v1.2.0 — Engineer system overhaul, agentmemory, docs site, command refactor

Overview

opencode-onboard 1.2.0 — Complete rework of the engineer creation system, agentmemory integration, documentation site, and a major command refactor.

---

Major changes since v1.1.0

Engineer System Overhaul

/make-engineer — full rewrite

• Persona-driven interactive design: 9 personas + custom option (Frontend, Backend, DevOps, Data, Security, Mobile, API/Integration, QA, Layout Designer)
• Deterministic skill discovery: mandatory npx skills find per detected signal, no skipping
• 5-10 skills per engineer (was 2-5), minimum = detected signals
• Quality filter: install count ≥100, prefer ≥1000, prefer official sources
• Post-install verification: every @skill-name checked against .agents/skills/ + skills-lock.json
• Global/personal skills from ~/.claude/skills/ and ~/.agents/skills/ are FORBIDDEN
• Template enforcement: frontmatter + one identity paragraph (2-3 sentences) + ## Abilities only
• FORBIDDEN list: no ## Core Expertise, ## Key Patterns, ## Domain Expertise, ## FSD Layer Rules, ## Testing, ## Conventions, ## File Responsibilities, ## When Spawned, ## Gate Order, ## i18n, ## What You Do NOT Touch sections
• Critical rules preamble at top of command + hard gate before file writing
• Structural validation: re-reads file, checks only ## Abilities heading exists, no model: field, categorized abilities
• Idempotency: checks if engineer exists before overwriting

Agent mode enforcement

• Persona engineer templates: always mode: primary (was mode: all)
• Tier variants: always mode: subagent (injected by ob-subagent-tiers plugin at startup)
• ob-subagent-tiers plugin now fixes wrong modes in-place on disk + in-memory
• fullstack-engineer.md excluded from variant generation (it is mode: primary, never spawned)

fullstack-engineer.md — identity + additive skills

• New identity paragraph: "You are the default engineer, mostly used by the user for architecture and planning..."
• Skills merged additively (dedup by name), never overwrites existing abilities
• Preserves existing frontmatter (model, color, permissions)

fullstack-engineer banned from task assignment

• /plan-propose Step 2: NEVER assign fullstack-engineer to a task — it is mode: primary, not a spawned worker
• /plan-apply Step 7: NEVER spawn fullstack-engineer
• AGENTS.md, README.md, repo-onboard.md, agents-content.json: all references updated
• If no specialist matches a task, the user is told to create one with /make-engineer

Agentmemory Integration

• Replaced basic-memory with agentmemory (local, no server required)
• All commands and skills updated: context notes, task-result notes, change-context notes
• Skills-lock.json updated with agentmemory dependency

Command Refactor

• Renamed all commands from ob-* to clearer names:
  • /ob-apply → /plan-apply
  • /ob-propose → /plan-propose
  • /ob-archive → /plan-archive
  • /ob-explore → /plan-explore
  • /ob-init → /repo-initialize
  • /ob-help → /repo-help
  • /ob-create-engineer → /make-engineer
  • /ob-create-architecture → /make-architecture
  • /ob-create-design → /make-design
  • /ob-set-model → /make-user-model
  • /ob-create-project-guardrails → /make-guardrails
  • /ob-pullrequest → /ops-review (feedback mode) + /ops-ship (ship mode)
  • /ob-autopilot → /plan-goal
  • /ob-main → /plan-quick
• New /repo-onboard command for onboarding new team members
• New /plan-quick for small in-session tasks without the full plan flow

Platform Skills as Commands

• ob-pullrequest-az/gh/gl skills consolidated into /ops-ship and /ops-review commands with platform-specific presets
• ob-userstory-* skills remain as skills (loaded by the lead when a work item URL is detected)

Documentation Site

• New landing page at docs/ with GitHub Pages deploy workflow
• Custom CSS, demo GIF, logo, feature highlights
• Contributing guide (CONTRIBUTING.md) and security policy (SECURITY.md)

Plugin Improvements

• ob-subagent-tiers: write deduplication (skip if unchanged), fullstack-engineer excluded from variants, enforces mode: primary on templates + mode: subagent on variants
• ob-subagent-monitor: trimmed dead code, debounced persist

