docs(cross-llm): coverage tables + final integration (PR-E) (#20)

* docs(cross-llm): coverage tables + final integration (PR-E)

- docs/cross-llm-coverage.md: host capability matrix — SKILL.md import,
  sub-agents, Agent Teams, background agents, MCP, slash commands,
  plan mode, task list, AGENTS.md, skill discovery paths, model tiers
  (Claude Code / Codex / OpenCode / Cursor with ✅/⚠️/❌ ratings)

- tests/cross-llm-coverage.md: per-skill host-conditional audit —
  17 skills classified as host-conditional (have <host:> blocks) or
  host-neutral (no forbidden tokens, no blocks needed)

- README.md: add links to both coverage files from Cross-LLM
  Compatibility section

Integration checks passed:
- tests/skill-content-grep.sh: PASS
- agents/team-conventions.md: coherent (Sequential Mode + Agent Teams)
- .codex/INSTALL.md: coherent (host declaration + <host:> explanation)
- .opencode/INSTALL.md: coherent (tool mapping + host declaration)

Marker smoke test (requesting-code-review, subagent-driven-development,
pr-monitoring): all <host: NAME> / </host> pairs syntactically uniform —
consistent space-after-colon form, consistent close tag.

* fix(cross-llm-pr-e): address 4 Copilot review findings on capability matrix

- Cursor SKILL.md import: update from stub to full plugin (skills/agents/commands/hooks + /plugin-add install)
- OpenCode host declaration: add correct path (~/.config/opencode/AGENTS.md)
- Model tier vocabulary: align with agents/model-tiers.md role-based vocabulary
- Cursor notes section: update to reflect actual plugin.json capability definition
- README.md: update Cursor row from 'stub' to documented marketplace install

* fix(cross-llm-coverage): update Cursor skill discovery path from unknown to via plugin

Agent-Logs-Url: https://github.com/GoCodeAlone/claude-superpowers/sessions/5435a1f1-c11d-4549-9f5e-6ba6ac6d84b8

Co-authored-by: intel352 <77607+intel352@users.noreply.github.com>

* fix(cross-llm-coverage): clarify Claude Code skill discovery vs marketplace install path

Agent-Logs-Url: https://github.com/GoCodeAlone/claude-superpowers/sessions/1582f604-1b7b-4edb-aac1-013a4b416909

Co-authored-by: intel352 <77607+intel352@users.noreply.github.com>

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: intel352 <77607+intel352@users.noreply.github.com>

Author: intel352

README.md

@@ -94,7 +94,10 @@ Superpowers skills run on any host that supports the SKILL.md format. Host-speci
 | Claude Code | `~/.claude/plugins/marketplace/superpowers/` | yes | Full Agent Teams support (experimental flag) |
 | Codex | `~/.agents/skills/superpowers/` | yes | Sequential sub-agent dispatch; `/plan` slash; `/agent` switching |
 | OpenCode | `~/.config/opencode/skills/superpowers/` | yes | Tool mapping documented in `.opencode/INSTALL.md` |
-| Cursor | manual reference | partial | Plugin manifest stub; install path TBD |
+| Cursor | `/plugin-add superpowers` | yes (via plugin) | Plugin manifest defines skills/agents/commands/hooks |
+
+Full capability matrix: [docs/cross-llm-coverage.md](docs/cross-llm-coverage.md)  
+Per-skill host-conditional audit: [tests/cross-llm-coverage.md](tests/cross-llm-coverage.md)
 
