rimuru-ai 1.19.0

package/.rimuru/improver/knowledge.md

@@ -0,0 +1,172 @@
+# Knowledge Base — OpenCode Improver
+
+Durable learnings, patterns, decisions, and their outcomes.
+
+---
+
+## 2026-06-13 — SAP Work Schedule Sync implementation
+
+- **Pattern**: Adding a new SAP sync = 4 changes: `sap.api.ts` (fetch method), `app.service.ts` (cron method + busy flag), `app.controller.ts` (POST endpoint), `AGENTS.md` (documentation).
+- **Delete+reseed+createMany** is the standard sync pattern for small datasets (<1000 rows). Works for work schedule, leaves, professional leaves.
+- **Prisma model naming**: `SAP_Workschedule` → client property is `sAP_Workschedule` (camelCase preserves case after prefix).
+- **Hardcoded fallbacks** in `sap.api.ts` are this project's standard (`.env` is git-tracked so secrets aren't stored there — only `DATABASE_URL` is in `.env`).
+- **Direct MSSQL CLI**: `pymssql` via pip works without root. Future projects should configure `connections.json` per-project (never hardcode DB credentials globally).
+
+## 2026-06-15 — Token-Optimized Subagent Design Patterns
+
+- **Key token-saving mechanisms for subagents**:
+  - `steps: 6-8` for narrow domain tasks (vs 25-50 for general coding agents) — saves ~60% on reasoning loops
+  - Short `description` (~15 words) — each description adds ~20-30 tokens to Task tool system prompt
+  - No `model` override — lets default model handle it; avoids hardcoding expensive models
+  - Minimal tool permissions — every extra tool adds schema tokens (~50-100 tokens per tool). Only allow what's needed.
+  - `bash: deny` for read-only agents (erp-architect) — saves tool schema tokens
+  - Inline command examples in prompts — avoids research websearch loops (~2-5k tokens per lookup)
+   - KV-cache friendly structure: stable prefix (role + methodology) before variable content
+
+## 2026-06-17 — MADAR → FastContext Migration Complete
+
+- **Migration**: Replaced [MADAR](https://github.com/lubab/madar) (graph-based code context via `madar generate`/`pack`/`query`) with [FastContext](https://github.com/microsoft/fastcontext) (delegated repo-exploration subagent from Microsoft Research).
+- **Why**: FastContext separates exploration from task solving — the main agent delegates a natural-language query to a dedicated subagent that returns compact file:line citations. This avoids the need to:
+  - Maintain a knowledge graph (`out/graph.json` + cache) that goes stale
+  - Remember to run `madar generate . --include-docs` after every config change
+  - Pay the overhead of full-file graph extraction
+- **Usage**: `fastcontext -q "<question>" --citation` for focused context (returns file:line ranges). `fastcontext -q "<question>" --max-turns 12 --citation` for complex multi-file traces.
+- **Skill installed**: `~/.agents/skills/fastcontext/SKILL.md`
+- **CLI installed**: `uv tool install` from github.com/microsoft/fastcontext
+- **Requires**: OpenAI-compatible endpoint via `BASE_URL`, `MODEL`, `API_KEY` env vars (not yet configured)
+- **Old `madar-graph` skill**: Deleted from `~/.agents/skills/madar-graph/`
+- **Old `out/` directory**: Deleted from `~/.config/opencode/out/`
+- **Old MADAR npm package** (`@lubab/madar`): Still globally installed but unused. Can be cleaned up later.
+
+## 2026-06-15 — DOCX Generation with npm `docx` Package
+
+- **Install locally** in the script's working directory: `npm install docx` (not `npm install -g` — global install doesn't resolve from arbitrary dirs like `/tmp/`)
+- **Key API changes in v9+**:
+  - `PageNumber` is now an object with constants (`CURRENT`, `TOTAL_PAGES`) — not a constructor
+  - **CORRECT way to add page numbers**: `new TextRun({ children: [PageNumber.CURRENT] })` inside a Paragraph
+  - `new PageNumberElement()` creates bare `<w:pgNum/>` XML — does NOT render a working page number. DO NOT use it.
+  - `PageNumber.CURRENT` inside a TextRun's `children` array produces the full field structure (`w:fldChar` begin/separate/end + `w:instrText` PAGE) that Word needs
+- **Diagram embedding**: Use `ImageRun` with PNG files on disk. Image dimensions in EMUs (dxa): `w * 914400 / dpi`, `h * 914400 / dpi`
+- **Structure**: Build doc in sections with `properties` (page margins, orientation), `headers`/`footers`, `children` array
+- **Output path**: `path.join(require('os').homedir(), 'Documents/...')` — always use `os.homedir()` for cross-platform
+
+## 2026-06-15 — DOCX Editing: Use the `docx` Skill (~/.agents/skills/docx/)
+
+- **Hard rule**: For ALL .docx file editing/creation, load the `docx` skill first (`skill({name: "docx"})`). It has the ooxml scripts, Document library, and proper workflows.
