@flitzrrr/agent-skills 1.1.0 → 1.2.0

package/.cursorrules

@@ -2,7 +2,7 @@
 # This file makes skills discoverable in Cursor IDE.
 # For detailed skill usage, see CHEATSHEET.md or individual SKILL.md files.
 
-## Available Skills (55 total from 19 sources)
+## Available Skills (55 total from 20 sources)
 
 ### Security & Code Quality
 - Use `skills/security-review/SKILL.md` for OWASP-based security audits

package/.github/copilot-instructions.md

@@ -1,6 +1,6 @@
 # Agent Skills Hub — VS Code / GitHub Copilot
 
-This repository contains 504 curated AI agent skills from 19 sources.
+This repository contains 480 curated AI agent skills from 19 sources.
 Each skill is a self-contained folder under `skills/` with a `SKILL.md` file.
 
 ## How to Use Skills

package/.lovable

@@ -1,7 +1,7 @@
 # Agent Skills Hub — Lovable Project Instructions
 # This file makes skills available in Lovable AI.
 
-This repository contains 504 curated AI agent skills for code review, security auditing,
+This repository contains 480 curated AI agent skills for code review, security auditing,
 document generation, frontend development, SEO, and project planning.
 
 ## How to Use

package/AGENTS.md

@@ -1,6 +1,6 @@
 # Agent Skills Hub
 
-This repository contains 55 curated AI agent skills from 19 sources.
+This repository contains 55 curated AI agent skills from 20 sources.
 Skills are located in the `skills/` directory — each contains a `SKILL.md` with instructions.
 
 ## Available Skill Categories

package/CHEATSHEET.md

@@ -53,7 +53,7 @@
 
 ## Skills by Source
 
-### Sentry (getsentry/skills) — 504 skills
+### Sentry (getsentry/skills) — 480 skills
 
 Security, code quality, and Git workflow. Battle-tested at scale.
 
@@ -123,7 +123,7 @@ Comprehensive SEO auditing toolkit with Playwright integration.
 
 ---
 
-### OpenCode Processing (DasDigitaleMomentum) — 504 skills
+### OpenCode Processing (DasDigitaleMomentum) — 480 skills
 
 Multi-session project planning and documentation framework.
 
@@ -141,7 +141,7 @@ Multi-session project planning and documentation framework.
 
 ---
 
-### Anthropic (anthropics/skills) — 504 skills
+### Anthropic (anthropics/skills) — 480 skills
 
 Document creation, creative design, developer tools, and enterprise workflows.
 
@@ -184,7 +184,7 @@ Document creation, creative design, developer tools, and enterprise workflows.
 
 ---
 
-### Vercel (vercel-labs/agent-skills) — 504 skills
+### Vercel (vercel-labs/agent-skills) — 480 skills
 
 React/Next.js performance, web design, and deployment.
 

package/CLAUDE.md

@@ -1,6 +1,6 @@
 # Agent Skills Hub
 
-This repository contains 55 curated AI agent skills from 19 sources.
+This repository contains 55 curated AI agent skills from 20 sources.
 Skills are located in the `skills/` directory — each contains a `SKILL.md` with instructions.
 
 ## Available Skill Categories

package/README.md

@@ -2,8 +2,8 @@
 
 [![npm version](https://img.shields.io/npm/v/@flitzrrr/agent-skills?style=flat-square&color=cb3837)](https://www.npmjs.com/package/@flitzrrr/agent-skills)
 [![npm downloads](https://img.shields.io/npm/dm/@flitzrrr/agent-skills?style=flat-square)](https://www.npmjs.com/package/@flitzrrr/agent-skills)
-[![Skills](https://img.shields.io/badge/skills-504-blue?style=flat-square)](skills/)
-[![Sources](https://img.shields.io/badge/sources-19-green?style=flat-square)](vendor/)
+[![Skills](https://img.shields.io/badge/skills-480-blue?style=flat-square)](skills/)
+[![Sources](https://img.shields.io/badge/sources-20-green?style=flat-square)](vendor/)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue?style=flat-square)](LICENSE)
 [![Weekly Discovery](https://img.shields.io/badge/auto--discovery-weekly-purple?style=flat-square)](.github/workflows/skill-discovery.yml)
 
