Skip to content

vTRKA/supervibe

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

63 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Supervibe

Compliance notice: This tool is designed exclusively for development assistance. By using it, you agree to comply with the Terms of Service (ToS) and Acceptable Use Policy (AUP) of all involved services, including Anthropic. Unauthorized automated usage, OAuth token abuse, or violation of third-party policies is the sole responsibility of the end user.

A plugin that turns Claude Code, Codex, and Gemini into a team of 83 specialist agents with a code graph, project memory, design intelligence, and confidence gates. Runs locally. No Docker.

v2.0 · MIT · Windows / macOS / Linux


What you get

Feature What it means
83 specialist agents ≥250 lines each: persona, decision tree, procedure, output contract, anti-patterns, verification
Code graph (10 languages) tree-sitter symbols and edges. Query --callers X, --callees Y, --neighbors Z --depth 2
Semantic code search multilingual-e5-small. Works offline. Speaks Russian, English, and 100 other languages
Project memory Five categories with FTS5 plus per-chunk embeddings. Decisions get reused, not rederived
Confidence engine Seventeen rubrics. Gate at score ≥9. Override rate above 5% triggers an audit
25 discipline rules operational-safety, compact agent-excellence-baseline validation, use-codegraph-before-refactor, single-question-discipline, design-system-governance, agent-install-profiles, anti-hallucination, and more
Auto-reindex A PostToolUse hook plus an mtime scan on session start. The memory:watch daemon is optional
Agent evolution loop Telemetry, underperformer detection, and /supervibe-strengthen with a user gate
Re-dispatch suggester When a Task finishes at confidence < 8.0, the hook checks past high-confidence runs on similar tasks and prints a [supervibe] dispatch-hint: with up to 3 alternative agents — never auto-dispatches
Autonomous loop /supervibe-loop turns a reviewed plan, PRD, epic, or validation request into a bounded, visible, cancellable agent loop with task graph scheduling, work-item templates, provider permission audit, side-effect ledger, and 9/10 confidence completion
Security audit loop /supervibe-security-audit runs read-only multi-agent AppSec/dependency/ops/AI security review, ranks vulnerabilities, then optionally plans, executes, and re-audits remediation to a 10/10 gate
Prompt AI engineering Optional prompt-ai-engineer add-on strengthens prompts, agent instructions, intent routing, structured outputs, tool policies, evals, and prompt-injection defenses
Network/router agent Optional network-router-engineer add-on handles routers, VPN, firewall, Wi-Fi, DNS/DHCP, and routing stability with read-only diagnostics first and scoped approval before mutations
Live preview server localhost:PORT with SSE hot reload, idle shutdown, and a max-server limit
Browser feedback channel 💬 click-to-comment overlay injected into preview pages — comments arrive as <system-reminder> on next user prompt via UserPromptSubmit hook (zero-dep WebSocket via node:net)
Design pipeline (5 targets) web · chrome-extension · electron · tauri · mobile-native — specialist designer per target, viewport presets, brandbook baselines, target-aware handoff adapters (RN / Flutter / MV3 / Electron renderer / Tauri webview)
Design intelligence (2.0) Internal RAG data pack for designer agents: memory + code + retrieval-backed style, UX, charts, decks, collateral, and stack UI evidence through existing /supervibe-design and /supervibe-audit flows
Component library bridges shadcn / MUI / Mantine / Radix / HeadlessUI — token bridge generated from approved design system
Pre-write prototype guard PreToolUse hook blocks writes to prototypes/<slug>/ until config.json exists AND blocks framework imports — prototypes stay native HTML/CSS/JS
Multi-CLI One installer wires Claude Code, Codex, and Gemini together

24 stacks supported: PHP (Laravel) · TypeScript / JavaScript (Next.js, Nuxt, Vue, Svelte, React, Express, NestJS) · Python (FastAPI, Django + DRF) · Ruby (Rails) · Java / Kotlin (Spring) · C# (ASP.NET) · Go · Mobile (Flutter, iOS, Android) · Browser Extensions (Chrome MV3 / WXT / Plasmo / Vite-CRXJS) · GraphQL · PostgreSQL · MySQL · MongoDB · Elasticsearch · Redis.


