Run Molmo live four-route agent matrix

.github/workflows/ci.yml

@@ -79,10 +79,10 @@ jobs:
       max-parallel: 3
       matrix:
         model_entry:
-          - kimi-k2.6
-          - mimo-v2.5
-          - kimi-k2.6-camera-raw
-          - mimo-v2.5-camera-raw
+          - claude-code-mimo-v2.5
+          - claude-code-kimi-k2.6
+          - agents-sdk-mimo-v2.5
+          - agents-sdk-kimi-k2.7-code
 
     steps:
       - uses: actions/checkout@v4
@@ -211,33 +211,33 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      - name: Download Molmo live Kimi artifact
+      - name: Download Molmo live Claude Code MiMo artifact
         continue-on-error: true
         uses: actions/download-artifact@v4
         with:
-          name: report-molmo-live-kimi-k2.6
-          path: molmo-live-src/kimi-k2.6
+          name: report-molmo-live-claude-code-mimo-v2.5
+          path: molmo-live-src/claude-code-mimo-v2.5
 
-      - name: Download Molmo live MiMo v2.5 artifact
+      - name: Download Molmo live Claude Code Kimi artifact
         continue-on-error: true
         uses: actions/download-artifact@v4
         with:
-          name: report-molmo-live-mimo-v2.5
-          path: molmo-live-src/mimo-v2.5
+          name: report-molmo-live-claude-code-kimi-k2.6
+          path: molmo-live-src/claude-code-kimi-k2.6
 
-      - name: Download Molmo live Kimi RAW_FPV artifact
+      - name: Download Molmo live Agents SDK MiMo artifact
         continue-on-error: true
         uses: actions/download-artifact@v4
         with:
-          name: report-molmo-live-kimi-k2.6-camera-raw
-          path: molmo-live-src/kimi-k2.6-camera-raw
+          name: report-molmo-live-agents-sdk-mimo-v2.5
+          path: molmo-live-src/agents-sdk-mimo-v2.5
 
-      - name: Download Molmo live MiMo v2.5 RAW_FPV artifact
+      - name: Download Molmo live Agents SDK Kimi artifact
         continue-on-error: true
         uses: actions/download-artifact@v4
         with:
-          name: report-molmo-live-mimo-v2.5-camera-raw
-          path: molmo-live-src/mimo-v2.5-camera-raw
+          name: report-molmo-live-agents-sdk-kimi-k2.7-code
+          path: molmo-live-src/agents-sdk-kimi-k2.7-code
 
       - name: Assemble Pages site
         run: |

just/molmo.just

@@ -584,8 +584,8 @@ household-world-impl driver="mcp-smoke" profile="smoke" seeds="7" output_dir="ou
           roboclaws_codex_provider_args codex_model_args
         elif [[ "$driver" == "openai-agents-live" ]]; then
           openai_agents_profile="$(roboclaws_code_agent_provider ROBOCLAWS_PROVIDER_PROFILE codex-router-responses)"
-          openai_agents_model="$(roboclaws_code_agent_model ROBOCLAWS_CODEX_MODEL ROBOCLAWS_PROVIDER_PROFILE codex-router-responses)"
-          openai_agents_profile_summary="$(roboclaws_code_agent_profile_summary ROBOCLAWS_PROVIDER_PROFILE ROBOCLAWS_CODEX_MODEL codex-router-responses)"
+          openai_agents_model="$(roboclaws_code_agent_model ROBOCLAWS_OPENAI_AGENTS_MODEL ROBOCLAWS_PROVIDER_PROFILE codex-router-responses)"
+          openai_agents_profile_summary="$(roboclaws_code_agent_profile_summary ROBOCLAWS_PROVIDER_PROFILE ROBOCLAWS_OPENAI_AGENTS_MODEL codex-router-responses)"
         elif [[ "$driver" == "claude-live" ]]; then
           claude_profile_summary="$(roboclaws_code_agent_profile_summary ROBOCLAWS_PROVIDER_PROFILE ROBOCLAWS_CLAUDE_MODEL auto-claude)"
           roboclaws_claude_provider_args claude_model_args claude_env_args
@@ -759,7 +759,7 @@ household-world-impl driver="mcp-smoke" profile="smoke" seeds="7" output_dir="ou
             HTTP_PROXY HTTPS_PROXY ALL_PROXY NO_PROXY \
             http_proxy https_proxy all_proxy no_proxy \
             DISPLAY XAUTHORITY MUJOCO_GL PYOPENGL_PLATFORM LIBGL_ALWAYS_SOFTWARE \
-            ROBOCLAWS_PROVIDER_PROFILE ROBOCLAWS_CODEX_MODEL \
+            ROBOCLAWS_PROVIDER_PROFILE ROBOCLAWS_CODEX_MODEL ROBOCLAWS_OPENAI_AGENTS_MODEL \
             ROBOCLAWS_CODEX_DISABLE_RESPONSES_WEBSOCKETS \
             ROBOCLAWS_PROVIDER_TIMING_PROXY \
             ROBOCLAWS_TIMING_PROXY_UPSTREAM_BASE_URL \

