Roadmap

Staged plan. Decisions and open questions are in ../DECISIONS.md.

Direction (decided 2026-10-06, after the adversarial review): evidence first, then trimmed scope (D-19, which supersedes the "both tracks in parallel" build order, S-01). The review's verdict was that the mechanism is sound but the plan was wider than its evidence. It also found that the benefit the pitch rests on (better skill routing, less clutter) was never measured.

Phase 0: evidence (before writing product code)#

1. Routing eval. Run 20 or more realistic prompts on a real plugin setup of the size the project targets (for example with claude plugin eval or -p runs). Compare the full set with a masked profile. Record which skill or plugin fires and whether it was the intended one. Fix the success criterion before running. The launcher is worth building only if the masked set measurably improves correct-first-try activation or reduces wrong-skill activations. If it does not, the launcher is documented as an alias recipe (dry-run prints the exact command) and not built as a product.
2. Adopt, build or contribute. Decide whether to contribute upstream instead of building a fourth tool. Evaluate fuzzyalej/claude-profile and edimuj/claude-rig against our gaps:
   • Default-deny masking on one shared plugin store.
   • Behavior under managed policy.
   • Concurrency.
   • Shared profiles from a private repo.
3. Bundle and masking prototype (T7). Does masking a plugin that an installed bundle depends on break the bundle? It needs a second config directory and a login only the user can do. It also settles whether profile = bundle stays the bridge between the launcher and the catalog.
4. Stage 0 on Linux and Windows (including windows-11-arm if a runner is used), repeating T1, T2 and the concurrency check, plus the MCP cmd /c question.
5. Measure MCP and connector removal through the settings keys (the one measurement that suggested a real saving).
6. Kill and obsolescence criteria, written down now and reviewed at every Claude Code minor release (K1 to K12 below).

Phase 1: MVP, if Phase 0 supports it#

• Launcher: run, show, dry-run, ls, dir sources only, extends, default-deny masking regenerated on every launch, settings validation before launch, exec on Unix and spawn on Windows. Closed schema from SR1. macOS and Linux first. Windows: compile-checked, and tested once Phase 0 item 4 passes.
• Catalog: lint (required fields from the marketplace entry plus a minimal sidecar: owner, status, when_to_use) and a generated CATALOG.md. Push "when to use" into native description and relevance blocks, which Claude Code itself reads. A static site only if people use the Markdown version.
• Project setup: employer approval for open-sourcing (the MIT LICENSE file already exists), SECURITY.md and a threat model, and CONTRIBUTING.md.
• Release gates, staged: github.com and GHE Cloud first. GHE Server and Windows arm64 are documented best-effort until tested (R1 and R2 stay the direction, not first-release gates). macOS, Linux and Windows amd64 are the first build targets. The real-claude smoke test runs nightly before any release.

Phase 2: deferred until someone asks#

• Accounts (--account).
• The trust lockfile and git sources. SR2 must land first, and the closure hash must be built before any sharing.
• compile and committed bundles (pending T7).
• recommend.
• shell-init for four shells.
• The static catalog site.
• Taxonomy.
• The Analytics join.
• The plugin source.
• A concierge.
• The OTel loop.
• The interactive wizards beyond the picker.

What changed from the earlier roadmap#

The shared core/ is built only as far as the MVP needs. The catalog track starts with CATALOG.md, not the site (Phase 1). Open source remains the direction, but the scaffolding is gated on employer approval.

Kill and obsolescence criteria (Phase 0.6, written 2026-10-06)#

Written down before the evidence is in, so that stopping or shrinking is a decision made against a rule and not against sunk cost. "Stop" means archive the launcher and keep nothing but a documented recipe. "Shrink" means keep the smaller part that still earns its place (usually the catalog). Each trigger names how to check it. Claims about what Claude Code does today are verified only where stage0.md verified them. Everything about future releases is unverified.

A. The launcher stops being worth building (stop or shrink to an alias recipe)#

B. The catalog stops being worth building#

C. Adoption and maintenance (applies to the whole project)#

What to watch and how#

Check at every Claude Code minor release, and at least monthly:

1. Release notes (code.claude.com/docs changelog and the GitHub releases of anthropics/claude-code) for the keywords above: profile, allowlist, default deny, enabledPlugins, skillOverrides, plugin enable, Discover, catalog, marketplace owner or status fields, managed.
2. claude --help and claude plugin --help, diffed against the saved previous output (new flags such as --plugins, --profile, --only-plugins).
3. The settings reference for new keys that affect plugin or skill loading, and the managed-settings reference for new keys that restrict --settings, --plugin-dir or --mcp-config.
4. The system/init event of a Stage 0 run (counts of plugins, skills, agents, tools, MCP servers) after each upgrade. A count that changes for the same inputs means the mechanism moved.
5. doctor --policy on a managed machine (O-01).
6. The tracker issues listed in landscape.md. The closed not_planned ones show demand without a plan. An open one with a milestone is the signal.

Review cadence#

• Every Claude Code minor release: items 1 to 4 above (15 minutes). Record the result as one line in the decision log only when something changed.
• Quarterly: K6 (tools), K8 to K11, and the adoption numbers. Write one dated line in DECISIONS.md either way. The entry "reviewed, no trigger" is valid.
• On the first trigger hit:
  1. Open a decision row with the evidence and the action.
  2. Discuss with the user before acting (stopping is the user's decision).
  3. Mark the superseded rows.

Non-goals (for now)#

Phase 2 lists the deferred items. Other non-goals:

• A registry server/DB.
• Vector search.
• A full-screen TUI dashboard (interactive prompts and pickers are in scope, see R6).
• Custom install path (bundles cover install).
• MCP gateway.
• Config-dir-per-profile.
• A concierge search tool before there is usage data.
• Anything that writes shared settings.

Checklist#

Tick items in the report to track them (saved in your browser). The checked state below is the committed state.

• Stage 0: launcher experiments on macOS (done 2026-10-06)
• Adversarial review round absorbed and facts corrected (done 2026-10-06)
• LICENSE added (MIT). Docs reorganized and the report generated from the Markdown
• Phase 0.1: routing eval on a real plugin set. Protocol and success criterion fixed (2026-10-06, routing-eval-protocol.md), runs pending the org's plugin set
• Phase 0.2: adopt, build or contribute: evaluated fuzzyalej/claude-profile and edimuj/claude-rig. Recommendation to build (D-33, adopt-or-build.md), pending user review
• Phase 0.3: bundle and masking prototype (T7). Needs a second config directory and a login
• Phase 0.4: Stage 0 on Linux and Windows
• Phase 0.5: measure MCP and connector removal through the settings keys (done 2026-10-06, saving about 5.7k tokens, 19%, see stage0.md)
• Phase 0.6: write kill and obsolescence criteria (done 2026-10-06, section above)
• Employer approval for open-sourcing (SECURITY.md and CONTRIBUTING.md are written)
• MVP launcher: run, show, dry-run, ls, dir sources, extends, default-deny masking, closed schema, settings validation
• MVP catalog: lint (owner, status, when_to_use) and a generated CATALOG.md
• Nightly real-claude smoke test before any release
• Built ahead of the evidence (see D-22): accounts, trust lockfile and git/plugin sources, compiled bundles, recommend, shell-init, static site, doctor, usage analytics.
• Not built yet: the taxonomy and the Analytics join beyond the OTel and usage-API reader, and the marketplace-source check for plugin sources (the hook exists, the launcher does not pass it yet)