@polderlabs/bizar 8.0.0 → 9.0.0

package/.claude/worktrees/kanban-overhaul/bizar-dash/skills/lightrag/SKILL.md

@@ -0,0 +1,86 @@
+---
+name: lightrag
+description: LightRAG integration with BizarHarness. Covers setup, cline-free defaults, querying, indexing, and the memory service integration.
+---
+
+# LightRAG Integration
+
+LightRAG provides fast full-text and keyword search over Obsidian vault content. It is the Phase 2 component of the Bizar Memory Service.
+
+## Overview
+
+- **LightRAG** - a lightweight RAG (Retrieval-Augmented Generation) server that indexes Markdown files
+- **Bizar Memory Service** - three-layer system: Markdown (truth) → Git (collaboration) → LightRAG (index)
+- **cline-free defaults** - LightRAG runs without cline, accessible to any HTTP client
+
+## Setup
+
+Start the LightRAG server:
+```bash
+npx lightrag-server --port 8724
+```
+
+Or via the Bizar dashboard: navigate to Settings → LightRAG and configure the server URL.
+
+## Configuration
+
+LightRAG settings are stored in `~/.config/cline/lightrag.json`:
+
+```json
+{
+  "serverUrl": "http://localhost:8724",
+  "indexRoot": "~/.config/cline/memory",
+  "autoIndex": true,
+  "embeddingModel": "bge-m3"
+}
+```
+
+## API Endpoints
+
+- `GET /api/lightrag/status` - server health and index stats
+- `POST /api/lightrag/index` - trigger re-indexing of vault
+- `GET /api/lightrag/search?q=` - full-text search over vault
+- `GET /api/lightrag/extract?entities=` - entity extraction
+
+## Search Example
+
+```bash
+curl "http://localhost:8724/search?q=authentication+flow&k=5"
+```
+
+Returns ranked snippets with source file paths.
+
+## Integration with Memory Service
+
+When both Obsidian vault and LightRAG are configured:
+1. The memory service writes new knowledge to the vault
+2. LightRAG indexes the vault continuously or on-demand
+3. `bizar memory search` routes natural-language queries to LightRAG
+
+## When to Use LightRAG vs Direct File Read
+
+| Query Type | Use |
+|---|---|
+| Natural-language question | LightRAG search |
+| Specific fact lookup | Direct file read |
+| Related concepts discovery | LightRAG search |
+| Precise location of text | grep + read |
+
+## Memory tab in the dashboard (v4.7.0+)
+
+The dashboard's **Memory** tab (sidebar entry between Skills and Settings) surfaces LightRAG controls without leaving the browser:
+
+- **LightRAG panel** — Start / Stop / Restart / Reindex all / Rebuild graph. Rebuild graph wipes the working dir and re-runs a full reindex (idempotent — safe to re-run).
+- **Stats** — Approximate indexed-chunk count (from `kv_store_*.json`), query count last 24h, and average response time. Stats are tracked in a sidecar JSONL at `<project>/.bizar/memory-cache/lightrag-query-log.jsonl`.
+- **Quick search** — Sends `GET /api/memory/query?q=...&topK=8` and shows the raw LightRAG response.
+- **Semantic search panel** — Cross-source search: runs both LightRAG and Obsidian vault lexical search and merges results, deduped by (source, relPath).
+
+Open the tab by setting `activeTab = 'memory'` or via the Overview's Memory status card. The LightRAG stats are fetched via `GET /api/memory/lightrag/stats`; the rebuild path uses `POST /api/memory/lightrag/rebuild-graph`.
+
+## Common Issues
+
+### Empty search results
+Cause: index is stale or the server is down. Fix: POST to `/api/lightrag/index` to re-index.
+
+### Server not responding
+Check that LightRAG is running: `curl http://localhost:8724/status`. If not, restart with `npx lightrag-server`.

