Skip to main content

Roadmap

Living roadmap for the magus plugin marketplace. Per-plugin entries with known in-flight work, plus cross-cutting initiatives and explicit unresolved decisions. Backed by source citations — items without citations are owner-populate placeholders. Updated as work lands. Last refreshed: 2026-05-28.

Reading order:

  1. Status legend — what the badges mean
  2. Cross-cutting initiatives — work spanning multiple plugins (read first; per-plugin sections refer back)
  3. Per-plugin roadmap — 18 published plugins + 3 source-only + 2 tools
  4. Decisions needed — conflicts and gaps the maintainer must resolve

For backward-looking history, see CHANGELOG.md and RELEASES.md. For per-feature design docs, see docs/plans/. (Note: ai-docs/sessions/ is git-ignored session scratch — durable artifacts live under docs/plans/ or ai-docs/ root, not under sessions/.)

Status legend​

BadgeMeaning
🟢 DesignedPlan doc exists and is approved; implementation not yet started
🟡 In-progressWork has started (commits, partial implementation, R&D harness)
🔴 BlockedWork is gated on an external factor (upstream fix, missing harness, undecided trade-off)
⚪ SurfacedPossibility raised but no decision made (could go either way; may be dropped)
➖ No signalPlugin has no in-flight work in any plan doc or recent session — owner to populate
🚫 Anti-roadmapExplicitly deferred, declined, or out-of-scope (named here so it doesn't accidentally creep back)

Cross-cutting initiatives​

Work that spans multiple plugins or affects the marketplace as a whole. Per-plugin sections below reference these by number.

CC-1 · Cross-harness distribution (Codex + Antigravity) 🟡​

Generate per-harness distribution manifests from the magus source marketplace so plugins ship to Codex and Antigravity, not just Claude Code. Codex generator + validator are working (17 plugins generated, install validated, hook trust gate validated on CLI, Desktop app-server, and live Desktop UI). Antigravity is blocked on local CLI availability.

  • Scope: all 18 published plugins, claudeup, distribution pipeline
  • Open work: KI-CODEX-005 (Codex task-management translation; generator must fail if Codex artifacts need Claude-only task tools), KI-CODEX-008 (commit Desktop/app-server smoke as a harness command, not R&D)
  • Sources: docs/plans/2026-05-27-harness-mapping.md, docs/plans/2026-05-27-cross-harness-known-issues.md, docs/plans/2026-05-26-hook-tool-conversion-map.md, docs/plans/2026-05-28-codex-plugin-distribution-spec.md

CC-2 · Cowork guided install lane 🟢​

Build a guided plugin install flow in claudeup TUI: list available plugins with versions from GitHub Release manifests, search/filter, download to deterministic cache, checksum verification, deep-link generation. Update flow with cached-version comparison + re-verify. 6-phase plan; implementation unverified from sources.

  • Scope: all plugins (artifact packaging), claudeup TUI, website, GitHub Releases pipeline
  • Source: docs/plans/2026-05-26-cowork-guided-plugin-install.md

CC-3 · Skill description budget remediation (T3 consolidation) 🟢​

Remediate Claude Code 2.1.105+'s skill listing budget. T1, T2, T4, T5 shipped (44% char reduction; 25 skills flagged disable-model-invocation: true; scripts/skill-budget-check.ts enforces on release). T3 (architectural consolidation) identifies 5 router consolidations across dev, multimodel, terminal, code-analysis — ~17 listing entries → 5. Estimated 1-week design effort.

  • Scope: dev, multimodel, terminal, code-analysis
  • Source: scripts/skill-budget-check.ts (enforces on release). Detailed T3 plan lives in session scratch from the 2026-05-06 skill-budget research/build sessions — needs promotion to docs/plans/ before further tracking.

CC-4 · best-practice-gates plugin 🟢​

New runtime plugin: rule-library + reranker + background validators. Rule library is loose YAML files curated by claudeup. Per-rule cross-model validators run as async: true hooks, deliver verdicts via additionalContext system reminders into actor's next turn. Bun + TypeScript end-to-end (no shell shims). Per-session state in .claude/best-practice-gates/session.{session_id}.json (no cross-session shared state by design). 8 design constants resolved; ready for implementation.