roboclaws/household/ci_live_reports.py

@@ -18,6 +18,12 @@
     "claude-version.txt",
     "claude-events.jsonl",
     "claude.stderr.log",
+    "driver.log",
+    "openai-agents-events.jsonl",
+    "openai-agents-trace.json",
+    "openai-agents-spans.jsonl",
+    "openai-agents-server.log",
+    "openai-agents-skill-context.json",
     "checker.log",
     "server.pid",
     "trace.jsonl",
@@ -34,6 +40,7 @@
 class MolmoLiveModelEntry:
     name: str
     label: str
+    agent_engine: str
     provider_profile: str
     model: str
     secret_env: str
@@ -42,36 +49,40 @@ class MolmoLiveModelEntry:
 
 MODEL_ENTRIES: tuple[MolmoLiveModelEntry, ...] = (
     MolmoLiveModelEntry(
-        name="kimi-k2.6",
-        label="Kimi K2.6",
-        provider_profile="kimi-anthropic",
-        model="kimi-k2.6",
-        secret_env="KIMI_API_KEY",
-        profile="world-public-labels",
-    ),
-    MolmoLiveModelEntry(
-        name="mimo-v2.5",
-        label="MiMo v2.5",
+        name="claude-code-mimo-v2.5",
+        label="Claude Code + MiMo v2.5",
+        agent_engine="claude-code",
         provider_profile="mimo-tp-anthropic",
         model="mimo-v2.5",
         secret_env="MIMO_TP_KEY",
         profile="world-public-labels",
     ),
     MolmoLiveModelEntry(
-        name="kimi-k2.6-camera-raw-fpv",
-        label="Kimi K2.6 RAW_FPV",
+        name="claude-code-kimi-k2.6",
+        label="Claude Code + Kimi K2.6",
+        agent_engine="claude-code",
         provider_profile="kimi-anthropic",
         model="kimi-k2.6",
         secret_env="KIMI_API_KEY",
-        profile="camera-raw-fpv",
+        profile="world-public-labels",
     ),
     MolmoLiveModelEntry(
-        name="mimo-v2.5-camera-raw-fpv",
-        label="MiMo v2.5 RAW_FPV",
-        provider_profile="mimo-tp-anthropic",
+        name="agents-sdk-mimo-v2.5",
+        label="OpenAI Agents SDK + MiMo v2.5",
+        agent_engine="openai-agents-sdk",
+        provider_profile="mimo-tp-openai-chat",
         model="mimo-v2.5",
         secret_env="MIMO_TP_KEY",
-        profile="camera-raw-fpv",
+        profile="world-public-labels",
+    ),
+    MolmoLiveModelEntry(
+        name="agents-sdk-kimi-k2.7-code",
+        label="OpenAI Agents SDK + Kimi K2.7 Code",
+        agent_engine="openai-agents-sdk",
+        provider_profile="kimi-openai-chat",
+        model="kimi-k2.7-code",
+        secret_env="KIMI_API_KEY",
+        profile="world-public-labels",
     ),
 )
 