package/.claude/worktrees/kanban-overhaul/bizar-dash/skills/minimax/SKILL.md

@@ -0,0 +1,80 @@
+---
+name: minimax
+description: MiniMax integration for the Bizar multi-agent system. Covers chat completions, usage tracking, multi-key rotation, rate limiting, and agent awareness of model limits and pricing.
+---
+
+# MiniMax Integration
+
+Bizar agents use MiniMax models as the primary paid tier. This skill covers the provider configuration, usage tracking, and common failure modes.
+
+## Model Reference
+
+| Model ID | Context | Cost (in/out) | Best for |
+|---|---|---|---|
+| `minimax/MiniMax-M2.7` | 32k | $0.30/M · $1.20/M | Mid-tier implementation, Git ops, moderate tasks |
+| `minimax/MiniMax-M3` | 32k | $0.30/M · $1.20/M | Complex features, deep debugging, architecture |
+
+The `MiniMax-M3` model is the primary high-tier agent (Tyr, Forseti). The `MiniMax-M2.7` model is used for Thor and Hermod.
+
+## Provider Configuration
+
+Do NOT set `baseURL` on the `minimax` provider - cline ships a built-in MiniMax provider that resolves the correct API endpoint. Adding an explicit baseURL is a common cause of 404s.
+
+```json
+{
+  "provider": "minimax",
+  "model": "minimax/MiniMax-M3"
+}
+```
+
+## Multi-Key Rotation
+
+When you have multiple MiniMax accounts and want to distribute load:
+
+```bash
+# Comma-separated (simplest)
+export MINIMAX_API_KEYS="key1,key2,key3"
+
+# Numbered env vars (explicit ordering)
+export MINIMAX_API_KEY="primary-key"
+export MINIMAX_API_KEY_2="backup-key"
+```
+
+**Rotation triggers:** 429 (rate limit), 402 (quota exhausted), 500/502/503/504 (server error).
+
+**Does NOT rotate on:** 401 (unauthorized - fix the key), 4xx client errors (the request is bad).
+
+**Round-robin on success:** the next request starts on the next key, spreading load evenly.
+
+## Usage Tracking
+
+The dashboard tracks token usage per session via `bizar-dash/src/server/minimax-usage-store.mjs`. Usage is recorded for:
+- Input tokens
+- Output tokens
+- Model used
+- Session ID
+- Timestamp
+
+View usage at `/usage` in the dashboard.
+
+## Common Failure Modes
+
+### 429 / Rate Limit
+Cause: single key hit rate limit. Fix: add more keys to rotation.
+
+### 402 / Quota Exhausted
+Cause: monthly spend limit reached on the key. Fix: add a different key with remaining quota.
+
+### 404 on Chat Completions
+Cause: wrong baseURL or malformed model ID. Fix: remove custom baseURL, use `minimax/MiniMax-M3` as model ID.
+
+### Empty Responses / Truncated Output
+Cause: output hit token limit or streaming was interrupted. Fix: increase `maxTokens` if possible, or split the request.
+
+## Agent Cost Awareness
+
+When routing work:
+- Free: Mimir, Heimdall (DeepSeek V4 Flash - no MiniMax cost)
+- Mid: Thor, Hermod (MiniMax-M2.7)
+- High: Tyr, Forseti (MiniMax-M3)
+- Ultra: Vidarr (GPT-5.5 via OpenAI-compatible endpoint)