+- **The `docx` skill recommends**: For editing existing documents → use the Document library (Python, via `ooxml/scripts/unpack.py` + Python script + `pack.py`). For creating new documents → use `docx-js` (npm `docx` package).
+- **python-docx alternative**: python-docx works for basic edits (images, headers, footers) but the skill's Document library gives full OOXML access for complex edits.
+- **Image `.undefined` fix**: PNG images with `.undefined` extension in docx zip → need two fixes: (1) rename to `.png`, (2) update `[Content_Types].xml` to map `Extension="undefined"` to `ContentType="image/png"`, (3) update all `.rels` and `.xml` references.
+- **Image sizing**: A4 page has ~5731510 EMU usable width. Images wider than this must be scaled down proportionally.
+
+## 2026-06-15 — CRITICAL: python-docx Page Dimension Unit Bug
+
+- **The Bug**: Setting `section.page_width = 11906` produces a page of **0.013 inches** — Word/LibreOffice crashes trying to render content on a page that tiny.
+- **Root Cause**: The script author confused **twips** (1in = 1440 twips, used in final XML) with **EMU** (1in = 914400 EMU, used by python-docx API). A comment saying "A4 width in DXA" added confusion — "DXA" is not a real OpenXML unit.
+- **The Fix**: NEVER use raw integer values for page dimensions or margins. ALWAYS use python-docx helper classes:
+  - `Cm(21.0).emu` for A4 width (21cm)
+  - `Cm(29.7).emu` for A4 height (29.7cm)
+  - `Cm(2.54).emu` (or `Inches(1).emu`) for 1-inch margins
+  - Or set directly: `section.page_width = Cm(21.0)` (python-docx accepts Length objects)
+- **How to verify**: Check the generated XML — `w:pgSz` values should be ~11907 x 16838 (twips), and `w:pgMar` should be 1440 for 1in margins. If numbers are tiny (19 x 27), the EMU→twips conversion is wrong.
+- **Detection script**: `python3 -c "import zipfile; z=zipfile.ZipFile('file.docx'); import re; s=re.search(r'<w:pgSz[^>]+/>', z.read('word/document.xml').decode()); print(s.group(0) if s else 'not found')"` — if `w:w` is < 100, page is corrupted.
+- **Hard rule**: Any python-docx script that creates a new document MUST verify page dimensions in the output before distributing. One wrong constant crashes every document viewer.
+
+## 2026-06-15 — User Communication Preference: English Only
+
+- **User prefers 100% English** in all communication. Do not mix languages (Chinese, Tamil, etc.) in responses, explanations, or examples — even when the topic or platform has multilingual content. All payloads, walkthroughs, and reasoning must be delivered in English only.
+- This preference is permanent — do not re-confirm or re-ask.
+
+## 2026-06-17 — Context-Aware Replacements: What Stays vs Changes When Renaming
+
+### `opencode` references that MUST NEVER be renamed (fork-safe)
+These are internal identifiers, config schemas, environment variables, or external project names — not "old brand name" references:
+- **Config files**: `opencode.json`, `opencode.jsonc` — these are the canonical config filenames
+- **Directory**: `.opencode/`, `~/.config/opencode/` — internal directory paths
+- **Environment variables**: `OPENCODE_HOME`, `OPENCODE_PROVIDER`, `OPENCODE_*` — all uppercase prefix vars
+- **Theme names**: `"opencode"` in theme config (it's a theme name, not the app name)
+- **npm package names**: `opencode-daytona`, `opencode-helicone-session` — actual published packages
+- **GitHub orgs**: `awesome-opencode` — real GitHub organization
+- **Community URLs**: `opencode.cafe` — real community site
+- **mDNS default**: `opencode.local` — hardcoded default in server code
+- **URL paths**: `.well-known/opencode` — hardcoded route path
+- **GitHub apps**: `opencode-agent` — real GitHub App name
+- **Web component tags**: `<opencode-language-select>` — custom HTML element name
+- **Social card CDN paths**: `opencode-docs/` in social card CDN URLs
+
+### `opencode` references that SHOULD rename to `rimuru-ai`
+- XDG data dir: `~/.local/share/opencode/` → `~/.local/share/rimuru-ai/`
+- XDG cache dir: `~/.cache/opencode/` → `~/.cache/rimuru-ai/`
+- XDG config dir body text: `$XDG_CONFIG_HOME/opencode/` → `$XDG_CONFIG_HOME/rimuru-ai/`
+- Windows app data paths: `%USERPROFILE%\\rimuru-ai` (was `.opencode`)
+- npm package name: `opencode-ai` → `rimuru-ai` (actual npm package)
+- ACP config examples: `command = "opencode"` → `command = "rimuru-ai"`, `name = "opencode"` → `name = "rimuru-ai"`
+- General prose references to the tool name
+
+### Batch replacement strategy
+- **Never use blanket `sed`** on forked codebases with overlapping names — always use context-aware scripts
+- Check 3+ chars before/after each match to determine context (config filename? env var? directory path? prose?)
