omnius 1.0.807 → 1.0.809

package/docs/DISCOVERY.json

@@ -5730,7 +5730,7 @@
       "operations": {
         "post": {
           "summary": "Stateful direct chat with explicit autonomous-agent opt-in",
-          "description": "Drop-in replacement for OpenAI /v1/chat/completions and Ollama /api/chat. Requests use direct backend inference by default. Set tools=true or use_tools=true to explicitly run the full Omnius autonomous-agent stack. Explicit agent runs have a daemon-enforced whole-run deadline and duplicate-work suppression.",
+          "description": "Drop-in replacement for OpenAI /v1/chat/completions and Ollama /api/chat. Requests use direct backend inference by default. Set tools=true or use_tools=true to explicitly run the full Omnius autonomous-agent stack. Explicit agent runs reject duplicate concurrent work and continue until completion, failure, explicit Stop, or daemon shutdown. They have no whole-run wall-clock deadline; legacy agent_timeout_s and OMNIUS_CHAT_AGENT_TIMEOUT_S are ignored.",
           "tags": [
             "Chat"
           ],
@@ -5798,10 +5798,8 @@
                     },
                     "agent_timeout_s": {
                       "type": "number",
-                      "minimum": 0.05,
-                      "maximum": 3600,
-                      "default": 180,
-                      "description": "Daemon-enforced whole-agent wall-clock deadline. Distinct from timeout_s."
+                      "deprecated": true,
+                      "description": "Ignored compatibility field. Agent runs have no whole-run wall-clock deadline."
                     }
                   }
                 }
@@ -6096,7 +6094,7 @@
               "$ref": "#/components/parameters/MinimumOmniusVersion"
             }
           ],
-          "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.",
+          "description": "Standard OpenAI chat-completion proxy by default. Set agent_loop=true to enter a 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 loops have no whole-run wall-clock deadline. Legacy agent_timeout_s and OMNIUS_AGENT_LOOP_TOTAL_TIMEOUT_MS are ignored. 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": {
@@ -6150,7 +6148,7 @@
                     },
                     "timeout_s": {
                       "type": "number",
-                      "description": "Q1 — per-request timeout in seconds (1-3600). Overrides OMNIUS_BACKEND_TIMEOUT_S."
+                      "description": "Per-backend-request liveness timeout in seconds, capped at 3600. Overrides OMNIUS_BACKEND_TIMEOUT_S."
                     },
                     "agent_loop": {
                       "type": "boolean",
@@ -6177,10 +6175,8 @@
                     },
                     "agent_timeout_s": {
                       "type": "number",
-                      "minimum": 5,
-                      "maximum": 600,
-                      "default": 45,
-                      "description": "Total server-side agent-loop deadline. Distinct from per-backend timeout_s."
+                      "deprecated": true,
+                      "description": "Ignored compatibility field. Agent loops have no whole-run wall-clock deadline."
                     },
                     "agent_max_tool_rounds": {
                       "type": "integer",
@@ -6285,7 +6281,7 @@
               }
             },
             "504": {
-              "description": "Backend round or total agent-loop deadline expired"
+              "description": "A backend request exceeded its liveness timeout"
             },
             "508": {
               "description": "The model repeated an identical daemon tool call"
@@ -29211,6 +29207,37 @@
         }
       ]
     },
+    {
+      "id": "guide.corpus-ocr",
+      "kind": "guide",
+      "title": "Explicit scanned-PDF corpus OCR",
+      "summary": "The public sourcecorpus tool can process an acquired PDF population with begin using filters: {channel: \"ocrtext\"} and optional ocrlanguage (default eng). Repeated process calls resume a durable per-source page cursor. Installed CPU Poppler and Tesseract are required. The default channel remains nativetext; completing native extraction of an empty scanned pa",
+      "keywords": [
+        "corpus",
+        "ocr",
+        "md"
+      ],
+      "maturity": "stable",
+      "audiences": [
+        "user",
+        "integrator",
+        "coding-agent"
+      ],
+      "layer": "documentation",
+      "interfaces": [
+        {
+          "type": "file",
+          "target": "docs/corpus-ocr.md"
+        }
+      ],
+      "references": [
+        {
+          "type": "documentation",
+          "target": "docs/corpus-ocr.md",
+          "relation": "canonical-artifact"
+        }
+      ]
+    },
     {
       "id": "guide.dedup-false-positive-meta-analysis",
       "kind": "guide",
@@ -29931,6 +29958,38 @@
         }
       ]
     },
+    {
+      "id": "guide.guides-pdf-collections",
+      "kind": "guide",
+      "title": "Large PDF collections",
+      "summary": "Ask Omnius to process the collection in ordinary language, specifying the result you need: search, ordering, original-page manipulation, summary, or extraction. For example: “Review every supplied PDF, extract each recorded duration with its source page, order the documents by the invoice date, and deliver the complete results.” A field's meaning, units, mis",
+      "keywords": [
+        "guides",
+        "pdf",
+        "collections",
+        "md"
+      ],
+      "maturity": "stable",
+      "audiences": [
+        "user",
+        "integrator",
+        "coding-agent"
+      ],
+      "layer": "documentation",
+      "interfaces": [
+        {
+          "type": "file",
+          "target": "docs/guides/pdf-collections.md"
+        }
+      ],
+      "references": [
+        {
+          "type": "documentation",
+          "target": "docs/guides/pdf-collections.md",
+          "relation": "canonical-artifact"
+        }
+      ]
+    },
     {
       "id": "guide.guides-realtime",
       "kind": "guide",
@@ -45026,6 +45085,62 @@
         }
       ]
     },
+    {
+      "id": "tool.pdf-transform",
+      "kind": "tool",
+      "title": "Pdf Transform",
+      "summary": "pdf_transform is a directly callable Omnius tool.",
+      "aliases": [
+        "pdf_transform"
+      ],
+      "keywords": [
+        "tool",
+        "pdf",
+        "transform"
+      ],
+      "maturity": "stable",
+      "layer": "execution",
+      "audiences": [
+        "coding-agent",
+        "integrator",
+        "service-agent"
+      ],
+      "use_when": [
+        "The caller needs this isolated operation and live metadata confirms direct-call availability"
+      ],
+      "avoid_when": [
+        "The live registry reports the tool unavailable or the caller lacks its required scope"
+      ],
+      "interfaces": [
+        {
+          "type": "rest-schema",
+          "target": "/v1/tools/pdf_transform"
+        },
+        {
+          "type": "rest-call",
+          "target": "/v1/tools/pdf_transform/call"
+        }
+      ],
+      "references": [
+        {
+          "type": "source",
+          "target": "packages/execution/src/tools/pdf-transform.ts",
+          "relation": "implementation"
+        }
+      ],
+      "direct_callable": true,
+      "live_metadata": "GET /v1/tools/pdf_transform",
+      "source_of_truth": [
+        "packages/execution/src/tools/pdf-transform.ts",
+        "GET /v1/tools/pdf_transform"
+      ],
+      "verification": [
+        {
+          "check": "Inspect GET /v1/tools/pdf_transform before use",
+          "expected": "Schema, security, exposure, and availability match the intended invocation"
+        }
+      ]
+    },
     {
       "id": "tool.pdf-visual-markdown",
       "kind": "tool",

