Task-shaped skills will draw their reference material (per-platform SDK
guides, per-signal concepts) from a single maintained library rather
than duplicating it. Each shipped skill must still be self-contained, so
the build copies a skill's declared references into it.
Add a CONTENT_ROOT env var (default: repo root) so the same build steps
can assemble an alternate skill tree, e.g. CONTENT_ROOT=skills-next.
hydrate-references.py walks a skills directory: for every skill with a
references.yml it brace-expands and globs the `needs:` patterns against
the library, copies the matches into the skill's references/ preserving
subpaths, and strips the manifest so it never ships. It runs via `uv run
--script` with inline PyYAML and braceexpand dependencies. build-common.sh
invokes it once per build, and only when some skill declares references,
so existing skills are unaffected and the common case spawns no
interpreter.
Factor the steps shared across all four agent builders into
build-common.sh (content-root resolution, skill copy+hydrate, commands,
SKILL_TREE), sourced by each plugin-src/<agent>/build.sh.
Aligns Claude's plugin.json with Cursor, Codex, and Grok which all
include ?utm_source=plugin in their distributed MCP configs, so
Claude-originated traffic is attributed under app.utm_source:plugin.
This repository is the single source of truth for skills and routing metadata,
but it is not itself an installable plugin, and each assistant (Claude Code,
Cursor, Codex, Grok) expects the plugin in a different on-disk shape. No plugin
ecosystem can consume a zip or release asset; all install from a git ref. So a
GitHub Actions workflow builds each agent's distribution from this repository
and deploys it to a dedicated repository whose root is exactly that agent's
plugin, which a marketplace points at as getsentry/plugin-<agent>.
Each agent has a `plugin-src/<agent>/` dir with its manifests, README, and a
`build.sh` that assembles the dist tree from shared repo content. Claude, Cursor,
and Grok consume the plugin at the repo root; Codex requires it under
`plugins/sentry/` and rejects the `disable-model-invocation` field the skill
tree relies on, so its build strips that field and emits a per-skill
`agents/openai.yaml` (see `hide-skills.py`). MCP config is built from the repo's
`mcp.json` source of truth: inline in Claude's manifest, and a file for the
others.
Cross-repo writes use a short-lived token from a dedicated GitHub App scoped to
contents:write on the plugin repos; the default GITHUB_TOKEN cannot push to
other repositories.
The root `.claude-plugin/`, `.cursor-plugin/`, and `.mcp.json` are retained so
existing installs that consumed the plugin from this repo's root keep working;
their READMEs announce removal on or after 2026-07-11.
Evan Purkhiser