package/.claude/worktrees/kanban-overhaul/bizar-dash/skills/obsidian/SKILL.md

@@ -0,0 +1,68 @@
+---
+name: obsidian
+description: Use the Bizar Memory Service for persistent project knowledge - local Obsidian-compatible Markdown + Git-shared sync. Read vault entries before non-trivial decisions; write after meaningful work. Three layers: Markdown (truth) → Git (collaboration) → LightRAG (derived index, Phase 2).
+---
+
+# Bizar Memory Service (Obsidian)
+
+Three-layer persistent memory for project knowledge.
+
+## Architecture
+
+1. **Markdown (truth)** - Obsidian-compatible `.md` files with YAML frontmatter, wikilinks, and callouts. Stored in per-project vaults under `~/.config/cline/memory/<project>/`.
+2. **Git (collaboration)** - Vaults are Git-tracked and synced. The vault path is unique per project, enabling team sharing via normal Git workflow.
+3. **LightRAG (derived index, Phase 2)** - When the LightRAG server runs, the memory service indexes all vault entries for fast full-text retrieval.
+
+## Vault Discovery
+
+At session start, call `obsidian_list_vaults` (via the Obsidian MCP tool or CLI) to find available vaults. Match by project name. The default vault is for cross-project knowledge only.
+
+## Vault Structure
+
+```
+~/.config/cline/memory/<project>/
+  .obsidian/         # Obsidian config (workspace.json etc.)
+  index.md           # Entry point: links to recent, important, pinned notes
+  projects/          # Project-specific notes
+  decisions/         # ADRs and architectural decisions
+  references/        # External knowledge captured for this project
+  daily/             # Daily notes (journal-style)
+```
+
+## Adding Knowledge
+
+After meaningful work, write a note:
+
+- Use frontmatter: `created:`, `updated:`, `tags:`, `project:`
+- Link related notes with `[[wikilinks]]`
+- Keep notes small and focused (one concept per note)
+- Update `index.md` if adding a significant new area
+
+## Querying Knowledge
+
+Use the memory service to find relevant context before making non-trivial decisions:
+- `bizar memory search "<topic>"` - full-text search across vault
+- `bizar memory status` - shows vault path, last sync, index state
+
+## LightRAG Integration
+
+When LightRAG is enabled (v4.5.0+), it indexes vault content automatically. Query with `bizar memory search` - this routes to LightRAG for natural-language queries. For simple lookups, direct file reading is faster.
+
+## Memory tab in the dashboard (v4.7.0+)
+
+The dashboard has a dedicated **Memory** tab (sidebar entry between Skills and Settings) for working with the vault without leaving the browser:
+
+- **Overview** — composite health score (vault + LightRAG + git + secrets + schema), with one-click navigation to each subsystem.
+- **Obsidian Vault** — folder tree on the left, note list (newest first), detail pane with body + frontmatter + backlinks. Inline create / edit (markdown editor modal) / delete.
+- **LightRAG** — Start / Stop / Restart / Reindex all / Rebuild graph. Quick search sends a natural-language query and shows the LightRAG response.
+- **Git Sync** — branch / ahead / behind / dirty / clean pills, working-tree diff, Pull / Push / Commit / Fetch + sync.
+- **Semantic Search** — single search input that queries both LightRAG and the Obsidian vault in one shot. Results deduplicated by (source, relPath) and sorted by score.
+- **Config** — grouped forms for LightRAG URL/models, Obsidian path/syncInterval, git repo/remote/branch/autoSync, plus a "Test git connection" button.
+
+Open the tab programmatically by setting `activeTab = 'memory'`. The Overview page now also surfaces a small Memory status card (health score + LightRAG + vault + git) with a single-click jump.
+
+## Common Gotchas
+
+- The vault is git-tracked but not auto-committed. Commit after meaningful knowledge additions.
+- Wikilinks work within Obsidian but are just plain links in plain Markdown viewers.
+- Heavy media files (images, PDFs) should be stored outside the vault or Git LFS-tracked.

package/.claude/worktrees/kanban-overhaul/bizar-dash/skills/providers/SKILL.md

@@ -0,0 +1,75 @@
+---
+name: providers
+description: How the provider subsystem works in BizarHarness - catalog management, backup keys, auto-add wizard, and provider store.
+---
+
+# Provider Subsystem
+
+The provider subsystem manages LLM API credentials and model routing. It is composed of the providers store, the config router, and the providers UI.
+
+## Core Concepts
+
+- **Provider:** A named API source (e.g., `minimax`, `openai`, `anthropic`).
+- **Catalog:** The list of all known/available providers.
+- **Active provider:** The currently selected provider for a given model tier.
+- **Backup keys:** Additional API keys for a provider, used in rotation on 429/402/5xx.
+
+## Provider Store
+
+`bizar-dash/src/server/providers-store.mjs` manages:
+- Provider credentials (API keys stored encrypted at rest)
+- Active/inactive state per provider
+- Priority ordering
+- Backup key configuration
+
+## Provider Config File
+
+Providers are configured via `~/.config/cline/providers.json` or via the dashboard UI:
+
+```json
+{
+  "providers": [
+    {
+      "id": "minimax",
+      "name": "MiniMax",
+      "apiKeys": ["sk-..."],
+      "active": true,
+      "priority": 1
+    }
+  ]
+}
+```
+
+## Backup Keys (Multi-Key Rotation)
+
+Configure additional keys for a provider:
+
+```bash
+# Via env var (comma-separated)
+export MINIMAX_API_KEYS="key1,key2,key3"
+
+# Via providers.json
+{
+  "providers": [{
+    "id": "minimax",
+    "apiKeys": ["key1", "key2", "key3"]
+  }]
+}
+```
+
+## Auto-Add Wizard
+
+When the active provider fails with 401/403, the system can prompt the user to add a new key. The wizard:
+1. Detects auth failure
+2. Opens a modal prompting for a new key
+3. Validates the key by making a test request
+4. Adds to backup keys or replaces the failed key
+
+## Provider API Endpoints
+
+- `GET /api/providers` - list all providers with status
+- `POST /api/providers` - add a new provider
+- `PUT /api/providers/:id` - update provider config
+- `DELETE /api/providers/:id` - remove a provider
+- `POST /api/providers/:id/keys` - add a backup key
+- `POST /api/providers/rotate` - force rotation to next key

