kuker24 · kuker24 · Sep 15, 2026 · Sep 15, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -2,8 +2,9 @@
 
 ## Unreleased
 
-Catalog update for AI LABS 8-repo wave adaptation and hygiene patch. Totals become **63** (47 model-invoked, 16 manual slash commands).
+Catalog update for AI LABS 8-repo wave adaptation, hygiene patch, and MarkItDown ingest. Totals become **64** (48 model-invoked, 16 manual slash commands).
 
+- Add model-invoked specialist `markitdown` as a Markdown ingest converter (Office/PDF/HTML/CSV/XLSX/PPTX/EPUB/ZIP). Optional MCP `opencode-bf markitdown enable` registers local stdio (`uvx --from markitdown-mcp markitdown-mcp`). `FOREIGN_ON_DEMAND`. SmartDoc keeps contract/QA/render. No npm plugin, no Azure keys, no docker bind-all.
 - Add first-party model-invoked specialist `img2threejs` for code-only procedural Three.js TypeScript model reconstruction from reference object images (editable Group factory, visible vs. inferred geometry spec, screenshot loop gate; no mesh blob downloads).
 - Add optional `reticle` MCP (`opencode-bf reticle enable` / `disable`) registering local perception server (`npx -y @reticlehq/server mcp`). `FOREIGN_ON_DEMAND`, server is FSL-1.1-ALv2, perception only, never an auto-implementer; default verification remains `playwright-qa` / `chrome-devtools-axi`.
 - Add optional `ui-skills` remote MCP (`opencode-bf ui-skills enable` / `disable`) registering `https://www.ui-skills.com/mcp` for design-skill lookup only. Product UI remains Design Bank + Impeccable + Design V2 atoms + shadcn; `BANK_MISS` never generates from random ui-skills documents.

diff --git a/README.md b/README.md
@@ -1,14 +1,14 @@
 # OpenCodeBestFriend
 
 Production-ready capability layer for OpenCode:
-63 routed skills (core + Wave 2/3 warehouse specialists), MCP, Codebase Memory,
+64 routed skills (core + Wave 2/3 warehouse specialists), MCP, Codebase Memory,
 Design Bank, Design Intelligence, SmartDoc, browser and verification tooling.
 
 OpenCodeBestFriend is an installer and runtime overlay for [OpenCode](https://opencode.ai). It is **not** Claude Code, **not** a model provider, and **not** a dump of a developer home directory.
 
 ## What it is
 
-- 63 skills: 47 model-invoked, 16 manual slash commands
+- 64 skills: 48 model-invoked, 16 manual slash commands
 - A thin `AGENTS.md` router (lazy, one primary specialist)
 - Core MCP: Codebase Memory, Context7, shadcn
 - Design Bank discovery or download (media is **not** in git)
@@ -61,7 +61,7 @@ Restart OpenCode after install. Config is not hot-reloaded.
         ┌───────────────────┼────────────────────┐
         ▼                   ▼                    ▼
       Skills               MCP                 Rules
-      47 automatic       Codebase Memory        Verification
+      48 automatic       Codebase Memory        Verification
      16 manual          Context7              Engineering
                        shadcn
         │
@@ -100,6 +100,7 @@ Default: repository evidence first. Then at most one specialist.
 | Procedural Three.js object from image | `img2threejs` |
 | Deterministic HTML composition video | `hyperframes` |
 | Documents (PDF/DOCX/answer/extract/review) | `smartdoc` |
+| File to Markdown ingest | `markitdown` |
 | Reusable book/module knowledge | `smartbook-ingest` |
 | Scholarly literature & manuscripts | `academic` |
 | Prose AI-tell removal & humanizing | `humanizer` (`/unslop`) |
@@ -129,6 +130,7 @@ Optional:
 - `stitch` — `opencode-bf stitch enable` registers Google Stitch as a remote comp/mock source. Not an owned core server and not a production UI implementer: hand screens to `found-this-design` or `impeccable` before shipping. Keys are never written into config, only referenced as `{env:STITCH_API_KEY}`, or omitted with `--oauth`. `opencode-bf stitch disable` removes only that server key. Absent is not a `doctor` failure; a malformed entry fails closed.
 - `reticle` — `opencode-bf reticle enable` registers Reticle as a local perception server (`npx -y @reticlehq/server mcp`). `FOREIGN_ON_DEMAND`. Server package is FSL-1.1-ALv2 (competing-use clause); SDK packages (Apache-2.0) are not vendored. Never an auto-implementer; default verification remains `playwright-qa` / `chrome-devtools-axi`. `opencode-bf reticle disable` removes only that server key. Absent is not a `doctor` failure; a malformed entry fails closed.
 - `ui-skills` — `opencode-bf ui-skills enable` registers UI Skills (`https://www.ui-skills.com/mcp`) as an optional remote MCP server. `FOREIGN_ON_DEMAND` for design-skill lookup only. Product UI remains Design Bank + Impeccable + Design V2 atoms + shadcn; `BANK_MISS` never generates from a random ui-skills document. `opencode-bf ui-skills disable` removes only that server key. Absent is not a `doctor` failure; a malformed entry fails closed.
+- `markitdown` — `opencode-bf markitdown enable` registers MarkItDown as a local stdio ingest converter (`uvx --from markitdown-mcp markitdown-mcp`). `FOREIGN_ON_DEMAND`. Local trusted agents only; never `--http` / `0.0.0.0` / docker bind-all. Output is Markdown data; SmartDoc keeps contract/QA/render. `opencode-bf markitdown disable` removes only that server key. Absent is not a `doctor` failure; a malformed entry fails closed.
 - `exa` — `FOREIGN_ON_DEMAND`; installer never adds, removes, or overwrites it
 