+- After every batch rename: build the project and verify — don't assume correctness
+- Always check external URLs after rename — GitHub orgs, npm packages, etc. are common breakage points
+- Run a post-build grep on rendered output to spot-check for stale visible references
+
+## 2026-06-17 — FastContext: Usage Patterns & Constraints
+
+- **Primary command**: `fastcontext -q "<natural language question>" --citation` — returns compact file:line ranges, ideal for pre-edit context gathering
+- **Deep trace**: `fastcontext -q "<question>" --max-turns 12 --citation` — for multi-file traces and architecture questions
+- **No graph to maintain**: Unlike MADAR, no `generate`/`pack`/`query` workflow. No `out/` directory. No `--include-docs` flag to remember.
+- **Token impact**: FastContext's subagent returns only file:line citations (~200-500 tokens) instead of full file contents (~3-8k tokens). The subagent itself consumes some tokens for exploration, but the main agent's context window is spared.
+- **Caveat**: FastContext runs read-only as a subagent — it can't modify files. It also requires an external LLM API endpoint (not yet configured). Until configured, the agent falls back to manual glob/grep/file-read sweeps.
+- **Backward compat**: MADAR is still globally installed (`@lubab/madar@0.28.0` via npm). No conflicts if both are present. Can uninstall later.
+- **The `fastcontext` skill** at `~/.agents/skills/fastcontext/SKILL.md` has full documentation. Load with `skill({name: "fastcontext"})`.
+
+## 2026-06-17 — Package Directory Rename: `@/*` Aliases Make It Safe
+
+- **When a codebase uses `@/*` path aliases** (tsconfig `paths: {"@/*": ["./src/*"]}`), all internal imports within a package resolve via the alias — NOT the directory name. This means:
+  - Renaming `packages/opencode/` → `packages/rimuru/` breaks ZERO import statements within the package
+  - Only external references to the directory path need updating: workspace config, turbo.json, build scripts
+- **Cross-package imports** use `@rimurucode-ai/package-name` (from other packages in the monorepo) — these are already decoupled from directory names
+- **No relative imports like `../../opencode/src/`** were found — the workspace resolution pattern prevented this antipattern entirely
+- **Always check** with `rg -rn 'packages/<oldname>'` before renaming to find all external references
+- **Turbo task names** in `turbo.json` use the `package.json#name` field, NOT the directory name. After rename, ensure task names match: `"rimuru-ai#test"` not `"rimuru#test"`
+
+## 2026-06-17 — Migration Complete: Phases 1-5 Done
+
+- **Final state**: All user-facing branding changed from `opencode` to `rimuru-ai`. Binary builds at `dist/rimuru-ai-linux-x64/bin/opencode` (internal binary name stays `opencode`).
+- **Preserved intentionally**:
+  - `ProviderV2.ID.opencode` — well-known provider ID for open-source models (OpenCode Go)
+  - `.opencode/` fallback path — backward compat with existing user configs
+  - External plugin names (`opencode-daytona`, `opencode-helicone-session`, etc.)
+  - `opencode-agent[bot]` — GitHub bot identity
+  - `opencode.json`/`opencode.jsonc` — config filenames (canonical)
+  - `OPENCODE_*` env vars — existing env variables
+- **Non-blocking remaining items**: `infra/*.ts` (SST resource names), `script/*.ts` (dev scripts), `patches/`, docs
+- **Build status**: `bun run build --single --skip-embed-web-ui --skip-install` passes with smoke test
+- **Version**: `1.0.0-dev-YYYYMMDDHHMM` (upgraded from `0.0.0-dev-`)
+
+## 2026-06-17 — Dual-Key Provider Registration Pattern
+
+- **Problem**: models.dev/api.json returns provider ID `"opencode"`, but rimuru custom handler only registered under `"rimuru"` key. Custom handler loop does `database[providerID]` lookup — `database["opencode"]` was undefined → handler silently skipped → models.dev free models never appeared.