Update Command Restored

• npx opencode-onboard update — non-interactive re-apply from saved config
• Protects user-owned files from force-overwrite

Other Improvements

• opencode-onboard.json flattened (no wizard wrapper, version 2)
• source-roots.json moved to .opencode/
• Guardrails split: ob-guardrails-generic (git, secrets, code quality) + ob-guardrails-project (project-specific rules)
• Optimization markers in commands (RTK, codegraph, memory, caveman)
• Skills installer with skills-lock.json tracking
• opencode-json.js — merge agent block into existing config without touching other keys
• ESLint config added

---

Getting started

Initialize a project:
npx opencode-onboard

New team member joining:
npx opencode-onboard join

Then in OpenCode:
/repo-initialize
/plan-propose
/plan-apply
/ops-ship
/plan-archive

For autonomous loop-engineering:
/plan-goal
/plan-goal push
/plan-goal pr

Create a custom engineer:
/make-engineer

---

Full changelog since v1.1.0

• Complete rewrite of /make-engineer: deterministic skill discovery, template enforcement, validation
• Agent mode enforcement: templates primary, variants subagent
• fullstack-engineer excluded from variants and banned from task assignment
• Agentmemory replaces basic-memory (local, no server)
• All commands renamed from ob-* to plan-* / ops-* / make-* / repo-*
• New /repo-onboard, /plan-quick, /ops-ship, /ops-review commands
• Platform pull request skills consolidated into /ops-ship + /ops-review commands
• Documentation site with GitHub Pages
• Contributing guide + security policy
• ob-subagent-tiers plugin: dedup writes, exclude fullstack, enforce modes
• ob-subagent-monitor plugin: debounced persist, dead code removal
• update command restored with file protection
• Flattened opencode-onboard.json (version 2)
• Guardrails split: generic + project
• Optimization markers in commands
• ESLint config added
• 162 tests passing

v1.1.0 — Multi-platform support, MCP markers, tier plugin improvements

Overview

opencode-onboard 1.1.0 — Multi-platform support, MCP optimization markers, and tier plugin improvements.

---

Major changes since v1.0.0

Split Platform Choice: Backlog + Repo

• Onboarding now asks two questions: "Where is your backlog?" and "Where is your code repository?"
• Supports mixed-platform setups (e.g. Jira backlog + GitHub repo, Azure backlog + GitLab repo)
• Stores wizard.backlogPlatform and wizard.repoPlatform in opencode-onboard.json
• wizard.platform kept as alias for backward compat
• Skills installer handles both platforms (installs matching skills for each)
• AGENTS.md patched with mixed-platform workflow content

New Platforms

Jira (Atlassian) — backlog-only

• CLI: acli (Atlassian CLI)
• Fetch work items: acli jira workitem view --key PROJ-123
• Transition state: acli jira workitem transition --key PROJ-123 --status "In Progress"
• New skill: ob-userstory-jira
• Auth: API token or OAuth via acli jira auth login

GitLab — repo-only

• CLI: glab (GitLab CLI)
• Create merge requests: glab mr create
• List MRs: glab mr list --state merged
• Read MR comments: glab mr note list
• New skill: ob-pullrequest-gl
• New archive preset: ob-archive-gl.md
• Auth: glab auth login (OAuth or PAT)

Others (Browser) — backlog-only

• No CLI required — uses opencode-browser plugin to navigate to work item URLs
• For when users have no API tokens or use tools without CLI support (Linear, Trello, Monday, etc.)
• New skill: ob-userstory-browser
• Opens URLs the user explicitly provides, reads page text + accessibility snapshot
• browser-automation skill updated with explicit external navigation exception for work item URLs
• Read-only — never modifies anything on the page

Full Platform Matrix

Backlog	Repo
GitHub	GitHub
GitHub	GitLab
Azure DevOps	Azure DevOps
Azure DevOps	GitLab
Azure DevOps	GitHub
Jira	GitHub
Jira	Azure DevOps
Jira	GitLab
Others (Browser)	GitHub
Others (Browser)	Azure DevOps
Others (Browser)	GitLab
None	None

