@polderlabs/bizar 3.20.0 → 3.20.8

package/cli/audit.mjs

@@ -80,7 +80,6 @@ export async function runAudit() {
         'opencode/deepseek-v4-flash-free',
         'minimax/MiniMax-M2.7',
         'minimax/MiniMax-M3',
-        'openai/gpt-5.5',
       ];
       if (!validModels.includes(model)) {
         warnings.push({ path, severity: 'WARN', msg: `Unknown model: ${model}` });

package/cli/bin.mjs

@@ -85,10 +85,12 @@ function showHelp() {
     update              Auto-update everything (opencode + bizar + dash + plugin)
     service             Manage the background service daemon
     bg <subcommand>     Manage background agents (list/view/kill/logs)
-    dash <subcommand>   Manage the dashboard (start/stop/status/tui)
+    dash <subcommand>   Manage the dashboard (start/stop/status/cleanup/tui)
+    browser-harness-up  Start Chromium for browser-harness (start/stop/status/restart)
     dev-link [src]      Symlink the local plugin source into opencode's plugin dir
     dev-unlink          Remove the dev symlink and restore the deployed copy
     doctor              Check the BizarHarness install for health issues
+    mod <subcommand>    Manage mods (install/upgrade/list via the dashboard API)
 
   Examples:
     bizar install
@@ -292,17 +294,168 @@ function showServiceHelp() {
   `);
 }
 
