omnius 1.0.639 → 1.0.641

package/dist/scripts/transcribe-file.py

@@ -8,16 +8,65 @@ stdout: JSON {"text": "...", "duration": null, "segments": []}
 
 ASR is GPU-first: CPU fallback is refused unless OMNIUS_ASR_ALLOW_CPU=1.
 """
-import sys, os, json, warnings
+import sys, os, json, math, struct, warnings, wave
 
 warnings.filterwarnings("ignore")
 os.environ.setdefault("PYTHONWARNINGS", "ignore")
 
-try:
-    import whisper
-except Exception as e:
-    print(json.dumps({"error": "deps_missing", "message": str(e)}))
-    sys.exit(0)
+PCM16_SILENCE_MAX_PEAK = 4
+PCM16_SILENCE_MAX_RMS = 1.0
+PCM16_ACTIVE_SAMPLE_THRESHOLD = 8
+
+
+def exact_pcm16_silence(path):
+    """Return typed no-speech evidence for exact PCM16/16k/mono WAV only.
+
+    This is intentionally not VAD. It reads user bytes but never rewrites,
+    normalizes, decodes another media format, or instantiates Whisper.
+    """
+    try:
+        with wave.open(path, "rb") as wav:
+            if (
+                wav.getcomptype() != "NONE"
+                or wav.getnchannels() != 1
+                or wav.getsampwidth() != 2
+                or wav.getframerate() != 16000
+            ):
+                return None
+            frames = wav.readframes(wav.getnframes())
+    except (wave.Error, OSError, EOFError):
+        return None
+    sample_count = len(frames) // 2
+    if sample_count == 0:
+        metrics = {
+            "sampleRateHz": 16000, "channels": 1, "bitsPerSample": 16,
+            "sampleCount": 0, "durationMs": 0.0, "peakPcm16": 0,
+            "rmsPcm16": 0.0, "activeSampleCount": 0, "activeSampleRatio": 0.0,
+        }
+    else:
+        samples = struct.iter_unpack("<h", frames[: sample_count * 2])
+        total_squares = 0
+        peak = 0
+        active = 0
+        for (sample,) in samples:
+            magnitude = abs(sample)
+            total_squares += sample * sample
+            peak = max(peak, magnitude)
+            if magnitude >= PCM16_ACTIVE_SAMPLE_THRESHOLD:
+                active += 1
+        metrics = {
+            "sampleRateHz": 16000, "channels": 1, "bitsPerSample": 16,
+            "sampleCount": sample_count, "durationMs": sample_count / 16.0,
+            "peakPcm16": peak, "rmsPcm16": math.sqrt(total_squares / sample_count),
+            "activeSampleCount": active, "activeSampleRatio": active / sample_count,
+        }
+    if (
+        metrics["peakPcm16"] <= PCM16_SILENCE_MAX_PEAK
+        and metrics["rmsPcm16"] <= PCM16_SILENCE_MAX_RMS
+        and metrics["activeSampleCount"] == 0
+    ):
+        return {"kind": "no_speech", "reason": "digital_silence", "text": "", "signal": metrics}
+    return None
 
 
 def main() -> None:
@@ -33,8 +82,14 @@ def main() -> None:
         print(json.dumps({"error": "file_not_found", "path": path or ""}))
         return
 
+    no_speech = exact_pcm16_silence(path)
+    if no_speech is not None:
+        print(json.dumps({**no_speech, "duration": no_speech["signal"]["durationMs"] / 1000.0, "segments": []}))
+        return
+
     model_name = data.get("model") or "tiny"
     try:
+        import whisper
         device = os.environ.get("OMNIUS_ASR_DEVICE", "cuda").strip().lower() or "cuda"
         allow_cpu = os.environ.get("OMNIUS_ASR_ALLOW_CPU", "").strip().lower() in ("1", "true", "yes", "on")
         if device == "auto":

package/dist/update-worker.js

@@ -293080,7 +293080,6 @@ var MAX_IMAGE_BYTES = 8 * 1024 * 1024;
 
 // packages/execution/dist/tools/ocr-image-advanced.js
 init_process_async();
-var OCR_PIPELINE_TIMEOUT_MS = 5 * 6e4;
 
