feat(harness): add OMP client support

README.md

@@ -105,7 +105,7 @@ Reference target: **RTX 3090 (Ampere sm_86)** — all headline numbers. Other NV
 
 [`harness/`](harness/) contains RTX 3090 client launchers and regression tests
 for Lucebox server compatibility. Run Lucebox inside Claude Code, Codex,
-OpenCode, Hermes, Pi, OpenClaw, or Open WebUI, or check if a server change
+OpenCode, Hermes, OMP, Pi, OpenClaw, or Open WebUI, or check if a server change
 still works with those clients.
 
 <table>
@@ -123,6 +123,7 @@ still works with those clients.
 | Codex | [`run_codex.sh`](harness/clients/run_codex.sh) |
 | OpenCode | [`run_opencode.sh`](harness/clients/run_opencode.sh) |
 | Hermes | [`run_hermes.sh`](harness/clients/run_hermes.sh) |
+| [OMP](https://omp.sh/) | [`run_omp.sh`](harness/clients/run_omp.sh) |
 | Pi | [`run_pi.sh`](harness/clients/run_pi.sh) |
 | OpenClaw | [`run_openclaw.sh`](harness/clients/run_openclaw.sh) |
 | Open WebUI | [`run_openwebui.sh`](harness/clients/run_openwebui.sh) |
@@ -149,7 +150,7 @@ Launcher scripts install missing real-client CLIs automatically under
 `.harness-work/`. To preinstall them yourself:
 
 ```bash
-python3 harness/client_test_runner.py install --clients codex,hermes,openwebui
+python3 harness/client_test_runner.py install --clients codex,hermes,omp,openwebui
 ```
 
 For direct TPS/TTFT numbers against a running server:

harness/README.md

@@ -10,7 +10,7 @@
 
 <p align="center">
   <strong>Client launchers and regression tests for Lucebox server compatibility.</strong><br/>
-  Run Lucebox from Claude Code, Codex, OpenCode, Hermes, Pi, OpenClaw, or Open WebUI.<br/>
+  Run Lucebox from Claude Code, Codex, OpenCode, Hermes, OMP, Pi, OpenClaw, or Open WebUI.<br/>
   RTX 3090 / 24 GB defaults are included for each client.
 </p>
 
@@ -31,6 +31,7 @@ cd lucebox-hub
 harness/clients/run_codex.sh
 harness/clients/run_claude_code.sh
 harness/clients/run_opencode.sh
+harness/clients/run_omp.sh
 ```
 
 The launchers default to the current repo, install/use client packages under
@@ -93,7 +94,7 @@ If you already have `dflash_server` running, use `probe`:
 ```bash
 python3 harness/client_test_runner.py probe \
   --url http://127.0.0.1:8000 \
-  --clients claude_code,codex,opencode,openwebui,pi \
+  --clients claude_code,codex,opencode,omp,openwebui,pi \
   --json-out /tmp/lucebox_harness_probe.json
 ```
 
@@ -103,7 +104,7 @@ client packages. Without it, the HTTP protocol probes still run.
 To preinstall real-client CLIs yourself:
 
