prompts ~17% + Judgment Day Skill

Parity with gentleman-ai v1.8.9.

Prompt Compaction

All shared convention files optimized for token efficiency:

• `engram-convention.md`, `openspec-convention.md`, `persistence-contract.md`, `sdd-phase-common.md`
• Tables → bullet lists, redundant emphasis removed, verbose sentences tightened
• All enforcement language preserved (MUST, CRITICAL, BREAKS, MANDATORY)

New Skill: judgment-day

Parallel adversarial dual-review protocol:

1. Launches two blind judge sub-agents simultaneously
2. Synthesizes findings into Confirmed / Suspect / Contradiction categories
3. Delegates fixes, re-judges — max 2 iterations before escalation

Validated with its own protocol — both judges passed clean on this release.

Full Changelog: v4.3.0...v4.3.1

v4.3.0 — Token optimization + executor boundary

~63% Token Reduction + Executor Boundary

Parity with gentleman-ai v1.8.3. Massive token savings by extracting shared boilerplate.

Token Optimization (-292 lines, ~12,900 fewer tokens/session)

Expanded `sdd-phase-common.md` with 4 shared sections that all 9 phase skills reference instead of duplicating:

• A — Skill Loading (registry discovery)
• B — Artifact Retrieval (engram two-step pattern)
• C — Artifact Persistence (engram/openspec/hybrid/none)
• D — Return Envelope (structured response format)

Executor Boundary

Sub-agent prompts now enforce: "You are an SDD executor, not the orchestrator. Do NOT delegate, Do NOT call task/delegate, Do NOT launch sub-agents."

Skill registry self-lookup is explicitly SKILL LOADING, not delegation.

Full Changelog: v4.2.1...v4.3.0

v4.2.1 — Self-sufficient sub-agents for skill discovery

Self-Sufficient Sub-Agents

Sub-agents now discover and load skills on their own via their SKILL.md files, which include a Step 1 for loading the skill registry:

1. `mem_search(query: "skill-registry", project: "{project}")` — engram first
2. Fallback: `.atl/skill-registry.md` — local file

The orchestrator still pre-resolves paths when possible (optimal). Sub-agents self-discover when no paths are provided (resilient fallback).

Changes

• `docs/sub-agents.md` — updated skill registry section
• Orchestrator prompt — removed mandatory SKILL LOADING injection block

Full Changelog: v4.2.0...v4.2.1

v4.2.0 — Per-agent model routing fix in delegate()

Plugin Fix: delegate() Now Passes Agent Model

The real fix for per-phase model routing. `session.prompt()` always accepted `model: { providerID, modelID }` — the plugin just never passed it.

The Fix

• `resolveAgentModel()` reads the agent's `model` from `opencode.json`, parses `"provider/model-id"` into the SDK shape, and passes it to `session.prompt()`
• Backward compatible — if no model is configured, behavior is unchanged
• Example configs now use `<your-provider/your-model>` placeholders instead of hardcoded values

Migration

Update the plugin file. If you already have `model` per agent in `opencode.json`, it works automatically.

Full Changelog: v4.1.1...v4.2.0

v4.1.1 — Per-agent model routing fix

Per-Agent Model Field + Routing Documentation

Reported by @BismarckCerda — sub-agents in OpenCode ignored their configured model and always used the orchestrator's model.

Root Cause

• Example configs lacked a `model` field per agent
• OpenCode SDK's `session.prompt()` wasn't receiving model information from the plugin

Changes

• Added `model` field to all 10 agents in both example configs
• Documented model routing behavior in `docs/sub-agents.md`

Note: This release added the config field. The actual plugin fix came in v4.2.0.

Full Changelog: v4.1.0...v4.1.1

v4.1.0 — Background agents plugin + unified configs + delegate-first

🚀 What's New in v4.1.0

Async background delegation for OpenCode, unified single/multi configs, and delegate-first orchestration policy.

---

🔌 Background Agents Plugin

Adds the background-agents plugin for OpenCode, enabling async sub-agent delegation:

Tool	What it does
delegate(prompt, agent)	Launch a sub-agent async — returns ID immediately
delegation_read(id)	Retrieve the full result of a completed delegation
delegation_list()	List all delegations (running + completed)

• Sub-agents run in the background while the orchestrator continues working
• Results persist to disk — survive compaction, crashes, and restarts
• <task-notification> arrives when delegation completes
• 15-minute timeout per delegation
• Adapted from kdcokenny/opencode-background-agents (MIT)

New files: examples/opencode/plugins/background-agents.ts + BACKGROUND-AGENTS-README.md

🔄 Unified Single/Multi Configs

opencode.single.json and opencode.multi.json are now structurally identical:

Config	Agents	Delegate tools	Purpose
single.json	10 (orchestrator + 9 SDD)	✅	Ready to use — all agents use default model
multi.json	10 (orchestrator + 9 SDD)	✅	Template — customize model per agent

Previously, single mode had 1 agent with mode: "all" and no sub-agents. Now both modes have the full agent team with async delegation support.