package/.claude/worktrees/kanban-overhaul/bizar-dash/skills/publishing/SKILL.md

@@ -0,0 +1,146 @@
+---
+name: publishing
+description: Cut and publish a release of BizarHarness (@polderlabs/bizar). Use when bumping version, updating CHANGELOG, committing release commits, tagging, pushing, or publishing to npm. Covers the 4 test runners, version bump rules, hotfixes, and 10 known pitfalls.
+---
+
+# Publishing BizarHarness
+
+Cut a release of `@polderlabs/bizar`. Read `/home/drb0rk/Projects/BizarHarness/docs/RELEASING.md` for the comprehensive guide; this skill is the action-oriented checklist.
+
+## Pre-flight (must all be true)
+
+```bash
+git status --short                                # empty
+git branch --show-current                         # master
+npm whoami                                        # drb0rk
+```
+
+## Test gate (all 4 runners must pass)
+
+```bash
+npx tsc --noEmit                                  # 0 errors
+npm test                                          # 388 node + 75 SDK + 295 bun + 7 smoke = 765 pass
+npm run build                                     # 0 errors
+```
+
+Counts above are post-v4.7.0 (commit `3005611`, tag `v4.7.0`).
+
+## Doc updates (in order)
+
+1. **`CHANGELOG.md`** — add `## vX.Y.Z — short title` at top, sections: Highlights / What's New / Tests / Upgrade.
+2. **`package.json`** — bump `"version"` field.
+3. **`.bizar/PROJECT.md`** — bump `Current Version` line.
+4. **`.bizar/AGENTS_SELF_IMPROVEMENT.md`** — append `### YYYY-MM-DD — vX.Y.Z` entry under `## Log`.
+
+## Release sequence
+
+```bash
+# Pre-flight + tests + build (see above)
+
+# Commit (hand off to Hermod for write-level git)
+git add -A
+git commit -m "vX.Y.Z: short title
+
+Highlights:
+- bullet 1
+- bullet 2
+- bullet 3
+
+Co-authored-by: Odin <noreply@polderlabs.dev>"
+
+# Tag
+git tag -a vX.Y.Z -m "vX.Y.Z — short title"
+
+# Push
+git push origin master --follow-tags
+
+# Publish (prepublishOnly runs build automatically)
+npm publish
+
+# Verify
+npm view @polderlabs/bizar@X.Y.Z version
+git ls-remote --tags origin | grep vX.Y.Z
+```
+
+## Versioning
+
+| Bump | When |
+|---|---|
+| Patch | Bug fixes only |
+| Minor | New features, behavior-preserving changes |
+| Major | Breaking changes requiring user action |
+
+When in doubt, ask the user.
+
+## Hotfix
+
+```bash
+git checkout -b hotfix/vX.Y.Z+1 vX.Y.Z
+# make minimal fix
+npx tsc --noEmit && npm test && npm run build
+# bump version, update CHANGELOG with "Hotfix" entry
+git add -A && git commit -m "vX.Y.Z+1: hotfix — ..."
+git tag -a vX.Y.Z+1 -m "vX.Y.Z+1 — Hotfix: ..."
+git push origin hotfix/vX.Y.Z+1 --follow-tags
+git checkout master && git merge --no-ff hotfix/vX.Y.Z+1
+git push origin master
+npm publish
+```
+
+## Rollback
+