+function showModHelp() {
+  console.log(`
+  bizar mod — Manage mods (via the dashboard's HTTP API)
+
+  Usage:
+    bizar mod upgrade <id> [--backup]   Upgrade an installed mod to the latest version
+    bizar mod install <id>              Install a mod from the registry
+    bizar mod list                       List installed mods
+    bizar mod registry                  Show registry URL + available mods
+
+  Description:
+    Subcommands call the running dashboard's HTTP API. If no dashboard is
+    reachable, you'll be told to run \`bizar dash start\` first.
+
+    \`bizar mod upgrade <id>\` will:
+      1. Snapshot the existing version of the mod
+      2. Optionally back up the folder (--backup)
+      3. Remove the existing copy (and its opencode-config instructions)
+      4. Install the latest version from the registry
+      5. Re-install the new mod's instruction files into opencode config
+      6. Print from-version → to-version
+`);
+}
+
+async function runModCommand(modArgs) {
+  const sub = modArgs[0];
+  const positional = modArgs.slice(1).filter((a) => !a.startsWith('-'));
+  const flags = modArgs.slice(1).filter((a) => a.startsWith('-'));
+
+  if (!sub || sub === '--help' || sub === '-h' || flags.includes('--help') || flags.includes('-h')) {
+    showModHelp();
+    return;
+  }
+
+  // Read dashboard port from the port file the dashboard writes on start.
+  const { readFileSync: rfs } = await import('node:fs');
+  const { join: joinPath } = await import('node:path');
+  const portFile = joinPath(getBizarConfigDir(), 'dashboard.port');
+  let port = null;
+  try {
+    port = parseInt(rfs(portFile, 'utf8').trim(), 10);
+    if (!Number.isFinite(port) || port <= 0) port = null;
+  } catch {
+    port = null;
+  }
+  if (!port) {
+    console.error(chalk.red('  ✗ Dashboard is not running (no port file at ' + portFile + ').'));
+    console.error(chalk.dim('  Start it first: `bizar dash start --bg`'));
+    process.exit(1);
+  }
+  const baseUrl = `http://127.0.0.1:${port}`;
+
+  async function postJson(path, body) {
+    const res = await fetch(`${baseUrl}${path}`, {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify(body || {}),
+    });
+    const text = await res.text();
+    let json = null;
+    try { json = text ? JSON.parse(text) : null; } catch { /* keep raw */ }
+    if (!res.ok) {
+      const msg = json?.message || json?.error || text || `HTTP ${res.status}`;
+      throw new Error(`${path} failed: ${msg}`);
+    }
+    return json;
+  }
+  async function getJson(path) {
+    const res = await fetch(`${baseUrl}${path}`);
+    const text = await res.text();
+    let json = null;
+    try { json = text ? JSON.parse(text) : null; } catch { /* keep raw */ }
+    if (!res.ok) {
+      const msg = json?.message || json?.error || text || `HTTP ${res.status}`;
+      throw new Error(`${path} failed: ${msg}`);
+    }
+    return json;
+  }
+
+  if (sub === 'install') {
+    const id = positional[0];
+    if (!id) {
+      console.error(chalk.red('  ✗ Missing mod id. Usage: bizar mod install <id>'));
+      process.exit(1);
+    }
+    try {
+      const m = await postJson('/api/mods', { id });
+      console.log(chalk.green(`  ✓ Installed "${m.id}" v${m.version}`));
+    } catch (err) {
+      console.error(chalk.red(`  ✗ ${err.message}`));
+      process.exit(1);
+    }
+  } else if (sub === 'upgrade') {
+    const id = positional[0];
+    if (!id) {
+      console.error(chalk.red('  ✗ Missing mod id. Usage: bizar mod upgrade <id> [--backup]'));
+      process.exit(1);
+    }
+    const backup = flags.includes('--backup') || flags.includes('-b');
+    try {
+      const r = await postJson(`/api/mods/${encodeURIComponent(id)}/upgrade`, { backup });
+      const note = r.backupPath ? chalk.dim(` (backup: ${r.backupPath})`) : '';
+      console.log(chalk.green(`  ✓ Upgraded "${id}" v${r.from} → v${r.to}`) + note);
+    } catch (err) {
+      console.error(chalk.red(`  ✗ ${err.message}`));
+      process.exit(1);
+    }
+  } else if (sub === 'list') {
+    try {
+      const r = await getJson('/api/mods');
+      const mods = r.mods || [];
+      if (mods.length === 0) {
+        console.log(chalk.dim('  (no mods installed)'));
+        return;
+      }
+      for (const m of mods) {
+        const state = m.enabled ? chalk.green('enabled ') : chalk.yellow('disabled');
+        console.log(`  ${m.id.padEnd(20)} v${m.version.padEnd(10)} ${state}  ${m.name || ''}`);
+      }
+    } catch (err) {
+      console.error(chalk.red(`  ✗ ${err.message}`));
+      process.exit(1);
+    }
+  } else if (sub === 'registry') {
+    try {
+      const r = await getJson('/api/mods/registry');
+      console.log(chalk.dim(`  Source: ${r.registry?.source || '(unknown)'}`));
+      console.log(chalk.dim(`  Updated: ${r.registry?.updatedAt || '(unknown)'}`));
+      console.log('');
+      const mods = r.mods || [];
+      if (mods.length === 0) {
+        console.log(chalk.dim('  (no mods in registry)'));
+        return;
+      }
+      for (const m of mods) {
+        const installed = m.installed ? chalk.green(`installed v${m.installedVersion || '?'}`) : chalk.dim('not installed');
+        const upgrade = m.upgradeAvailable ? chalk.yellow(` ↑ v${m.upgradeAvailable} available`) : '';
+        console.log(`  ${m.id.padEnd(20)} v${(m.latest || '?').padEnd(10)} ${installed}${upgrade}  ${m.name || ''}`);
+      }
+    } catch (err) {
+      console.error(chalk.red(`  ✗ ${err.message}`));
+      process.exit(1);
+    }
+  } else {
+    console.error(chalk.red(`  ✗ Unknown mod subcommand: ${sub}`));
+    showModHelp();
+    process.exit(1);
+  }
+}
+
 function showDashHelp() {
   console.log(`
   bizar dash — Manage the Bizar dashboard
 
   Usage:
-    bizar dash <subcommand> [options]
+    bizar dash <subcommand> [options]　
 
   Subcommands:
     start [--bg] [--port N]   Start the dashboard (default port 4321)
     stop                       Stop the running dashboard
     status                     Show dashboard port and URL
+    cleanup                    Find + kill zombie/orphan dashboards
     tui [--no-web]             Launch the TUI
 