 ```bash
-python3 harness/client_test_runner.py install --clients codex,hermes,openwebui
+python3 harness/client_test_runner.py install --clients codex,hermes,omp,openwebui
 ```
 
 For a GPU sweep, let the runner start Lucebox for each profile:

harness/benchmarks/README.md

@@ -6,7 +6,7 @@ deterministic prompts.
 
 Use this when you want to know whether a server change affects output quality or
 decode speed. Use `harness/clients/` when you want to know whether Codex,
-OpenCode, Open WebUI, Pi, and the other clients still work.
+OpenCode, OMP, Open WebUI, Pi, and the other clients still work.
 
 ## Bench suites (HumanEval, GSM8K, Math500, Agent)
 

harness/client_test_runner.py

@@ -101,6 +101,14 @@ class ClientSpec:
         binary="opencode",
         protocol="openai_chat",
     ),
+    "omp": ClientSpec(
+        name="omp",
+        install="omp",
+        package="https://raw.githubusercontent.com/can1357/oh-my-pi/main/scripts/install.sh",
+        binary="omp",
+        protocol="responses",
+        notes="OMP uses a custom openai-responses provider pointed at Lucebox.",
+    ),
     "pi": ClientSpec(
         name="pi",
         install="npm",
@@ -281,6 +289,8 @@ def client_bin(work_dir: Path, spec: ClientSpec) -> Path:
         return pip_venv(work_dir, spec.name) / "bin" / spec.binary
     if spec.install == "hermes":
         return hermes_home(work_dir) / ".local" / "bin" / spec.binary
+    if spec.install == "omp":
+        return work_dir / "clients" / spec.name / "bin" / spec.binary
     raise HarnessError(f"unknown installer {spec.install}")
 
 
@@ -339,6 +349,24 @@ def install_client(work_dir: Path, spec: ClientSpec) -> dict[str, Any]:
             timeout=1800,
             stream=True,
         )
+    elif spec.install == "omp":
+        # Install the latest release, matching the repo's unpinned npm/pip
+        # client policy. The --version smoke below records the exact binary
+        # in the install report for reproducibility.
+        root = work_dir / "clients" / spec.name
+        bin_dir = root / "bin"
+        root.mkdir(parents=True, exist_ok=True)
+        bin_dir.mkdir(parents=True, exist_ok=True)
+        script_path = root / "install.sh"
+        urllib.request.urlretrieve(spec.package, script_path)
+        env = os.environ.copy()
+        env["PI_INSTALL_DIR"] = str(bin_dir)
+        result = run_cmd(
+            ["sh", str(script_path), "--binary"],
+            env=env,
+            timeout=900,
+            stream=True,
+        )
     else:
         raise HarnessError(f"unknown installer {spec.install}")
 

harness/clients/README.md

@@ -20,7 +20,7 @@ If a client CLI is missing, the launcher installs it automatically. Set
 To preinstall real-client CLIs yourself:
 
 ```bash
-python3 harness/client_test_runner.py install --clients codex,hermes,openwebui
+python3 harness/client_test_runner.py install --clients codex,hermes,omp,openwebui
 ```
 
 The launcher will start `server/build/dflash_server` by default, or the path in
@@ -52,7 +52,8 @@ harness/clients/run_codex.sh
 
 The C++ server is expected to handle the same client protocol shapes covered by
 these launchers and probes: OpenAI Chat Completions, streaming chunks, tool
-metadata, OpenAI Responses for Codex, Anthropic Messages for Claude Code, and
+metadata, OpenAI Responses for Codex and OMP, Anthropic Messages for Claude
+Code, and
 Open WebUI model metadata.
 
 ## Defaults
@@ -66,6 +67,7 @@ The defaults below are the current RTX 3090 starting points for
 | Codex | `run_codex.sh` | `MAX_CTX=32768 BUDGET=22 VERIFY_MODE=ddtree EXTRA_SERVER_ARGS=--lazy-draft` |
 | OpenCode | `run_opencode.sh` | `MAX_CTX=86016 BUDGET=22 VERIFY_MODE=ddtree EXTRA_SERVER_ARGS=--lazy-draft` |
 | Hermes Agent | `run_hermes.sh` | `MAX_CTX=98304 BUDGET=22 VERIFY_MODE=ddtree EXTRA_SERVER_ARGS=--lazy-draft` |
+| OMP | `run_omp.sh` | `MAX_CTX=65536 BUDGET=22 VERIFY_MODE=ddtree EXTRA_SERVER_ARGS=--lazy-draft OMP_TIMEOUT=3600` |
 | Pi | `run_pi.sh` | `MAX_CTX=65536 BUDGET=22 VERIFY_MODE=ddtree EXTRA_SERVER_ARGS=--lazy-draft PI_TIMEOUT=3600` |
 | OpenClaw | `run_openclaw.sh` | `MAX_CTX=204800 BUDGET=22 VERIFY_MODE=ddtree EXTRA_SERVER_ARGS=--lazy-draft` |
 | Open WebUI chat | `run_openwebui.sh` | `MAX_CTX=262144 BUDGET=22 VERIFY_MODE=ddtree EXTRA_SERVER_ARGS=--lazy-draft` |
@@ -79,6 +81,17 @@ PROMPT='Explain the repo and end with lucebox-client-ok' harness/clients/run_ope
 PROMPT_FILE=harness/clients/prompts/repo_inspection.txt harness/clients/run_hermes.sh
 ```
 