  • Scope: new plugin plugins/best-practice-gates/; depends on claudish + mnemex + multimodel
  • Acceptance gate: madbench trajectory-replay of the two evidenced failures (R1 fork-vs-extend; R2 mock-coverage-as-validation) — control variant (no bpg) reproduces failure, treatment variant (bpg installed) catches and fixes. Per-rule madbench suites land in v1.1 alongside the starter rule pack.
  • Claudeup integration: plugin install via existing claudeup plugin flow; new "Rules" tab for browsing/toggling rule packs; static curated rule-pack manifest shipped with claudeup (real registry deferred to v1.1).
  • Rule sources: wraps select existing magus skills (e.g., dev:verification-before-completion, dev:enforcement) as validator prompts; receives software-architecture gates being authored in the user's cowork project as a downstream supplier (no schema/naming forced now).
  • Source: docs/plans/2026-05-28-best-practice-gates-design.md

CC-5 · Tasks system migration (TodoWrite → TaskCreate/Update/List/Get) 🟡​

Migrate all plugins from legacy TodoWrite to the new TaskCreate/TaskUpdate/TaskList/TaskGet task system. Approved 2026-01-30; originally scoped at ~115 files across 12 plugins (dev: 29 files HIGH priority; multimodel: 7; seo: 13; etc.). Scope shrank by conductor's 11 files when it was retired at magus v8.0.0. Completion status not verified by mining.

  • Conflict with CC-1 — see "Decisions needed" below.
  • Source: docs/plans/2026-01-30-tasks-migration-design.md

CC-6 · Multi-target distribution (magus / magus-alpha / magus-marketing) 🟡​

Per-plugin distTargets field in marketplace.json routes plugins to distribution channels. magus-alpha live (autolinear shipped via it). magus-marketing target slot exists (v0.1.0) with zero plugins assigned.

  • Scope: all plugins via opt-in distTargets
  • Sources: .claude-plugin/marketplace.json targetMetadata block; CLAUDE.md "Alpha Marketplace" section

CC-7 · Dependency-graph correctness (claudish + mnemex extraction) ✅ shipped 2026-05-09​

Code-analysis, dev, multimodel, designer, seo now declare claudish / mnemex as dependencies per Anthropic's documented pattern (Claude Code v2.1.110+). Listed here for context, not for further work.

CC-8 · Marketplace.json schema evolution + Claude --strict drift ⚪​

Source marketplace.json uses Magus-only fields (targetMetadata, distTargets) that fail Claude --strict validation. Generated per-target distribution manifests strip them. No schema unification planned in sources; could be revisited.

  • Source: docs/plans/2026-05-27-cross-harness-known-issues.md KI-CLAUDE-002

CC-9 · Documentation drift (CLAUDE.md / AGENTS.md) ✅ DONE​

First addressed at magus v8.0.0, then regressed and fixed properly on 2026-08-08.

The v8.0.0 pass restructured both files around the three distribution channels and hand-corrected the versions. Because nothing gated them, they drifted again: 39 wrong version numbers, three plugins that never existed (Agent Development in both files at two different versions, Nanobanana in AGENTS.md), and bunjs missing from AGENTS.md entirely.

The v8.0.0 pass also fixed AGENTS.md's four broken skill references (code-analysis:claudemem-search → mnemex-search, claudemem-orchestration → mnemex-orchestration, the nonexistent architect-detective → investigate, and code-analysis:claudish-usage → multimodel:), its .Codex-plugin/ find-replace artifacts, and its unrunnable --model Codex-sonnet-4-6 examples.

The 2026-08-08 fix removed the version columns outright rather than regenerating them: marketplace.json is the authority and validate-versions.js already gates it, so a second copy could only ever be wrong. scripts/check-doc-plugin-lists.ts now enforces the plugin names and rejects a version column growing back, and scripts/check-doc-references.ts validates every plugin/agent/skill/command reference across the repo. Retired-plugin mentions in RELEASES.md/CHANGELOG.md are left as historical record.

Standing risk: this drifts every release because nothing regenerates it. The plugin tables are hand-maintained against marketplace.json. A scripts/ check that diffs the doc tables against the manifest would make it self-policing.