MCP Optimization Markers in Commands

• Command .md files now use marker pairs (OB-CMD-CODEGRAPH, OB-CMD-MEMORY, OB-CMD-RTK) around MCP-specific instructions
• During onboarding, the optimization step injects or strips these sections based on what the user selected (codegraph, basic-memory, RTK)
• No more "if available" hedging — the CLI controls whether MCP instructions exist at all
• All MCP references explicitly say "MCP tool, NOT CLI command" to prevent agents from running codegraph or basic-memory as bash commands

ob-subagent-tiers Plugin Improvements

• Writes physical *-engineer.<tier>.md variant files on disk AND injects in-memory
• Hardened file filter (^[\w-]+-engineer\.md$) prevents cascading duplicates on restart
• Error logging to stderr — visible feedback on what was injected

Basic-Memory Reliability

• experimental.mcp_timeout: 300000 added to opencode.json — fixes -32001 Request timed out on tool requests
• MCP command changed to basic-memory mcp (direct binary, no uvx resolution delay)

.gitignore npm Fix

• Source renamed to _gitignore so npm includes it in the package (npm strips .gitignore files during publish)
• Merge logic in copy step handles existing files

Other Improvements

• /ob-create-project-guardrails — new command extracting rules from ARCHITECTURE.md into a skill
• /ob-create-architecture and /ob-create-design — incremental update mode using timestamps + git log
• All three create commands now use codegraph MCP and basic-memory MCP with markers
• Autopilot modes: default (merge to main), push (push branch only), pr (push + create PR)
• /ob-set-model — user prefix for machine-local model overrides
• Model suggestions updated to match live models.dev IDs; added OpenCode Go provider
• Removed obsolete stampAgentModels call
• source-roots.json moved to .opencode/
• Join command: 10 steps covering all local installs and checks

---

Getting started

Initialize a project:
npx opencode-onboard

New team member joining:
npx opencode-onboard join

Then in OpenCode:
/ob-init
/ob-propose
/ob-apply
/ob-pullrequest
/ob-archive

For autonomous loop-engineering:
/ob-autopilot
/ob-autopilot push
/ob-autopilot pr

---

Full changelog since v1.0.0

• Split platform choice into backlog + repo for mixed-platform support
• Add Jira (Atlassian) as backlog-only platform with acli CLI
• Add GitLab as repo-only platform with glab CLI
• Add Others (Browser) as backlog-only platform using opencode-browser plugin
• MCP optimization markers in command files (codegraph, memory, RTK)
• All MCP references clarify "MCP tool, NOT CLI command"
• ob-subagent-tiers plugin: physical files + in-memory injection, hardened filter
• basic-memory: mcp_timeout fix, direct binary command
• _gitignore npm fix
• /ob-create-project-guardrails command
• /ob-create-architecture and /ob-create-design: incremental update mode
• Autopilot output modes (default/push/pr)
• /ob-set-model user prefix for local overrides
• Model suggestions updated, OpenCode Go added
• Join command: 10-step overhaul
• README fully updated with all platform combinations and skills

v1.0.0 — First stable release

Overview

opencode-onboard 1.0.0 — the first stable release. Everything you need to turn any codebase into a multi-agent development workflow powered by OpenCode native subagents.

---

What is opencode-onboard?

A single interactive wizard that prepares any codebase for AI-driven development. It wires together:

• OpenCode — the agent runtime and TUI
• OpenSpec — spec-driven change management
• codegraph — code intelligence for agents
• basic-memory — persistent knowledge graph across agent sessions
• Native parallel subagents — multi-agent execution via OpenCode's built-in task tool

Supports GitHub, Azure DevOps, or no platform at all.

---

Major features in this release

Multi-Agent Tier System (ob-subagent-tiers plugin)

• Agent files (*-engineer.md) are templates with no model: field
• On startup, the plugin reads wizard.models from config and injects tier variants (backend-engineer.build, basic-engineer.fast, etc.) into the live config — both in-memory and as physical .md files
• Model resolution: opencode-onboard.user.json (user override, gitignored) then opencode-onboard.json (team shared)
• /ob-set-model [user] — set models per tier. Use user prefix for machine-local overrides. Use current keyword for the active session model

