Personal, self-contained OpenCode configuration. LOTR-themed agents, custom orchestration runtime, dynamic context pruning, and a tuned set of skills/commands for Wpromote engineering work.

Recent shift: This config used to layer on top of the @wpro-eng/opencode-config team plugin. That dependency has been removed — everything below is now defined locally in ~/.config/opencode/.

~/.config/opencode/
├── opencode.json                # Model, instructions, hidden agents, permissions, npm plugins
├── dcp.jsonc                    # Dynamic Context Pruning config (currently in trial/manual mode)
├── package.json                 # npm deps (@opencode-ai/plugin) — Node 24+, npm 11.x
├── package-lock.json            # tracked for reproducible installs
├── .nvmrc                       # pins Node version (24)
├── agent/                       # Local agent definitions (LOTR-named)
│   ├── gandalf.md               # Primary orchestrator
│   ├── legolas.md               # Codebase exploration subagent
│   ├── radagast.md              # External docs/OSS research subagent (GPT-5.5)
│   ├── treebeard.md             # Read-only planner/reviewer subagent
│   ├── celebrimbor.md           # Implementation subagent
│   └── elrond.md                # Architecture/tradeoff advisor subagent
├── command/                     # Slash commands that drive skills
│   ├── ticket.md, plan.md, ac-quality.md, impl-plan.md
│   ├── review.md, check-ac.md, review-plan.md, sonar.md
│   ├── bug-hunt.md
│   └── update-opencode-deps.md  # check & update opencode config deps
├── skill/                       # Local skill library (11 skills)
│   ├── bug-hunter/, chrome-devtools/, figma/, gh-fetch-pr-comments/
│   ├── github-review-analyzer/, jira-enhance/, jira-ticket/, pr-review/
│   ├── sonarcloud/, tdd/, ticket-plan/
├── instruction/                 # Auto-loaded into every agent's context
│   ├── repo-context.md, codebase-map.md, orchestration-runtime.md
│   ├── script-usage.md
├── mcp/                         # Reference-only per-server JSON snippets (NOT auto-loaded)
│   ├── chrome-devtools/, context7/, exa/, figma/  # source-of-truth lives in opencode.json
├── plugins/                     # Local TypeScript plugins
│   ├── orchestration.ts         # Task tracking + runtime commands
│   └── vision-tool.ts           # Large-image / PDF / video vision via Gemini
└── logs/                        # DCP debug logs (gitignored)

All agents are defined locally under agent/. The default OpenCode build and plan agents are explicitly hidden in opencode.json so the LOTR roster is the surface area.

Why Radagast uses a different model: GPT-5.5 with xhigh reasoning effort has been stronger in practice for web research and synthesizing external docs across many fetched pages. Every other agent stays on Claude Opus 4.7 for code-centric reasoning consistency. This is the only intentional model split — revisit if/when a Claude release closes the research-quality gap.

Treebeard is locked down: permission denies write and edit, and bash is ask by default with only git diff, git log, and grep whitelisted. He's the safety-rails agent — never runs anything destructive.

Commands live in command/*.md and are thin wrappers around skills.

MCP servers are defined inline in opencode.json under the top-level mcp key — that is the only location OpenCode reads. The mcp/<name>/mcp.json files in this repo are reference snippets only (kept for diffability and as paste-ready blocks); they are not auto-discovered. Any change must be made in opencode.json to take effect.

Chrome and Figma both require local processes. The chrome-devtools skill teaches the launch/check workflow on demand (auto-running ~/code/scripts/agent/chrome_mcp.sh when needed); the figma skill prefers the desktop MCP and falls back to navigating Figma in Chrome.

Verify registration after edits with opencode mcp list.

Three plugins are loaded — two local TypeScript files and one npm package.

Lightweight task tracking and timeout enforcement for subagent work. Provides the /tasks, /task, /diagnostics, /continue, /stop runtime commands. Configuration is inline in the file:

• TASK_TIMEOUT_MS = 180_000 (3 minutes per subagent task)
• MAX_TRACKED_TASKS = 50 (in-memory ring buffer)

This is a stripped-down replacement for the old Wpromote orchestration plugin — intentionally omits tmux integration, persistent storage, provider fallback, and notification buffering. Add back as needed.