   Options:
@@ -511,10 +664,32 @@ async function main() {
     // v3.11.1 — Background agent manager (list / view / kill / logs).
     const { runBg } = await import('./bg.mjs');
     await runBg(args[1], args.slice(2));
+  } else if (args[0] === 'browser-harness-up') {
+    // v3.20.7 — Browser-harness daemon: start / stop / status / restart
+    // Chromium with remote debugging so the browser-harness Python tool
+    // (from https://github.com/browser-use/browser-harness) can connect.
+    const { execFileSync } = await import('node:child_process');
+    const sub = args[1] || 'start';
+    const scriptPath = join(import.meta.dirname || process.cwd(), 'browser-harness-up.sh');
+    try {
+      const out = execFileSync('bash', [scriptPath, sub], {
+        encoding: 'utf8',
+        stdio: 'inherit',
+      });
+      if (out) process.stdout.write(out);
+    } catch (err) {
+      console.error(chalk.red(`  ✗ browser-harness-up ${sub} failed (exit ${err.status ?? 1})`));
+      process.exit(err.status || 1);
+    }
   } else if (args[0] === 'providers' && args[1] === 'detect') {
     // v3.16.0 — Auto-detect provider API keys from env + opencode.json.
     const { runProvidersDetect } = await import('./providers-detect.mjs');
     await runProvidersDetect(args.slice(2));
+  } else if (args[0] === 'mod') {
+    // v3.20.5 — Mod manager. Talks to the running dashboard over HTTP
+    // (the dashboard owns the actual mod install/upgrade logic and the
+    // opencode-config install/uninstall of instruction files).
+    await runModCommand(args.slice(1));
   } else if (args[0] === 'dash' || args[0] === 'dashboard') {
     // `bizar dashboard` is a deprecated alias for `bizar dash`
     if (args[0] === 'dashboard') {
@@ -628,6 +803,11 @@ async function runDash(dashArgs) {
     case 'status':
       await dashModule.status(subOpts);
       break;
+    case 'cleanup':
+      // v3.20.7 — Find + kill zombie/orphan dashboards.
+      // Pass through --force and --kill flags to the dash module.
+      await dashModule.cleanup(subOpts);
+      break;
     case 'tui':
       await dashModule.tui(subOpts);
       break;