Install

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/vTRKA/supervibe/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/vTRKA/supervibe/main/install.ps1 | iex

The installer auto-detects every supported AI CLI on your machine and registers the plugin. Re-running it is a clean managed reinstall: tracked local edits stop the install, while stale untracked/ignored files from older plugin versions are removed before dependencies and generated registries are rebuilt.

Claude Code (auto-detect):

curl -fsSL https://raw.githubusercontent.com/vTRKA/supervibe/main/install.sh | bash

Cursor:

/add-plugin supervibe

Or search for "supervibe" in Cursor's plugin marketplace.

OpenCode: Add to your opencode.json:

{
  "plugin": ["supervibe@git+https://github.com/vTRKA/supervibe.git"]
}

GitHub Copilot CLI:

copilot plugin marketplace add vTRKA/supervibe-marketplace
copilot plugin install supervibe@supervibe-marketplace

Gemini CLI:

gemini extensions install https://github.com/vTRKA/supervibe

Codex CLI: Use the one-line installer above. For Codex it registers the official plugin cache at ~/.codex/plugins/cache/supervibe-marketplace/supervibe/local, enables [plugins."supervibe@supervibe-marketplace"] in ~/.codex/config.toml, keeps a legacy ~/.codex/plugins/supervibe link for older wrappers, and links ~/.agents/skills/supervibe for native skill discovery in Codex/Zed ACP sessions. Current Codex supports this skills/config surface; plugin slash-command, agent, and hook manifest fields are not advertised to Zed by codex-acp.

Restart your AI CLI. On the next session you should see:

[supervibe] welcome — plugin v2.0.12 initialized for this project
[supervibe] code RAG ✓ N files / M chunks (fresh)
[supervibe] code graph ✓ N symbols / M edges (X% resolved)

Check multi-host readiness at any time:

npm run supervibe:doctor -- --host all

Requirements: Node.js 22.5+ and Git. The installer checks node:sqlite before registration; if Node is missing or too old, it asks for explicit consent to install or upgrade Node and only continues after SQLite/RAG/CodeGraph can run. Git LFS is optional, but the ONNX embedding model is not: the installer prepares it before registration, using Git LFS when available and a direct HuggingFace download fallback otherwise. No Docker, no Python, no native compile step.

For unattended installs, set SUPERVIBE_INSTALL_NODE=1 to allow Node bootstrap or SUPERVIBE_INSTALL_NODE=0 to fail fast with manual instructions.

Release integrity evidence is documented in release security, install integrity, and third-party licenses. The default installer follows main; strict installs can set SUPERVIBE_REF, SUPERVIBE_EXPECTED_COMMIT, and SUPERVIBE_EXPECTED_PACKAGE_SHA256 before running the installer.

Update

Three ways, pick whichever fits your context:

One-liner (matches the install style):

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/vTRKA/supervibe/main/update.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/vTRKA/supervibe/main/update.ps1 | iex

From the AI CLI session:

/supervibe-update

Auto-update policy:

Supervibe checks upstream in the background from Claude Code SessionStart when the plugin root is available. The default mode is managed: managed installer checkouts under ~/.claude/plugins/marketplaces/supervibe-marketplace auto-apply safe git updates in the background, while dev/manual/IDE checkouts stay notify-only so local work is not pulled unexpectedly.

Override with:

SUPERVIBE_AUTO_UPDATE=apply npm run supervibe:auto-update -- --refresh
SUPERVIBE_AUTO_UPDATE=check  npm run supervibe:auto-update -- --refresh
SUPERVIBE_AUTO_UPDATE=off    npm run supervibe:auto-update -- --status

Host coverage: Claude Code gets host autoUpdate registration plus the SessionStart background check/apply path; OpenCode follows the git source on restart; Codex gets official plugin cache/config registration plus native skill links for Zed ACP sessions; Gemini symlink/include installs follow the managed checkout after it updates; Cursor and other IDE/manual installs rely on their host marketplace where available and npm run supervibe:doctor -- --host all to surface drift.