CC-10 · Marketplace deletion bug — discovery still breaks 🔴​

Plugin discovery in /doctor still breaks during Claude Code's cacheMarketplaceFromGit() delete-then-clone race. Hooks survive it, because the loader reads content from the immutable cache. Discovery needs an upstream Claude Code fix; not tracked.

No workaround was ever shipped. This item used to claim one ("git-subdir sources"). Magus uses plain string sources — all 20 of them — and publish-dist.sh:306 converts git-subdir entries away on publish. Cache-reading is the loader's normal behaviour for every source type except directory marketplaces, so nothing was done to earn it. Corrected 2026-08-06.

  • Source: CLAUDE.md "Marketplace directory deletion bug" section

Per-plugin roadmap​

Listed in .claude-plugin/marketplace.json order. Each section: in-flight items and surfaced/undecided items.

No version numbers here. They belong in .claude-plugin/marketplace.json, which is the authority and is gated by scripts/validate-versions.js. These headings used to carry a (current: vX.Y.Z) label; 11 of 17 had drifted, dev by two minor versions, and this file is published to users. scripts/check-doc-plugin-lists.ts now rejects the label.

code-analysis​

  • 🟡 CC-1 · PreToolUse:^Bash$ hook ("Block accidental git add of private files") needs Bun adapter and Codex generated-plugin trust gate before shipping for Codex
  • ⚪ CC-3 · mnemex-* skill cluster consolidation (2 entries → 1 router) deferred to T3
  • ⚪ CC-1 · Desktop UI plugin browse path still needs validation; Codex marketplace already includes 17 plugins

claudish (runtime plugin, new in v7.5.0)​

  • 🟡 CC-1 · Codex MCP smoke harness: end-to-end claudish session polling + Codex-native MCP progress/elicitation/reporting must be recorded before claiming Codex parity (research complete, smoke not committed)
  • ⚪ Backward-compat shim: --channels plugin:code-analysis@magus kept as alias for new canonical --channels plugin:claudish@magus. Long-term shim removal unscheduled.
  • ⚪ Marketplace registration for non-magus consumers (Codex manifests have no dependencies field; claudish@magus-codex install must be driven explicitly)

mnemex (runtime plugin, new in v7.5.0)​

  • 🟡 CC-1 · System-wide mnemex executable must be installed/repaired by claudeup before installing/validating mnemex@magus-codex (KI-CODEX-009)

multimodel​

  • 🟡 CC-1 · PreToolUse:Task / PreToolUse:Bash hooks need Codex redesign — Codex has no Task hookable tool; switch to SubagentStart / SubagentStop
  • ✅ agent-enforcement's unrecognized triggers: frontmatter removed 2026-08-08, along with 8 other unread keys (version, updated, plugin, keywords, tags, namespace, globs, skills) across 56 skill files. Only keys the matcher actually reads remain
  • ⚪ CC-3 · Skill router consolidations deferred to T3 (names below are proposed, not yet created): multimodel:orchestration (delegate-patterns + hierarchical-coordinator + multi-agent-coordination + task-orchestration → 1); multimodel:claudish (3 → 1)

seo​

  • 🟡 CC-1 · SessionStart hook needs Antigravity first-run PreInvocation shim before shipping

video-editing​

  • ➖ No in-flight signal found in plan docs or sessions — owner to populate

image-generate​

  • ➖ No in-flight signal found in plan docs or sessions — owner to populate

conductor — RETIRED​

  • ⚫ Removed at magus v8.0.0: outdated and unused. Marketplace entry and plugins/conductor/ deleted; the conductor-missing-for-multi-session-feature coaching rule was dropped with it. Source recoverable from git history.

dev​

  • 🟡 CC-1 · SessionStart, PreToolUse, Stop, TaskUpdate hooks need cross-harness redesign (Codex has no TaskCreate/TaskUpdate hook parity)
  • 🟡 CC-1 · Transcript-based coaching hook: transcript field availability needs fixture validation for Codex + Antigravity before shipping
  • 🟢 CC-5 · TodoWrite → Tasks migration: HIGH priority, 29 files (12 commands + 14 agents + 3 discipline skills) — see Decisions Needed for the CC-1 conflict
  • 🟢 E2E test suite for the feature-development loop — two-layer architecture (deterministic unit + claudish 6-model integration, 30 cases). Approved 2026-02-28 against the then-current /dev:loop, which v3.0.0 folded into /dev:dev; the plan needs retargeting before implementation. Status unverified
  • 🟡 Phase instruction file loading enforcement in /dev:dev Full depth (commit e030c2e) — active development
  • ⚪ CC-3 · dev:discipline skill router consolidation (8 → 1) deferred to T3 — proposed name, not yet created