package/cli/browser-harness-up.sh

@@ -0,0 +1,193 @@
+#!/usr/bin/env bash
+#
+# cli/browser-harness-up.sh — start Chromium with remote debugging enabled
+# and register the browser-harness skill. Survives parent shell exit
+# via setsid + nohup. Idempotent: re-running is a no-op if already up.
+#
+# v3.20.7 — Required by the browser-harness Bizar agent for any
+# browser-driven E2E verification. The agent runs `browser-harness
+# skill` calls; this script makes sure Chrome is alive on CDP 9222
+# with the DevToolsActivePort file browser-harness auto-detects.
+#
+# Usage:
+#   cli/browser-harness-up.sh           # start if not running
+#   cli/browser-harness-up.sh status    # print status
+#   cli/browser-harness-up.sh stop      # kill chrome
+#   cli/browser-harness-up.sh restart   # stop + start
+#
+# Environment overrides:
+#   BH_CHROME_BIN  — path to a chrome-headless-shell / chrome binary
+#                     (default: chrome-headless-shell from puppeteer cache)
+#   BH_PORT        — CDP port (default: 9222)
+#   BH_PROFILE     — chrome user-data-dir (default: ~/.config/chromium)
+#
+# Exit codes:
+#   0   started / already running / status clean
+#   1   failed to start (missing binary, port busy, etc.)
+#
+set -euo pipefail
+
+BH_PORT="${BH_PORT:-9222}"
+BH_PROFILE="${BH_PROFILE:-$HOME/.config/chromium}"
+BH_DEVTOOLS_FILE="$BH_PROFILE/DevToolsActivePort"
+
+# Find the chrome binary. Prefer chrome-headless-shell from the puppeteer
+# cache (it's smaller, has no crashpad, and starts cleanly under daemon).
+# Fall back to the system chromium / chrome / google-chrome.
+find_chrome_bin() {
+  if [ -n "${BH_CHROME_BIN:-}" ] && [ -x "$BH_CHROME_BIN" ]; then
+    echo "$BH_CHROME_BIN"
+    return 0
+  fi
+  local puppeteer_bin
+  puppeteer_bin="$(find "$HOME/.cache/puppeteer" -name chrome-headless-shell -type f 2>/dev/null | head -1)"
+  if [ -n "$puppeteer_bin" ] && [ -x "$puppeteer_bin" ]; then
+    echo "$puppeteer_bin"
+    return 0
+  fi
+  for cand in chromium google-chrome chrome; do
+    local p
+    p="$(command -v "$cand" 2>/dev/null || true)"
+    if [ -n "$p" ] && [ -x "$p" ]; then
+      echo "$p"
+      return 0
+    fi
+  done
+  return 1
+}
+
+is_chrome_running() {
+  # The DevToolsActivePort file's format varies across Chrome versions:
+  #   - older:  PID\nPORT\nWS_PATH
+  #   - newer:  PORT\nWS_PATH
+  # We can't reliably extract a PID, so we check the TCP port directly.
+  # That's the only invariant that matters for browser-harness: it
+  # connects via CDP on this port.
+  if (echo >"/dev/tcp/127.0.0.1/$BH_PORT") 2>/dev/null; then
+    if [ -f "$BH_DEVTOOLS_FILE" ]; then
+      echo "port:$BH_PORT"
+      return 0
+    fi
+  fi
+  return 1
+}
+
+# Find the chrome PID by scanning /proc (used by `do_stop`). Linux only.
+find_chrome_pid() {
+  local pid
+  for pid in $(pgrep -f 'chrome.*--remote-debugging-port=' 2>/dev/null); do
+    # Confirm it's our chrome (matches our port)
+    local cmdline
+    cmdline="$(tr '\0' ' ' </proc/"$pid"/cmdline 2>/dev/null)"
+    if echo "$cmdline" | grep -q -- "--remote-debugging-port=$BH_PORT"; then
+      echo "$pid"
+      return 0
+    fi
+  done
+  return 1
+}
+
+do_status() {
+  if is_chrome_running >/dev/null; then
+    local pid
+    pid="$(find_chrome_pid 2>/dev/null || echo '?')"
+    echo "  ✓ Chromium is running (pid $pid, port $BH_PORT, profile $BH_PROFILE)"
+    if [ -f "$BH_DEVTOOLS_FILE" ]; then
+      echo "  ✓ DevToolsActivePort: $(head -1 "$BH_DEVTOOLS_FILE" 2>/dev/null)"
+    fi
+    if command -v browser-harness >/dev/null 2>&1; then
+      echo "  ✓ browser-harness: $(browser-harness --version 2>&1 | head -1)"
+      if [ -f "$HOME/.opencode/skills/browser-harness/SKILL.md" ]; then
+        echo "  ✓ Skill registered at ~/.opencode/skills/browser-harness/SKILL.md"
+      else
+        echo "  ! Skill not yet registered — run: browser-harness skill > ~/.opencode/skills/browser-harness/SKILL.md"
+      fi
+    else
+      echo "  ! browser-harness not installed — run: uv tool install --python 3.12 --upgrade --force browser-harness"
+    fi
+    return 0
+  else
+    echo "  ✗ Chromium is NOT running on port $BH_PORT (profile $BH_PROFILE)"
+    return 1
+  fi
+}
+
+do_start() {
+  if is_chrome_running >/dev/null; then
+    local pid
+    pid="$(find_chrome_pid 2>/dev/null || echo '?')"
+    echo "  ✓ Chromium already running (pid $pid, port $BH_PORT)"
+    return 0
+  fi
+
+  local bin
+  bin="$(find_chrome_bin)" || {
+    echo "  ✗ No chromium / chrome binary found." >&2
+    echo "    Set BH_CHROME_BIN or install one of: chromium, google-chrome, chrome," >&2
+    echo "    or place chrome-headless-shell in ~/.cache/puppeteer/." >&2
+    return 1
+  }
+
+  mkdir -p "$BH_PROFILE"
+  echo "  Starting $bin on port $BH_PORT (profile $BH_PROFILE)"
+
+  # setsid + nohup + redirect stdio so the daemon survives the parent
+  # shell exit (the Bizar background-process footgun — see AGENTS.md
+  # rule #25). The parent shell can exit cleanly while chrome keeps
+  # running on its own session.
+  setsid bash -c "nohup '$bin' \
+    --headless \
+    --disable-gpu \
+    --no-sandbox \
+    --disable-crashpad \
+    --remote-debugging-port=$BH_PORT \
+    --remote-debugging-address=127.0.0.1 \
+    --user-data-dir=$BH_PROFILE \
+    </dev/null >/tmp/chromium-bh.log 2>&1 & disown"
+
+  # Wait for the port to come up (up to 8s)
+  for _ in $(seq 1 40); do
+    if (echo >"/dev/tcp/127.0.0.1/$BH_PORT") 2>/dev/null; then
+      sleep 0.3  # give chrome time to write DevToolsActivePort
+      if is_chrome_running >/dev/null; then
+        local pid
+        pid="$(find_chrome_pid 2>/dev/null || echo '?')"
+        echo "  ✓ Chromium started (pid $pid, port $BH_PORT)"
+        return 0
+      fi
+    fi
+    sleep 0.2
+  done
+  echo "  ✗ Failed to start Chrome on port $BH_PORT (see /tmp/chromium-bh.log)" >&2
+  return 1
+}
+
+do_stop() {
+  local pid
+  if pid="$(find_chrome_pid 2>/dev/null)"; then
+    echo "  Stopping Chromium (pid $pid)"
+    kill "$pid" 2>/dev/null || true
+    sleep 1
+    if kill -0 "$pid" 2>/dev/null; then
+      kill -9 "$pid" 2>/dev/null || true
+    fi
+    rm -f "$BH_DEVTOOLS_FILE"
+    echo "  ✓ Stopped"
+  else
+    echo "  (Chromium was not running)"
+  fi
+}
+
+do_restart() {
+  do_stop
+  do_start
+}
+
+# Entry point
+case "${1:-start}" in
+  status)  do_status ;;
+  start)   do_start ;;
+  stop)    do_stop ;;
+  restart) do_restart ;;
+  *)       echo "Usage: $0 {start|stop|restart|status}" >&2; exit 1 ;;
+esac

