Recommended skills for projects and repositories
Open Sessions → Session settings, select a project group or repository and
choose Recommended skills (s in TUI). The repository's Project workflows
also has a separate Recommended skills section. TUI, web and desktop use the
same panel, with real checkboxes in browser clients.
Use Scan manifests to request the existing bounded scan on the execution host. No local repository mirror is required. Recommendations show the actual manifest path and package field that caused the match. React and Nest are detected independently from direct dependencies; Python packaging and the other supported ecosystem parsers use the same saved technology profiles. Unsupported or dynamic configurations retain their existing partial/unknown status. No repository program is executed to infer a framework.
Select and save recommendations at project-group level to make them defaults for its owned repositories. A repository can add recommendations or exclude an inherited choice. The selection survives catalog refreshes. Existing manual skill bindings remain separate and retain precedence, including explicit session disables. These optional choices do not change locked review rules.
Saving a selection does not download code. Install selected skill explicitly loads maintained files or downloads the complete public GitHub bundle from its pinned revision. The downloader restricts origin, paths, redirects, file count, bytes, depth, requests and duration; rejects links, submodules and unsupported binary encodings; verifies Git blob hashes; then checks frontmatter and local references. No scripts or hooks run during installation. API rate limits, unavailable sources or incomplete bundles fail visibly without inventing an installation. Native compatibility and actual execution remain unverified.
Installed definitions are private to the signed-in user. Recommended skills are resolved before existing manual bindings and delivered through the normal session agent configuration. New sessions inherit the effective selection; existing sessions show the normal configuration-update indicator. Installation does not restart an agent. Unselecting a recommendation does not delete its installed definition or unrelated manual bindings. A catalog refresh never silently updates installed bundle contents. A newer curated source revision is shown separately from the installed revision and offers an explicit update.
The language-organized source data and maintenance command are documented in skill-recommendations. Empty mappings for an ecosystem are explicit: detecting a technology does not establish that a useful specialized skill exists.
CLI and MCP
devctl technology skills show repository REPOSITORY_ID --json
devctl technology skills show group PROJECT_GROUP_ID --json
devctl technology skills save repository REPOSITORY_ID --file selection.json
devctl technology skills install repository REPOSITORY_ID --file install.json
selection.json contains the observed revision, explicit selected reference
IDs and inherited excluded IDs. install.json contains referenceId, the
saved selectionRevision and the displayed mappingDigest. Stale revisions
return a conflict. Install after saving the selection and reviewing the source.
Session MCP's existing get_skill_recommendations returns the same curated
links, evidence, selected/inherited choices and installation status in
curated, alongside its existing catalog and review-coverage fields. Catalog suggestions
are paged with limit (default 10, maximum 25), offset, total and
nextOffset; the full public catalog is never attached to a single default
recommendation response. Reading
recommendations grants no installation authority. The controller's existing
operator MCP discovery also exposes the real scoped HTTP operations and their
validated input schemas:
GET /v1/recommended-skills/:kind/:idPUT /v1/recommended-skills/:kind/:idPOST /v1/recommended-skills/:kind/:id/install
kind is group or repository. Each handler derives the owner from the
authenticated request and checks the exact scope. Supplying another user's ID
does not grant access.