Phase 12: Universal agent runtime + LLM-driven tool selection + live Vercel dashboard

Replaced hardcoded shell scripts with a universal LLM-driven agent runtime.
The contract defines WHAT to do — the LLM agent decides HOW, including
which tools to use. One runtime script for all stages.

Universal agent-runtime.sh:
- Reads /context/brief.md (contract instructions)
- Sends to LLM: "plan your approach, here are your tools"
- LLM returns JSON array of shell commands
- Executes step by step with self-repair on failure
- No hardcoded tool usage — LLM decides everything

Validated end-to-end: 3-stage oil & gas market dashboard pipeline
- Stage 1 (researcher): LLM generated market data, news, analysis
- Stage 2 (developer): LLM wrote Python script to generate Chart.js dashboard
- Stage 3 (deployer): deployed to Vercel

Live dashboard: https://app-theta-umber.vercel.app

Changes:
- Created sandbox/scripts/agent-runtime.sh (universal LLM-driven runtime)
- Deleted hardcoded scripts (researcher.sh, writer.sh, publisher.sh, developer.sh, deployer.sh)
- Pipeline always uses agent-runtime.sh — LLM plans tool usage from contract
- ContextBridge tags with contract:{name}, stage:{stage} for scoped memory
- Pipeline.build_env injects vercel_token from env when contract declares it
- market-dashboard.yaml updated with vercel_token credential