package/cli/install.mjs

@@ -375,7 +375,7 @@ export async function runInstaller() {
   console.log(chalk.dim('  You can configure API keys now or later via /connect in opencode.'));
   const keys = await promptApiKeys();
 
-  if (keys.opencodeZen || keys.minimax || keys.openai |) {
+  if (keys.opencodeZen || keys.minimax || keys.openai) {
     console.log(chalk.dim('\n  Keys noted. Add them to your opencode.json or run /connect in opencode.\n'));
   }
 

package/cli/prompts.mjs

@@ -10,7 +10,7 @@ const AGENTS_LIST = [
   { name: 'Thor ᚦ  — MiniMax-M2.7 (Mid impl.)', value: 'thor.md', checked: true },
   { name: 'Baldr ᛒ  — MiniMax-M2.7 (Design)', value: 'baldr.md', checked: true },
   { name: 'Tyr ᛏ  — MiniMax-M3 (Complex impl.)', value: 'tyr.md', checked: true },
-  { name: 'Vidarr ᛉ — GPT-5.5 (Last resort)', value: 'vidarr.md', checked: false },
+  { name: 'Vidarr ᛉ — MiniMax-M3 (Last resort)', value: 'vidarr.md', checked: false },
   { name: 'Forseti ᚨ — MiniMax-M3 (Plan audit)', value: 'forseti.md', checked: true },
 ];
 
@@ -102,7 +102,7 @@ export async function promptApiKeys() {
     {
       type: 'password',
       name: 'openai',
-      message: 'OpenAI API key (GPT-5.5 — optional for Vidarr):',
+      message: 'OpenAI / opencode-zen API key (optional — for non-default agents):',
       mask: '*',
     },
     {

package/config/AGENTS.md

@@ -195,7 +195,7 @@ Odin (`@odin`) is the All-Father and primary/default agent. He analyzes each req
 
 ### Vidarr
 
-- **Model**: `openai/gpt-5.5` (via OpenAI ChatGPT subscription)
+- **Model**: `minimax/MiniMax-M3` (via MiniMax (direct), reasoning enabled)
 - **Use for**: The ultimate fallback — when Tyr stalls, debugging is stuck, or novel insight is needed
 - **Cost**: Highest — use very sparingly, last resort only
 
@@ -285,7 +285,7 @@ From the project root, the user (or heimdall via `/init` or any other agent prom
 
 ## General Agent Baseline — Always-On Behavior
 
-This section is the single source of truth for every Bizar agent's behavior. It is **adapted from the upstream system prompt and translated to Bizar**. Every Claude-specific reference has been mapped to the Bizar equivalent (BizarHarness, opencode, obsidian, Semble, Skills CLI, agent-browser, the opencode tool set). All agents **MUST** follow these rules at all times.
+This section is the single source of truth for every Bizar agent's behavior. It is **adapted from the upstream system prompt and translated to Bizar**. Every Claude-specific reference has been mapped to the Bizar equivalent (BizarHarness, opencode, obsidian, Semble, Skills CLI, browser-harness, the opencode tool set). All agents **MUST** follow these rules at all times.
 
 > **Tool name translation table** (used throughout this baseline):
 >
@@ -424,7 +424,17 @@ Load the SKILL.md via the `skill` tool before writing code or making changes cov
 
 #### Browser interaction
 
-For browser-driven E2E validation, use **agent-browser** (the `agent_browser_*` tools). Common operations: `open`, `snapshot`, `click`, `type`, `fill`, `press`, `screenshot`, `eval`, `wait_for_*`. Do **not** install headless Chrome via raw shell commands when agent-browser is available.
+For browser-driven E2E validation, use **browser-harness** (the Python tool from https://github.com/browser-use/browser-harness). Invoke via `bash` heredoc:
+
+```bash
+browser-harness <<'PY'
+new_tab("https://example.com")
+wait_for_load()
+print(page_info())
+PY
+```
+
+Common operations: `new_tab`, `goto_url`, `wait_for_load`, `page_info`, `click_at_xy`, `fill_input`, `press_key`, `scroll`, `capture_screenshot`, `js`, `cdp("Domain.method", ...)`. Read the SKILL.md at `~/.opencode/skills/browser-harness/SKILL.md` on first use. Do **not** install headless Chrome via raw shell commands when browser-harness is available.
 
