Editing Skills in opencode-swarm

Skills in this repo exist across parallel trees (.opencode/skills/, .claude/skills/, .agents/skills/) governed by explicit mirror contracts. Editing a SKILL.md without knowing its contract silently desynchronizes the OpenCode and Claude Code surfaces. Classify first, then edit.

Step 1 — Classify the skill

Look the slug up in src/config/skill-mirrors.ts:

• MIRRORED_ARCHITECT_MODE_SKILLS (brainstorm, specify, clarify-spec, swarm-resume, clarify, discover, consult, pre-phase-briefing, council, deep-dive, deep-research, issue-ingest, plan, critic-gate, design-docs): .opencode and .claude copies must stay byte-identical. Any edit is a dual-tree edit — apply the identical change to both files.
• ADAPTER (execute, phase-wrap, swarm-pr-review, swarm-pr-feedback, swarm-pr-subscribe, swarm-ci-monitor): .opencode is canonical; .claude and .agents where present are thin shims that must keep the exact relative reference ../../../.opencode/skills/<slug>/SKILL.md (drift-check verifies the string). swarm-implement follows the same operational adapter pattern, but is currently classified as an additional divergent contract because the additional-contract registry does not yet model non-architect adapter shims.
• DIVERGENT: both trees exist, content intentionally differs — a single-tree edit is fine. codebase-review-swarm is in DIVERGENT_ARCHITECT_MODE_SKILLS; engineering-conventions is a divergent entry in ADDITIONAL_SKILL_MIRROR_CONTRACTS. swarm-implement and writing-tests are classified divergent only because the additional-contract model does not yet have an adapter kind; operationally, .opencode is canonical and .claude delegates to it.
• OPENCODE-ONLY: loop (OPENCODE_ONLY_ARCHITECT_MODE_SKILLS) and running-tests (an opencode-only ADDITIONAL contract) — do not create .claude mirrors (a .claude/skills/loop would shadow Claude Code's built-in /loop).
• ADDITIONAL contracts: commit-pr is divergent (#1692). The .claude copy is the repo-INTERNAL publication protocol (CI's pr-standards workflow declares it the source of truth); the .opencode copy is the PORTABLE, project-agnostic version bundled into end-user projects, so the two trees intentionally differ — do not sync them byte-for-byte. The discovery shims in .agents/skills/commit-pr/ and .github/skills/commit-pr/ point at the repo-internal .claude file as canonical.
• Adapter skills (issue-tracer): .opencode/skills/<slug>/SKILL.md is the canonical protocol; .claude/skills/<slug>/SKILL.md and .agents/skills/<slug>/SKILL.md are thin shims that reference it via the relative path. See src/config/skill-mirrors.ts for the kind: adapter contract.
• No skill-mirrors.ts entry (qa-sweep, research-first, swarm, unswarm, tech-debt-ci-review, rust-crate-ci, orchestrating-subagents, durable-session-state, editing-skills, …): the .claude file is the source protocol and there is no .opencode copy, no CI gate, and no npm shipment. But most are not single-file: many have a Codex adapter shim in .agents/skills/<slug>/ that reads ".claude/skills/<slug>/SKILL.md is the source protocol" (check ls .agents/skills/). Content edits to the .claude file are safe — shims delegate — but renaming, moving, or removing one of these skills silently orphans its .agents shim with zero drift-check coverage. Update or remove the shim in the same change.
• Pre-flight for new skills (blocking): Before authoring any SKILL.md content for a skill that should reach npm users, the slug MUST be registered in all three places: src/config/skill-mirrors.ts (classification), src/config/bundled-skills.ts (BUNDLED_PROJECT_SKILLS), and package.json#files. Creating content first and registering later is a common source of drift-check failures and missing shipments.

Step 2 — Know what ships where

• Only canonical package sources under .opencode/skills/ are published: package.json#files lists the bundled source directories and BUNDLED_PROJECT_SKILLS (src/config/bundled-skills.ts) drives a private runtime sync to .swarm/bundled-skills/<slug>/ at plugin init. Never sync into a target project's native skill roots.
• .claude/skills/ is repo-internal — it configures Claude Code sessions in this repository only.
• A new skill that should reach npm users must be added to all three: .opencode/skills/<slug>/, BUNDLED_PROJECT_SKILLS, and package.json#files — the drift checker flags incomplete combinations.
• A new cross-tree pair (same slug in both trees) must be classified in src/config/skill-mirrors.ts, or drift-check reports it "unclassified".

Step 3 — MODE skills are not triggered by descriptions

The mirrored architect skills carry descriptions like "Full execution protocol for MODE: PLAN". These are loaded on demand by the architect stubs in src/agents/architect.ts when a [MODE: …] signal fires — they are not meant to match natural user language, and in a plain Claude Code session (no swarm runtime tools) they are protocol documentation, not executable workflows. Do not "fix" their descriptions to natural language, and do not expect them to auto-trigger.

Step 4 — Frontmatter and description conventions

• Every tracked static skill declares audience: swarm-plugin. Consumer repos may instead use a domain tag such as ragappv3 and optionally add a runner constraint (runner:opencode, runner:claude, or runner:codex). Keep byte-identical mirrors' audience lines byte-identical. Runtime-generated skills are intentionally allowed to omit the field (legacy match-all).
• description is how Claude Code auto-selects skills: third person, the core use case and concrete trigger keywords first, under ~1024 chars. Vague descriptions ("helps with X") never trigger.
• disable-model-invocation: true for slash-command-style skills the model must not self-invoke (swarm, unswarm, tech-debt-ci-review).
• Supported optional fields used in this repo: effort (low…max), context: fork + agent: Explore|Plan|general-purpose|<custom> (runs the skill body in an isolated subagent — the body then cannot rely on conversation history or on spawning further subagents), allowed-tools (space- or comma-separated), argument-hint.
• Keep SKILL.md under ~500 lines; move deep material into references/ files linked one level from the SKILL.md (see issue-tracer for the pattern).

Step 5 — Validate and publish

1. If the edit changed content wording (not just whitespace/formatting), read file:.swarm/bundled-skills/skill-edit-validation/SKILL.md and run its content-assertion sweep to catch stale test assertions before proceeding.
2. bun run drift:check --enforce — CI runs the same enforcement mode; blocking drift findings fail the job, while the full report is posted as a PR comment.
3. If you touched a byte-identical pair, diff the two files to prove identity.
4. Skill-only diffs (no src/, scripts/, or test changes) count as docs/meta-only for PR gating — no release-notes fragment required. The moment you also touch src/config/skill-mirrors.ts, bundled-skills.ts, or package.json, a docs/releases/pending/<slug>.md fragment becomes mandatory (see commit-pr skill, Step 2).
5. Skill edits are still "changed work" under the swarm-mode contract: independent reviewer + critic gates apply before completion.