v0.6.0 — Agent Plugins on the host wall, Grok Build, a curated roster
Agent Plugins on the host wall, a new xAI host, and a curated roster
v0.6.0 makes the public host list tell the truth: it marks which hosts receive the Agent Plugins 1.0.0 bundle, adds xAI's Grok Build, retires a dead host, and corrects a set of upstream facts that had quietly rotted.
Agent Plugins marker
Every coverage card now carries an AP marker, filterable by a new Spec row. Three evidence-backed states:
| State | Meaning | Hosts |
|---|---|---|
delivered |
we package the spec bundle and a plugin verb we drive installs it | Codex · GitHub Copilot CLI · Kiro · Hermes |
delegated |
same bundle, no headless verb of its own | VS Code Copilot · JetBrains Copilot |
client |
listed AP client, but its own format carries hooks/commands | Cursor · OpenClaw |
The roster was re-verified against agent-plugins.org/compatible-clients, which lists nine clients. Two near-collisions are deliberately not marked: xAI's "Grok Bot" is not the community grok-cli, and Nanoco's "NanoClaw" is not NVIDIA's NemoClaw. delivered ∪ delegated is drift-guarded to equal what MARKETPLACE_FORMAT_BY_PLATFORM routes to agent-plugin.
New host: Grok Build (xAI)
xai-org/grok-build, Apache-2.0, json-stdio, full surface coverage. Every path byte-confirmed against upstream:
- MCP — TOML at
$GROK_HOME/config.toml,[mcp_servers.<name>] - Hooks — JSON at
$GROK_HOME/hooks/*.json - Skills / commands / subagents under
.grok/, memory viaAGENTS.md
It also fixes a pre-existing bug: grok-cli and grok-build both default to ~/.grok, and the old adapter detected on the bare directory — so on a Grok-Build-only machine it reported Grok CLI as installed and would have written a config that host never reads. Both now bow out on the sibling's exclusive marker.
Upstream renames, each verified against source
A full sweep of every tracked repo found six that had moved. Two were real product renames — and only one of them actually changed a config contract, which is exactly why each was checked instead of assumed.
Mux → Xum (coder/mux → coder/xum) — the config home really did move, .mux → .xum, root key servers unchanged. Xum still reads the old home, so the adapter now resolves the way the host itself does: prefer .xum, keep an existing .mux home, default new installs to .xum. Writing .xum unconditionally would have stranded every current user's servers in a file the adapter no longer touches. All three cases are pinned by tests.
Codebuff → Freebuff (CodebuffAI/codebuff → CodebuffAI/freebuff) — the config contract is unchanged: .agents/mcp.json, root key mcpServers, same search order. No adapter path moved. The npm package and binary did (codebuff → freebuff), and the host's verification lane was re-checked live against the new binary: freebuff --help exposes only login, so the placement-only conclusion holds — now sourced from the binary users actually install.
Both platform ids stay mux and codebuff; only the display names change. Renaming an id would break existing connector configs and telemetry keys for a change nobody sees.
Renamed with no config impact: sst/opencode → anomalyco/opencode, All-Hands-AI/OpenHands → OpenHands/OpenHands, block/goose → aaif-goose/goose, OpenInterpreter/open-interpreter → openinterpreter/openinterpreter.
Roo Code retired
RooCodeInc/Roo-Code is archived upstream (last push 2026-05-15) and the team pointed users at Cline. The adapter, its usage reader, its tests, the PlatformId member and every doc reference are gone — 43 adapters become 42.
Breaking: "roo-code" is no longer a valid PlatformId. A config that targets it will fail to typecheck and be reported as an unknown platform.
The public wall is curated, not just star-gated
Three rules now govern what appears:
- Archived upstream is never public.
- Independent OSS needs 5,000 stars, not 1,000. A vendor's flagship agent is exempt from the raised floor only — never the base one — because a star count measures repo popularity, not product reach, and under-reports first-party agents whose public repo is an issues tracker. AWS ships Kiro (4.2k) and Amazon Q (2.0k) to millions.
- An editorial opt-out for closed hosts, which have no star signal at all. Their public repos measure feedback-tracker engagement: VS Code Copilot's sits near 1k against Warp's 65k, which says nothing true about reach.
Adapters, drivers and install behaviour are untouched — this governs public pages only.
Terminology: "agent harness" researched, not adopted
The term is now real vendor vocabulary, but it names the session engine, not the config-owning surface our adapters target. VS Code carries both words: it supports "multiple agent harnesses", where "an agent harness coordinates an agent session", while its separate Agent Host owns the session. The three specs we implement all say client and never "harness". The Platforms reference gains a Host, client, harness section instead of a rename.
Docs staleness swept, and guarded
ARCHITECTURE.md paradigm counts said 18/8/9 against a real 23/8/12; an AGENTS.md tally's denominator matched nothing. Every hand-maintained count that could be derived is now pinned by a drift test, and the one that could not be derived cleanly was de-numbered rather than guessed.
Full changelog: see CHANGELOG.md — 0.6.0.