@@ -104,10 +115,11 @@ def base_status(
         "schema": STATUS_SCHEMA,
         "entry": entry.name,
         "label": entry.label,
+        "agent_engine": entry.agent_engine,
         "provider_profile": entry.provider_profile,
         "model": entry.model,
         "secret_env": entry.secret_env,
-        "driver": "claude",
+        "driver": entry.agent_engine,
         "profile": profile,
         "seed": seed,
         "generated_mess_count": generated_mess_count,
@@ -294,7 +306,7 @@ def write_live_index(root: Path, statuses: list[dict[str, Any]] | None = None) -
                 "<h1>MolmoSpaces Live Cleanup Reports</h1>",
                 '<p class="sub">Published by GitHub Actions for main-only / opt-in '
                 "MolmoSpaces cleanup runs. Each row links to the report or diagnostics "
-                "for one provider-backed Claude Code run.</p>",
+                "for one provider-backed live-agent run.</p>",
                 "<table>",
                 "<thead><tr><th>Model</th><th>Status</th><th>Profile</th><th>Report</th>"
                 "<th>Provider</th><th>Rerun locally</th><th>Reason</th></tr></thead>",
@@ -313,6 +325,7 @@ def _live_index_row(status: dict[str, Any]) -> str:
     label = html.escape(str(status.get("label") or status.get("entry") or "Molmo live run"))
     state = html.escape(str(status.get("status") or "unknown"))
     provider = html.escape(str(status.get("provider_profile") or "unknown provider"))
+    agent_engine = html.escape(str(status.get("agent_engine") or status.get("driver") or "unknown"))
     model = html.escape(str(status.get("model") or "unknown model"))
     profile = html.escape(str(status.get("profile") or "unknown profile"))
     reason = html.escape(str(status.get("reason") or ""))
@@ -335,7 +348,7 @@ def _live_index_row(status: dict[str, Any]) -> str:
         f"<td><code>{state}</code></td>"
         f"<td><code>{profile}</code></td>"
         f"<td>{report}</td>"
-        f"<td><code>{provider}</code></td>"
+        f"<td><code>{agent_engine}</code><br><code>{provider}</code></td>"
         f"<td><code>{rerun_command}</code></td>"
         f"<td>{reason}</td>"
         "</tr>"

scripts/molmo_cleanup/run_ci_live_cleanup_matrix.py

@@ -51,11 +51,7 @@ def parse_args(argv: list[str] | None = None) -> argparse.Namespace:
         "--generated-mess-count",
         type=int,
         default=None,
-        help=(
-            "Generated mess count override. Defaults to 5 for world-public-labels entries "
-            "and 10 for camera-raw-fpv entries so the RAW_FPV success gate can require "
-            "7 accepted placements."
-        ),
+        help="Generated mess count override. Defaults to 5 for world-public-labels entries.",
     )
     parser.add_argument(
         "--profile",
@@ -108,7 +104,7 @@ def _run_preflight_or_statuses(
         return None
     try:
         if not args.skip_version_check:
-            _version_checks(args)
+            _version_checks(args, entries)
         if not args.skip_uv_sync:
             _run_checked([args.uv_bin, "sync", "--extra", "dev", "--extra", "molmospaces"])
         if not args.skip_prewarm:
@@ -181,7 +177,7 @@ def _run_entry(
             "rerun_command": rerun_command,
             "env": {
                 "ROBOCLAWS_PROVIDER_PROFILE": entry.provider_profile,
-                "ROBOCLAWS_CLAUDE_MODEL": entry.model,
+                _model_env_key(entry): entry.model,
                 PROVIDER_TIMING_PROXY_ENV: _default_provider_timing_proxy_value(),
             },
             "cache_roots": [
@@ -205,7 +201,7 @@ def _run_entry(
 
     env = os.environ.copy()
     env["ROBOCLAWS_PROVIDER_PROFILE"] = entry.provider_profile
-    env["ROBOCLAWS_CLAUDE_MODEL"] = entry.model
+    env[_model_env_key(entry)] = entry.model
     env.setdefault(PROVIDER_TIMING_PROXY_ENV, "1")
     env["ROBOCLAWS_REPORT_RERUN_COMMAND"] = rerun_command
     try:
@@ -258,8 +254,6 @@ def _entry_profile(entry: MolmoLiveModelEntry, args: argparse.Namespace) -> str:
 def _entry_generated_mess_count(entry: MolmoLiveModelEntry, args: argparse.Namespace) -> int:
     if args.generated_mess_count is not None:
         return int(args.generated_mess_count)
-    if _entry_profile(entry, args) == "camera-raw-fpv":
-        return 10
     return 5
 
 
@@ -284,7 +278,7 @@ def _live_command(
         "world=molmospaces/val_0",
         "backend=mujoco",
         "intent=cleanup",
-        "agent_engine=claude-code",
+        f"agent_engine={entry.agent_engine}",
         f"provider_profile={entry.provider_profile}",
         f"evidence_lane={profile}",
         f"seed={args.seed}",
@@ -307,7 +301,7 @@ def _live_report_rerun_command(entry: MolmoLiveModelEntry, args: argparse.Namesp
         "world=molmospaces/val_0",
         "backend=mujoco",
         "intent=cleanup",
-        "agent_engine=claude-code",
+        f"agent_engine={entry.agent_engine}",
         f"provider_profile={entry.provider_profile}",
         f"evidence_lane={profile}",
         f"seed={args.seed}",
@@ -317,12 +311,20 @@ def _live_report_rerun_command(entry: MolmoLiveModelEntry, args: argparse.Namesp
     ]
     return (
         f"ROBOCLAWS_PROVIDER_PROFILE={entry.provider_profile} "
-        f"ROBOCLAWS_CLAUDE_MODEL={entry.model} "
+        f"{_model_env_key(entry)}={entry.model} "
         f"{PROVIDER_TIMING_PROXY_ENV}={_default_provider_timing_proxy_value()} "
         f"{shell_join(command)}"
     )
 