Exposes a vision tool that bypasses Claude's 8000px image limit by routing files to a Gemini agent. Supports JPEG/PNG/GIF/WebP/HEIC/BMP, PDFs, MP4/MOV/AVI/WebM, and WAV/MP3/OGG. Use it instead of read for any media file.

Dynamic Context Pruning. Configured via dcp.jsonc — see next section for trial details.

Currently running in trial / manual mode while behavior is validated. Settings in dcp.jsonc:

Tuned for GitHub Copilot's ~128K effective context (defaults assume 200K+):

• maxContextLimit: 80000 — strong compression nudges above this
• minContextLimit: 40000 — no nudges below this
• nudgeFrequency: 5, iterationNudgeThreshold: 15, nudgeForce: "soft"
• protectUserMessages: false — allows compression of large pasted content

Trial graduation checklist (post-validation):

1. Flip manualMode.enabled → false.
2. Flip debug → false.
3. Soften pruneNotification (or move to pruneNotificationType: "log").
4. Set compress.showCompression → false.
5. Revisit experimental.allowSubAgents and experimental.customPrompts.
6. Consider raising maxContextLimit if Copilot effective window grows.

Default model: github-copilot/claude-opus-4.7. No auto-update — run opencode models and edit opencode.json when a newer Opus ships.

Hidden built-in agents: build and plan are hidden via opencode.json. The LOTR roster covers their roles (celebrimbor for build, treebeard for plan).

External directory permissions: opencode.json allowlists ~/code/wpromote/* and ~/code/scripts/* so agents can operate across all team repos and the shared scripts dir without per-call prompts.

LOTR identity in prompts: Each agent's system prompt opens with both names ("You are Gandalf, the orchestrator") so the role is unmistakable regardless of how the agent is invoked.

Skill/agent activation cost: All assets are enabled at all times; skills only inject context when triggered, and agents only consume tokens when invoked. Keeping the full set on has no idle cost.

This config uses npm (not Bun) and pins all dependency versions for reproducibility. Recent OpenCode versions install plugin/config dependencies via npm/Arborist internally, so npm is the right tool here.

# 1. Install Node 24 (LTS) via nvm
nvm install --lts          # installs 24
cd ~/.config/opencode
nvm use                    # picks up .nvmrc

# 2. Enable Corepack (one-time, system-level) so the packageManager field takes effect
corepack enable

# 3. Install dependencies
npm install

engines.node requires Node 24+ and packageManager pins npm to 11.12.1. Corepack will shim the right npm version automatically once enabled.

All external packages are pinned to exact versions:

• package.json → @opencode-ai/plugin (the plugin SDK)
• opencode.json → @tarquinen/opencode-dcp (DCP plugin)
• opencode.json → chrome-devtools-mcp (in the MCP command array)

Floating @latest references were intentionally removed — historical Bun-era issues (e.g. cache-induced stale resolution of @latest) and current Arborist install failures are both avoided by pinning. To check for and apply updates use /update-opencode-deps.

npm audit currently flags 3 moderate severity vulnerabilities transitively from @opencode-ai/plugin (uuid < 14, via effect). The audit fix --force path downgrades the plugin to a pre-1.4.5 release, which is worse than current. Leave as-is until upstream publishes a patched plugin release.

Things that are intentionally pending or under iteration:

• DCP trial graduation — see checklist above. Currently running ~1–2 weeks of validation in manual mode before flipping to autonomous.
• Shell-layer hardening — bolster ~/code/scripts/ with tests and stronger error handling. Plan tracked locally (gitignored).
• Radagast model parity — re-evaluate whether Claude Opus catches up to GPT-5.5 xhigh on web-research quality; consolidate to a single model if so.
• Orchestration plugin features — tmux integration, persistent task storage, provider fallback, and notification buffering were intentionally stripped from the local rewrite. Re-add only if a real need surfaces.
• Subagent context pruning — experimental.allowSubAgents is off until DCP behavior on the primary thread is fully trusted.
• MCP key rotation — Exa is currently on the free tier; add an API key via opencode.json → mcp.exa.headers["x-api-key"] if rate limits start hurting.

Per-project overrides go in <repo>/.opencode/ and "local wins" — a local skill/command of the same name supersedes the global one.