 // packages/execution/dist/tools/browser-action.js
 init_process_lifecycle();

package/docs/DISCOVERY.json

@@ -4171,7 +4171,7 @@
           ],
           "responses": {
             "200": {
-              "description": "Transcription with actual engine/model identity and supported structured metadata"
+              "description": "Transcription with actual engine/model identity; exact digital PCM silence returns text='' plus typed noSpeech evidence without model inference"
             },
             "400": {
               "description": "Empty or tiny audio"
@@ -5399,7 +5399,7 @@
           "tags": [
             "Audio"
           ],
-          "description": "Admin-only. Requires kind=acoustic|speaker|semantic in the query (canonical) or JSON body. acoustic reuses the pinned JetPack YAMNet/TensorRT setup; speaker installs the isolated CPU-only WeSpeaker CAM++ ONNX Runtime and pinned model, then installs only checksummed wheels into ~/.omnius/runtimes/audio/speaker/venv: CPython 3.10/aarch64 Torchaudio 2.2.0 and the self-contained typing_extensions support dependency, each with --no-deps/--no-index. It retains --system-site-packages and links the selected OMNIUS_AUDIO_PYTHON JetPack Torch 2.2/CUDA 12.2 provider without writing to it; setup proves the Torch/Torchaudio ABI and CPU kaldi.fbank before readiness. An operator may replace the Torchaudio wheel only by supplying both OMNIUS_SPEAKER_TORCHAUDIO_WHEEL and OMNIUS_SPEAKER_TORCHAUDIO_WHEEL_SHA256. semantic installs the isolated JetPack CUDA CLAP dependencies and pinned model. This is the only REST operation allowed to provision. It activates and warms only the requested role worker; no role is substituted for another.",
+          "description": "Admin-only. Requires kind=acoustic|speaker|semantic in the query (canonical) or JSON body. acoustic reuses the pinned JetPack YAMNet/TensorRT setup; speaker creates a private CPython 3.10 CPU-only WeSpeaker CAM++ venv, installs only checksum-pinned NumPy, ONNX Runtime, and its locked CPU support wheels with --no-deps/--no-index, then downloads the pinned model. Its 80-bin Kaldi-compatible NumPy fbank is validated with a versioned cross-platform numeric-invariant probe before readiness. The speaker path never imports, links, replaces, or otherwise depends on JetPack Torch/Torchaudio, so an incompatible generic Torchaudio wheel cannot affect CUDA-enabled Egg Torch. semantic installs the isolated JetPack CUDA CLAP dependencies and pinned model. This is the only REST operation allowed to provision. It activates and warms only the requested role worker; no role is substituted for another.",
           "parameters": [
             {
               "name": "kind",
@@ -6070,7 +6070,7 @@
               "$ref": "#/components/parameters/MinimumOmniusVersion"
             }
           ],
-          "description": "Standard OpenAI chat-completion proxy by default. Set agent_loop=true to enter a synchronous server-side loop that auto-executes daemon tools (web_search, web_fetch, file_read, ocr, etc.) and re-prompts the model — collapsing N round-trips into 1. include_daemon_tools=['read','run'] auto-merges daemon tools matching those scopes into the offered tools[]. timeout_s overrides OMNIUS_BACKEND_TIMEOUT_S (default 120s, max 1h). prompt_template='factual-first' prepends a web_search-first policy. SSE streaming honors tool_calls in delta as standard OpenAI shape: delta.tool_calls=[{id,type:'function',index,function:{name,arguments:<JSON-string>}}]. parallel_tool_calls is forwarded to the backend.",
+          "description": "Standard OpenAI chat-completion proxy by default. Set agent_loop=true to enter a bounded synchronous server-side loop that auto-executes selected daemon tools and re-prompts the model. Ollama targets use native /api/chat tool messages. timeout_s bounds each backend call; agent_timeout_s bounds the whole loop (default 45s, max 600s). daemon_tool_names explicitly limits the offered catalog; factual-first implicitly limits it to web_search/web_fetch when no names are supplied. Identical repeated calls fail with HTTP 508 instead of looping. SSE streaming outside agent_loop honors OpenAI tool_calls deltas.",
           "requestBody": {
             "required": true,
             "content": {
@@ -6140,7 +6140,28 @@
                           "admin"
                         ]
                       },
-                      "description": "Q3 — scopes to merge from daemon tool catalog into tools[]."
+                      "description": "Scopes from which Omnius may offer its bounded core daemon-tool catalog. Use daemon_tool_names to request other exact tools."
+                    },
+                    "daemon_tool_names": {
+                      "type": "array",
+                      "items": {
+                        "type": "string"
+                      },
+                      "description": "Optional exact allowlist of daemon tool names. Use this to keep local-model tool prompts small."
+                    },
+                    "agent_timeout_s": {
+                      "type": "number",
+                      "minimum": 5,
+                      "maximum": 600,
+                      "default": 45,
+                      "description": "Total server-side agent-loop deadline. Distinct from per-backend timeout_s."
+                    },
+                    "agent_max_tool_rounds": {
+                      "type": "integer",
+                      "minimum": 1,
+                      "maximum": 8,
+                      "default": 1,
+                      "description": "Maximum daemon-tool planning rounds. The default executes one tool round, then removes daemon schemas for lower-latency final synthesis."
                     },
                     "max_turns": {
                       "type": "integer",
@@ -6160,7 +6181,7 @@
           },
           "responses": {
             "200": {
-              "description": "OpenAI chat.completion shape, SSE if stream=true. agent_loop responses include _agent_loop:{turns,log,done,reason}."
+              "description": "OpenAI chat.completion shape, SSE if stream=true. agent_loop responses include _agent_loop:{turns,log,done,reason,elapsed_ms,backend_transport}."
             },
             "400": {
               "description": "Invalid request",
@@ -6231,6 +6252,12 @@
                   }
                 }
               }
+            },
+            "504": {
+              "description": "Backend round or total agent-loop deadline expired"
+            },
+            "508": {
+              "description": "The model repeated an identical daemon tool call"
             }
           }
         }
@@ -18540,7 +18567,7 @@
           "tags": [
             "Voice"
           ],
-          "description": "Returns synthesized audio and automatically enables/warms the selected runtime. Format `wav` (default) ships a complete WAV with header; `pcm` ships raw Int16 LE PCM. Set model=voxtral-4b-tts-2603 and voice=<preset> for the managed CUDA Voxtral renderer; the OpenAI-compatible alias also accepts model=mistralai/Voxtral-4B-TTS-2603. Other model transactions remain serialized and restore the persistent renderer. Headers report X-Voice-Model, X-Voice-Backend, and X-Sample-Rate.",
+          "description": "Returns synthesized audio and automatically enables/warms the selected runtime. An explicit model has no silent fallback. Format `wav` (default) ships a complete WAV with header; `pcm` ships raw Int16 LE PCM. Set model=voxtral-4b-tts-2603 and voice=<preset> for the managed CUDA Voxtral renderer; the OpenAI-compatible alias also accepts model=mistralai/Voxtral-4B-TTS-2603. Other model transactions remain serialized and restore the persistent renderer. Headers report X-Voice-Model, X-Voice-Backend, and X-Sample-Rate.",
           "responses": {
             "200": {
               "description": "Audio bytes (audio/wav or audio/L16). Headers: X-Voice-Model, X-Sample-Rate."