Manually from the plugin checkout:

cd ~/.claude/plugins/marketplaces/supervibe-marketplace
npm run supervibe:upgrade

All three do the same thing: refuse tracked edits in the plugin checkout, clean stale untracked/ignored files, then git pull --ff-only + LFS pull + npm ci + rebuild generated registry.yaml + run all tests + run the install lifecycle doctor + refresh the upstream-check cache. Restart the AI CLI afterwards.


Workflows

Brainstorm -> Plan -> Review -> Atomize -> Safe Run

The trigger-safe path is explicit and chainable:

  1. /supervibe-brainstorm <topic> writes the approved spec, then asks: Следующий шаг - написать план. Переходим?
  2. /supervibe-plan <spec-path> writes the plan, then asks for the review loop before execution.
  3. /supervibe-plan --review <plan-path> reviews plan quality, safety, missing checks, and README impact.
  4. /supervibe-loop --from-plan --atomize <plan-path> splits the reviewed plan into atomic work items and an epic.
  5. /supervibe-loop --guided --max-duration 3h runs in the current session after provider-safe preflight, explicit approval, side-effect ledger setup, and stop/resume/status controls. Worktree is optional: add --worktree only when you want isolated or parallel sessions.

Diagnostics are first-class: use /supervibe --diagnose-trigger when a phrase did not route as expected, and /supervibe --why-trigger to explain the selected command, selected skill, confidence, missing artifacts, and safety blockers. The router also has a semantic intent layer for implicit needs: "I cannot see epics/tasks", "old tasks are cluttering memory", "agents do not use tools", "RAG/codegraph wastes tokens", "docs has internal TODO garbage", and "Figma tokens drift from code" all route to the nearest safe command without requiring slash-command phrasing. Long-running work stays visible through stop/resume/status commands and never attempts provider bypass, hidden background execution, or policy evasion.

Unreleased capability label: the durable autonomous loop is implemented in this workspace and remains opt-in until the release gate publishes it. Autonomous execution is opt-in, not the default. The default path is read-only planning, review, atomization preview, status, diagnostics, and dry-run artifacts.

Copy-paste path from brainstorm -> reviewed plan -> atomized epic -> safe execution:

/supervibe-brainstorm "idea"
/supervibe-plan --from-brainstorm docs/specs/example.md
/supervibe-plan --review docs/plans/example.md
/supervibe-loop --atomize-plan docs/plans/example.md --plan-review-passed
/supervibe-loop --guided --max-duration 3h
/supervibe-loop --epic example-epic --worktree --max-duration 3h
/supervibe-loop --epic example-epic --worktree --assigned-task T1 --assigned-write-set src/auth.ts --max-duration 3h
/supervibe-loop --status --epic example-epic
/supervibe-loop --resume .claude/memory/loops/example-run/state.json
/supervibe-loop --stop example-run

Execution modes are explicit: --dry-run, --guided, --manual, --fresh-context --tool codex|claude|gemini|opencode, and optional worktree-backed execution with --worktree. A single current-session run is valid when the user does not need isolation or parallel sessions. Provider prompts, rate limits, network/MCP approvals, secrets, billing, deploys, production mutations, and credential changes are never bypassed. Missing credentials, missing provider permissions, CI/external access failures, worktree conflicts, policy stops, stale claims, and sync drift become blocked states with a next safe action.

For parallel work on one epic, scope each worktree session with --assigned-task and --assigned-write-set. The registry is lock-protected and status output shows each active session's wave, task IDs, write-set, agents, and path so separate sessions do not silently claim the same zone.