 ## The Basic Workflow
 

docs/cross-llm-coverage.md

@@ -0,0 +1,38 @@
+# Cross-LLM Capability Coverage
+
+Host-by-host capability matrix for the Superpowers skills system.
+
+`✅` = fully supported  `⚠️` = partial / workaround required  `❌` = not supported
+
+| Capability | Claude Code | Codex CLI | OpenCode | Cursor |
+|---|---|---|---|---|
+| SKILL.md import | ✅ native | ✅ native | ✅ native | ✅ plugin manifest defines skills/agents/commands/hooks; install via `/plugin-add superpowers` |
+| Sub-agent dispatch | ✅ `Agent` tool | ✅ natural language | ⚠️ `@mention` to peer sessions | ❌ not documented |
+| Agent Teams (persistent multi-agent DM) | ✅ experimental flag | ❌ | ❌ | ❌ |
+| Background agents | ✅ `run_in_background` | ⚠️ thread-based; no explicit background flag | ❌ not documented | ❌ not documented |
+| MCP servers | ✅ | ✅ `config.toml` | ✅ | ⚠️ partial |
+| Slash commands | ✅ | ✅ 30+ built-ins incl. `/plan`, `/agent`, `/review` | ✅ | ✅ |
+| Plan mode | ✅ `EnterPlanMode` + Shift-Tab | ✅ `/plan` slash | ⚠️ not documented; use prose planning | ⚠️ built-in Composer; not slash-invokable |
+| Task list / TodoWrite | ✅ built-in | ❌ no documented equivalent | ⚠️ `update_plan` mapping (see `.opencode/INSTALL.md`) | ⚠️ unknown |
+| AGENTS.md / project context | CLAUDE.md | AGENTS.md (+ `.override.md`) | AGENTS.md | n/a |
+| Host declaration for skill conditionals | `Host: claude-code` in CLAUDE.md | `Host: codex` in `~/.codex/AGENTS.md` | `Host: opencode` in `~/.config/opencode/AGENTS.md` | n/a |
+| Skill discovery path (user scope) | `~/.claude/skills/` (personal skills); superpowers installed to `~/.claude/plugins/marketplace/superpowers/` via marketplace | `~/.agents/skills/` | `~/.config/opencode/skills/` | via plugin (no manual symlink) |
+| Model tier vocabulary | role names → `haiku`/`sonnet`/`opus` (see `agents/model-tiers.md`) | role names → `gpt-5.4-mini`/`gpt-5.4`/`gpt-5.5` | role names → host-pass-through | role names → host-pass-through |
+
+## Notes
+
+**Sub-agent dispatch (Codex):** Codex uses natural-language spawn ("spawn one agent per X") rather than an explicit `Agent` tool call. The `<host: codex>` blocks in skills provide the correct phrasing.
+
+**Agent Teams:** The `TeamCreate` / `SendMessage` persistent-chat pattern is exclusive to Claude Code (experimental flag). Skills fall back to **Sequential Mode** (one sub-agent at a time) on all other hosts — see `skills/subagent-driven-development/SKILL.md`.
+
+**Task list (Codex):** No built-in task-tracking tool is documented in Codex CLI. Skills that reference `TodoWrite` wrap those references in `<host: claude-code>` blocks; the host-neutral path uses prose checklists.
+
+**Cursor:** The `.cursor-plugin/plugin.json` manifest defines `skills`, `agents`, `commands`, and `hooks`. Installation is via `/plugin-add superpowers` in the Cursor agent chat (same marketplace mechanism as Claude Code). Skill discovery path (user scope) is managed through the plugin; no manual symlink required.
+
+## Related files
+
+- `tests/cross-llm-coverage.md` — per-skill host-conditional vs host-neutral audit
+- `tests/skill-content-grep.sh` — CI guard: fails if forbidden tokens appear outside `<host: claude-code>` blocks
+- `.codex/INSTALL.md` — Codex setup instructions
+- `.opencode/INSTALL.md` — OpenCode setup instructions
+- `agents/model-tiers.md` — role-to-model-name resolution table (fast / balanced / frontier / coding-specialist)

tests/cross-llm-coverage.md

@@ -0,0 +1,34 @@
+# Cross-LLM Skill Coverage
+
+A snapshot of which skills have host-conditional content and which are
+host-neutral. Updated whenever a skill changes.
+
+| Skill | Claude Code | Codex | OpenCode | Cursor | Notes |
+|---|---|---|---|---|---|
+| alignment-check | host-conditional | host-conditional | host-conditional | host-conditional | spawn block in `<host: claude-code>`; prose fallback outside |
+| brainstorming | host-conditional | host-conditional | host-conditional | host-conditional | `AskUserQuestion` in `<host: claude-code>`; numbered-list fallback in `<host: codex, opencode, cursor>` |
+| dispatching-parallel-agents | host-neutral | host-neutral | host-neutral | host-neutral | generic parallel-dispatch pattern; no tool-specific refs |
+| executing-plans | host-conditional | host-conditional | host-conditional | host-conditional | tool-use block in `<host: claude-code>`; prose fallback in `<host: codex, opencode, cursor>` |
+| finishing-a-development-branch | host-neutral | host-neutral | host-neutral | host-neutral | audited clean — no forbidden tokens; bash-based throughout |
+| pr-monitoring | host-conditional | host-conditional | host-conditional | host-conditional | Agent spawn block in `<host: claude-code>`; poll-loop prose outside |
+| receiving-code-review | host-neutral | host-neutral | host-neutral | host-neutral | audited clean — pattern-based guidance, no tool refs |
+| requesting-code-review | host-neutral | host-neutral | host-neutral | host-neutral | already portable (Group I) |
+| runtime-launch-validation | host-neutral | host-neutral | host-neutral | host-neutral | already portable (Group I) |
+| subagent-driven-development | host-conditional | host-conditional | host-conditional | host-conditional | Agent Teams setup in `<host: claude-code>`; Sequential Mode is host-neutral default |
+| systematic-debugging | host-neutral | host-neutral | host-neutral | host-neutral | already portable (Group I) |
+| test-driven-development | host-neutral | host-neutral | host-neutral | host-neutral | already portable (Group I) |
+| using-git-worktrees | host-neutral | host-neutral | host-neutral | host-neutral | already portable (Group I) |
+| using-superpowers | host-neutral | host-neutral | host-neutral | host-neutral | host-access phrasing is prose-based ("In Claude Code: … In other environments: …"); no forbidden tokens |
+| verification-before-completion | host-neutral | host-neutral | host-neutral | host-neutral | already portable (Group I) |
+| writing-plans | host-neutral | host-neutral | host-neutral | host-neutral | Plan Mode reference is prose-based ("If you are running in Claude Code…"); no `<host:>` blocks needed |
+| writing-skills | host-conditional | host-conditional | host-conditional | host-conditional | `TodoWrite` checklist and tier-brand names wrapped in `<host: claude-code>` blocks |
+
+## Audit cadence
+
+Re-run `./tests/skill-content-grep.sh` and update this table whenever a skill
+is added or rewritten. The grep guard catches forbidden tokens; this table
+records intent.
+
+## Related
+
+- `docs/cross-llm-coverage.md` — host capability matrix (what each host supports natively)