 ### skills_mandatory_read
 
@@ -433,7 +443,7 @@ Before writing any code, creating any file, or running any computer tool, **scan
 Concrete triggers:
 - Frontend/React work → `frontend-design` or framework-specific skill
 - Backend/API work → framework-specific skill
-- Browser E2E → `agent-browser` SKILL.md
+- Browser E2E → `browser-harness` SKILL.md
 - Skill creation → `skill-creator` SKILL.md
 - BizarHarness-specific work → `~/.opencode/skills/bizar/` SKILL.md (always)
 - Self-improvement logging → `~/.opencode/skills/self-improvement/` SKILL.md (always)
@@ -516,7 +526,7 @@ For Bizar-internal claims (citing files, lines, tool results), use file:line ref
 ### images_and_visual_content
 
 - Bizar does not have an `image_search` tool. Do not assume one exists.
-- For local screenshots and image inspection, use `agent-browser` (`agent_browser_screenshot` + `agent_browser_eval`).
+- For local screenshots and image inspection, use `browser-harness` (`capture_screenshot(path="...")` + `js(...)` inside a `<<'PY' ... PY` heredoc).
 - For image generation, dispatch to `@baldr` (design) or use a user-supplied image-generation MCP server if connected.
 - Never claim to inspect or edit an image that isn't actually available.
 

package/config/agents/_shared/AGENT_BASELINE.md

@@ -5,7 +5,7 @@ description: Always-on rules for every Bizar agent. Loaded automatically by open
 
 # Agent Baseline — Always-On Rules
 
-Every Bizar agent follows these rules at all times. They are translated from the upstream Claude Fable 5 system prompt, with every Claude-specific tool / function / directory mapped to the BizarHarness equivalent (opencode tools, Semble, Skills CLI, Obsidian vault, agent-browser, dashboard artifact pipeline).
+Every Bizar agent follows these rules at all times. They are translated from the upstream Claude Fable 5 system prompt, with every Claude-specific tool / function / directory mapped to the BizarHarness equivalent (opencode tools, Semble, Skills CLI, Obsidian vault, browser-harness, dashboard artifact pipeline).
 
 ---
 
@@ -384,7 +384,7 @@ The following rules apply to every agent at all times. They are the single sourc
 
 ### Knowledge Cutoff and Research-First
 
-- Bizar does not have a single knowledge cutoff shared by all models. Subagents may run on DeepSeek V4 Flash, MiniMax M2.7 / M3, or GPT-5.5, each with their own training window.
+- Bizar does not have a single knowledge cutoff shared by all models. Subagents may run on DeepSeek V4 Flash (opencode-zen, free tier) or MiniMax M2.7 / M3, each with their own training window.
 - For facts that change quickly (current positions, prices, breaking news) or anything that could have changed recently, **search before answering**: use `websearch` and `webfetch` or delegate to `@mimir` for deep research.
 - For stable technical knowledge (language semantics, well-established APIs, mathematical truths), answer directly without search.
 - Default to reading `.obsidian/INDEX.md` at session start to retrieve prior project context before answering anything project-specific.
@@ -402,7 +402,7 @@ Bizar can connect to external tools via MCP servers. Always check what's connect
 
 **Domain skills:** see section 3 above.
 
-**Browser interaction:** for browser-driven E2E validation, use **agent-browser** (the `agent_browser_*` tools). Do **not** install headless Chrome via raw shell commands when agent-browser is available.
+**Browser interaction:** for browser-driven E2E validation, use **browser-harness** (the Python tool from https://github.com/browser-use/browser-harness). It exposes raw CDP via a `bash` heredoc: `browser-harness <<'PY' ... PY`. Do **not** install headless Chrome via raw shell commands when browser-harness is available.
 