Implemented loop operations now include graph inspection, doctor/repair, fresh-context prime summaries, context packs, PRD/story intake, visual local control, reversible GC, and safe export/import bundles:

/supervibe-loop --from-prd docs/specs/checkout.md --dry-run
/supervibe-loop --atomize-plan docs/plans/example.md --dry-run
/supervibe-loop --tracker-sync-push --file .claude/memory/work-items/example-epic/graph.json
/supervibe-loop graph --file .claude/memory/loops/<run-id>/state.json --format text
/supervibe-loop doctor --file .claude/memory/loops/<run-id>/state.json
/supervibe-loop prime --file .claude/memory/loops/<run-id>/state.json
/supervibe-ui --file .claude/memory/work-items/example-epic/graph.json
/supervibe-gc --all --dry-run
/supervibe-loop export --file .claude/memory/loops/<run-id>/state.json --out .claude/memory/bundles/<run-id>

Work-item status uses the same vocabulary in CLI, reports, and query answers: ready, blocked, claimed, stale, orphan, drift, review, and done. State and evidence live under .claude/memory/loops/, atomized epics under .claude/memory/work-items/, external tracker mappings in task-tracker-map.json, and archived/exported run bundles under .claude/memory/bundles/.

The localhost UI is universal across IDEs: run /supervibe-ui or npm run supervibe:ui -- --file <graph.json>, then open the printed 127.0.0.1 URL in a browser or IDE webview. It shows epics, tasks, selected context packs, loop state.json, waves, gates, SLA reports, GC previews, and RAG/memory/codegraph health tabs. The Overview phase rail is derived from real graph/run state and marks plan, atomize, execute, verify, close, and archive as complete/current/blocked/pending from task status, gates, waves, and archive markers. The Kanban view groups real work items by ready/claimed/blocked/ deferred/review/done and shows each task's epic, agent, blockers, verification count, and write scope. Context packs and fresh-context loop prompts include a compact workflowSignal with the same phase, epic, task, claim, gate, and next action signal, so agents act from the same state that the UI shows. Mutating actions are local-only and require a preview plus explicit apply confirmation. npm run supervibe:ide-bridge -- --out .supervibe/ide-bridge.json writes a portable descriptor that IDE webviews can consume without becoming a separate task store.

Long-lived projects can prune clutter without deleting evidence: /supervibe-gc and npm run supervibe:gc -- --all --dry-run preview completed epic archival and memory cleanup; --apply moves candidates into reversible .archive/ folders with JSONL audit logs. npm run supervibe:status -- --gc-hints prints the same cleanup signal inside the normal status flow.

Production-prep may complete autonomously when evidence is complete, but production mutation, destructive migration, credential mutation, billing, DNS, account, access-control, and remote server changes remain blocked without an exact approval lease.

Policy profiles make those boundaries explicit per project and role. Use /supervibe-status --policy to inspect the active profile, /supervibe-status --role to inspect team governance, and /supervibe-loop --policy-profile guided|contributor|maintainer|CI-readonly|CI-verify to run with a named local profile. Approval receipts live in a scoped expiring local ledger and can be listed with /supervibe-loop --approval-receipts. Details are in policy profiles.

Semantic anchors are optional local hints for large files. They let agents find stable regions, file-local contracts, invariants, and per-file change summaries without requiring heavy markup everywhere. Use /supervibe-status --anchors --file src/example.ts to inspect one file and /supervibe-loop --anchor-doctor to check derived anchor drift. Details are in semantic anchors.

Multi-agent orchestration can be inspected before any fan-out. Use /supervibe-loop --plan-waves docs/plans/example.md to see safe parallel waves, /supervibe-loop --assign-ready --explain --file <state.json> to see worker/reviewer reasoning, and /supervibe-status --assignment <task-id> --file <state.json> to answer why a task was assigned or serialized. Details are in multi-agent orchestration.

Three named flows cover most of the day-to-day use. Each has an explicit slash-command entry point — no need to remember the right phrase to make the AI pick the right skill.

Brainstorm → Plan → Execute

For any new feature, component, or behavior change.