bunjs​

  • 🔴 Highest priority · The six data widgets accept no layout props while Panel does. Meter, Sparkline, HeatRow and StackedBar take a numeric width and expose no flexShrink / minWidth; Panel got a PanelLayout forwarding pass and they did not. A row one column over budget inside the flexDirection="row" composition the aesthetics reference prescribes makes Yoga squeeze the FIRST child, rendering a badge as a single 1-column cell of coloured background. That is invisible in captureCharFrame() — a 1-column bg stub reads as a space — so every automated test the skill ships still passes; only a raw-ANSI dump reveals it. Fix shape: forward flexShrink={0} / minWidth on the data widgets exactly as Panel does, so an over-budget row fails loudly (overflow) instead of silently mangling its first child.
  • 🔴 scripts/check-surface.ts:118 warns on every ordinary app project: it prints WARNING 0 fenced blocks found — extraction may be broken over a normal bun init tree, whose README.md/CLAUDE.md carry no ts fences. SKILL.md lists the script under Acceptance for app projects, so the noise lands exactly where it is recommended. Fix shape: a --docs mode scoping the "scanned nothing" alarm to markdown expected to hold snippets, keeping exit-2 for the skill's own tree where it is the correct guard.
  • 🔴 Vertically adjacent same-coloured badges merge: 24 green UP chips padded to a fixed label width fused into one solid green rectangle down the column. Per-cell output is correct, the visual is wrong. Fix shape: one line in aesthetics-and-color.md — pad with a plain filler span outside the chip so the background does not run between rows.
  • ⚪ Right-aligning a group in a one-row header/footer is undocumented. The aesthetics reference prefers gap over spacer boxes but never says what pushes a group to the last column; the answer is justifyContent="space-between" over two row groups, which both fresh-consumer builds worked out independently.
  • ⚪ HeatRow's global-max mandate makes low-volume rows near-black. widgets.tsx mandates ONE max across all rows, so a 5xx row at ~10% of 2xx renders almost black — correct per the cross-row-comparability contract, but it reads as a dead panel. Options: an opt-in per-row max, or document the trade-off and suggest splitting scales when magnitudes differ by an order of magnitude.
  • 🔴 CI job test-bunjs-skill written but not committed (~12 lines in .github/workflows/test-plugins.yml: working-directory: plugins/bunjs/skills/opentui-tui, then install / test / typecheck / check-surface). Withheld because the render tests drive the real native Zig renderer — green on macOS arm64, UNVERIFIED on ubuntu-latest — and committing it blind risks turning repo CI red. Add it when someone can watch one run.
  • Provenance: all six found and confirmed during the v0.1.0 build (two fresh-consumer builds, three code reviews, one execution verifier) and consciously deferred; none blocks release. Whether dev's four bunjs* skills migrate here is a separate open question — see D-6.

statusline​

  • ➖ No in-flight signal found beyond shipped bugfixes (reset countdowns, diff chip split, memory display) — owner to populate

browser-use​

  • 🟡 CC-1 · Python/pip browser-use upstream runtime treated as external dep; validate install/doctor output

designer​

  • ➖ No in-flight signal found beyond skill-budget T1.1 (already shipped) — owner to populate

terminal​

  • 🟡 CC-1 · PreToolUse:^Bash$ destructive-tmux-kill-server hook needs Bun adapter and Codex generated-plugin trust gate
  • ⚪ CC-3 · terminal:run skill router consolidation (5 → 1) deferred to T3

gtd​

  • 🟡 CC-1 · GTD task workflow has no Codex parity — TaskCreate/TaskUpdate hooks must be redesigned; generator must fail if Codex artifacts require Claude-only task tools. Codex smoke for task-heavy workflow required.

kanban​

  • ⚪ Legacy .claude/gtd/tasks.json migration helper — currently NOT auto-migrated by design (users re-add via /kanban:add). Auto-migration helper undecided.