+- **Solution**: Register the same handler under TWO keys: `"rimuru"` (for user config block) and `"opencode"` (for models.dev data). This is cleaner than renaming either side.
+- **Pattern**: `custom(dep, [key1, key2], handler)` — Provider.ts registration supports arrays of aliases.
+- **Session handler matching**: `isRimuruProvider` must check BOTH `startsWith("rimuru-ai")` AND `=== ProviderV2.ID.opencode` because opencode free models route through the same rimuru endpoint but have a different provider ID.
+- **Native runtime allowlist**: Must explicitly include `ProviderV2.ID.opencode` — it's not automatically covered by `startsWith("rimuru-ai")`.
+- **Key insight**: `ProviderV2.ID.opencode` as `"opencode"` (not `"rimuru-ai"`) is correct — models.dev always returns `"opencode"`, and the catalog/plugin system looks up by this constant. Changing it to something else would require renaming in the API response parser too.
+- **Server run pattern**: `setsid <binary> < /dev/null` to prevent shell timeout killing the process in non-interactive environment.
+- **Binary rebuild**: `bun run build --single --skip-install` from `packages/rimuru/` directory (not root).
+- **Binary install**: `cp dist/rimuru-ai-linux-x64/bin/rimuru-ai ~/.local/bin/` (no sudo needed in non-interactive env).
+
+### 2026-06-18 — TUI Logo Component: logo.right Must Match logo.left Length
+
+- **Bug**: `Array.from(undefined)` crash in `component/logo.tsx:856` when `logo.right` is `[]` but `logo.left` has 6 elements.
+- **Root cause**: `Logo` component iterates `<For each={ctx.shape.left}>` and accesses `ctx.shape.right[index()]` — if `right` is shorter than `left`, the excess indices return `undefined`, and `renderLine(line)` calls `Array.from(undefined)`.
+- **Fix**: Always ensure `logo.left.length === logo.right.length`. For the Rimuru logo: `logo.right = rimuruLines.map(() => "")`.
+- **Empty `go` arrays**: Safe for `bg-pulse-render.ts` because `go.left.map()` returns `[]`, so `Array.from(line)` in `flatMap` never executes.
+
+### 2026-06-18 — Rimuru SVG Icon Generation Toolchain
+
+- **Tool**: `rsvg-convert` for SVG→PNG rendering, ImageMagick `magick convert` for padding/resizing/ICO assembly.
+- **Command**: `rsvg-convert -h 2048 rimuru.svg -o base-raw.png && magick base-raw.png -background none -gravity center -extent 2048x2048 base.png` to pad landscape SVG to square with transparent background.
+- **ICO generation**: `magick base.png \( -clone 0 -resize 256x256 \) \( -clone 0 -resize 64x64 \) ... -delete 0 icon.ico` — layers from largest to smallest.
+- **ICNS (Linux fallback)**: ImageMagick can write a simple ICNS but `iconutil` (macOS-only) is preferred. On Linux, the fallback is a single-resolution ICNS.
+- **All channels**: Must copy icons to `icons/dev/`, `icons/beta/`, `icons/prod/`.
+- **Source SVG**: CC0-licensed Rimuru slime from Wikimedia Commons. 2KB, 3900×2869 viewBox, single `<path>` + 2 ellipses + eye detail.

package/.rimuru/improver/plugins.md

@@ -0,0 +1,21 @@
+# Plugins Log - OpenCode Improver
+
+Log of plugins evaluated/installed, with notes on what they do and whether they helped.
+
+---
+
+## 2026-06-15 — Anthropic Skills Repository (github.com/anthropics/skills)
+
+- **What**: Official Anthropic skills repo — 17 skills covering creative, document, development, enterprise, and meta domains
+- **How**: Cloned repo, copied all `skills/*/` directories to `~/.agents/skills/`. OpenCode auto-discovers skills from SKILL.md files.
+- **Python deps**: Installed `pypdf`, `pdfplumber`, `reportlab`, `openpyxl`, `python-pptx`, `markitdown` for document skill scripts
+- **Notable sizes**: `canvas-design` (83 files, 5.6 MB), `docx` (61 files, 1.2 MB), `pptx` (59 files, 1.2 MB), `xlsx` (54 files, 1.2 MB), `claude-api` (49 files, 756 KB)
+- **Verdict**: ✅ All installed and verified. Some skills (docx, pdf, pptx, xlsx) are proprietary/source-available per Anthropic's LICENSE.txt.