/supervibe-brainstorm payment idempotency
  ↓ collaborative dialogue, kill criteria, decision matrix
  ↓ saves docs/specs/2026-04-28-payment-idempotency-design.md
  ↓ score ≥9 against requirements rubric
/supervibe-plan docs/specs/2026-04-28-payment-idempotency-design.md
  ↓ phased TDD plan, parallelization batches, risk register
  ↓ saves docs/plans/2026-04-28-payment-idempotency.md
  ↓ score ≥9 against plan rubric
  ↓ choose: subagent-driven OR inline execution

You can skip /supervibe-brainstorm if you already have an approved spec, or skip /supervibe-plan for trivial one-line changes.

Design pipeline → Live preview → Browser feedback

For any visual surface — web landing, in-product flow, browser extension, Electron / Tauri desktop, or mobile native.

/supervibe-design landing in the style of Linear, focused on dev-tool buyers
  ↓ Stage 0: target surface (web | chrome-extension | electron | tauri | mobile-native)
  ↓ creative-director: brand direction (mood-board, tokens, animation library, graphics medium)
  ↓ brandbook: target-aware baselines + Section 6.5 component library decision
  ↓ optional: component-library-integration (shadcn / MUI / Mantine / Radix / HeadlessUI bridge)
  ↓ specialist designer per target (extension-ui-designer / electron / tauri / mobile / ux-ui)
  ↓ copywriter: every visible string nailed
  ↓ prototype-builder: 1:1 HTML/CSS in prototypes/<slug>/ (native only — pre-write hook enforces)
  ↓ AUTO: supervibe:preview-server spawns http://localhost:NNNN with hot reload + 💬 feedback overlay
  ↓ ui-polish-reviewer + accessibility-reviewer in parallel
  ↓ feedback loop (✅ / ✎ / 🔀 / 📊 / 🛑) — never silent
  ↓ on approval: handoff bundle with target-specific adapter (RN / Flutter / Electron / Tauri / MV3)
  ↓ score ≥9 against prototype rubric

Browser feedback in real time: click the 💬 button in the preview, select any region, type a comment. Hits .claude/memory/feedback-queue.jsonl; the UserPromptSubmit hook injects new entries as <system-reminder> on your next prompt — the supervibe:browser-feedback skill triages and dispatches to the right designer.

Manage running servers with /supervibe-preview --list / --kill <port>. Disable the overlay with --no-feedback.

Refactor with safety

For any rename / move / extract / delete on a public symbol.

ask: who calls processPayment?
  ↓ AI runs supervibe:code-search --callers "processPayment"
  ↓ shows N callers with file:line
if N > 10:
  ↓ rule use-codegraph-before-refactor escalates → architect-reviewer
  ↓ migration ADR
refactoring-specialist makes the rename in one PR
  ↓ verifies --callers "processPayment" returns 0
  ↓ score ≥9, no missed call sites

This flow has no slash command — you trigger it by asking the question. The graph + the discipline rules handle the rest.


Commands

Slash commands (run inside an AI CLI session). The normal user path is intentionally short; advanced commands stay available for diagnostics and plugin maintenance.

Primary

Command What it does
/supervibe Auto-router: picks genesis, design, security audit, network diagnostics, audit, strengthen, adapt, score, or update based on project state
/supervibe-genesis First-time scaffold of .claude/ for your stack
/supervibe-brainstorm <topic> Explicit entry to the brainstorming flow; produces an approved spec
/supervibe-plan [<spec-path>] Turn an approved spec into a phased TDD implementation plan
/supervibe-execute-plan [<plan-path>] Execute a plan with explicit 10/10 confidence gates. Supports --dry-run and --resume
/supervibe-loop --request/--plan/--from-prd Bounded autonomous loop with graph scheduler, status/resume/stop, doctor, graph export, and policy gates
/supervibe-security-audit Read-only multi-agent security audit, prioritized vulnerability backlog, optional remediation plan, execute, and re-audit loop to 10/10
/supervibe-ui Local browser/IDE-webview control plane with Kanban for epics, tasks, agent claims, loop state, waves, context packs, reports, and safe local actions
/supervibe-gc Reversible dry-run-first cleanup for completed work-item graphs and stale/superseded memory
/supervibe-design <brief> End-to-end design pipeline with memory/code/design-intelligence preflight: brand → spec → prototype → live preview → approval
/supervibe-presentation <brief> Presentation pipeline: storyboard → slide preview → feedback → approved .pptx → Google Drive handoff
/supervibe-preview Manage live preview servers
/supervibe-update Update the plugin itself. Idempotent, with rollback on failed checks
/supervibe-adapt Pull upstream agent/rule/skill improvements into the current project after plugin updates