package/docs/rest/endpoints/chat.md

@@ -43,7 +43,10 @@ Important body fields:
 | `parallel_tool_calls` | boolean | Forwarded to backend |
 | `timeout_s` | number | Per-request timeout |
 | `agent_loop` | boolean | Run server-side tool loop |
-| `include_daemon_tools` | array | Merge daemon tools by scope: `read`, `run`, `admin` |
+| `include_daemon_tools` | array | Permit the bounded core daemon-tool catalog by scope: `read`, `run`, `admin` |
+| `daemon_tool_names` | array | Exact daemon-tool allowlist; recommended for local models |
+| `agent_timeout_s` | number | Whole-loop deadline, default 45 seconds and maximum 600 |
+| `agent_max_tool_rounds` | integer | Daemon tool rounds before forced final synthesis; default 1 |
 | `max_turns` | integer | Server-side agent loop turn cap |
 | `prompt_template` | string | Optional template such as `factual-first` |
 
@@ -126,3 +129,5 @@ For ASR/TTS systems that only need the text brain, use `/realtime` or `/v1/realt
 ## Server-Side Agent Loop
 
 `/v1/chat/completions` can run an internal tool loop when `agent_loop: true`. This lets clients collapse multiple model/tool round trips into one daemon request. Daemon tool calls execute inline; client-owned tool calls can still be yielded in OpenAI-compatible shape.
+
+Ollama-backed loops use its native `/api/chat` tool protocol. `timeout_s` applies to each backend round, while `agent_timeout_s` bounds the complete loop and defaults to 45 seconds. The planning turn is capped at 96 output tokens. By default Omnius executes one daemon-tool round, then removes daemon schemas for the final synthesis; set `agent_max_tool_rounds` only when a workflow genuinely needs deeper tool chaining. Omnius returns a typed HTTP 504 when the total budget expires and HTTP 508 when a model repeats the same daemon tool with identical arguments. Tool results are capped at 6,000 characters before the next prompt. Without `daemon_tool_names`, Omnius offers only a compact core catalog permitted by `include_daemon_tools`; `prompt_template: "factual-first"` narrows it further to `web_search` and `web_fetch`.

package/docs/rest/endpoints/voice-vision.md

@@ -115,7 +115,7 @@ pipeline payload in `result.data`.
 ```bash
 curl -sS -X POST http://127.0.0.1:11435/v1/ocr/advanced \
   -H 'content-type: application/json' \
-  -d '{"args":{"image":"/data/invoice.png","language":"eng","psm":6},"timeout_ms":300000}'
+  -d '{"args":{"image":"/data/invoice.png","language":"eng","psm":6},"timeout_ms":90000}'
 ```
 
 Inference never invokes sudo, apt, pip, or venv creation. If readiness is
@@ -126,6 +126,22 @@ version results without requiring particular dpkg package names. The caller
 needs `run` scope because optional
 `output_dir`, batch, and debug modes write OCR artifacts.
 