+OMP uses a generated `models.yml` with Lucebox as a keyless custom
+`openai-responses` provider. `OMP_TIMEOUT` controls the launcher's wall-clock
+deadline, while `OMP_STREAM_IDLE_TIMEOUT_MS` defaults to one hour so long
+prefills are not interrupted by OMP's stream watchdog; set it to `0` to
+disable that watchdog. Override either when needed:
+
+```bash
+OMP_TIMEOUT=0 OMP_STREAM_IDLE_TIMEOUT_MS=7200000 \
+  harness/clients/run_omp.sh
+```
+
 `PI_TIMEOUT` is Pi's total wall-clock limit in seconds. Its one-hour default
 allows long-context prefill and long generations to finish; set
 `PI_TIMEOUT=0` to run without a launcher deadline. The launcher also disables
@@ -92,8 +105,8 @@ put the same setting in `~/.pi/agent/settings.json`:
 
 The other real-client launchers also allow one hour by default. Override their
 deadlines with `CLAUDE_TIMEOUT`, `CODEX_TIMEOUT`, `OPENCODE_TIMEOUT`,
-`HERMES_TIMEOUT`, `OPENCLAW_TIMEOUT`, or (for Open WebUI's curl probe)
-`CURL_MAX_TIME`. The CLI launcher timeouts accept `0` to disable the outer
+`OMP_TIMEOUT`, `HERMES_TIMEOUT`, `OPENCLAW_TIMEOUT`, or (for Open WebUI's
+curl probe) `CURL_MAX_TIME`. The CLI launcher timeouts accept `0` to disable the outer
 deadline. OpenCode's provider-level request and chunk deadlines default to one
 hour too and can be changed with `OPENCODE_REQUEST_TIMEOUT_MS` and
 `OPENCODE_CHUNK_TIMEOUT_MS`. The server independently sends SSE heartbeat
@@ -119,9 +132,9 @@ CLIENT=opencode PROMPT_FILE=harness/clients/prompts/repo_inspection.txt \
   harness/clients/run_backend_pair.sh
 ```
 
-OpenAI Chat Completions clients can call llama.cpp directly. Claude Code and
-Codex use `llamacpp_compat_proxy.py` so their real Anthropic Messages and
-Responses requests can be compared too.
+OpenAI Chat Completions clients and OMP's Responses client can call llama.cpp
+directly. Claude Code and Codex use `llamacpp_compat_proxy.py` so their real
+Anthropic Messages and Responses requests can be compared too.
 
 ## Notes
 

harness/clients/run_backend_pair.sh

@@ -18,6 +18,7 @@ case "$CLIENT" in
   openclaw) CLIENT_SCRIPT="$SCRIPT_DIR/run_openclaw.sh" ;;
   openwebui) CLIENT_SCRIPT="$SCRIPT_DIR/run_openwebui.sh" ;;
   openwebui_tools) CLIENT_SCRIPT="$SCRIPT_DIR/run_openwebui_tools.sh" ;;
+  omp) CLIENT_SCRIPT="$SCRIPT_DIR/run_omp.sh" ;;
   pi) CLIENT_SCRIPT="$SCRIPT_DIR/run_pi.sh" ;;
   *)
     echo "unknown CLIENT=$CLIENT" >&2

harness/clients/run_omp.sh

@@ -0,0 +1,89 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+: "${MAX_CTX:=65536}"
+: "${BUDGET:=22}"
+: "${VERIFY_MODE:=ddtree}"
+: "${EXTRA_SERVER_ARGS:=--lazy-draft}"
+: "${OMP_TOOLS:=read,grep,glob}"
+: "${OMP_TIMEOUT:=3600}"
+: "${OMP_STREAM_IDLE_TIMEOUT_MS:=3600000}"
+if [[ ! "$OMP_TIMEOUT" =~ ^[0-9]+$ ]]; then
+  echo "OMP_TIMEOUT must be a non-negative integer (seconds; 0 disables it)" >&2
+  exit 2
+fi