 NVIDIA SkillEvaluator is `FOREIGN_ON_DEMAND` in the same sense: a maintainer may run it externally for embedding-based overlap scoring or live catalog evaluation. Caliper is `FOREIGN_ON_DEMAND` similarly: a maintainer may `pipx install caliper-eval` off-tree for prompt/agent benchmark evaluation. Neither is vendored into `lib/`, the installer never adds them, `doctor` does not fail when they are absent, and a malformed MCP entry fails closed like any other schema violation.

diff --git a/docs/architecture.md b/docs/architecture.md
@@ -10,7 +10,7 @@
         ┌───────────────────┼────────────────────┐
         ▼                   ▼                    ▼
       Skills               MCP                 Rules
-      47 automatic       Codebase Memory        Verification
+      48 automatic       Codebase Memory        Verification
       16 manual          Context7              Engineering
                        shadcn
         │

diff --git a/docs/mcp.md b/docs/mcp.md
@@ -16,6 +16,7 @@ Optional:
 - `stitch` — `opencode-bf stitch enable` (remote comp/mock source only; auth via `{env:STITCH_API_KEY}` or `--oauth`)
 - `reticle` — `opencode-bf reticle enable` (local stdio via `npx -y @reticlehq/server mcp`; perception only, never auto-implementer)
 - `ui-skills` — `opencode-bf ui-skills enable` (remote HTTP `https://www.ui-skills.com/mcp`; design-skill lookup only)
+- `markitdown` — `opencode-bf markitdown enable` (local stdio via `uvx --from markitdown-mcp markitdown-mcp`; Markdown ingest only)
 - `exa` — foreign; never add/remove/overwrite
 
 Merge is parse-aware. Comment-free JSON is rewritten with `json.dumps`. JSONC with comments is patched surgically (owned MCP keys only). If surgical merge cannot be verified, install fails closed instead of destroying comments.
@@ -29,3 +30,5 @@ Doctor reports `CONFIGURED` for owned MCP entries present in config. That is not
 `opencode-bf reticle enable` configures Reticle as an optional local perception MCP server (`npx -y @reticlehq/server mcp`). It is `FOREIGN_ON_DEMAND`. The server package is FSL-1.1-ALv2 (competing-use clause); SDK packages (Apache-2.0) are not vendored. Reticle is never an auto-implementer; after a feature is done, default verification remains `playwright-qa` or `chrome-devtools-axi`. Reticle is extra perception if the user enabled it. `opencode-bf reticle disable` surgically removes only the reticle server key. Absent is not a doctor failure; a malformed entry fails closed.
 
 `opencode-bf ui-skills enable` configures UI Skills as an optional remote MCP server (`https://www.ui-skills.com/mcp`). It is `FOREIGN_ON_DEMAND` for design-skill lookup only (`list_skills`, `get_skill`). Product UI remains Design Bank + Impeccable + Design V2 atoms + shadcn; `BANK_MISS` never generates from a random ui-skills document. `opencode-bf ui-skills disable` surgically removes only the ui-skills server key. Absent is not a doctor failure; a malformed entry fails closed.
+
+`opencode-bf markitdown enable` configures MarkItDown as an optional local stdio ingest MCP (`uvx --from markitdown-mcp markitdown-mcp`). It is `FOREIGN_ON_DEMAND`. Official server is for local trusted agents only; never `--http`, never bind `0.0.0.0`, never docker bind-all. The converter is not vendored into `lib/`. Missing `uvx` is documented in the skill (CLI/`pipx`/`enable`); enable still writes the stdio command like reticle. `opencode-bf markitdown disable` surgically removes only the markitdown server key. Absent is not a doctor failure; a malformed entry (including `--http` / `0.0.0.0`) fails closed.
diff --git a/docs/routing.md b/docs/routing.md
@@ -26,7 +26,7 @@ UI direction from the bank routes to `found-this-design` first, which stops befo
 
