Reference

Skills

Skills are reusable step-by-step workflow documents under .opencode/skills/<name>/SKILL.md. Agent definitions reference them by name and load them on demand; the skill list is also surfaced to the model automatically. There are 18 skills, grouped by owning agent:

Agent	Skills
architect (9)	`architect-task-intake`, `architect-doc-intake`, `architect-ux-intake`, `architect-local-repo-clone`, `architect-github-repo-intel`, `architect-adr-authoring`, `architect-progress-review`, `architect-outcome-verifier`, `architect-code-feedback`
developer (5)	`developer-worktree-bootstrap`, `developer-ux-intake`, `developer-implementation`, `developer-test-validation`, `developer-change-report`
staff (4)	`staff-doc-compare`, `staff-repo-audit`, `staff-task-triage`, `staff-risk-assessment`

Skills whose source category is not bound in opencode (e.g. architect-task-intake → task, architect-ux-intake → ux) degrade gracefully: the agent states the source is unavailable and proceeds with what is accessible rather than fabricating data.

MCP availability

opencode is configured with the GitHub MCP server (opencode.json → mcp.github, a remote endpoint at api.githubcopilot.com/mcp/ with a GITHUB_TOKEN bearer header). Tool names are prefixed github_*.

The task, documentation, and ux source categories are declared in opencode.json (mcp.task / mcp.documentation / mcp.ux, disabled by default) and in .mcp.json for the parallel Claude Code path. No provider is bound by default — bind one per category to enable the corresponding skills; agents discover at runtime which category servers are connected and proceed with local repos, GitHub, and user-provided content when a category is unbound.

Shared assets

Path	Purpose
`rules/`	Cross-cutting policies: approval-gate.md (ADR human approval gate), cleanup.md (`.tmp/` lifecycle), design-conflict.md (UX-design-vs-ADR conflicts), README.md (index)
`schemas/`	JSON Schemas: handoff.v1.json (+ non-secret example), state.v1.json — see Handoff
`scripts/`	`new-worktree.sh` (canonical clone + per-task/role worktree), `clone-repo-for-analysis.sh` (shallow clone into `.tmp/<task-id>/repos/`), `slack-server.sh` (headless opencode server with the Slack bridge), `healthcheck.sh` (branch/commit/date), `push-metrics.sh` (Claude Code Stop-hook metrics — Claude-only)
`metrics/`	Prometheus + Grafana stack for Claude Code usage/cost dashboards (Claude-only)
`opencode-slack-bridge`	npm package (external repo) providing the Slack bridge plugin; loaded via `opencode.json` → `plugin` (see Slack integration)

Guardrails and configuration

Permissions (opencode.json): read/edit/glob/grep/list/task are allowed; bash has an explicit allowlist (safe read-only commands, git/gh read + write commands, npm/node, scripts/*.sh) with ask as the default; destructive commands (rm, git push --force, git reset --hard, git clean, branch deletion, worktree removal) are denied.
Environment: cp .env.example .env, then export GITHUB_TOKEN=... (opencode reads env vars; it does not load .env automatically — with one exception: the Slack plugin, see Slack integration).
.tmp/, .repos/, .worktrees/, .env, metrics/data/ are gitignored.

Two harnesses in parallel

AGENTS.md / .opencode/ — the opencode harness described on this site.
CLAUDE.md / .claude/ — the original Claude Code harness (commands, prompts, skills, hooks). It supports the task / documentation / ux source categories via .mcp.json bindings, which opencode does not read.