Full Command Suite

• /ob-init — Initialize project (greenfield or brownfield)
• /ob-explore — Think through an idea before committing to a plan
• /ob-propose — Parse a work item or idea into a structured plan with agent + tier annotations
• /ob-apply — Implement tasks via parallel subagent waves
• /ob-autopilot — Fully autonomous pipeline. Default: merge to main + delete branch. push keyword: push branch only. pr keyword: push + create PR
• /ob-pullrequest — Create a PR or process PR feedback
• /ob-archive — Archive a completed OpenSpec change
• /ob-main — Quick direct implementation, no OpenSpec, no waves
• /ob-create-engineer — Create a custom specialist engineer from a description + skills.sh
• /ob-create-architecture — Generate ARCHITECTURE.md from the codebase
• /ob-create-design — Generate DESIGN.md from the design system
• /ob-create-project-guardrails — Generate a project-guardrails skill from ARCHITECTURE.md + project config
• /ob-set-model — Set model for a tier (team or user-local override)
• /ob-help — Show all commands

Onboarding Wizard

• Greenfield vs brownfield detection
• Platform selection (GitHub / Azure DevOps / None)
• Source scope selection for monorepos
• Model selection from 75+ providers via models.dev (including OpenCode Zen and OpenCode Go)
• Optimization toggles: RTK, codegraph, basic-memory, browser extension
• OpenSpec initialization
• Bootstrap AGENTS.md with platform-specific workflow injection

Join Command

• One-command setup for new team members joining an existing onboarded project
• 10 steps: platform CLI, plugins, skills, OpenSpec, basic-memory (uv tool install), codegraph index, RTK, browser extension, model tier guidance, .gitignore check
• Shows team model configuration and guides user to set local overrides

Reliability

• experimental.mcp_timeout: 300000 in opencode.json — fixes basic-memory timeout errors
• basic-memory MCP command uses direct binary (basic-memory mcp), no uvx delay
• .gitignore source renamed to _gitignore so npm includes it in the package
• Tier variant file filter hardened to prevent cascading duplicates on restart

---

Getting started

Initialize a project:
npx opencode-onboard

New team member joining:
npx opencode-onboard join

Then in OpenCode:
/ob-init
/ob-propose
/ob-apply
/ob-pullrequest
/ob-archive

For autonomous loop-engineering:
/ob-autopilot
/ob-autopilot push
/ob-autopilot pr

---

Full changelog since v0.8.6

• /ob-create-project-guardrails — new command extracting rules from ARCHITECTURE.md into a skill
• Join command overhaul — 10 steps covering all local installs and checks
• basic-memory — install via uv tool install, execute via basic-memory mcp
• ob-subagent-tiers plugin — writes physical variant files AND injects in-memory; hardened filter
• Autopilot modes — default (merge to main), push (push branch only), pr (push + create PR)
• /ob-set-model — user prefix for machine-local overrides
• Model suggestions updated to match live models.dev IDs; added OpenCode Go provider
• Removed obsolete stampAgentModels call
• .gitignore npm fix (_gitignore)
• source-roots.json moved to .opencode/

---

1.0.0 means the API, command surface, and config format are now stable.

Huge thanks to everyone who tested, reported bugs, and provided feedback throughout the 0.x cycle.

v0.8.6 — Fix .gitignore stripped by npm

Overview

Fixes a bug where .opencode/.gitignore was never copied to projects during onboarding.

Root Cause

npm automatically strips .gitignore files during npm publish. The file content/.opencode/.gitignore (containing node_modules, .ob-run.json, opencode-onboard.user.json, source-roots.json, *-engineer.*.md) was present in the repo but missing from the published package. The merge code ran but found no source file, so it silently skipped the write.

Fix

• Renamed content/.opencode/.gitignore → content/.opencode/_gitignore (npm includes it normally).
• Updated src/steps/copy/index.js to read from _gitignore and write as .gitignore in the destination project.
• Verified _gitignore appears in npm pack output.

v0.8.5 — OpenCode Go, tier plugin fixes, basic-memory reliability, updated model suggestions

Overview

Version 0.8.5 brings the ob-subagent-tiers plugin to a working state, adds OpenCode Go as a provider, fixes basic-memory MCP timeouts, and updates all model suggestions to match current models.dev IDs.