+`npm unpublish` is almost never right (breaks existing installs, 72h limit). Prefer:
+
+```bash
+npm deprecate @polderlabs/bizar@X.Y.Z "Critical bug; use X.Y.Z-1 instead"
+# then cut a new patch with the fix
+```
+
+## 10 pitfalls (avoid these)
+
+1. **Auth is NOT needed in v1 routes.** Tailscale handles it. Don't add auth during a release refactor.
+2. **`--help` global vs per-subcommand.** `bizar --help` ≠ `bizar <cmd> --help`. Test both.
+3. **Pre-existing merge conflict markers.** `grep -c '<<<<<<<\|>>>>>>>' <file>` before declaring a file clean. TS1185 will fail otherwise.
+4. **Source maps.** `vite.config.ts` must have `sourcemap: 'hidden'`. Don't ship 3 MB of maps.
+5. **Empty catch blocks.** Replace `catch { /* ignore */ }` with `catch (err) { logger.warn('swallowed in <ctx>:', err.message); }`.
+6. **Sibling collisions.** Define `Owns (read+write)` and `Siblings READ-ONLY` before dispatching parallel streams.
+7. **Test imports after refactor.** `grep -rn 'from.*<old-path>' <test-dirs>` to find stale imports.
+8. **`npm publish` requires version bump.** Re-publishing the same version is rejected.
+9. **`prepublishOnly` runs build.** Don't run build separately before publish — it runs anyway.
+10. **`npm view` propagation delay.** New version may 404 for ~5s after publish. Retry.
+
+## Agent routing
+
+| Agent | Release role |
+|---|---|
+| Odin | Decompose release into parallel prep streams |
+| Heimdall | Mechanical edits (bump `package.json`, append to AGENTS_SELF_IMPROVEMENT.md) |
+| Mimir | Research prior releases, summarize changes |
+| Thor | CHANGELOG drafting, test additions |
+| Tyr | Full release orchestration, breaking-change analysis |
+| **Hermod** | **The only agent that runs `git commit`, `git tag`, `git push`, `npm publish`** |
+| Forseti | Review the release plan (when called by Odin) |
+
+If a non-Hermod agent tries to push or publish, stop — route to Hermod via handoff.
+
+## Post-release verification
+
+```bash
+npm view @polderlabs/bizar version                          # X.Y.Z
+npm view @polderlabs/bizar@X.Y.Z dist.tarball                # registry URL
+git ls-remote --tags origin | grep vX.Y.Z                   # tag visible
+
+# Smoke test in a clean dir
+mkdir /tmp/bizar-smoke && cd /tmp/bizar-smoke
+npm install -g @polderlabs/bizar@X.Y.Z
+bizar --version                                             # X.Y.Z
+bizar doctor                                                # exit 0
+```
+
+## Reference
+
+- Comprehensive guide: `docs/RELEASING.md` (~600 lines, all details)
+- Most recent release: **v4.7.0** (commit `3005611`, tag `v4.7.0`, npm `@polderlabs/bizar@4.7.0`)
+- Diff stats: 62 files, +7,361 / −4,083 lines
+- Test counts: 388 node + 75 vitest + 295 bun + 7 smoke = **765 pass, 0 fail**
+- Build: 372 KB main + 476 KB mobile JS bundles