+if [[ ! "$OMP_STREAM_IDLE_TIMEOUT_MS" =~ ^[0-9]+$ ]]; then
+  echo "OMP_STREAM_IDLE_TIMEOUT_MS must be a non-negative integer (0 disables OMP's stream watchdog)" >&2
+  exit 2
+fi
+source "$SCRIPT_DIR/common.sh"
+
+CLIENT_OUT="$LOG_DIR/omp.out"
+OMP_BIN="${OMP_BIN:-$CLIENT_WORK_DIR/clients/omp/bin/omp}"
+require_client_binary "OMP" "$OMP_BIN" "omp" "OMP_BIN"
+HOME_DIR="$LOG_DIR/omp-home"
+AGENT_DIR="$HOME_DIR/agent"
+mkdir -p "$AGENT_DIR" "$HOME_DIR/sessions"
+
+cat > "$AGENT_DIR/models.yml" <<YAML
+providers:
+  lucebox:
+    baseUrl: "$BASE_URL/v1"
+    auth: none
+    api: openai-responses
+    compat:
+      supportsDeveloperRole: false
+      supportsReasoningEffort: false
+      supportsUsageInStreaming: true
+      maxTokensField: max_tokens
+      streamIdleTimeoutMs: $OMP_STREAM_IDLE_TIMEOUT_MS
+    models:
+      - id: "$MODEL_ID"
+        name: "Lucebox DFlash"
+        reasoning: false
+        input: [text]
+        contextWindow: $MAX_CTX
+        maxTokens: $MAX_TOKENS
+        cost:
+          input: 0
+          output: 0
+          cacheRead: 0
+          cacheWrite: 0
+YAML
+
+start_lucebox_server
+trap stop_lucebox_server EXIT
+wait_lucebox_server
+
+omp_env=(
+  "HOME=$HOME_DIR"
+  "PI_CODING_AGENT_DIR=$AGENT_DIR"
+  "PI_CODING_AGENT_SESSION_DIR=$HOME_DIR/sessions"
+  # An inherited OMP_PROFILE/PI_PROFILE would redirect OMP to a profile
+  # agent dir and ignore the generated models.yml above.
+  "OMP_PROFILE="
+  "PI_PROFILE="
+)
+omp_cmd=(
+  "$OMP_BIN"
+  --model "lucebox/$MODEL_ID"
+  --print
+  --mode json
+  --tools "$OMP_TOOLS"
+  --no-session
+  --no-extensions
+  --no-skills
+  --no-rules
+  --no-title
+  "$PROMPT"
+)
+
+set +e
+run_with_timeout "$OMP_TIMEOUT" env "${omp_env[@]}" "${omp_cmd[@]}" \
+  < /dev/null > "$CLIENT_OUT" 2>&1
+RC=$?
+set -e
+
+finish_report "$CLIENT_OUT" "$RC"
+exit "$RC"

harness/clients/summarize_backend_pair.py

@@ -28,6 +28,7 @@
     "openclaw.out",
     "openwebui.out",
     "openwebui-tools.out",
+    "omp.out",
     "pi.out",
 }
 MARKERS = ("OK_DONE", "lucebox-client-ok", "OPENWEBUI_TOOL_OK")
@@ -129,6 +130,36 @@ def first_json_value(text: str):
     return value
 
 
+def extract_omp_json_text(text: str) -> str:
+    """Pull assistant text out of OMP's `--mode json` event stream.
+
+    OMP echoes the user prompt and tool results as message events too, so only
+    `role == "assistant"` message content counts as generated output.
+    """
+    parts: list[str] = []
+    for line in text.splitlines():
+        line = line.strip()
+        if not line.startswith("{"):
+            continue
+        try:
+            event = json.loads(line)
+        except json.JSONDecodeError:
+            continue
+        if not isinstance(event, dict) or event.get("type") != "message_end":
+            continue
+        message = event.get("message")
+        if not isinstance(message, dict) or message.get("role") != "assistant":