instantly​

  • 🟡 CC-1 · SessionStart hook needs Antigravity first-run PreInvocation shim before shipping

autolinear (magus-alpha)​

  • 🟡 Autonomous webhook-triggered pickup — receiver wired but queue → Claude Code dispatch is a TODO stub
  • 🟡 CC-1 · SessionStart hook (Linear/webhook readiness) + PreToolUse:Task (Linear state before delegation) need Codex redesign
  • 🟡 Deferred from 2026-04-15 rename: landing-page marketing (D). The claudeup-core test fixture item (B) is moot — tools/claudeup-core/ was deleted when claudeup collapsed to a single package.

Source-only plugins (not in magus marketplace)​

stats (source-only, current: v0.1.0)​

  • 🟡 CC-1 · All-tool capture hook (SessionStart, PreToolUse, PostToolUse, Stop) has Codex partial coverage
  • 🔴 stats/skills/stats-reading/SKILL.md has no parseable description — open follow-up
  • ⚪ Marketplace publication path undecided

dingo (source-only, current: v1.0.0)​

  • ➖ No in-flight signal; marketplace publication path undecided

go (source-only, current: unknown — no plugin.json)​

  • 🔴 Plugin scaffolding status unclear (only knowledge/ subdir; no plugin.json or .claude-plugin/)

Tools​

claudeup (current: v4.17.0+, npm package)​

  • 🟢 CC-2 · Cowork TUI screen (Phase 4) — plugin list from GitHub Release manifests, search/filter, deterministic cache, checksum verification, deep-link generation
  • 🟢 CC-2 · Cowork install update flow (Phase 5) — cached version comparison, redownload + re-verify, guided update prompt
  • 🟡 CC-1 · Mnemex repair before mnemex@magus-codex install (per KI-CODEX-009)
  • 🟢 Raise the @opentui/core pin. tools/claudeup pins ^0.1.75 (published 2026-01-25) and 0.1.107 is verified compile-safe: bun build --compile, no --external, binary launches, exit 0, 66,534,800 B — 32 patch versions of headroom inside the safe line. 0.4.x still breaks the compiled binary at launch (resolveFallbackFilePath → normalizeLoadedFilePath, native lib resolution inside /$bunfs/root/), so the 0.1.x ceiling stands, but the pin does not have to stay at 0.1.75. Surfaced by the bunjs v0.1.0 build.
  • ⚪ CC-2 · Optional helper command surface (claudeup cowork list/install/update non-interactive)
  • ⚪ Trust action for Codex Desktop hooks — security trade-off: "may guide users to the Hooks screen or offer an explicit 'trust these installed hooks' action only if the user asks for it"

claude-desktop-profiles (separate tool, current: v1.2.2)​

  • 🟢 Path 2 · Developer-ID broker daemon for cryptographic keychain isolation (~500-800 LOC, 1-2 weeks). Iteration-3 research determined ad-hoc partition lists do NOT enforce isolation; v1.2.0 shipped Path 1 (snapshot/restore). Path 2 user decision required before starting.

Candidate plugins​

Categories raised as worth covering, none of them designed or committed. Moved here from README.md, which describes what ships today rather than what might. ⚪ Surfaced — each could go either way, and none has an owner or a plan doc.

CategoryShape
Testing toolsE2E testing, visual regression, performance testing
UI componentsDesign system tooling, component generators
Backend developmentNode.js, API design, database tooling
DevOpsDocker, Kubernetes, CI/CD automation
DocumentationGenerated docs, API reference, guides

Community suggestions arrive as GitHub issues labelled plugin-request. A category earns a row in "Per-plugin roadmap" only once it has a plan doc, per the maintenance protocol below.

Decisions needed​

Items where the maintainer must choose before downstream work can proceed. Not just "TBD" — these are real conflicts or blocking choices.

D-1 · Tasks migration vs Codex distribution conflict​

CC-5 (TodoWrite → Tasks migration, approved 2026-01-30) targets the dev and gtd plugins. CC-1 (cross-harness distribution) explicitly says Codex has no TaskCreate / TaskUpdate hookable tools, and the Codex generator must fail if artifacts require them (docs/plans/2026-05-27-cross-harness-known-issues.md KI-CODEX-005). The two plans are both approved; the reconciliation path is not stated in any source.