 Browser verification follows four explicit doors: exploratory application UI routes to `playwright-qa`, persistent multi-account sessions route to `browser-act`, observed Chromium cause routes to `chrome-devtools-axi`, and button handler sequential undo / shared-store side effects route to `click-path-audit`.
 
-Documents (answer, create, transform, extract, review, PDF/DOCX) route to `smartdoc`. Reusable book/module knowledge routes to `smartbook-ingest`. `/docx` and `/pdf` are missing aliases; nearest is `smartdoc`. `/pptx` is NOT_APPLICABLE. Do not add `commands/pdf.md` or `commands/docx.md`. Impeccable `document` remains DESIGN.md generation.
+Documents (answer, create, transform, extract, review, PDF/DOCX) route to `smartdoc`. File-to-Markdown ingest routes to `markitdown`. Reusable book/module knowledge routes to `smartbook-ingest`. `/docx` and `/pdf` are missing aliases; nearest is `smartdoc`. `/pptx` is NOT_APPLICABLE. Do not add `commands/pdf.md` or `commands/docx.md`. Impeccable `document` remains DESIGN.md generation.
 
 Prose AI-tell removal and natural tone polishing route to `humanizer` (`/unslop` is its manual alias). Scholarly research, academic manuscripts, and structured peer critique route to `academic`. Deterministic HTML composition rendered to video routes to `hyperframes`. Editorial technical diagrams (HTML/SVG) route to `diagram-design`.
 

diff --git a/docs/skills.md b/docs/skills.md
@@ -2,10 +2,10 @@
 
 Policy: `vendor/skill-policy.json` plus `vendor/skill-allowlist.txt`.
 
-- 47 model-invoked skills live under `skills/` and install to `~/.config/opencode/skills/` (core + Wave 2/3 warehouse specialists)
+- 48 model-invoked skills live under `skills/` and install to `~/.config/opencode/skills/` (core + Wave 2/3 warehouse specialists)
 - 16 manual skills live under `manual-skills/` and install to `~/.config/opencode/bestfriend/skills/` plus `commands/`
 
-`smartdoc` is per-job document intelligence. `smartbook-ingest` compiles reusable local knowledge. `humanizer` cleans user-facing prose tells (`/unslop` is its manual alias). `academic` manages scholarly research, writing, and peer review. `hyperframes` handles deterministic HTML-to-MP4 video composition. `diagram-design` crafts editorial HTML/SVG diagrams. `img2threejs` reconstructs procedural Three.js models from reference images. Warehouse diagnostics include `agent-architecture-audit` (agent stack layers), `cost-aware-llm-pipeline` (token budgeting), `eval-harness` (benchmarks), `prompt-optimizer` (prompt refinement), and `skill-stocktake` (catalog hygiene). Wave 3 adds `api-design`, `contract-first`, `automation-audit-ops`, `code-tour`, and `click-path-audit`. Handwriting is a SmartDoc renderer, not a skill.
+`smartdoc` is per-job document intelligence. `markitdown` converts Office/PDF/HTML/CSV/XLSX/PPTX/EPUB/ZIP to Markdown for ingest; SmartDoc keeps contract/QA/render. `smartbook-ingest` compiles reusable local knowledge. `humanizer` cleans user-facing prose tells (`/unslop` is its manual alias). `academic` manages scholarly research, writing, and peer review. `hyperframes` handles deterministic HTML-to-MP4 video composition. `diagram-design` crafts editorial HTML/SVG diagrams. `img2threejs` reconstructs procedural Three.js models from reference images. Warehouse diagnostics include `agent-architecture-audit` (agent stack layers), `cost-aware-llm-pipeline` (token budgeting), `eval-harness` (benchmarks), `prompt-optimizer` (prompt refinement), and `skill-stocktake` (catalog hygiene). Wave 3 adds `api-design`, `contract-first`, `automation-audit-ops`, `code-tour`, and `click-path-audit`. Handwriting is a SmartDoc renderer, not a skill.
 