+
+## 2026-06-15 — `npm docx` package (v9+)
+
+- **What**: Node.js library for creating Word (.docx) documents programmatically
+- **Why**: Used to generate a professional R&D proposal document (467 KB, 10+ sections, 4 embedded flow diagrams, tables, headers/footers, page numbers)
+- **Install**: `npm install docx` (prefer local install, global `-g` doesn't resolve from `/tmp/`)
+- **Verdict**: ✅ Works well. Much more control than LibreOffice headless for complex structured docs. Only dependency needed for DOCX generation.
+- **Caveat**: API changed at v9 — `PageNumber` is no longer a constructor. Use `new TextRun({ children: [PageNumber.CURRENT] })` for page numbers. `PageNumberElement` creates bare XML that doesn't render. Check the installed version's API before writing scripts.

package/.rimuru/improver/skills.md

@@ -0,0 +1,60 @@
+# Skills Log — OpenCode Improver
+
+Log of skills created/modified and their purpose.
+
+---
+
+## 2026-06-15 — Installed 17 skills from github.com/anthropics/skills
+
+- **Source**: https://github.com/anthropics/skills (cloned to /tmp/anthropic-skills)
+- **Location**: All installed to `~/.agents/skills/<name>/`
+- **Skills installed**:
+  - **Creative**: `algorithmic-art` (p5.js generative art), `brand-guidelines` (Anthropic branding), `canvas-design` (PNG/PDF visual art, 83 files), `theme-factory` (10 preset themes for artifacts), `slack-gif-creator` (animated GIFs for Slack)
+  - **Document**: `docx` (Word creation/edit, 61 files — Python scripts + npm docx), `pdf` (PDF read/edit/fill/OCR, 12 files), `pptx` (PowerPoint creation/edit, 59 files), `xlsx` (Excel creation/edit, 54 files)
+  - **Development**: `claude-api` (Anthropic API reference, 49 files), `mcp-builder` (MCP server creation guide), `webapp-testing` (Playwright for local web apps), `web-artifacts-builder` (React/shadcn HTML artifacts), `frontend-design` (UI design guidance)
+  - **Enterprise**: `internal-comms` (templates for status reports, newsletters, etc.), `doc-coauthoring` (structured doc writing workflow)
+  - **Meta**: `skill-creator` (create/modify/evaluate skills)
+- **Dependencies installed**: `pypdf`, `pdfplumber`, `reportlab`, `openpyxl`, `python-pptx`, `markitdown` (via pip3 --break-system-packages) — needed by docx/pdf/pptx/xlsx skills for unpacking, validation, and conversion
+- **Conflict check**: Zero naming conflicts with existing skills
+- **Verification**: All 17 SKILL.md files present and parseable
+
+## 2026-06-15 — DOCX editing workflow (used docx skill content)
+
+- **Lesson learned**: The `docx` skill at `~/.agents/skills/docx/` should be loaded FIRST when editing/creating .docx files. It contains:
+  - `ooxml/scripts/unpack.py` / `pack.py` — zip-level docx manipulation (critical for fixing `.undefined` image extensions)
+  - `SKILL.md` with detailed OOXML reference (table styles, numbering, headers/footers XML patterns)
+  - Python Document library for high-level editing
+- **python-docx limitations**: Can't handle malformed content-types. Must pre-fix `.undefined` image extensions via zipfile before python-docx can open a document.
+- **Workflow for future**: (1) Load docx skill → (2) Fix zip-level issues → (3) Use skill's Document library or python-docx → (4) Validate with `unpack.py`
+
+## 2026-06-15 — Document Generation via `docx` npm package
+
+- **Discovered/Used**: `npm docx` package (`npm install -g docx`) for programmatic Word document generation
+- **Capability**: Creates professional DOCX with tables, headers/footers, embedded PNG diagrams (via `ImageRun`), page numbers, styled paragraphs, color-coded sections, Table of Contents scaffolding
+- **When to use**: Any task requiring a formatted Word document proposal, report, or analysis
+- **How**: Write a Node.js script using `docx`'s `Document`, `Packer`, `Paragraph`, `TextRun`, `Table`, `ImageRun`, etc. Then `node script.js` produces the `.docx` file
+- **Caveats**: 
+  - `PageNumber()` is deprecated in v9+ — use `new TextRun({ children: [PageNumber.CURRENT] })` instead
+  - `new PageNumberElement()` creates bare XML — does NOT render a working page number
+  - Install locally (not globally) for reliable resolution from scripts in `/tmp/`
+  - Diagrams must be PNG files on disk; `ImageRun` embeds them inline
+- **Related**: `document-prep` subagent handles LibreOffice-based document tasks; use `docx` npm for programmatic generation (more control, no GUI dep)
+
+---
+
+## 2026-06-15 — Created 3 new subagents (not skills, but domain-specific agents)
+
+- Created subagents as global agent `.md` files (not skills). Rationale: subagents run autonomously with tool access; skills just inject instructions. These domains need tool autonomy.