Major Changes

1. ob-subagent-tiers Plugin — Working In-Memory Injection

• Fixed the plugin to reliably inject tier-suffixed agent variants (backend-engineer.build, basic-engineer.fast, etc.) into cfg.agent via the config hook.
• Removed silent catch {} that was hiding errors — now logs to stderr exactly what was injected or why it failed.
• Template agent files (*-engineer.md) remain clean templates with no model: field.

2. OpenCode Go Provider Added

• Added opencode-go to the model provider list in presets/models.json.
• 19 models available (deepseek-v4-pro, qwen3.6-plus, kimi-k2.7-code, glm-5.1, etc.).
• Suggestions per tier: build → opencode-go/deepseek-v4-pro, fast → opencode-go/qwen3.6-plus.

3. Basic-Memory MCP Reliability

• Added experimental.mcp_timeout: 300000 to opencode.json — fixes -32001 Request timed out errors on tool calls (the per-server timeout only covered the connect handshake, not tool requests).
• Changed MCP command from uvx basic-memory mcp to uv tool run basic-memory mcp — uses the pre-installed binary directly, near-instant startup instead of 15-20s uvx resolution.

4. Model Suggestions Updated

• All suggestions now verified against live models.dev/api.json IDs.
• Fixed GitHub Copilot provider value: github → github-copilot (filter was broken — model IDs use github-copilot/ prefix).
• Stale suggestions fixed:
  • opencode/big-pickle → opencode/glm-4.7
  • github/gpt-4o-mini → github-copilot/gpt-5.4-nano
  • deepseek/deepseek-v3 → deepseek/deepseek-v4-pro
  • github/claude-sonnet-4-5 → github-copilot/claude-sonnet-4.5
• Added OpenCode Zen suggestions for plan and build tiers (were null before).

5. Removed Obsolete stampAgentModels

• The stampAgentModels call in the models step wrote model: to agent files during onboarding — obsolete now that the tier plugin handles model injection at startup.
• Agent files stay as clean templates; the plugin does the work.

6. .gitignore Copy Fix

• Fixed .opencode/.gitignore not being copied when a .gitignore already exists in the destination (opencode creates one for node_modules).
• Now merges both files (dedup lines) after the bulk copy step.

7. Source Roots Moved to .opencode/

• source-roots.json relocated from .agents/ to .opencode/ for consistency.
• All command references updated (/ob-create-architecture, /ob-create-design, AGENTS.md bootstrap).

v0.8.2 — Tier-based subagent models, autopilot, and user-local model overrides

Overview

Version 0.8.2 introduces a plugin-driven tier model system for subagents, an autonomous end-to-end pipeline command, and user-local model overrides — all designed to make multi-agent execution more flexible without mid-session file hacking.

Major Changes

1. Tier-Based Subagent Models via Plugin (ob-subagent-tiers.js)

• New plugin that injects tier-suffixed agent variants (backend-engineer.build, backend-engineer.fast, backend-engineer.plan) into the live config at startup via the config hook.
• Template agent files (*-engineer.md) no longer carry a model: field — they are pure templates.
• The plugin reads wizard.models from the config and bakes the right model into each variant at startup. No mid-session file editing, no git churn.
• Models are resolved from opencode-onboard.user.json (user override) first, then opencode-onboard.json (team shared).

2. /ob-set-model with User Overrides

• New user prefix: /ob-set-model user fast opencode/big-pickle writes to opencode-onboard.user.json (gitignored, machine-local).
• Without prefix: writes to opencode-onboard.json (team shared, git-tracked).
• No more agent file re-stamping — the plugin rebuilds tier variants on next restart.

3. /ob-autopilot — Autonomous Pipeline

• New command: branch off main → propose → apply → archive → merge back, with zero human interaction.
• Each phase produces its own commit. Stops only on hard failures (stall, verification failure, merge conflict), leaving the branch unmerged.
• Built for loop-engineering / unattended runs.

4. Updated /ob-propose and /ob-apply

• /ob-propose now annotates tasks with tier-suffixed agent names (e.g. backend-engineer.build, basic-engineer.fast).
• /ob-apply spawns the tier-suffixed name directly — the model is already resolved at startup by the plugin.
• Fallback: if a tier variant is missing, strips the suffix and uses the base template (inherits lead model).

