oh-my-opencode-slim

# oh-my-opencode-slim Configuration Skill You help users configure and safely improve oh-my-opencode-slim, including small prompt/config changes that make future runs better. ## When to Use Use when the user asks about, or a session reveals repeatable friction in: - plugin config (`oh-my-opencode-slim.json[c]`), presets, models/variants - delegation behavior, specialist prompts, custom agents (`agents.<name>`) - skills, MCPs, tool access, disabled agents; background/multiplexer/deepwork Suggest config changes only for recurring behavior, not one-offs. ## Files - User config: `~/.config/opencode/oh-my-opencode-slim.jsonc` (preferred) or `.json`; project override: `<project>/.opencode/oh-my-opencode-slim.json`. - Prompt tuning: `~/.config/opencode/oh-my-opencode-slim/{agent}[_append].md` (prefer `_append.md`; preset variant under `{preset}/`). - Skill shadow: `~/.config/opencode/skills/<skill-name>/SKILL.md` shadows the bundled skill. - Agent names: `orchestrator|oracle|explorer|designer|fixer|observer|council`. Lookup: with an active preset, check `{preset}/` first, then the base dir. `{agent}.md` replaces the built-in prompt; `{agent}_append.md` appends to it. Each works alone; neither present → unchanged built-in prompt. Common customizations: - Presets/models: switch preset; cheaper models for `explorer`/`fixer`, stronger ones for `orchestrator`/`oracle`/`designer`. - Permissions: per-agent `skills`/`mcps` (`['*']`, names, or `['*', '!x']`); enable optional agents by removing them from `disabled_agents`. - Custom agents under `agents.<name>` with `prompt` plus `orchestratorPrompt` (trigger and non-use). - Prompt tuning via `{agent}_append.md` for recurring behavior or house style. Important schema boundary: - Built-in agents (`orchestrator`, `oracle`, `explorer`, `designer`, `fixer`, `observer`, `council`) can set models, variants, skills, MCPs, options, and display names in config. - Built-in agent `prompt` and `orchestratorPrompt` fields are **not** supported in `oh-my-opencode-slim.json[c]`; use markdown prompt override files instead. - Unknown keys under top-level `agents` are custom agents. Custom agents may use `prompt` and `orchestratorPrompt` directly in config. ## Config Shapes Tune built-ins at `presets.<preset>.<agent>` (model/variant/skills/mcps). Append tuning via `{agent}_append.md` (preset variant under `{preset}/`). Full `{agent}.md` replacement is rare — it must restate all essential behavior. ## Safe Improvement Rules Configuration changes affect future agent behavior, so treat them as user-owned. 1. **Ask before changing config or prompts.** - Explain the proposed improvement briefly. - State which file would change. - Ask for confirmation unless the user explicitly requested the exact edit. 2. **Prefer narrow changes.** - Do not rewrite large prompts when a small rule solves the problem. - Do not add custom agents for one-off tasks. 3. **Preserve existing user settings.** - Merge with current config rather than regenerating from scratch. - Keep comments and formatting where practical for JSONC files. 4. **Avoid hidden behavior changes.** - Mention cost, permissions, or delegation changes before applying them. - Be explicit if a model/provider change may increase spend. 5. **Tell the user about restart requirements.** - Config, prompt, agent, skill, MCP, or plugin changes may need a restart. - Phrase it as: "This should apply on the next OpenCode run; restart OpenCode if you need it immediately." ## Configuration Workflow 1. Inspect: read current config, active `preset`, relevant agent blocks, project overrides. 2. Minimize: pick the smallest change (model/preset, prompt rule, custom agent, or permission). 3. Confirm: show a concise proposal with the target file path. 4. Edit: preserve unrelated settings, JSONC formatting, and exact agent/skill/MCP names. 5. Validate: ensure the file still parses; run repo config checks if available. 6. Activate: state whether it applies immediately or on next OpenCode run/restart. ## Custom Agent Pattern Custom agents suit narrow, repeatable jobs the Orchestrator can route. Use this shape: ```jsonc { "agents": { "api-reviewer": { "model": "openai/gpt-6", "variant": "high", "prompt": "You review API contracts, error semantics, and migration risk. Return concise findings with file references.", "orchestratorPrompt": "Delegate API contract, public SDK, or backwards-compatibility reviews to @api-reviewer. Do not use it for routine implementation.", "skills": [], "mcps": [] } } } ``` Good custom agents have: - a specific job; - clear trigger conditions in `orchestratorPrompt`; - explicit non-use conditions; - only the skills and MCPs they actually need; - a model appropriate to the task's judgment/cost needs. Avoid custom agents that duplicate existing specialists: - codebase scouting and external docs/research → `explorer` - architecture/debugging/review → `oracle` - UI/UX polish → `designer` - scoped mechanical implementation → `fixer` ## Prompt Tuning Pattern Prompt edits suit recurring behavior across many sessions. Good reasons to tune a prompt: - The Orchestrator repeatedly over- or under-delegates for this workflow. - A specialist repeatedly misses a project convention. - The user wants a stable communication or verification style. - The team has a recurring review checklist or deployment rule. Poor reasons to tune a prompt: - A one-off task failed once. - The current problem can be solved by normal instruction in this session. - The change would make the agent worse for general use. When suggesting a prompt improvement, say: ```text I noticed this is recurring. I can add a small rule to <agent/config path> so future runs handle it automatically. Want me to make that config change? ``` ## Final Checklist - [ ] Did the user confirm config/prompt edits, unless explicitly requested? - [ ] Did the edit preserve existing settings? - [ ] Is the active preset still valid? - [ ] Are skill/MCP/tool permissions intentional and minimal? - [ ] Did you mention OpenCode restart/next-run behavior?