package/.claude/worktrees/kanban-overhaul/bizar-dash/skills/sdk/SKILL.md

@@ -0,0 +1,138 @@
+---
+name: sdk
+description: @polderlabs/bizar-sdk usage - building AI agents on Cloudflare Workers using the Agents SDK. Covers Agent class, state management, callable RPC, Workflows, durable execution, queues, retries, observability, and React hooks.
+---
+
+# @polderlabs/bizar-sdk
+
+The Bizar SDK lets you build AI-powered agents on Cloudflare Workers using the Agents SDK. It runs on Cloudflare's global network with automatic scaling and durable state.
+
+## Installation
+
+```bash
+npm install @polderlabs/bizar-sdk
+```
+
+## Core Concepts
+
+### Agent Class
+
+```typescript
+import { Agent } from '@polderlabs/bizar-sdk';
+
+const agent = new Agent({
+  name: 'my-agent',
+  model: 'minimax/MiniMax-M3',
+  systemPrompt: 'You are a helpful assistant.',
+  env: { MINIMAX_API_KEY },
+});
+
+export default agent;
+```
+
+### State Management
+
+Agents maintain durable state across requests using Cloudflare Durable Objects:
+
+```typescript
+const state = await agent.state.get();
+// { count: 0, lastSeen: Date }
+await agent.state.patch({ count: state.count + 1 });
+```
+
+### Callable RPC
+
+Agents expose RPC methods callable via HTTP:
+
+```typescript
+agent.method('greet', async (name: string) => {
+  return `Hello, ${name}!`;
+});
+```
+
+Call from the dashboard or any HTTP client:
+```bash
+curl -X POST https://my-agent.workers.dev/rpc/greet \
+  -H "Content-Type: application/json" \
+  -d '{"name": "World"}'
+```
+
+## Workflows
+
+For long-running tasks, use Workflows (durable execution):
+
+```typescript
+import { Workflow } from '@polderlabs/bizar-sdk';
+
+const workflow = new Workflow({
+  name: 'data-processor',
+  retry: { maxAttempts: 3, backoff: 'exponential' },
+});
+
+workflow.step('fetch', async ({ input }) => {
+  return await fetch(input.url);
+});
+
+workflow.step('process', async ({ input, state }) => {
+  const data = await input.json();
+  return { ...data, processed: true };
+});
+
+export default workflow;
+```
+
+## Queues
+
+For async message processing:
+
+```typescript
+import { Queue } from '@polderlabs/bizar-sdk';
+
+const queue = new Queue('my-queue', {
+  maxRetries: 3,
+  batchSize: 10,
+});
+
+queue.processor(async (messages) => {
+  for (const msg of messages) {
+    console.log('Received:', msg.body);
+  }
+});
+```
+
+## Observability
+
+Built-in tracing to Cloudflare Analytics:
+- Request logs
+- Token usage
+- Error rates
+- Latency percentiles
+
+## React Hooks
+
+For building chat interfaces:
+
+```typescript
+import { useAgent, useSession } from '@polderlabs/bizar-sdk/react';
+
+function ChatComponent({ agentName }: { agentName: string }) {
+  const { messages, send, isLoading } = useAgent(agentName);
+
+  return (
+    <div>
+      {messages.map((m) => (
+        <div key={m.id}>{m.role}: {m.content}</div>
+      ))}
+      <input onSend={send} disabled={isLoading} />
+    </div>
+  );
+}
+```
+
+## Deployment
+
+```bash
+npx wrangler deploy
+```
+
+Workers are deployed to Cloudflare's global network. No cold starts, automatic HTTPS.

package/.claude/worktrees/kanban-overhaul/bizar-dash/skills/self-improvement/SKILL.md

@@ -0,0 +1,53 @@
+---
+name: self-improvement
+description: Use when setting up, configuring, or debugging the project-level self-improvement system. Every task records lessons learned to .bizar/AGENTS_SELF_IMPROVEMENT.md for agent behavior improvement across sessions.
+---
+
+# Self-Improvement Protocol
+
+Records lessons learned after every meaningful task to `.bizar/AGENTS_SELF_IMPROVEMENT.md`. This enables agents to avoid repeating mistakes and build on past successes.
+
+## How It Works
+
+After completing a non-trivial task, the agent appends a structured entry to `.bizar/AGENTS_SELF_IMPROVEMENT.md` in the project root. Entries are written in Markdown with frontmatter for easy querying.
+
+## Entry Format
+
+```markdown
+## [YYYY-MM-DD] <one-line summary>
+
+**Task:** <what was asked>
+
+**Approach taken:** <key decisions and why>
+
+**What worked:** <specific things that went well>
+
+**What could be better:** <mistakes, friction points, missed opportunities>
+
+**Will do differently:** <concrete rule or habit for next time>
+
+**Tags:** #<tag1> #<tag2>
+```
+
+## When to Write an Entry
+
+Write an entry when:
+- A bug took more than 30 minutes to find or fix
+- A significant architectural decision was made
+- A known pattern or idiom was rediscovered
+- A previous mistake was identified and corrected
+- A task involved non-obvious tool usage or workflow
+
+## Reading Before New Tasks
+
+Before starting a non-trivial task in an unfamiliar area, agents should search the self-improvement file:
+```
+grep -A5 "tag:debugging" .bizar/AGENTS_SELF_IMPROVEMENT.md
+grep -A5 "tag:typescript" .bizar/AGENTS_SELF_IMPROVEMENT.md
+```
+
+This surfaces prior lessons that apply to the current work.
+
+## Skill File Location
+
+The skill instructions are at `~/.cline/skills/self-improvement/SKILL.md`. Load with the `skill` tool when setting up or debugging the system.