Six plugins address six recurring Claude Code jobs
Toolshed packages six plugins that can be installed separately. Codebase Mapper writes a project map; Live Rules stores conditional instructions; Sidequest supplies a local work board; Model Gateway routes selected subscription models; Observability records usage metadata; Quartermaster recommends and checks the other pieces. A team can add one behavior without accepting every local service, hook, or settings edit in the repository.
Codebase Mapper and Live Rules are the easiest pair to inspect. The mapper writes under .claude/.codebase-info/ and may commit its updates unless the user prohibits commits. Live Rules keeps one Markdown file per rule under .claude/live-rules/ and generates a manifest. Both leave CLAUDE.md alone. Review those generated files because stale maps and broad rules become instructions for later sessions.
Sidequest has a local handoff model and current dispatch defects
Sidequest gives Claude a backlog with dependencies, claims, routes, verification, review, and integration. A code change hands back a candidate revision; other work can return a verified project snapshot. That is more precise than asking several agents to edit the same checkout. Its dashboard and ticket database stay local.
Issue 49 reports Sidequest 5.1.1 arming a routed ticket, then hanging until the host backgrounds the call at 120 seconds without returning the launch token. The held write lock also blocked the board. Issue 51 describes remote delivery staying pending when a candidate reached origin/develop but local develop was stale. The reports concern core dispatch and delivery paths, though they do not prove every installation fails.
Model Gateway adds subscription models through a local proxy
Model Gateway requires Claude Code 2.1.129 or newer for gateway rows in /model. It keeps Claude models on Anthropic and sends selected gateway identifiers through supported ChatGPT, Codex, or Grok subscriptions. Setup checks login, starts the gateway, edits project settings, and verifies wiring. A process-level ANTHROPIC_BASE_URL takes precedence and can make project settings ineffective.
The plugin maintains a shim supervisor, proxy discovery, recovery records, model-window aliases, and route metadata. SessionStart waits up to 12 seconds inside a 30-second hook budget, while model-picker discovery requires a full process restart. The README says route logs omit prompts, bodies, tools, credentials, and arbitrary headers. Administrators still need a retention policy for the local JSONL records.
Observability does not enforce opt-in at ingestion
Observability says hook events can reach its shared spool and ingest path before it verifies that the repository opted in. Teams handling sensitive repositories should stop there until the boundary matches the policy. Records are designed to hold identifiers and usage metadata without prompt text, response text, code, file contents, tool arguments, credentials, or environment values. That narrower record does not repair the opt-in gap.
The observer keeps a 128 MiB writable reserve below a 4 GiB database limit and normally retains 30 days. Under pressure it can remove older whole days sooner. A local dashboard usually listens on 127.0.0.1:3000; remote sinks are optional. Cost panels use API list-price equivalents, while unpriced models remain visible as token counts without a made-up dollar total.
What happened when we ran it
Our sandbox tested plugins/model-gateway/ at commit c6ce202. Npm installed 6 packages in 26 seconds and used 3 MB on disk. There was no build script or target, so the build step was skipped. Npm audit reported 0 known vulnerabilities. This covers Model Gateway's repository mechanics, not subscription login, model selection, Sidequest dispatch, or Observability services.
The test command returned exit code 1 after 198 seconds. Node reported 276 passed and 5 failed out of 281. The supplied log tail does not show the five failures; its final lines mark compatibility migration and SessionStart cases ok, including SQ-1901. We will not guess what failed. The checkout had 855 files, roughly 192,378 source lines, 4 CI workflows, a tests directory, and no Dockerfile.
Quartermaster asks before changing a workspace
Quartermaster inspects a project and bounded summaries of recent Claude Code sessions, then suggests one setup or maintenance change at a time. Its README says installs, file edits, and settings changes wait for approval. The mining script uses clipped prompts, path segments, counts, and short correction evidence instead of loading raw transcripts into model context. Its default command looks back 30 days and considers up to 40 sessions.
After approved installs and file writes, Quartermaster stops for a plugin reload or Claude Code restart before verification continues. Updates follow the recorded user, project, or local scope, even if the request starts in another project. Read the proposed targets before approving an updater run.
September 10 activity comes with five open items
GitHub recorded a push on September 10, 2026, the day release v3.541.0 was published. The repository had 124 stars and 5 open issues and pull requests when fetched. Four open reports shown by GitHub concerned Sidequest dispatch, review, or delivery. Frequent releases indicate active maintenance, but an open Claude Code process can lag installed files until plugins reload or the process restarts.
Toolshed is best treated as six separate decisions. Codebase Mapper and Live Rules leave inspectable artifacts. Sidequest owns workflow state, Model Gateway owns a local routing path, and Observability has an acknowledged privacy gap. The failed 281-test run makes a local check part of adoption. Start with one plugin and keep it only if the recovered time exceeds the extra operational work.