 
+def _model_env_key(entry: MolmoLiveModelEntry) -> str:
+    if entry.agent_engine == "claude-code":
+        return "ROBOCLAWS_CLAUDE_MODEL"
+    if entry.agent_engine == "openai-agents-sdk":
+        return "ROBOCLAWS_OPENAI_AGENTS_MODEL"
+    raise ValueError(f"unsupported Molmo live agent engine: {entry.agent_engine}")
+
+
 def _default_provider_timing_proxy_value() -> str:
     return os.environ.get(PROVIDER_TIMING_PROXY_ENV, "1")
 
@@ -344,8 +346,14 @@ def _prewarm(args: argparse.Namespace, *, generated_mess_count: int) -> None:
     )
 
 
-def _version_checks(args: argparse.Namespace) -> None:
-    for binary in ("codex", "claude"):
+def _version_checks(
+    args: argparse.Namespace,
+    entries: tuple[MolmoLiveModelEntry, ...],
+) -> None:
+    required_binaries = (
+        {"claude"} if any(entry.agent_engine == "claude-code" for entry in entries) else set()
+    )
+    for binary in sorted(required_binaries):
         resolved = shutil.which(binary)
         if not resolved:
             raise RuntimeError(f"{binary} command not found")

scripts/reports/write_pages_index.py

@@ -100,6 +100,9 @@ def _molmo_live_section(site_dir: Path) -> str:
     items = []
     for entry in entries:
         label = html.escape(str(entry.get("label") or entry.get("entry") or "Molmo live run"))
+        agent_engine = html.escape(
+            str(entry.get("agent_engine") or entry.get("driver") or "unknown")
+        )
         model = html.escape(str(entry.get("model") or "unknown model"))
         provider = html.escape(str(entry.get("provider_profile") or "unknown provider"))
         status = html.escape(str(entry.get("status") or "unknown"))
@@ -116,7 +119,7 @@ def _molmo_live_section(site_dir: Path) -> str:
         else:
             title = f"<span>{label}</span>"
         desc = (
-            f'      <div class="desc">Claude Code live cleanup via '
+            f'      <div class="desc">Live cleanup via <code>{agent_engine}</code> and '
             f"<code>{provider}</code> / <code>{model}</code> using "
             f"<code>{profile}</code>. Status: "
             f"<code>{status}</code>"

tests/contract/dev_tools/test_task_agent_just_recipes.py

@@ -37,9 +37,7 @@
 CODING_AGENT_DOCKER = REPO_ROOT / "scripts" / "dev" / "coding_agent_docker.sh"
 LIVE_CODEX_RUNNER = REPO_ROOT / "scripts" / "molmo_cleanup" / "run_live_codex_cleanup.py"
 LIVE_CLAUDE_RUNNER = REPO_ROOT / "scripts" / "molmo_cleanup" / "run_live_claude_cleanup.py"
-LIVE_OPENAI_AGENTS_RUNNER = (
-    REPO_ROOT / "scripts" / "molmo_cleanup" / "run_live_openai_agents_cleanup.py"
-)
+LIVE_OPENAI_AGENTS_RUNNER = REPO_ROOT / "scripts/molmo_cleanup/run_live_openai_agents_cleanup.py"
 AGIBOT_MAP_BUILD_CODEX_RUNNER = (
     REPO_ROOT / "scripts" / "molmo_cleanup" / "run_live_codex_agibot_map_build.py"
 )
@@ -848,6 +846,8 @@ def test_openai_agents_sdk_cleanup_route_stays_private_non_default() -> None:
     assert "--model-thinking-mode" in molmo_text
     assert "ROBOCLAWS_OPENAI_AGENTS_THINKING_MODE" in molmo_text
     assert "--context-soft-limit-tokens" in molmo_text
+    assert "ROBOCLAWS_OPENAI_AGENTS_MODEL ROBOCLAWS_PROVIDER_PROFILE" in molmo_text
+    assert "ROBOCLAWS_PROVIDER_PROFILE ROBOCLAWS_OPENAI_AGENTS_MODEL" in molmo_text
     assert "openai-agents-live" not in trace_household_cleanup_run("codex")
 
     plan = resolve_surface_launch(