 ### Mandatory Skill Read
 
@@ -412,7 +412,7 @@ Concrete triggers:
 
 - Frontend/React work → `frontend-design` or framework-specific skill
 - Backend/API work → framework-specific skill
-- Browser E2E → `agent-browser` SKILL.md
+- Browser E2E → `browser-harness` SKILL.md
 - Skill creation → `skill-creator` SKILL.md
 - BizarHarness-specific work → `~/.opencode/skills/bizar/SKILL.md` (always)
 - Self-improvement logging → `~/.opencode/skills/self-improvement/SKILL.md` (always)
@@ -494,7 +494,7 @@ For Bizar-internal claims (citing files, lines, tool results), use `file:line` r
 ### Images and Visual Content
 
 - Bizar does not have an `image_search` tool. Do not assume one exists.
-- For local screenshots and image inspection, use `agent-browser` (`agent_browser_screenshot` + `agent_browser_eval`).
+- For local screenshots and image inspection, use `browser-harness` (`capture_screenshot(path="...")` + `js(...)` inside a `<<'PY' ... PY` heredoc).
 - For image generation, dispatch to `@baldr` (design) or use a user-supplied image-generation MCP server if connected.
 - Never claim to inspect or edit an image that isn't actually available.
 

package/config/agents/browser-harness.md

@@ -1,7 +1,7 @@
 ---
 description: browser-harness — Primary agent for browser-driven E2E verification. No-edit permissions. Drives Chromium via CDP for end-to-end testing of web apps.
 mode: primary
-model: minimax/MiniMax-M2.7
+model: openrouter/minimax/minimax-m2.7
 color: "#84cc16"
 permission:
   read: allow
@@ -26,8 +26,19 @@ You are browser-harness — the silent observer. You drive a real browser via CD
 
 ## Tools Available
 
-- `agent_browser_*` tools (`open`, `snapshot`, `click`, `type`, `fill`, `press`, `screenshot`, `eval`, `wait_for_*`)
-- read, glob, grep
+- **Primary: `browser-harness` (Python via uv)** — installed at `~/.local/bin/browser-harness` from https://github.com/browser-use/browser-harness. Use heredoc syntax:
+  ```bash
+  browser-harness <<'PY'
+  new_tab("https://example.com")
+  wait_for_load()
+  print(page_info())
+  PY
+  ```
+  Pre-imported helpers: `new_tab`, `goto_url`, `page_info`, `click_at_xy`, `type_text`, `fill_input`, `press_key`, `scroll`, `capture_screenshot`, `list_tabs`, `current_tab`, `switch_tab`, `js`, `cdp`, `wait_for_load`, `ensure_real_tab`, `ensure_daemon`. The daemon auto-starts and attaches to Chrome at `~/.config/chromium/` on CDP port 9222.
+
+- Setup: if Chrome is not running, `cli/browser-harness-up.sh start` (or `bizar browser-harness-up start`).
+- The browser-harness SKILL.md lives at `~/.opencode/skills/browser-harness/SKILL.md` — read it on first use.
+- Read, glob, grep
 - bash for `npx bizar dev` to start the dev server, `curl` for health checks
 - webfetch, websearch
 - edit/write **denied** — you cannot modify the project
@@ -35,12 +46,13 @@ You are browser-harness — the silent observer. You drive a real browser via CD
 ## Workflow
 
 1. **Start the app if needed.** `npx bizar dev` or the project's dev command. Wait for the port to be ready.
-2. **Open the URL.** `agent_browser_open <url>`.
-3. **Take a snapshot.** `agent_browser_snapshot` to see the DOM.
-4. **Interact.** `agent_browser_click`, `agent_browser_fill`, `agent_browser_press` — use the accessibility tree, not pixel coordinates.
-5. **Capture state.** `agent_browser_screenshot` for visual evidence.
-6. **Evaluate.** `agent_browser_eval` to run JS in the page context.
-7. **Report.** What you did, what you saw, what passed, what failed.
+2. **Open the URL.** `new_tab("https://...")` inside a `browser-harness <<'PY' ... PY` heredoc, or `goto_url(...)` if a tab is already open.
+3. **Wait for load.** `wait_for_load()` after every navigation.
+4. **Inspect.** `js("document.querySelector(...)...")` for DOM extraction.
+5. **Interact.** `click_at_xy(x, y)` after a screenshot — the helpers are pre-imported.
+6. **Capture state.** `capture_screenshot(path="/tmp/...png")` for visual evidence.
+7. **Evaluate.** `js(...)` to run arbitrary JS in the page context.
+8. **Report.** What you did, what you saw, what passed, what failed.
 