+The managed pipeline is bounded: it starts with high-yield preprocessing and
+expands variants only for low-evidence or larger images. Small crops avoid the
+former all-variant/all-PSM explosion. The REST default is 90 seconds (maximum
+180 seconds), while the worker has an 80-second internal deadline so it can
+return diagnostics. A timeout or cancellation returns
+`result.data.schema=omnius.ocr-diagnostic.v1` with code `ocr_timeout` or
+`ocr_cancelled`; cancellation terminates the Python/Tesseract process group
+with TERM followed by KILL.
+
+OCR text is evidence-gated before it is returned. Strong text with adequate
+confidence is accepted. An empty or tiny non-substantive result is a successful
+`omnius.ocr-evidence.v1` `low_information` observation with diagnostic code
+`ocr_low_information`, not a fabricated transcript. High-volume low-confidence
+or symbol-heavy output is suppressed and returned as
+`ocr_evidence_rejected`; its raw text is not exposed as alternate OCR evidence.
+
 ## TTS
 
 `POST /v1/voice/tts` returns audio bytes. `format` can be `wav` or `pcm`. `X-Sample-Rate` reports the sample rate.
@@ -251,28 +267,23 @@ startup follows `OMNIUS_AUDIO_AUTO_SETUP` for acoustic and speaker; semantic
 startup additionally requires `OMNIUS_SEMANTIC_AUDIO_AUTO_SETUP=1`.
 
 On JetPack, speaker setup deliberately leaves `OMNIUS_AUDIO_PYTHON` (including
-an Egg project `.venv`) unchanged. It creates
-`~/.omnius/runtimes/audio/speaker/venv` with `--system-site-packages`, links
-that selected JetPack Torch 2.2/CUDA 12.2 provider from the managed venv only,
-and installs the exact checksummed upstream CPython 3.10/aarch64
-`torchaudio-2.2.0` wheel there with `--no-deps --no-index`. It never invokes a
-generic Torch/Torchaudio pip resolver or changes host/Egg Torch. Before it is
-ready, setup proves the effective Torch/Torchaudio 2.2 ABI and runs the
-CPU-only `torchaudio.compliance.kaldi.fbank` 80-bin probe. If an operator needs
-a vendor-specific wheel, both an HTTPS/local-wheel
-`OMNIUS_SPEAKER_TORCHAUDIO_WHEEL` and immutable
-`OMNIUS_SPEAKER_TORCHAUDIO_WHEEL_SHA256` must be supplied; an incomplete or
-failed ABI override remains unready with a remediation error rather than
-falling back to generic pip.
-
-The managed speaker venv also carries the checksummed `typing_extensions`
-support wheel required by JetPack Torch 2.2. This matters when a project venv
-can import that module only from its Python user site: speaker inference keeps
-`PYTHONNOUSERSITE=1`, so it never borrows mutable user-site packages at runtime.
-The bootstrap step inspects only CPython/site-package location before this
-repair; the managed-venv Torch import is authoritative. If another dependency
-is missing, readiness and setup preserve the exact missing module name instead
-of incorrectly reporting that the selected JetPack Torch is invalid.
+an Egg project `.venv`) unchanged. It creates a private, non-system-site
+`~/.omnius/runtimes/audio/speaker/venv`, then installs only a locked,
+checksum-verified CPython 3.10/aarch64 CPU dependency set: NumPy 1.26.4,
+ONNX Runtime 1.17.3, and the exact CPU support wheels required by that ONNX
+Runtime release. Each wheel is materialized and SHA-256 verified before the
+managed venv installs it with `--no-deps --no-index`.
+
+The worker implements the WeSpeaker CAM++ 80-bin Kaldi configuration in pure
+NumPy: 25 ms / 10 ms Hamming frames, dither disabled, Kaldi pre-emphasis and
+mel bank behavior, then full-clip CMN without CVN. Readiness invokes a
+versioned CPU preprocessing probe. It checks the fixed Kaldi configuration,
+feature shape, CMN, energy range, and expected mel-band profile with explicit
+cross-platform numeric tolerances before it can report ready. The speaker path neither imports
+nor links Torch or Torchaudio, so the generic `torchaudio-2.2.0` ABI mismatch
+cannot bind against, replace, or otherwise affect JetPack's CUDA-enabled Egg
+Torch. Inference remains install-free and network-free; failed package imports
+or calibration leave the role unready with the exact probe error.
 
 The dedicated route returns typed HTTP 503 whenever its requested runtime is
 unprovisioned or unavailable: `audio_classifier_not_ready` (acoustic),

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.639",
+  "version": "1.0.641",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.639",
+      "version": "1.0.641",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.639",
+  "version": "1.0.641",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/library.js",