+- **`ethical-hacking.md`**: Offensive/defensive security agent. Uses `bash: ask` for tool execution, `webfetch: allow` for CVE/exploit research.
+- **`document-prep.md`**: Document automation via LibreOffice headless. References `soffice` commands inline. Uses `bash: ask` for document generation/conversion.
+- **`erp-architect.md`**: ERP architecture consultant. Read-only agent (`bash: deny`) that advises on SAP/Oracle/Odoo/D365 architecture.
+- All subagents omit `model` override → use default provider model, cheaper than hardcoding an expensive model.
+
+## 2026-06-17 — MADAR → FastContext Skill Swap
+
+- **Deleted**: `~/.agents/skills/madar-graph/` — MADAR knowledge-graph skill (replaced by FastContext)
+- **Created**: `~/.agents/skills/fastcontext/SKILL.md` — FastContext delegation-based repo explorer from [github.com/microsoft/fastcontext](https://github.com/microsoft/fastcontext)
+- **Usage**: `fastcontext -q "<question>" --citation` for focused file:line citations, or `--max-turns 12` for complex multi-file traces
+- **Requirements**: Requires `BASE_URL`, `MODEL`, `API_KEY` env vars (OpenAI-compatible endpoint) — not yet configured
+
+## 2026-06-13 — mssql skill: added connections.json template

package/.rimuru/improver/token-audit.md

@@ -0,0 +1,21 @@
+# Token Audit — OpenCode Improver
+
+## 2026-06-15 — MADAR Graph Was Empty Due to Missing `--include-docs`
+
+- **Issue**: All `madar generate .` calls omitted `--include-docs`. Since `.config/opencode/` is ~90% markdown, the extractor produced 0 nodes → `pack`/`query` returned nothing → agents fell back to glob/grep/file sweeps (wasting thousands of tokens).
+- **Fix**: Added `--include-docs` to all `madar generate` calls in `codeflux.md`, `codeflux-god.md`, and `madar-graph/SKILL.md`.
+- **Result**: 94 nodes, 124 edges from 16 files. Pack/query now work and can replace expensive file sweeps.
+- **Token impact**: Each successful MADAR `pack` replaces ~2-5k tokens of glob/grep/file reads. Estimated savings: ~3-8k tokens per lookup.
+- **SUPERSEDED 2026-06-17**: MADAR replaced entirely by FastContext. Now uses `fastcontext -q "..." --citation` instead of `madar pack`/`madar query`. No graph to maintain; exploration is delegated to a subagent that returns compact file:line citations (~200-500 tokens per lookup vs ~3-8k per file sweep).
+
+---
+
+## 2026-06-15 — Subagent token optimization applied
+
+- **Ethical-hacking agent**: 8 steps instead of default 25+ → saves ~200-400 tokens per subagent invocation (fewer reasoning loops). `bash: ask` prevents runaway tool execution.
+- **Document-prep agent**: 6 steps + inline soffice commands → saves ~2-5k tokens by avoiding websearch for "how to convert ODT to PDF with LibreOffice" on every use.
+- **ERP-architect agent**: 6 steps + `bash: deny` + `webfetch: allow` → read-only by design. No bash tool schema overhead (~100 tokens saved per request).
+- **All new subagents**: No `model` override → inherits default model, no hardcoded expensive provider tokens.
+- **codeflux.json / opencode.json agent entries**: Added `permission.task: {"*": "allow"}` to formalize subagent access (was implicit before, now explicit — same token cost, clearer intent).