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: MagnetonIO

sandbox/scripts/agent-runtime.sh

@@ -0,0 +1,222 @@
+#!/bin/sh
+# Universal Agent Runtime — LLM-driven plan+execute+fix loop
+#
+# This is the ONLY script that runs inside microVMs. It reads the contract
+# instructions from /context/brief.md, asks the LLM to plan which tools
+# to use, executes the plan step by step, and self-repairs on failure.
+#
+# The contract defines WHAT. The LLM decides HOW.
+set -e
+
+# Install base tools (Alpine, ~3 seconds)
+apk add --no-cache -q curl jq git bash nodejs npm python3 2>/dev/null || true
+
+MAX_ITERATIONS=10
+LLM_PROXY="http://localhost:4000/api/v1/vm/llm/chat"
+
+echo "=== Agent Runtime Starting ==="
+
+# 1. Read the brief (contract instructions for this stage)
+BRIEF=""
+if [ -f /context/brief.md ]; then
+  BRIEF=$(cat /context/brief.md)
+else
+  echo "ERROR: No /context/brief.md found"
+  exit 1
+fi
+
+# 2. Discover available context files
+CONTEXT_FILES=$(ls -1 /context/ 2>/dev/null | grep -v brief.md | head -20)
+CONTEXT_CONTENTS=""
+for f in $CONTEXT_FILES; do
+  if [ -f "/context/$f" ]; then
+    SIZE=$(wc -c < "/context/$f" | tr -d ' ')
+    if [ "$SIZE" -lt 8000 ]; then
+      CONTEXT_CONTENTS="$CONTEXT_CONTENTS
+--- /context/$f ---
+$(cat /context/$f)"
+    else
+      CONTEXT_CONTENTS="$CONTEXT_CONTENTS
+--- /context/$f (${SIZE} bytes, showing first 2000) ---
+$(head -c 2000 /context/$f)"
+    fi
+  fi
+done
+
+# 3. Discover available tools
+TOOLS="curl, jq, git, bash, node, npm, python3"
+if command -v vercel > /dev/null 2>&1; then TOOLS="$TOOLS, vercel"; fi
+if command -v pdflatex > /dev/null 2>&1; then TOOLS="$TOOLS, pdflatex"; fi
+
+# Check for injected credentials
+CREDS=""
+if [ -n "$GH_TOKEN" ]; then
+  CREDS="$CREDS GitHub (GH_TOKEN set),"
+  git config --global user.email "agent-os@agenthero.work"
+  git config --global user.name "Agent-OS Pipeline"
+  git config --global init.defaultBranch main
+  git config --global url."https://x-access-token:${GH_TOKEN}@github.com/".insteadOf "https://github.com/"
+fi
+if [ -n "$VERCEL_TOKEN" ]; then
+  CREDS="$CREDS Vercel (VERCEL_TOKEN set),"
+  npm install -g vercel 2>/dev/null || true
+fi
+
+echo "Tools: $TOOLS"
+echo "Credentials: ${CREDS:-none}"
+echo "Context files: $(echo $CONTEXT_FILES | tr '\n' ', ')"
+
+# 4. Ask LLM to plan the approach
+echo ""
+echo "=== Planning ==="
+
+PLAN_PROMPT="You are an autonomous agent running inside an isolated Linux container.
+
+YOUR TASK:
+$BRIEF
+
+AVAILABLE TOOLS: $TOOLS
+AVAILABLE CREDENTIALS: ${CREDS:-none}
+OUTPUT DIRECTORY: /shared/output/ (write ALL output files here)
+
+CONTEXT FILES AVAILABLE:
+$CONTEXT_CONTENTS
+
+CRITICAL RULES:
+- You have NO access to external APIs or websites. Do NOT use curl to fetch data from APIs.
+- Generate all content from YOUR OWN KNOWLEDGE. You are an expert — write the data yourself.
+- For JSON data files, write them directly using cat/heredoc or a python/node script.
+- For markdown files, write them directly using cat/heredoc.
+- For HTML/CSS/JS dashboards, write a python script that generates the complete file.
+- Use real, specific, accurate data from your training knowledge.
+- If credentials are available (GH_TOKEN, VERCEL_TOKEN), use git/vercel CLI tools for deployment.
+
+INSTRUCTIONS:
+1. Plan your approach step by step
+2. For each step, provide the EXACT shell command to run
+3. Use the context files to inform your work
+4. Write all output to /shared/output/
+5. If you need to create directories, use mkdir -p
+6. For large file generation, write a python3 script to /tmp/gen.py then run it
+
+OUTPUT FORMAT — respond with ONLY a JSON array of steps:
+[
+  {\"description\": \"what this step does\", \"command\": \"the exact shell command\"},
+  {\"description\": \"next step\", \"command\": \"next command\"},
+  ...
+]
+
+IMPORTANT:
+- Output ONLY the JSON array, no markdown fences, no explanation
+- Each command must be a single shell command (use && to chain if needed)
+- For multi-line file writes, use heredoc: cat > /shared/output/file.md << 'HEREDOC' ... HEREDOC
+- For complex file generation, write a python3 or node script to /tmp/ first, then execute it
+- NEVER try to curl/fetch external URLs for data — generate everything from your knowledge"
+
+PLAN_RESPONSE=$(curl -s -X POST "$LLM_PROXY" \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer $JOB_TOKEN" \
+  -d "$(jq -n \
+    --arg prompt "$PLAN_PROMPT" \
+    '{messages: [{role: "user", content: $prompt}], model: "gpt-4o", max_tokens: 8192, temperature: 0.2}')" 2>&1)
+
+# Extract content from LLM response
+PLAN=""
+if echo "$PLAN_RESPONSE" | jq -e '.content' > /dev/null 2>&1; then
+  PLAN=$(echo "$PLAN_RESPONSE" | jq -r '.content')
+else
+  echo "ERROR: LLM proxy not available or returned invalid response"
+  echo "Response: $(echo "$PLAN_RESPONSE" | head -c 500)"
+  exit 1
+fi
+
+# Strip markdown code fences if present
+PLAN=$(echo "$PLAN" | sed 's/^```json//' | sed 's/^```//' | sed 's/```$//')
+
+# Validate JSON
+if ! echo "$PLAN" | jq -e '.' > /dev/null 2>&1; then
+  echo "ERROR: LLM returned invalid JSON plan"
+  echo "Plan: $(echo "$PLAN" | head -c 500)"
+  exit 1
+fi
+
+STEP_COUNT=$(echo "$PLAN" | jq 'length')
+echo "Plan: $STEP_COUNT steps"
+
+# 5. Execute plan step by step
+echo ""
+echo "=== Executing ==="
+
+ITERATION=0
+STEP_INDEX=0
+
+while [ "$STEP_INDEX" -lt "$STEP_COUNT" ] && [ "$ITERATION" -lt "$MAX_ITERATIONS" ]; do
+  ITERATION=$((ITERATION + 1))
+  DESC=$(echo "$PLAN" | jq -r ".[$STEP_INDEX].description")
+  CMD=$(echo "$PLAN" | jq -r ".[$STEP_INDEX].command")
+
+  echo ""
+  echo "--- Step $((STEP_INDEX + 1))/$STEP_COUNT: $DESC ---"
+  echo "\$ $CMD"
+
+  # Execute the command
+  OUTPUT=$(eval "$CMD" 2>&1) || {
+    EXIT_CODE=$?
+    echo "FAILED (exit $EXIT_CODE): $OUTPUT"
+
+    # Self-repair: ask LLM to fix
+    echo "Asking LLM to fix..."
+
+    FIX_PROMPT="A command failed during autonomous execution.
+
+Step description: $DESC
+Command: $CMD
+Error output: $OUTPUT
+Exit code: $EXIT_CODE
+
+Available tools: $TOOLS
+Working directory: $(pwd)
+
+Provide a FIXED command that accomplishes the same goal. Output ONLY the shell command, nothing else."
+
+    FIX_RESPONSE=$(curl -s -X POST "$LLM_PROXY" \
+      -H "Content-Type: application/json" \
+      -H "Authorization: Bearer $JOB_TOKEN" \
+      -d "$(jq -n \
+        --arg prompt "$FIX_PROMPT" \
+        '{messages: [{role: "user", content: $prompt}], model: "gpt-4o", max_tokens: 2048, temperature: 0.1}')" 2>&1)
+
+    if echo "$FIX_RESPONSE" | jq -e '.content' > /dev/null 2>&1; then
+      FIXED_CMD=$(echo "$FIX_RESPONSE" | jq -r '.content' | sed 's/^```sh//' | sed 's/^```bash//' | sed 's/^```//' | sed 's/```$//' | tr -d '\n')
+      echo "Fix: $FIXED_CMD"
+      eval "$FIXED_CMD" 2>&1 || echo "Fix also failed — continuing"
+    else
+      echo "Could not get fix from LLM — continuing"
+    fi
+  }
+
+  # Show truncated output
+  if [ -n "$OUTPUT" ]; then
+    echo "$(echo "$OUTPUT" | head -5)"
+    LINES=$(echo "$OUTPUT" | wc -l | tr -d ' ')
+    if [ "$LINES" -gt 5 ]; then echo "... ($LINES lines total)"; fi
+  fi
+
+  STEP_INDEX=$((STEP_INDEX + 1))
+done
+
+# 6. Verify output
+echo ""
+echo "=== Output Verification ==="
+if [ -d /shared/output ]; then
+  echo "Files in /shared/output/:"
+  find /shared/output -type f | while read f; do
+    SIZE=$(wc -c < "$f" | tr -d ' ')
+    echo "  $f ($SIZE bytes)"
+  done
+else
+  echo "WARNING: /shared/output/ is empty"
+fi
+
+echo ""
+echo "=== Agent Runtime Complete ($ITERATION iterations) ==="