 OpenCode 1.18.x has no `disable-model-invocation` field. Manual skills must not be copied into the discovered skills directory.
 

diff --git a/docs/warehouse-inventory.md b/docs/warehouse-inventory.md
@@ -367,3 +367,19 @@ Evaluation and disposition contract for the AI LABS 8-repo wave (procedural 3D,
 | `swiftui-skills` | **DEFER** | `-` | Apple platform / Xcode 26 ecosystem deferred. Target platform gate remains Linux x86_64 and OpenCode 1.18.x. |
 | `caliper` | **FOREIGN_ON_DEMAND** | `-` | Benchmark CLI runner (`caliper-eval`). Maintainer may run off-tree via pipx; zero `lib/` vendor coupling. |
 | `anti-slop` | **MERGE** | `install-anti-slop`, `impeccable`, `rules/03-prose-discipline.md` | Anti-pattern guardrails merged into existing taste, prose, and linting references; zero extra catalog skills. |
+
+---
+
+## Wave MarkItDown
+
+Microsoft MarkItDown as an ingest converter, not a second document OS. SmartDoc keeps contract/QA/render.
+
+| Candidate / Repo | Decision | BestFriend Target | Reason |
+| :--- | :---: | :--- | :--- |
+| `microsoft/markitdown` CLI/lib | **NEW** | `skills/markitdown` | Thin first-party skill: convert Office/PDF/HTML/CSV/XLSX/PPTX/EPUB/ZIP to Markdown, then hand off. |
+| `markitdown-mcp` official | **FOREIGN_ON_DEMAND** | `mcp.markitdown` | Optional local stdio (`uvx --from markitdown-mcp markitdown-mcp`). Local trusted agents only. |
+| `opencode-markitdown` npm plugin | **REJECT** | `-` | Config-hook mutation forbidden. |
+| community `trsdn-markitdown-mcp` | **REJECT** | `-` | Not Microsoft. |
+| Azure Document Intelligence / Content Understanding | **DEFER** | `-` | No keys in config. |
+| youtube / audio extras | **DEFER** | `-` | Out of document lane. |
+| Duplicate SmartDoc modes | **REJECT** | `-` | SmartDoc keeps contract/QA/render. |
diff --git a/lib/cli.py b/lib/cli.py
@@ -24,6 +24,8 @@
     cmd_install,
     cmd_restore,
     cmd_restore_list,
+    cmd_markitdown_disable,
+    cmd_markitdown_enable,
     cmd_reticle_disable,
     cmd_reticle_enable,
     cmd_serena_enable,
@@ -111,6 +113,9 @@ def build_parser() -> argparse.ArgumentParser:
     ret = sub.add_parser("reticle", help="optional Reticle local perception MCP")
     ret.add_argument("action", choices=["enable", "disable"])
 
+    md = sub.add_parser("markitdown", help="optional MarkItDown local ingest MCP")
+    md.add_argument("action", choices=["enable", "disable"])
+
     uis = sub.add_parser("ui-skills", help="optional UI Skills remote MCP")
     uis.add_argument("action", choices=["enable", "disable"])
 
@@ -177,6 +182,10 @@ def main(argv: list[str] | None = None) -> int:
         if args.action == "enable":
             return cmd_reticle_enable()
         return cmd_reticle_disable()
+    if cmd == "markitdown":
+        if args.action == "enable":
+            return cmd_markitdown_enable()
+        return cmd_markitdown_disable()
     if cmd == "ui-skills":
         if args.action == "enable":
             return cmd_ui_skills_enable()