@@ -130,7 +130,7 @@ New sources are discovered weekly via [GitHub Actions](.github/workflows/skill-d
 ```
 agent-skills/
   skills/            504 entries (symlinks to vendor + local forks)
-  vendor/          19 Git submodules (upstream sources)
+  vendor/          20 Git submodules (upstream sources)
   bin/               CLI and build scripts
   docs/              GitHub Pages catalog and project documentation
   .github/           CI workflows and copilot-instructions.md
@@ -153,17 +153,62 @@ Skills are namespaced by source to avoid collisions: `anthropic-pdf`, `tob-stati
 
 ---
 
-## MCP Execution Backend (l4l)
+## MCP Execution Backend (l4l-oci)
 
 The [execute-work-package](skills/execute-work-package) skill supports three transport modes for delegating implementation to a sub-agent:
 
 | Transport | Mechanism | When to use |
 | --- | --- | --- |
-| **MCP via l4l** (default) | [l4l](https://github.com/flitzrrr/l4l) exposes 5 MCP tools for the gated precheck-approve-execute lifecycle | Any IDE with MCP support |
-| **Fresh Agent** | IDE spawns a new sub-agent per step (no external server) | Fallback when l4l is not configured |
+| **MCP via l4l-oci** (default) | [l4l-oci](https://github.com/DasDigitaleMomentum/l4l-oci) exposes 10 MCP tools for the gated blueprint-gate-execute lifecycle | Any IDE with MCP support |
+| **Fresh Agent** | IDE spawns a new sub-agent per step (no external server) | Fallback when l4l-oci is not configured |
 | **Stateful Session** | Sub-agent session resumed across steps | OpenCode only |
 
-When l4l MCP tools are available (`precheck_new`, `precheck_iterate`, `approve_blueprint`, `execute`, `handle_report`), the skill uses them automatically. See the [l4l setup guide](https://github.com/flitzrrr/l4l/blob/main/docs/CLAUDE_CODE_SETUP.md) for configuration.
+When l4l-oci MCP tools are available (`create_handle`, `generate_blueprint`, `submit_gate`, `execute_handle`, `get_digest`), the skill uses them automatically. The bundled `scripts/start-l4l-oci.sh` auto-starts the server if needed. See the [l4l-oci setup guide](https://github.com/DasDigitaleMomentum/l4l-oci/blob/main/docs/setup.md) for configuration.
+
+### Lifecycle
+
+```mermaid
+sequenceDiagram
+    participant User
+    participant Primary as Primary Agent<br/>(IDE)
+    participant MCP as l4l-oci MCP Server
+    participant Sub as Sub-LLM<br/>(configurable)
+
+    User->>Primary: Task with scope and DoD
+
+    rect rgb(240, 245, 250)
+    Note over Primary,Sub: Blueprint Phase
+    Primary->>MCP: create_handle(project_root)
+    MCP-->>Primary: handle_id
+    Primary->>MCP: generate_blueprint(handle_id, prompt)
+    MCP->>Sub: Generate Execution Blueprint
+    Sub-->>MCP: Blueprint (ordered steps)
+    Primary->>MCP: get_blueprint(handle_id)
+    MCP-->>Primary: Blueprint for review
+    end
+
+    Primary->>User: Present Blueprint for review
+
+    User->>Primary: Approve
+
+    rect rgb(240, 250, 240)
+    Note over Primary,Sub: Gate
+    Primary->>MCP: submit_gate(handle_id, "accept")
+    MCP-->>Primary: Gated
+    end
+
+    rect rgb(250, 245, 240)
+    Note over Primary,Sub: Execute Phase
+    Primary->>MCP: execute_handle(handle_id)
+    MCP->>Sub: Implement Blueprint steps
+    Sub->>Sub: Edit files, run verification
+    Sub-->>MCP: Execution Digest
+    Primary->>MCP: get_digest(handle_id)
+    MCP-->>Primary: Digest (outcome, files, verify result)
+    end
+
+    Primary->>User: Report results
+```
 
 ---
 

package/bin/build-catalog.js

@@ -155,7 +155,7 @@ function main() {
       const stats = fs.lstatSync(entryPath);