Advanced

Command What it does
/supervibe-audit Read-only health check across agents, rules, memory, indexes, design evidence, and project overrides
/supervibe-strengthen [agent_id] Strengthen a weak agent from telemetry; without arguments auto-detects flagged agents
/supervibe-score [--record] <artifact> Score an artifact against its rubric; --record also updates telemetry. This is the preferred scoring/evaluation command

Internal command specs for diagnostics, plugin QA, low-level GC internals, legacy aliases, and override logging live in references/internal-commands/. They are intentionally outside the published commands/ directory so they do not add slash-command noise.

Shell scripts (run inside the plugin directory ~/.claude/plugins/marketplaces/supervibe-marketplace/):

Command What it does
npm run supervibe:status Health check across every index
npm run supervibe:loop -- --help Local no-tty help for loop status, graph, doctor, prime, export/import, and execution modes
npm run supervibe:ui -- --file <graph.json> Local visual control plane for work items, loop state, RAG/memory/codegraph health, reports, context packs, and safe actions
npm run supervibe:ide-bridge -- --out .supervibe/ide-bridge.json Webview descriptor for wrapping the local UI in any IDE
npm run supervibe:gc -- --all --dry-run Reversible cleanup preview for work-item graphs and memory
npm run supervibe:context-pack -- --file <graph.json> --item T1 Compact high-signal context pack for one active task
npm run supervibe:context-eval -- --case-file <cases.json> Retrieval/context-pack evals for required memory, evidence, anchors, and token budgets
npm run supervibe:happy-path -- --plan <plan.md> Ralph-style happy path: PRD/plan -> atomize -> execute -> verify -> archive
npm run supervibe:docs-audit User-facing docs relevance audit; flags internal dev files if they drift into docs/
npm run supervibe:install-doctor Post-install lifecycle audit: package versions, registry, stale files, and host registration state
npm run supervibe:upgrade clean checkout, git pull, lfs pull, npm ci, rebuild registry, run all tests, run install doctor
npm run supervibe:upgrade-check Manually query upstream for new commits
npm run code:index Full reindex
npm run code:search -- --query "..." Semantic search
npm run code:search -- --context "..." Agent-ready RAG + graph + anchor context
npm run code:search -- --symbol-search "Symbol" Graph: ranked symbol lookup
npm run code:search -- --callers "Symbol" Graph: who calls this symbol
npm run code:search -- --impact "Symbol" --depth 2 Graph: inbound blast radius before refactor
npm run code:search -- --files "src" Graph: indexed files with language and symbol counts
npm run presentation:build -- --input presentations/<slug>/deck.json --output presentations/<slug>/export/<slug>.pptx Export an approved deck spec to PPTX
npm run memory:watch Optional watcher daemon
npm run migrate:prototype-configs One-shot: backfill config.json for legacy prototype directories (also runs auto on SessionStart)
npm run check All 773 tests plus manifest, frontmatter, design-skill, question-discipline, spec-artifact, plan-artifact, agent-footer, knip, confidence-gate, package, and release-security validation

Troubleshooting

No banner after install. Re-run the installer — it is idempotent and refreshes the three Claude config files. Then fully restart the AI CLI (close the desktop app, do not just open a new chat).

The installer now writes .supervibe/audits/install-lifecycle/latest.json; if the banner is still absent, check that report for stale files or missing host registration.

Not visible in VS Code or Zed. Claude-backed IDE sessions read the same ~/.claude/ as the terminal. If the banner appears in the terminal, restart the IDE. If still nothing, re-run the installer.