diff --git a/lib/doctor.py b/lib/doctor.py
@@ -118,11 +118,11 @@ def mcp_status_map() -> dict[str, str]:
         try:
             data = jsonc.load_path(cfg)
         except (OSError, json.JSONDecodeError, ValueError):
-            return {k: "FAIL" for k in ("codebase-memory-mcp", "context7", "shadcn", "serena", "stitch", "reticle", "ui-skills", "exa")}
+            return {k: "FAIL" for k in ("codebase-memory-mcp", "context7", "shadcn", "serena", "stitch", "reticle", "ui-skills", "markitdown", "exa")}
     mcp = data.get("mcp") or {}
     owned = {"codebase-memory-mcp", "context7", "shadcn"}
-    optional = {"serena", "stitch", "reticle", "ui-skills", "exa"}
-    for name in ("codebase-memory-mcp", "context7", "shadcn", "serena", "stitch", "reticle", "ui-skills", "exa"):
+    optional = {"serena", "stitch", "reticle", "ui-skills", "markitdown", "exa"}
+    for name in ("codebase-memory-mcp", "context7", "shadcn", "serena", "stitch", "reticle", "ui-skills", "markitdown", "exa"):
         spec = mcp.get(name)
         if spec is None:
             out[name] = "OPTIONAL_ABSENT" if name in optional else "FAIL"
@@ -165,6 +165,21 @@ def mcp_status_map() -> dict[str, str]:
                 continue
             out[name] = "CONFIGURED"
             continue
+        if name == "markitdown":
+            typ = spec.get("type")
+            cmd = spec.get("command")
+            if typ != "local" or not isinstance(cmd, list) or not cmd:
+                out[name] = "FAIL"
+                continue
+            joined = " ".join(str(part) for part in cmd)
+            if "--http" in joined or "0.0.0.0" in joined:
+                out[name] = "FAIL"
+                continue
+            if cmd[0] != "uvx" or "markitdown-mcp" not in cmd:
+                out[name] = "FAIL"
+                continue
+            out[name] = "CONFIGURED"
+            continue
         if name not in owned:
             out[name] = "FOREIGN"
             continue

diff --git a/lib/install.py b/lib/install.py
@@ -872,7 +872,7 @@ def take(src: Path, dest: Path) -> None:
         "modelInvokedSkills": meta["model"],
         "manualSkills": meta["manual"],
         "ownedMcp": list(OWNED_MCP),
-        "optionalMcp": ["serena", "stitch", "reticle", "ui-skills", "exa"],
+        "optionalMcp": ["serena", "stitch", "reticle", "ui-skills", "markitdown", "exa"],
         "designBank": {
             "root": bank_root,
             "source": bank_source,
@@ -1402,6 +1402,19 @@ def cmd_reticle_disable() -> int:
     return _optional_mcp_disable("reticle")
 
 
+def cmd_markitdown_enable() -> int:
+    spec: dict[str, object] = {
+        "type": "local",
+        "command": ["uvx", "--from", "markitdown-mcp", "markitdown-mcp"],
+        "enabled": True,
+    }
+    return _optional_mcp_enable("markitdown", spec)
+
+
+def cmd_markitdown_disable() -> int:
+    return _optional_mcp_disable("markitdown")
+
+
 def cmd_ui_skills_enable() -> int:
     spec: dict[str, object] = {
         "type": "remote",

diff --git a/rules/00-routing.md b/rules/00-routing.md
@@ -37,7 +37,7 @@ Do not infer a model provider from a logical model name. Treat custom-gateway al
 - Throwaway evidence for one design question: `/prototype`. Not for production UI; skip ordinary implementation, ADHD, and `/arena`.
 - Unknown / hard bugs, regressions, measured slowdown: `/diagnosing-bugs`. Skip typos, known-cause, and test-first known fixes (`/tdd`).
 - Authoring SKILL.md / AGENTS.md / skill descriptions / context pointers: `/writing-for-agents` (distinct from `/skill-stocktake` which audits catalog hygiene). Workflow choice stays `/ask-matt`.
-- Documents (answer, create, transform, extract, review, PDF/DOCX): `/smartdoc`. Reusable book/module knowledge: `/smartbook-ingest`. SmartDoc may read an existing SmartBook; that is not a second implementation specialist. Impeccable `document` stays DESIGN.md.
+- Documents (answer, create, transform, extract, review, PDF/DOCX): `/smartdoc`. File to Markdown ingest: `/markitdown`. Reusable book/module knowledge: `/smartbook-ingest`. SmartDoc may read an existing SmartBook; that is not a second implementation specialist. Impeccable `document` stays DESIGN.md.
 - Prose AI-tell removal and natural tone polishing: `/humanizer`. Manual `/unslop` is an alias to the same specialist body. Do not auto-apply on ordinary code or diffs. Technical documentation structure stays `/technical-writing`. Code linting stays `/install-anti-slop`. Prompt structure stays `/prompt-optimizer`.
 - Editorial HTML and inline SVG diagrams (architecture, sequence, ER, flowcharts, Wardley): `/diagram-design`. Mermaid/draw.io are inputs to redraw, not final output. Frontend UI implementation stays `/impeccable`. Code-level seams stay `/codebase-design`.
 - Agent architecture diagnosis, autonomous loop failures, context leakage, wrapper regressions: `/agent-architecture-audit`. Load only when user names the agent stack job. Defensive security and secrets auditing stay `/full-audit-keamanan`.