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

[intel352]

Summary

Closing PR for the cross-LLM portability track. Three changes:

1. docs/cross-llm-coverage.md — Host capability matrix: SKILL.md import, sub-agent dispatch, Agent Teams, background agents, MCP, slash commands, plan mode, task list, AGENTS.md hierarchy, skill discovery paths, model tier vocabulary — rated ✅/⚠️/❌ for Claude Code / Codex / OpenCode / Cursor with explanatory notes on each gap.

2. tests/cross-llm-coverage.md — Per-skill host-conditional audit: all 17 skills classified as host-conditional (have formal <host:> blocks) or host-neutral (no forbidden tokens, no blocks needed), with notes explaining each choice.

3. README.md — Added links to both coverage files from the existing Cross-LLM Compatibility section.

Integration pass

• tests/skill-content-grep.sh: PASS (all skills clean)
• agents/team-conventions.md: coherent — Sequential Mode + Agent Teams Mode documented
• .codex/INSTALL.md: coherent — host declaration + <host:> block explanation
• .opencode/INSTALL.md: coherent — tool mapping + host declaration

Marker smoke test

Verified <host:> syntax is uniform across 3 representative skills (requesting-code-review, subagent-driven-development, pr-monitoring):

• All open tags: <host: NAME> (space after colon, no stray spaces)
• All close tags: </host> (consistent)
• Multi-host form: <host: codex, opencode, cursor> (comma-separated, consistent)

Test plan

• ./tests/skill-content-grep.sh passes clean
• Both coverage files reviewed for accuracy against actual skill content
• README links verified
• Rebased on main (includes all PR-A through PR-D merges)

🤖 Generated with Claude Code

[intel352]

@copilot review

[Copilot]

@copilot review

Code Review

CI / grep guard: ./tests/skill-content-grep.sh passes clean. All 17 skills are correctly classified in tests/cross-llm-coverage.md — verified by diffing directory listing against table rows and checking <host:> block presence in each skill file.

---

Issues

1. Claude Code skill-discovery path inconsistency (accuracy)

docs/cross-llm-coverage.md row:

| Skill discovery path (user scope) | `~/.claude/skills/` | …

README.md and docs/plans/ consistently say:

~/.claude/plugins/marketplace/superpowers/

These are different paths. If ~/.claude/skills/ is the actual discovery root and ~/.claude/plugins/marketplace/superpowers/ is the install target, the table should either match the README or explain the distinction. As-is, a reader following the coverage doc to set up the skill path will end up in the wrong directory.

2. Model tier vocabulary row mixes brand names and role names (consistency)

| Model tier vocabulary | Haiku / balanced / frontier | gpt-5.x / gpt-5.x-mini / gpt-5.x-codex | …

Haiku is a brand name (maps to the fast role), while balanced and frontier are role names. Per agents/model-tiers.md the Claude Code brand names are: haiku / sonnet / opus. The row should read Haiku / Sonnet / Opus (brand names, consistent with the Codex column style).

3. Codex model names use placeholder x (minor)

gpt-5.x / gpt-5.x-mini / gpt-5.x-codex don't match the pinned versions in agents/model-tiers.md (gpt-5.4, gpt-5.4-mini, gpt-5.5, gpt-5.3-codex). Using x as a placeholder is fine if intentional for future-proofing, but should be flagged as a deliberate choice (e.g., a note that the table rows show schema, not pinned versions) to avoid reader confusion against the model-tiers table.

---

Looks good

• tests/cross-llm-coverage.md: all 17 skills correctly classified; classification matches <host:> block presence in every skill file.
• README.md: links are valid relative paths and land in the right place.
• agents/team-conventions.md: Sequential Mode / Agent Teams Mode description is consistent with docs/cross-llm-coverage.md.
• docs/cross-llm-coverage.md Notes section is accurate (Agent Teams, TodoWrite, Cursor caveat all match actual skill content).
• No new forbidden tokens introduced outside <host: claude-code> blocks in skills/ or agents/.