       if (stats.isSymbolicLink()) {
         const linkTarget = fs.readlinkSync(entryPath);
-        const normalized = linkTarget.replace(/^\.\.\/\.\.\//, "");
+        const normalized = linkTarget.replace(/^(\.\.\/)+/, "");
         realDir = path.join(ROOT, normalized);
       } else if (stats.isDirectory()) {
         realDir = entryPath;

package/bin/rebuild-symlinks.js

@@ -0,0 +1,161 @@
+#!/usr/bin/env node
+
+/**
+ * Rebuilds all skill symlinks from vendor/ submodules.
+ *
+ * Each SKILL.md found in vendor/<source>/ gets a symlink in skills/ with the
+ * naming convention: <prefix>-<skill_name>
+ *
+ * Prefix mapping is defined explicitly for known sources, with a fallback
+ * to the first segment of the vendor directory name.
+ *
+ * Run: node bin/rebuild-symlinks.js
+ */
+
+const fs = require("fs");
+const path = require("path");
+
+const ROOT = path.resolve(__dirname, "..");
+const SKILLS_DIR = path.join(ROOT, "skills");
+const VENDOR_DIR = path.join(ROOT, "vendor");
+
+// Explicit prefix mapping: vendor directory name → skill prefix
+const PREFIX_MAP = {
+  "trailofbits-skills": "tob",
+  "anthropic-skills": "anthropic",
+  "anthropic-finance": "finance",
+  "vercel-agent-skills": "vercel",
+  "getsentry-skills": "sentry",
+  "google-stitch-skills": "google",
+  "cloudflare-skills": "cloudflare",
+  "stripe-skills": "stripe",
+  "expo-skills": "expo",
+  "hashicorp-skills": "hashicorp",
+  "supabase-skills": "supabase",
+  "callstack-skills": "callstack",
+  "scientific-skills": "scientific",
+  "opencode-processing-skills": "opencode",
+  "agentic-seo-skill": "seo",
+  "marketingskills": null, // no prefix, use skill name directly
+  "addyosmani-agent-skills": "addyosmani",
+  "itsmostafa-aws-agent-skills": "aws",
+  "MoizIbnYousaf-Ai-Agent-Skills": "MoizIbnYousaf",
+  "JackyST0-awesome-agent-skills": "JackyST0",
+};
+
+function getPrefix(vendorName) {
+  if (vendorName in PREFIX_MAP) return PREFIX_MAP[vendorName];
+  // Fallback: first segment before first hyphen
+  return vendorName.split("-")[0];
+}
+
+function findSkillMds(dir, maxDepth, results) {
+  if (maxDepth <= 0) return;
+  let entries;
+  try {
+    entries = fs.readdirSync(dir, { withFileTypes: true });
+  } catch {
+    return;
+  }
+  for (const e of entries) {
+    if (e.name === ".git" || e.name === "node_modules") continue;
+    const full = path.join(dir, e.name);
+    if (e.isFile() && e.name.toLowerCase() === "skill.md") {
+      results.push(path.dirname(full));
+    } else if (e.isDirectory()) {
+      findSkillMds(full, maxDepth - 1, results);
+    }
+  }
+}
+
+function main() {
+  // Preserve manually-created skill directories (non-symlinks)
+  const manualSkills = new Set();
+  if (fs.existsSync(SKILLS_DIR)) {
+    for (const entry of fs.readdirSync(SKILLS_DIR)) {
+      const entryPath = path.join(SKILLS_DIR, entry);
+      const stat = fs.lstatSync(entryPath);
+      if (stat.isDirectory() && !stat.isSymbolicLink()) {
+        manualSkills.add(entry);
+      }
+    }
+  }
+
+  // Remove all existing symlinks
+  if (fs.existsSync(SKILLS_DIR)) {
+    for (const entry of fs.readdirSync(SKILLS_DIR)) {
+      const entryPath = path.join(SKILLS_DIR, entry);
+      if (fs.lstatSync(entryPath).isSymbolicLink()) {
+        fs.unlinkSync(entryPath);
+      }
+    }
+  } else {
+    fs.mkdirSync(SKILLS_DIR);
+  }
+
+  let created = 0;
+  let skipped = 0;
+  const vendors = fs.readdirSync(VENDOR_DIR).filter((d) => {
+    const p = path.join(VENDOR_DIR, d);
+    return !d.startsWith(".") && fs.statSync(p).isDirectory();
+  });
+
+  for (const vendor of vendors) {