+
+## 2026-06-13 — Initial observations

package/.rimuru/opencode.jsonc

@@ -0,0 +1,140 @@
+{
+  "$schema": "https://opencode.ai/config.json",
+
+  "provider": {},
+
+  "permission": {},
+
+  "references": {
+    "effect": {
+      "repository": "github.com/Effect-TS/effect-smol",
+      "description": "Use for Effect v4 and effect-smol implementation details"
+    },
+    "rimuru-local": {
+      "path": "~/.local/share/opencode",
+      "description": "Contains opencode logs and data"
+    },
+    "rimuru-docs": {
+      "path": "./packages/docs",
+      "description": "Rimuru documentation site built with Astro"
+    },
+    "rimuru-core": {
+      "path": "./packages/core",
+      "description": "Core shared library for rimuru-ai"
+    }
+  },
+
+  "mcp": {
+    "figma": {
+      "type": "local",
+      "command": ["npx", "-y", "@nexus2520/figma-mcp-server"],
+      "enabled": true,
+      "environment": {
+        "FIGMA_ACCESS_TOKEN": "figd_hnE1be8CLqwEb2ZKciIr4y8bokJs38c-wTasSVQj"
+      }
+    },
+    "filesystem": {
+      "type": "local",
+      "command": ["npx", "-y", "@modelcontextprotocol/server-filesystem", "/home/gowdaman"],
+      "enabled": true
+    },
+    "postman": {
+      "type": "local",
+      "command": ["npx", "-y", "@postman/postman-mcp-server@latest", "--region", "us", "--full"],
+      "enabled": true,
+      "environment": {
+        "POSTMAN_API_KEY": "PMAK-6a27e753d8c40900019727e5-35d5902e704787a6b5a23328fbf0e7922d"
+      }
+    },
+    "chrome-devtools": {
+      "type": "local",
+      "command": ["npx", "-y", "chrome-devtools-mcp@latest"],
+      "enabled": true
+    },
+    "browser-use": {
+      "type": "local",
+      "command": ["npx", "-y", "browser-use", "--mcp"],
+      "enabled": true
+    },
+  },
+
+  "tools": {
+    "github-triage": false,
+    "github-pr-search": false
+  },
+
+  "agent": {
+    "rimuru-build": {
+      "mode": "primary",
+      "permission": {
+        "task": { "*": "allow" },
+        "tool": { "*": "allow" }
+      }
+    },
+    "rimuru-docs": {
+      "mode": "primary",
+      "permission": {
+        "task": { "*": "allow" },
+        "tool": { "*": "allow" }
+      }
+    },
+    "rimuru-codebase": {
+      "mode": "primary",
+      "permission": {
+        "task": { "*": "allow" },
+        "tool": { "*": "allow" }
+      }
+    },
+    "veldora": {
+      "mode": "primary",
+      "permission": {
+        "task": { "*": "allow" },
+        "tool": { "*": "allow" }
+      }
+    },
+    "veldora-pro": {
+      "mode": "primary",
+      "permission": {
+        "task": { "*": "allow" },
+        "tool": { "*": "allow" }
+      }
+    },
+    "veldora-doc": {
+      "mode": "primary",
+      "permission": {
+        "task": { "*": "allow" },
+        "tool": { "*": "allow" }
+      }
+    }
+  },
+
+  "plugin": [
+    "opencode-optimal-model-temps",
+    "opencode-codebase-index",
+    "opencode-helicone-session",
+    "opencode-goal-plugin"
+  ],
+
+  "command": {
+    "goal": {
+      "description": "Set a session-scoped goal and auto-continue until complete.",
+      "template": "$ARGUMENTS",
+      "agent": "build"
+    },
+    "build-package": {
+      "description": "Build a specific package with turbo",
+      "template": "bun turbo build --filter=$ARGUMENTS",
+      "agent": "build"
+    },
+    "typecheck": {
+      "description": "Run typecheck across the project",
+      "template": "bun turbo typecheck",
+      "agent": "build"
+    },
+    "lint": {
+      "description": "Run linter",
+      "template": "oxlint",
+      "agent": "build"
+    }
+  }
+}

package/.rimuru/plugins/smoke-theme.json

@@ -0,0 +1,223 @@
+{
+  "$schema": "https://github.com/gowdaman-dev/rimuru-ai/theme.json",
+  "defs": {
+    "nord0": "#2E3440",
+    "nord1": "#3B4252",
+    "nord2": "#434C5E",
+    "nord3": "#4C566A",
+    "nord4": "#D8DEE9",
+    "nord5": "#E5E9F0",
+    "nord6": "#ECEFF4",
+    "nord7": "#8FBCBB",
+    "nord8": "#88C0D0",
+    "nord9": "#81A1C1",
+    "nord10": "#5E81AC",
+    "nord11": "#BF616A",
+    "nord12": "#D08770",
+    "nord13": "#EBCB8B",