[Copilot]

Pull request overview

Adds Cross-LLM coverage documentation to close out the portability track, and links it from the main README so users can quickly see host capabilities and per-skill portability status.

Changes:

• Introduces a host capability matrix covering key Superpowers features across Claude Code / Codex CLI / OpenCode / Cursor.
• Adds a per-skill audit table classifying each skill as host-neutral vs host-conditional, with brief rationale.
• Updates README Cross-LLM Compatibility section to link to both new coverage documents.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
docs/cross-llm-coverage.md	New host capability matrix (✅/⚠️/❌) with explanatory notes and related links.
tests/cross-llm-coverage.md	New per-skill portability audit table (host-neutral vs host-conditional) and update guidance.
README.md	Adds links to the two new coverage docs from the Cross-LLM Compatibility section.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[intel352]

@copilot review

[intel352]

@copilot review

[Copilot]

@copilot review

Code Review

All 4 inline suggestions from the automated reviewer have been applied (commit c6386ad):

Issue	Fix
Cursor SKILL.md row "plugin manifest stub; install path TBD"	→ ✅ "plugin manifest defines skills/agents/commands/hooks; install via /plugin-add superpowers"
OpenCode host declaration missing path	→ Host: opencode in ~/.config/opencode/AGENTS.md
Model tier vocabulary mixing brand/role names	→ role names → resolved IDs (e.g., haiku/sonnet/opus), consistent with agents/model-tiers.md
Cursor notes section outdated	→ reflects actual plugin.json capability and marketplace install mechanism

One additional fix in 9190cc8: the "Skill discovery path" row had unknown for Cursor while the Notes section immediately below already said "managed through the plugin; no manual symlink required" — resolved the internal inconsistency to via plugin (no manual symlink).

./tests/skill-content-grep.sh still passes clean. No remaining issues found.

[Copilot]

@copilot review

The one open issue from the previous review pass — Claude Code skill-discovery path inconsistency — is now fixed in commit 6c7aaeb.

The table cell previously showed only ~/.claude/skills/, while the README shows ~/.claude/plugins/marketplace/superpowers/. These are distinct things: ~/.claude/skills/ is the user-scope skill discovery root (where personal skills go, per skills/writing-skills/SKILL.md), and ~/.claude/plugins/marketplace/superpowers/ is the superpowers-specific marketplace install target. The cell now shows both to prevent reader confusion:

~/.claude/skills/ (personal skills); superpowers installed to ~/.claude/plugins/marketplace/superpowers/ via marketplace

./tests/skill-content-grep.sh still passes clean. No further issues found.

[intel352]

@copilot review

[Copilot]

@copilot review

Code Review

./tests/skill-content-grep.sh PASS. No remaining issues found.

Summary of all fixes across the 4 commits

Issue	Commit	Status
Cursor SKILL.md row was "plugin manifest stub; install path TBD"	c6386ad	✅ fixed
OpenCode host declaration missing ~/.config/opencode/AGENTS.md path	c6386ad	✅ fixed
Model tier vocabulary mixed brand/role names	c6386ad	✅ fixed → role names → resolved IDs, per agents/model-tiers.md
Cursor notes section contradicted README install instructions	c6386ad	✅ fixed
Cursor skill discovery path was "unknown" (inconsistent with Notes)	9190cc8	✅ fixed → "via plugin (no manual symlink)"
Claude Code skill discovery path ambiguous vs README marketplace path	6c7aaeb	✅ fixed → cell now shows both user-scope root and marketplace install target

Looks good

• All 17 skills correctly classified in tests/cross-llm-coverage.md; classifications match <host:> block presence in every skill file.
• README.md links are valid relative paths (docs/cross-llm-coverage.md, tests/cross-llm-coverage.md) and the Cursor row reflects current state.
• docs/cross-llm-coverage.md model tier row now consistently uses role-name → resolved-ID form and links to agents/model-tiers.md for the full table (including gpt-5.3-codex for the coding-specialist tier).
• No forbidden tokens outside <host: claude-code> blocks in skills/ or agents/.