+    const vendorPath = path.join(VENDOR_DIR, vendor);
+    const prefix = getPrefix(vendor);
+    const skillDirs = [];
+    findSkillMds(vendorPath, 5, skillDirs);
+
+    // Special case: seo skill — single top-level link
+    if (vendor === "agentic-seo-skill") {
+      const linkName = "seo";
+      const target = `../vendor/${vendor}`;
+      if (!manualSkills.has(linkName)) {
+        fs.symlinkSync(target, path.join(SKILLS_DIR, linkName));
+        created++;
+      }
+      continue;
+    }
+
+    for (const skillDir of skillDirs) {
+      const skillName = path.basename(skillDir);
+      const linkName = prefix ? `${prefix}-${skillName}` : skillName;
+
+      // Don't overwrite manual skills
+      if (manualSkills.has(linkName)) {
+        skipped++;
+        continue;
+      }
+
+      // Relative path from skills/ to the skill directory
+      const relPath = path.relative(SKILLS_DIR, skillDir);
+      const target = `../vendor/${path.relative(VENDOR_DIR, skillDir)}`;
+
+      try {
+        // Check for name collisions
+        const linkPath = path.join(SKILLS_DIR, linkName);
+        if (fs.existsSync(linkPath) || fs.lstatSync(linkPath).isSymbolicLink()) {
+          // Already exists, skip
+          skipped++;
+          continue;
+        }
+      } catch {
+        // lstatSync throws if path doesn't exist — that's what we want
+      }
+
+      try {
+        fs.symlinkSync(target, path.join(SKILLS_DIR, linkName));
+        created++;
+      } catch (err) {
+        // Name collision from a different vendor
+        skipped++;
+      }
+    }
+  }
+
+  console.log(`Rebuilt symlinks: ${created} created, ${skipped} skipped (manual or collision)`);
+  console.log(`Manual skill dirs preserved: ${[...manualSkills].join(", ") || "(none)"}`);
+  console.log(`Total skills/ entries: ${fs.readdirSync(SKILLS_DIR).length}`);
+}
+
+main();

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@flitzrrr/agent-skills",
-  "version": "1.1.0",
-  "description": "504 AI agent skills from 19 sources — multiplatform-ready for VS Code (GitHub Copilot), Claude Code, Codex, Cursor, Antigravity, OpenCode, Lovable, and more.",
+  "version": "1.2.0",
+  "description": "480 AI agent skills from 20 sources — multiplatform-ready for VS Code (GitHub Copilot), Claude Code, Codex, Cursor, Antigravity, OpenCode, Lovable, and more.",
   "scripts": {
     "test": "node bin/test-cli.js"
   },

package/skills/execute-work-package/SKILL.md

@@ -78,19 +78,40 @@ Do **not** use this skill to:
 
 ### Transport
 
-This skill supports three transport mechanisms. **If l4l MCP tools are available** (check: do you have `precheck_new`, `approve_blueprint`, `execute` tools?), **always use Option A.** Only fall back to B/C if MCP is not configured.
+This skill supports three transport mechanisms. **If l4l-oci MCP tools are available** (check: do you have `create_handle`, `generate_blueprint`, `submit_gate`, `execute_handle`, `get_digest` tools?), **always use Option A.** Only fall back to B/C if MCP is not configured.
 
-#### Option A: MCP via l4l (Default)
+#### Pre-flight: Ensure l4l-oci server is running
 
-Use the l4l Sub-Agent MCP server. The primary calls MCP tools directly — no Agent spawning needed:
+Before using Option A, run the bundled helper:
 
-1. `precheck_new(intent, scope_paths, output_format="skill")` → returns Execution Blueprint
-2. `approve_blueprint(handle_id)` → explicit gate
-3. `execute(handle_id, output_format="skill")` → returns Execution Digest
+```bash
+skills/execute-work-package/scripts/start-l4l-oci.sh
+```
+
+Behavior:
+- Uses `L4L_OCI_ROOT` (default: `$HOME/git/l4l-oci`)
+- Starts `python -m l4l_oci` in background if needed
+- Waits for MCP `health` to return `ok`
+- Exports `L4L_OCI_DEFAULT_MODEL=qwen-openrouter` for startup unless already set
+
+Prerequisites:
+- `OPENROUTER_API_KEY` (or `QWEN_API_KEY`) must be available in the environment
+
+#### Option A: MCP via l4l-oci (Default)
+
+Use the l4l-oci MCP server. The primary calls MCP tools directly — no Agent spawning needed:
+
+1. `create_handle(project_root)` → returns `handle_id`
+2. `generate_blueprint(handle_id, prompt)` → async; poll with `get_status(handle_id)` until complete
+3. `get_blueprint(handle_id, format="markdown")` → review the Execution Blueprint
+4. `submit_gate(handle_id, decision="accept"|"reject", notes=...)` → explicit gate
+5. `execute_handle(handle_id)` → async; poll with `get_status(handle_id)` until complete
+6. `get_digest(handle_id, format="markdown")` → returns Execution Digest
+7. `cleanup_handle(handle_id)` → optional cleanup
 
 Benefits: state persistence across restarts, model decoupling (cheap Sub-LLM), scope enforcement, iterative blueprint refinement.
 
-Setup: See l4l's `docs/CLAUDE_CODE_SETUP.md`.
+Additional tools: `health()` (service health check), `list_handles()` (list all active handles).
 
 #### Option B: Stateful Subagent (OpenCode)
 
@@ -147,7 +168,7 @@ Before delegating:
 
 ### 1) MODE: BLUEPRINT (Execution Blueprint)
 
-> **Option A (MCP):** Call `precheck_new(intent=..., scope_paths=..., output_format="skill")`. The blueprint is returned directly — no Agent spawn needed.
+> **Option A (MCP):** Call `create_handle(project_root)`, then `generate_blueprint(handle_id, prompt)`. Poll `get_status(handle_id)` until complete, then `get_blueprint(handle_id, format="markdown")` to review.
 
 Primary delegates to `implementer` with a prompt based on `tpl-implementer-preflight-prompt.md`.
 
@@ -167,7 +188,7 @@ If the user requests changes, the step list must be revised and re-approved with
 
 ### 2) Execute (new Agent)
 
-> **Option A (MCP):** Call `approve_blueprint(handle_id)` then `execute(handle_id, output_format="skill")`. The digest is returned directly — no Agent spawn needed.
+> **Option A (MCP):** Call `submit_gate(handle_id, decision="accept")` then `execute_handle(handle_id)`. Poll `get_status(handle_id)` until complete, then `get_digest(handle_id, format="markdown")` for the digest.
 
 Primary spawns a **new** `implementer` Agent with a prompt based on `tpl-implementer-execute-prompt.md`. The prompt includes the full approved step list, all references, and the verify command. Do NOT use SendMessage to resume the BLUEPRINT agent — spawn a fresh Agent instead.
 
@@ -179,7 +200,7 @@ The execute prompt MUST start with a clear mode indicator:
 
 and MUST include the approval token.
 
-> **Option A (MCP):** Skip Steps 2–3 above. The `execute` MCP tool returns the digest directly.
+> **Option A (MCP):** Skip the Agent spawn above. Call `submit_gate` → `execute_handle` → poll `get_status` → `get_digest` directly via MCP tools.
 
 ### 3) Digest back to Primary
 

package/skills/execute-work-package/scripts/start-l4l-oci.sh

@@ -0,0 +1,75 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+L4L_OCI_ROOT="${L4L_OCI_ROOT:-$HOME/git/l4l-oci}"
+L4L_OCI_URL="${L4L_OCI_URL:-http://127.0.0.1:8765/mcp}"
+L4L_OCI_DEFAULT_MODEL="${L4L_OCI_DEFAULT_MODEL:-qwen-openrouter}"
+L4L_OCI_LOG_FILE="${L4L_OCI_LOG_FILE:-$L4L_OCI_ROOT/logs/l4l-oci-service.log}"
+L4L_OCI_PID_FILE="${L4L_OCI_PID_FILE:-$L4L_OCI_ROOT/.l4l-oci/l4l-oci.pid}"
+
+PY="$L4L_OCI_ROOT/.venv/bin/python"
+if [ ! -x "$PY" ]; then