 ## Output Style
 

package/config/agents/vidarr.md

@@ -1,7 +1,7 @@
 ---
-description: Vidarr — The ultimate fallback using GPT-5.5. For the hardest problems when Tyr stalls, debugging is stuck, or novel insight is needed. Use sparingly — highest cost.
+description: Vidarr — The ultimate fallback via MiniMax M3. For the hardest problems when Tyr stalls, debugging is stuck, or novel insight is needed. Use sparingly — highest cost.
 mode: subagent
-model: openai/gpt-5.5
+model: minimax/MiniMax-M3
 color: "#0ea5e9"
 permission:
   read: allow

package/config/opencode.json.template

@@ -62,7 +62,7 @@
       "permission": {
         "read": "allow",
         "list": "allow",
-        "question": "allow",,
+        "question": "allow"
       }
     },
     "frigg": {
@@ -77,7 +77,7 @@
         "grep": "allow",
         "bash": "allow",
         "webfetch": "allow",
-        "websearch": "allow",,,
+        "websearch": "allow",
         "question": "allow"
       }
     },
@@ -203,9 +203,9 @@
       }
     },
     "vidarr": {
-      "description": "Vidarr — Ultimate fallback via GPT-5.5.",
+      "description": "Vidarr — Ultimate fallback via MiniMax M3.",
       "mode": "subagent",
-      "model": "openai/gpt-5.5",
+      "model": "minimax/MiniMax-M3",
       "color": "#dc2626",
       "permission": {
         "read": "allow",
@@ -237,6 +237,23 @@
         "websearch": "allow"
       }
     },
+    "browser-harness": {
+      "description": "browser-harness — Primary agent for browser-driven E2E verification. No-edit permissions. Drives Chrome via CDP for end-to-end testing of web apps.",
+      "mode": "primary",
+      "model": "minimax/MiniMax-M2.7",
+      "color": "#84cc16",
+      "permission": {
+        "read": "allow",
+        "bash": "allow",
+        "glob": "allow",
+        "grep": "allow",
+        "list": "allow",
+        "webfetch": "allow",
+        "websearch": "allow",
+        "edit": "deny",
+        "write": "deny"
+      }
+    },
     "semble-search": {
       "description": "Code search agent for exploring any codebase.",
       "mode": "subagent",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@polderlabs/bizar",
-  "version": "3.20.0",
+  "version": "3.20.8",
   "description": "Norse-pantheon multi-agent system for opencode — 13 agents across 4 cost tiers with cost-aware routing, plans, and a configurable agent harness.",
   "type": "module",
   "bin": {
@@ -19,7 +19,7 @@
     "build:sdk": "npm run build -w @polderlabs/bizar-sdk",
     "test:sdk": "npm run test -w @polderlabs/bizar-sdk",
     "test:sdk:watch": "npm run test:watch -w @polderlabs/bizar-sdk",
-    "test": "npm run typecheck && npm run test:sdk && (cd plugins/bizar && bun test tests/loop.test.ts tests/block.test.ts tests/stall-think.test.ts tests/tools/bg-get-comments.test.ts tests/tools/bg-spawn-delegation.test.ts tests/tools/opencode-runner.test.ts tests/settings.test.ts tests/commands.test.ts tests/commands-impl.test.ts tests/tools/plan-action.test.ts tests/tools/wait-for-feedback.test.ts tests/reasoning-clean.test.ts tests/key-rotation.test.ts) && node bizar-dash/tests/smoke-v2.mjs && node --test bizar-dash/tests/path-safe.test.mjs bizar-dash/tests/tmux-wrap.test.mjs bizar-dash/tests/opencode-runner.test.mjs",
+    "test": "npm run typecheck && npm run test:sdk && (cd plugins/bizar && bun test tests/loop.test.ts tests/block.test.ts tests/stall-think.test.ts tests/tools/bg-get-comments.test.ts tests/tools/bg-spawn-delegation.test.ts tests/tools/opencode-runner.test.ts tests/settings.test.ts tests/commands.test.ts tests/commands-impl.test.ts tests/tools/plan-action.test.ts tests/tools/wait-for-feedback.test.ts tests/reasoning-clean.test.ts tests/key-rotation.test.ts) && node bizar-dash/tests/smoke-v2.mjs && node --test bizar-dash/tests/path-safe.test.mjs bizar-dash/tests/tmux-wrap.test.mjs bizar-dash/tests/opencode-runner.test.mjs bizar-dash/tests/mod-instructions.node.test.mjs bizar-dash/tests/mod-upgrade.node.test.mjs bizar-dash/tests/graphify-mod-spawn.node.test.mjs bizar-dash/tests/no-agent-browser.node.test.mjs",
     "build": "npm run build:sdk"
   },
   "keywords": [