+            continue
+        content = message.get("content")
+        if isinstance(content, str):
+            parts.append(content)
+        elif isinstance(content, list):
+            for item in content:
+                if isinstance(item, dict) and isinstance(item.get("text"), str):
+                    parts.append(item["text"])
+    return "\n".join(part for part in parts if part)
+
+
 def extract_generated_text(text: str) -> str:
     parts: list[str] = []
     value = first_json_value(text)
@@ -164,6 +195,9 @@ def extract_generated_text(text: str) -> str:
         return ""
 
     if not parts:
+        omp_text = extract_omp_json_text(text)
+        if omp_text:
+            return omp_text
         for line in text.splitlines():
             line = line.strip()
             if not line:
@@ -205,6 +239,10 @@ def tool_call_ok(text: str) -> bool:
     return "tool_call:" in text or '"tool_calls"' in text
 
 
+def omp_tool_call_ok(text: str) -> bool:
+    return '"toolCall"' in text or '"stopReason":"toolUse"' in text or '"tool_execution_start"' in text
+
+
 def preview(text: str, limit: int = 180) -> str:
     compact = re.sub(r"\s+", " ", text).strip()
     return compact[:limit]
@@ -230,7 +268,11 @@ def summarize_backend(pair_dir: Path, backend: str) -> dict:
     m = re.search(r"^rc=(\d+)$", backend_out, flags=re.M)
     if m:
         rc = m.group(1)
-    observed_tool_call = tool_call_ok(generated_text) or any(call.get("finish") == "tool_calls" for call in calls)
+    observed_tool_call = (
+        tool_call_ok(generated_text)
+        or omp_tool_call_ok(client_text)
+        or any(call.get("finish") == "tool_calls" for call in calls)
+    )
     return {
         "backend": backend,
         "run_dir": str(run_dir),

harness/tests/test_client_launcher_timeouts.sh

@@ -117,6 +117,7 @@ launcher_cases=(
   'run_hermes.sh|HERMES_BIN|HERMES_TIMEOUT|1'
   'run_openclaw.sh|OPENCLAW_BIN|OPENCLAW_TIMEOUT|2'
   'run_opencode.sh|OPENCODE_BIN|OPENCODE_TIMEOUT|1'
+  'run_omp.sh|OMP_BIN|OMP_TIMEOUT|1'
 )
 for launcher_case in "${launcher_cases[@]}"; do
   IFS='|' read -r script client_var timeout_var expected_calls <<<"$launcher_case"
@@ -159,6 +160,8 @@ grep -Fq ': "${HERMES_TIMEOUT:=3600}"' "$CLIENTS/run_hermes.sh"
 grep -Fq 'run_with_timeout "$HERMES_TIMEOUT"' "$CLIENTS/run_hermes.sh"
 grep -Fq ': "${CLAUDE_TIMEOUT:=3600}"' "$CLIENTS/run_claude_code.sh"
 grep -Fq 'run_with_timeout "$CLAUDE_TIMEOUT"' "$CLIENTS/run_claude_code.sh"
+grep -Fq ': "${OMP_TIMEOUT:=3600}"' "$CLIENTS/run_omp.sh"
+grep -Fq 'run_with_timeout "$OMP_TIMEOUT"' "$CLIENTS/run_omp.sh"
 grep -Fq ': "${OPENCLAW_TIMEOUT:=3600}"' "$CLIENTS/run_openclaw.sh"
 grep -Fq 'run_with_timeout "$OPENCLAW_TIMEOUT"' "$CLIENTS/run_openclaw.sh"
 grep -Fq 'openclaw_cmd+=(--timeout "$OPENCLAW_TIMEOUT")' "$CLIENTS/run_openclaw.sh"