+  echo "ERROR: Python not found at $PY" >&2
+  exit 1
+fi
+
+if [ -z "${OPENROUTER_API_KEY:-}" ] && [ -z "${QWEN_API_KEY:-}" ]; then
+  echo "ERROR: Neither OPENROUTER_API_KEY nor QWEN_API_KEY is set." >&2
+  exit 1
+fi
+
+health_ok() {
+  "$PY" - <<'PY'
+import asyncio
+import os
+from fastmcp import Client
+
+url = os.environ["L4L_OCI_URL"]
+
+async def main() -> int:
+    try:
+        async with Client(url) as client:
+            res = await client.call_tool("health", {})
+        status = (res.data or {}).get("status")
+        return 0 if status == "ok" else 1
+    except Exception:
+        return 1
+
+raise SystemExit(asyncio.run(main()))
+PY
+}
+
+export L4L_OCI_URL
+
+if health_ok; then
+  echo "l4l-oci already reachable at $L4L_OCI_URL"
+  exit 0
+fi
+
+if [ -f "$L4L_OCI_PID_FILE" ]; then
+  OLD_PID="$(cat "$L4L_OCI_PID_FILE" 2>/dev/null || true)"
+  if [ -n "$OLD_PID" ] && ! kill -0 "$OLD_PID" 2>/dev/null; then
+    rm -f "$L4L_OCI_PID_FILE"
+  fi
+fi
+
+mkdir -p "$(dirname "$L4L_OCI_LOG_FILE")" "$(dirname "$L4L_OCI_PID_FILE")"
+
+(
+  cd "$L4L_OCI_ROOT"
+  export L4L_OCI_DEFAULT_MODEL
+  nohup "$PY" -m l4l_oci >>"$L4L_OCI_LOG_FILE" 2>&1 &
+  echo "$!" >"$L4L_OCI_PID_FILE"
+)
+
+for _ in $(seq 1 30); do
+  if health_ok; then
+    echo "l4l-oci started at $L4L_OCI_URL"
+    exit 0
+  fi
+  sleep 0.5
+done
+
+echo "ERROR: l4l-oci did not become healthy. Last log lines:" >&2
+tail -n 40 "$L4L_OCI_LOG_FILE" >&2 || true
+exit 1

package/skills/execute-work-package/tpl-implementer-execute-prompt.md

@@ -28,7 +28,7 @@ Execution invariants (must):
   - Concrete reason
   - What input is missing
 
-<!-- IF MCP: This prompt is used for Options B/C (Stateful/Fresh Agent). For Option A (MCP), call approve_blueprint() then execute() — this template is not needed. -->
+<!-- IF MCP: This prompt is used for Options B/C (Stateful/Fresh Agent). For Option A (MCP), run `skills/execute-work-package/scripts/start-l4l-oci.sh` first, then call `submit_gate()` and `execute_handle()`. -->
 
 ## Approved Step List
 {{approved_steps}}
@@ -43,7 +43,7 @@ Execution invariants (must):
 - Modules (optional): {{docs_modules_ref}}
 - Features (optional): {{docs_features_ref}}
 
-<!-- IF MCP: When using l4l MCP (Option A), the execute tool runs verification automatically. The output_format="skill" parameter returns a digest matching tpl-execution-digest.md format. -->
+<!-- IF MCP: When using l4l-oci MCP (Option A), poll `get_status()` and fetch digest via `get_digest()`. -->
 
 ## Verify Command
 {{verify_command}}

package/skills/execute-work-package/tpl-implementer-preflight-prompt.md

@@ -24,7 +24,7 @@ In BLUEPRINT mode:
 - Do NOT run commands.
 - Do NOT claim you changed code.
 
-<!-- IF MCP: This prompt is used for Options B/C (Stateful/Fresh Agent). For Option A (MCP), call precheck_new() directly — this template is not needed. -->
+<!-- IF MCP: This prompt is used for Options B/C (Stateful/Fresh Agent). For Option A (MCP), run `skills/execute-work-package/scripts/start-l4l-oci.sh` first, then call `create_handle()` and `generate_blueprint()`. -->
 
 ## Work Packet
 
@@ -53,7 +53,7 @@ In BLUEPRINT mode:
 
 If `{{verify_command}}` is empty/unknown, propose exactly **one** verify command.
 
-<!-- IF MCP: When using l4l MCP (Option A), the precheck_new tool accepts plan_ref, phase_ref, impl_plan_ref, and docs_refs parameters directly — no need to embed them in the prompt. -->
+<!-- IF MCP: When using l4l-oci MCP (Option A), pass planning/doc refs through the `generate_blueprint` prompt and retrieve results with `get_blueprint()`. -->
 
 ## Output