5. /ob-create-engineer Simplified

• Agent files are templates with no model: field.
• Removed the model resolution step — the plugin handles it.
• Tier is still recorded for documentation and AGENTS.md.

6. ob-subagent-monitor Updated

• Now resolves models from wizard.models (user first, then team) instead of reading model: from agent files.
• Handles tier-suffixed agent names (e.g. backend-engineer.build → looks up wizard.models.build).

7. Source Roots Moved to .opencode/

• source-roots.json relocated from .agents/ to .opencode/ for consistency.
• All command references updated (/ob-create-architecture, /ob-create-design, AGENTS.md bootstrap).
• Added to .opencode/.gitignore.

Migration Notes

• After updating, run /ob-set-model <tier> <model> for each tier (plan, build, fast) and restart opencode — the plugin will rebuild all tier variants.
• Existing agent files with model: fields still work as before (the model is kept), but new agents created via /ob-create-engineer will not have model: in the file.

v0.7.0

Overview

Version 0.7.0 is a workflow shift focused on replacing opencode-ensemble with native OpenCode subagent waves, plus model management and archive-flow improvements.

Major Changes

1. Native Subagent Waves (Core Architecture Change)

• Replaced ensemble/worktree execution with native task-based parallel subagents.
• Updated apply flow to run dependency-aware waves with conflict-safe grouping.
• Added progress/recovery state handling through run-state tracking.

2. Command Workflow Redesign

• Reworked apply command around wave orchestration.
• Propose now enriches tasks with:
  • assigned agent
  • depends_on metadata
  • touches metadata
• Added pre-checks for unarchived changes before propose/explore.
• Updated help/main command language to match wave-based execution.
• Archive flow now prioritizes oldest eligible unarchived merged change.

3. Model Management Overhaul

• Removed ensemble model config flow.
• Moved to per-agent model stamping by tier (plan/build/fast).
• Added set-model command to update tier models and re-stamp engineers.
• Preserved onboarding runtime config as load-bearing metadata.

4. Onboarding and Join Improvements

• Join now performs stronger local setup checks:
  • plugin install/update
  • skills install
  • basic-memory prerequisites
• Improved platform detection and join messaging.
• Default max concurrent agents changed from 4 to 3.
• Concurrency now clamped to 1..5.

5. Plugin and Package Changes

• Removed ensemble plugin dependency.
• Added notifier and TUI support for native subagent visibility.

6. Reliability and Test Coverage

• Added significant tests across copy, metadata, models, optimization, and openspec flows.
• Added process exit wrapper usage to improve testability.
• Added codegraph timeout safeguards under parallel load.

Behavior Impact for Users

• Parallel execution is now native and wave-based instead of ensemble-based.
• Model assignment is tied to agent files/tier config, not ensemble mapping.
• Archiving and proposal flows are stricter and safer for sequencing.
• Setup and re-setup for team members is more automated and robust.

0.6.2 Initial usable version

Big update to the AI agent workflow setup. Here's what changed:
New /ob-* commands — cleaner workflow, no more opsx- confusion:

• /ob-propose — parses a GitHub/ADO work item or idea, generates a plan, and annotates every task with the best agent + model before you confirm
• /ob-apply — implements via parallel engineers using ensemble
• /ob-pullrequest — creates a PR or reads/classifies review comments
• /ob-explore — thinking partner before committing to a plan
• /ob-archive — close out a completed change
• /ob-help — shows all commands and when to use each
  Smarter task assignment — /ob-propose now reads every *-engineer.md in .opencode/agents/ (description + abilities) and picks the best specialist per task. Heavy tasks get the build model, light work (i18n, config, verification) gets the fast model.
  codegraph — code intelligence for agents — agents use codegraph_search and codegraph_impact to understand your codebase before spawning. No more agents guessing file locations or breaking things they didn't know were connected.
  basic-memory — shared context across sessions — agents store decisions, architecture notes, and findings. The next session picks up where the last one left off. Cross-agent context during /ob-apply means engineers share knowledge without stepping on each other.
  Model picker UX — 3 models (plan / build / fast), pick provider first then search. ← Change provider to go back if you pick wrong.
  Run again to get updates:

npx opencode-onboard@latest