⚡ Delegate-First Orchestration

New delegation policy across all platforms (Claude Code, Gemini CLI, Antigravity, OpenCode):

Situation	Tool
Sub-agent work where you can continue	delegate — always
Parallel phases (e.g., spec + design)	delegate × N — launch all at once
You MUST have the result before your next step	task — only exception
User is waiting and there's nothing else to do	task — acceptable

The default is delegate. You need a REASON to use task.

🐛 Bug Fixes

• install.sh: VS Code path pointed to .copilot/ instead of .vscode/ — fixed
• install.sh: branch-pr and issue-creation skills were silently skipped during install — fixed
• install_test.sh: Updated expected skill count from 9 to 12, all 41 tests passing

📦 Setup Script Changes

• Plugin + unique-names-generator npm dependency installed automatically for both modes
• No manual plugin setup required — setup.sh and setup.ps1 handle everything

---

Full Changelog: v4.0.0...v4.1.0

v4.0.0 — Issue-first enforcement, token optimization, and Hard Stop Rule

🚀 What's New in v4.0.0

Major release with three pillars: contribution enforcement, token optimization, and delegation discipline.

---

🔒 Issue-First Enforcement System

Every contribution now follows a strict workflow: Issue → Approval → PR → Review → Merge.

• YAML issue templates with required fields, pre-flight checkboxes, and environment dropdowns
• Blank issues disabled — contributors must use Bug Report or Feature Request templates
• PR validation CI (pr-check.yml) — 4 automated checks gate every PR:
  • Issue reference (Closes #N)
  • Issue has status:approved label
  • PR has exactly one type:* label
  • Shellcheck on modified scripts
• Label system: type:* (6 types), status:* (needs-review, approved), priority:* (high/medium/low)
• CONTRIBUTING.md with full workflow documentation
• PR template with automated checks table, changes table, and notes for reviewers

⚡ Token Economics Optimization (-38% per SDD pipeline)

Three independent analyses measured real token costs. Six optimizations implemented:

Optimization	Savings
Remove redundant persistence-contract reads	~22,000T
Artifact size budgets (word limits)	~14,000T
Skill registry pre-resolution by orchestrator	~11,400T
Common boilerplate extraction	~4,200T
Orchestrator doc compression	~4,200 chars
Parallel engram reads	~800T

Full analysis: docs/token-economics.md

🛑 Hard Stop Rule

New zero-exception delegation enforcement:

• Orchestrator NEVER uses Read/Edit/Write/Grep on source files
• "It's just a small change" is explicitly NOT a valid excuse
• Anti-patterns restored as 4 explicit DO NOT bullets for AI compliance

📋 New Skills

• issue-creation — Full issue template reference, CLI examples, maintainer approval workflow
• branch-pr — PR workflow with branch naming regex, commit message regex, automated checks reference

📖 AGENTS.md Skill Index

New root-level file mapping all 12 skills to their triggers and file paths.

🔧 Standards

• Branch names: ^(feat|fix|chore|docs|style|refactor|perf|test|build|ci|revert)\/[a-z0-9._-]+$
• Commits: ^(build|chore|ci|docs|feat|fix|perf|refactor|revert|style|test)(\([a-z0-9\._-]+\))?!?: .+

---

Breaking Changes

• Skill registry ownership changed: The orchestrator now pre-resolves skill paths and passes them to sub-agents. Sub-agents no longer search for the skill registry themselves. Update your CLAUDE.md/GEMINI.md if you override the orchestrator instructions.
• persistence-contract.md no longer required reading: Sub-agents have inline instructions. The file remains as human reference only.

Files Changed

• 6 new files: .github/ templates + workflows, CONTRIBUTING.md, AGENTS.md, docs/token-economics.md, skills/_shared/sdd-phase-common.md
• 2 new skills: skills/issue-creation/, skills/branch-pr/
• Modified: all 8 SDD skills, all 5 example configs, README.md, persistence-contract.md

v3.4.1 — Skill loading for all delegated tasks

What's new

Skill loading for non-SDD tasks

Sub-agents delegated for any task (not just SDD phases) now receive skill registry awareness in their prompt. They know where to find coding and workflow skills (mem_search("skill-registry") or .atl/skill-registry.md) and can load them based on task context.

Before: Only SDD phases had skill loading instructions. A sub-agent asked to write React code wouldn't know that react-19, typescript, or tailwind-4 skills existed.

After: Every sub-agent is informed that skills may be available and where to find them. It checks the registry, sees which skills match its task, and loads them.

Multi-model subagent prompts updated

Subagent prompts in multi-model mode now explicitly mention the skill registry step: "It includes a Step 1 for loading the skill registry — do NOT skip it."

README: real model format

Multi-model examples now use the actual provider/model-id format (e.g., anthropic/claude-sonnet-4-6, google/gemini-2.5-flash) instead of generic placeholders.