Zed with Codex ACP does not show Supervibe after typing /. Current codex-acp advertises only its own built-in commands to Zed. Supervibe follows the Codex-supported route instead: Codex sees the plugin through ~/.codex/plugins/cache/supervibe-marketplace/supervibe/local plus ~/.codex/config.toml, and Zed/Codex ACP sessions get Supervibe behavior through native skills linked at ~/.agents/skills/supervibe. Re-run the installer, restart the Zed external-agent session, then check npm run supervibe:doctor -- --host codex --strict.

Protobuf parsing failed. The embedding model is missing or still an LFS pointer. Re-run the current installer; it verifies the ONNX file, tries bounded Git LFS, then downloads the model directly from HuggingFace before registration.

Install hangs at git-lfs filter-process. Re-run with the current installer. Clone/checkout disables LFS smudge, and the required ONNX setup uses bounded Git LFS plus direct HuggingFace fallback so the plugin is not registered until the model is ready.

Windows install starts in WSL. If install.sh runs under C:\Windows\System32\bash.exe, it uses WSL $HOME and WSL Node, not the Windows Codex/Claude/Gemini profile. Use PowerShell install.ps1 for Windows, or set SUPERVIBE_ALLOW_WSL_INSTALL=1 only when you intentionally want a separate WSL install.

SQLite errors. Node.js 22.5+ is required for the built-in node:sqlite used by semantic RAG, code graph, project memory, and agent task memory. Re-run the installer and approve the Node upgrade prompt, or install Node.js 22.5+ manually and then re-run.

Stale code index. The mtime scan on session start catches most external edits. For a full rebuild: rm .claude/memory/code.db && npm run code:index from your project directory.

Windows. If PowerShell rejects the installer with an Execution Policy error: Set-ExecutionPolicy -Scope Process Bypass. The Codex symlink needs Developer Mode — without it, the installer falls back to a directory copy.


Uninstall

# macOS / Linux
rm -rf ~/.claude/plugins/marketplaces/supervibe-marketplace
rm -rf ~/.codex/plugins/cache/supervibe-marketplace/supervibe
rm -f  ~/.codex/plugins/supervibe
rm -rf ~/.agents/skills/supervibe
# Also remove [plugins."supervibe@supervibe-marketplace"] from ~/.codex/config.toml if present.

node -e "
const fs=require('fs'),p=process.env.HOME+'/.claude/plugins/installed_plugins.json';
if(!fs.existsSync(p))process.exit(0);
const d=JSON.parse(fs.readFileSync(p,'utf8'));
delete d.plugins['supervibe@supervibe-marketplace'];
fs.writeFileSync(p,JSON.stringify(d,null,2)+'\n');
"

node -e "
const fs=require('fs'),p=process.env.HOME+'/.claude/plugins/known_marketplaces.json';
if(!fs.existsSync(p))process.exit(0);
const d=JSON.parse(fs.readFileSync(p,'utf8'));
delete d['supervibe-marketplace'];
fs.writeFileSync(p,JSON.stringify(d,null,2)+'\n');
"

node -e "
const fs=require('fs'),p=process.env.HOME+'/.claude/settings.json';
if(!fs.existsSync(p))process.exit(0);
const d=JSON.parse(fs.readFileSync(p,'utf8'));
if(d.enabledPlugins) delete d.enabledPlugins['supervibe@supervibe-marketplace'];
if(d.extraKnownMarketplaces) delete d.extraKnownMarketplaces['supervibe-marketplace'];
fs.writeFileSync(p,JSON.stringify(d,null,2)+'\n');
"

sed -i.bak '/<!-- supervibe-plugin-include: do-not-edit -->/,/<!-- supervibe-plugin-include: do-not-edit -->/d' ~/.gemini/GEMINI.md 2>/dev/null || true

Windows equivalent: replace rm -rf with Remove-Item -Recurse -Force and run the same node -e blocks (paths via $HOME work in PowerShell too).

Project indexes are your data — remove only when sure: rm -rf .claude/memory/code.db .claude/memory/memory.db.