+    "nord14": "#A3BE8C",
+    "nord15": "#B48EAD"
+  },
+  "theme": {
+    "primary": {
+      "dark": "nord10",
+      "light": "nord9"
+    },
+    "secondary": {
+      "dark": "nord9",
+      "light": "nord9"
+    },
+    "accent": {
+      "dark": "nord7",
+      "light": "nord7"
+    },
+    "error": {
+      "dark": "nord11",
+      "light": "nord11"
+    },
+    "warning": {
+      "dark": "nord12",
+      "light": "nord12"
+    },
+    "success": {
+      "dark": "nord14",
+      "light": "nord14"
+    },
+    "info": {
+      "dark": "nord8",
+      "light": "nord10"
+    },
+    "text": {
+      "dark": "nord6",
+      "light": "nord0"
+    },
+    "textMuted": {
+      "dark": "#8B95A7",
+      "light": "nord1"
+    },
+    "background": {
+      "dark": "nord0",
+      "light": "nord6"
+    },
+    "backgroundPanel": {
+      "dark": "nord1",
+      "light": "nord5"
+    },
+    "backgroundElement": {
+      "dark": "nord2",
+      "light": "nord4"
+    },
+    "border": {
+      "dark": "nord2",
+      "light": "nord3"
+    },
+    "borderActive": {
+      "dark": "nord3",
+      "light": "nord2"
+    },
+    "borderSubtle": {
+      "dark": "nord2",
+      "light": "nord3"
+    },
+    "diffAdded": {
+      "dark": "nord14",
+      "light": "nord14"
+    },
+    "diffRemoved": {
+      "dark": "nord11",
+      "light": "nord11"
+    },
+    "diffContext": {
+      "dark": "#8B95A7",
+      "light": "nord3"
+    },
+    "diffHunkHeader": {
+      "dark": "#8B95A7",
+      "light": "nord3"
+    },
+    "diffHighlightAdded": {
+      "dark": "nord14",
+      "light": "nord14"
+    },
+    "diffHighlightRemoved": {
+      "dark": "nord11",
+      "light": "nord11"
+    },
+    "diffAddedBg": {
+      "dark": "#36413C",
+      "light": "#E6EBE7"
+    },
+    "diffRemovedBg": {
+      "dark": "#43393D",
+      "light": "#ECE6E8"
+    },
+    "diffContextBg": {
+      "dark": "nord1",
+      "light": "nord5"
+    },
+    "diffLineNumber": {
+      "dark": "nord2",
+      "light": "nord4"
+    },
+    "diffAddedLineNumberBg": {
+      "dark": "#303A35",
+      "light": "#DDE4DF"
+    },
+    "diffRemovedLineNumberBg": {
+      "dark": "#3C3336",
+      "light": "#E4DDE0"
+    },
+    "markdownText": {
+      "dark": "nord4",
+      "light": "nord0"
+    },
+    "markdownHeading": {
+      "dark": "nord8",
+      "light": "nord10"
+    },
+    "markdownLink": {
+      "dark": "nord9",
+      "light": "nord9"
+    },
+    "markdownLinkText": {
+      "dark": "nord7",
+      "light": "nord7"
+    },
+    "markdownCode": {
+      "dark": "nord14",
+      "light": "nord14"
+    },
+    "markdownBlockQuote": {
+      "dark": "#8B95A7",
+      "light": "nord3"
+    },
+    "markdownEmph": {
+      "dark": "nord12",
+      "light": "nord12"
+    },
+    "markdownStrong": {
+      "dark": "nord13",
+      "light": "nord13"
+    },
+    "markdownHorizontalRule": {
+      "dark": "#8B95A7",
+      "light": "nord3"
+    },
+    "markdownListItem": {
+      "dark": "nord8",
+      "light": "nord10"
+    },
+    "markdownListEnumeration": {
+      "dark": "nord7",
+      "light": "nord7"
+    },
+    "markdownImage": {
+      "dark": "nord9",
+      "light": "nord9"
+    },
+    "markdownImageText": {
+      "dark": "nord7",
+      "light": "nord7"
+    },
+    "markdownCodeBlock": {
+      "dark": "nord4",
+      "light": "nord0"
+    },
+    "syntaxComment": {
+      "dark": "#8B95A7",
+      "light": "nord3"
+    },
+    "syntaxKeyword": {
+      "dark": "nord9",
+      "light": "nord9"
+    },
+    "syntaxFunction": {
+      "dark": "nord8",
+      "light": "nord8"
+    },
+    "syntaxVariable": {
+      "dark": "nord7",
+      "light": "nord7"
+    },
+    "syntaxString": {
+      "dark": "nord14",
+      "light": "nord14"
+    },
+    "syntaxNumber": {
+      "dark": "nord15",
+      "light": "nord15"
+    },
+    "syntaxType": {
+      "dark": "nord7",
+      "light": "nord7"
+    },
+    "syntaxOperator": {
+      "dark": "nord9",
+      "light": "nord9"
+    },
+    "syntaxPunctuation": {
+      "dark": "nord4",
+      "light": "nord0"
+    }
+  }
+}