The decision: does dev/gtd ship the Tasks-migration changes as Claude-only features (with Codex distribution permanently lagging), or does the migration get rolled back / refactored to use a Codex-compatible mechanism?

D-2 · Four plugins with no in-flight roadmap​

video-editing, image-generate, statusline, designer have shipped bugfixes recently but have no forward-looking roadmap items in any plan doc or session. Either they're in steady-state (correct → mark "maintained, no active roadmap") or they have unwritten direction (owner must populate). conductor was the fifth — resolved by retiring it at magus v8.0.0.

D-3 · claude-desktop-profiles Path 2 commitment​

Iteration-3 research from the 2026-05-09 electron-keychain-isolation session (still in session scratch — needs promotion to docs/plans/ before further tracking) explicitly posed the cryptographic isolation question as a maintainer decision. Path 1 (snapshot/restore) shipped in v1.2.0; Path 2 (broker daemon) is 1-2 weeks of work and needs an explicit go/no-go.

D-4 · Source-only plugin marketplace publication​

stats, dingo, go live in plugins/ but don't ship via the marketplace. Each needs a yes/no on publication path. go additionally lacks any plugin scaffolding.

D-5 · CLAUDE.md inventory drift​

Trivial but high-impact: documentation says "12 published plugins"; reality is 18. Combined with stale AGENTS.md references (retired plugins, old versions, old names). One-shot maintainer pass to fix.

D-6 · Do dev's four bunjs* skills migrate into the bunjs plugin?​

plugins/dev/skills/backend/ holds bunjs (839), bunjs-architecture (840), bunjs-production (988) and bunjs-apidog (855) — 3,522 lines, all server-side, all disable-model-invocation: true. The new plugin was deliberately named bunjs and described Bun-broad so they COULD move without a rename. They were not moved in v0.1.0: 3,522 lines of unrequested churn, and it would change dev's surface.

The decision: move them or leave them. The status quo is the awkward option — a bunjs plugin holding only a TUI skill while dev holds four bunjs* skills. This is a product call for the repo owner, not an engineering task; nothing downstream is blocked on it.

Anti-roadmap (explicitly NOT planned)​

Named here so they don't accidentally creep back as roadmap items in future planning.

  • 🚫 Cowork silent plugin install for individual users — admin/org deployment only; explicit "Product Boundary" non-goal
  • 🚫 Public Cowork deep link that opens Customize > Plugins directly / clicks upload / bypasses trust
  • 🚫 Claude Code --scope project as a Cowork install mechanism
  • 🚫 Codex project-scoped plugin enablement — codex plugin add is user-scoped only; don't promise project scope in docs
  • 🚫 Claude-style realtime/dev-channel parity in Codex — realtime_conversation disabled; claudish in Codex runs through MCP only
  • 🚫 Generic conversation normalization layer (cross-harness V1) — explicit V1 non-goal
  • 🚫 claude-desktop-profiles iter-2 helper-rename recipe — superseded by iter-3 finding
  • 🚫 Auto-migration of legacy GTD tasks into kanban v1.6.0 — explicit non-goal in v1.6.0
  • 🚫 Tightening SKILL_BUDGET_FAIL_TOTAL from 16000 → 12000 — deferred until T3 makes it achievable
  • 🚫 claudeup writing trusted hook hashes for normal users (Codex Desktop) — security rule; automation may write trust only in isolated smoke tests

Maintenance protocol​

This file is the single source of truth for direction. Updated as work lands.

  • When shipping a plugin update: if it completes a roadmap item, move the item to CHANGELOG.md and remove it from here.
  • When approving a new design plan in docs/plans/: add the corresponding roadmap entry (🟢 Designed) with a link to the plan doc.
  • When work starts: flip 🟢 → 🟡.
  • When blocked: flip to 🔴 with a one-line reason.
  • When deciding to drop something: move it to the Anti-roadmap section with the reason.

Per-plugin entries without citations are placeholder slots for the plugin owner to populate.

Provenance: initial inventory mined from 8 plan docs, CHANGELOG, RELEASES, marketplace.json, recent session reports, and 90 days of git log on 2026-05-28.