package/docs/DISCOVERY.md

@@ -420,6 +420,7 @@ Daemon equivalents are `GET /v1/discovery/bootstrap`, `GET /v1/discovery?q=<inte
 | `guide.asd-ste100-communication-audit-uppercase` | ASD-STE100 Communication Audit | Use ASD-STE100 Issue 9 for all Omnius natural-language communication. Apply the rule to external replies and internal agent messages. Preserve exact machine content without a change. |
 | `guide.concept-relational-language` | Concept Relational Language (CRL) — Token-Efficient Concept Communication | &gt; Experimental hyper-compressed symbolic notation for mind-mapped logical flow tracking |
 | `guide.context-management-medium-models-proposal` | Context Management for Medium Models (30-40B) — Implementation Proposal | Date: 2026-04-25 Problem: Medium-tier models (~35B params) exhibit excessive repetition at ~35% context fill despite 256K context windows Root Cause: Monolithic context structure + attention degradation + tool schema bloat |
+| `guide.corpus-ocr` | Explicit scanned-PDF corpus OCR | The public sourcecorpus tool can process an acquired PDF population with begin using filters: {channel: "ocrtext"} and optional ocrlanguage (default eng). Repeated process calls resume a durable per-source page cursor. Installed CPU Poppler and Tesseract are required. The default channel remains nativetext; completing native extraction of an empty scanned pa |
 | `guide.dedup-false-positive-meta-analysis` | Meta-Analysis: False Positive Duplicate Tool Call Detection | Date: 2026-05-14 Scope: packages/orchestrator/src/agenticRunner.ts — proactivePrune() + buildResourceKey() |
 | `guide.discovery-agent-map` | Agent Map | Bundled Omnius documentation artifact: docs/discovery/agent-map.json. |
 | `guide.discovery-catalog-overrides` | Discovery Catalog Curated Source | Curated aliases, relationships, invocation modes, and conceptual entries merged into the generated catalog. |
@@ -442,6 +443,7 @@ Daemon equivalents are `GET /v1/discovery/bootstrap`, `GET /v1/discovery?q=<inte
 | `guide.guides-long-horizon-feature-workflow` | Long-horizon feature work | Implementation, acceptance cases and delivery records are tracked in WO-01. |
 | `guide.guides-media-generation` | Media Generation | Omnius exposes media generation through TUI commands, tools, Telegram creative workflows, and sponsor media endpoints. |
 | `guide.guides-osint-research` | Categorized OSINT research | Omnius ships a local, searchable snapshot of the OSINT4ALL Start.me page. It uses the same compact discovery pattern as AIWG skills: |
+| `guide.guides-pdf-collections` | Large PDF collections | Ask Omnius to process the collection in ordinary language, specifying the result you need: search, ordering, original-page manipulation, summary, or extraction. For example: “Review every supplied PDF, extract each recorded duration with its source page, order the documents by the invoice date, and deliver the complete results.” A field's meaning, units, mis |
 | `guide.guides-realtime` | Realtime Conversations | Realtime mode is for short, natural, back-and-forth spoken conversation behind ASR and TTS. |
 | `guide.guides-sponsor-and-cohere` | Sponsor And COHERE Mesh | Sponsor and COHERE let Omnius share inference capacity across peers. Sponsor mode is explicit provider/consumer sharing. COHERE is a distributed cognitive inference mesh. |
 | `guide.guides-system-tray` | System Tray Indicator | Omnius ships a native per-login tray indicator for Linux, macOS, and Windows on x64 hosts. It is separate from the REST daemon: closing the indicator never stops the daemon, and the daemon remains usable in headless sessions. |
@@ -830,6 +832,7 @@ Daemon equivalents are `GET /v1/discovery/bootstrap`, `GET /v1/discovery?q=<inte
 | `tool.osint-search` | OSINT Search | Search the bundled OSINT4ALL snapshot locally and return a bounded ranked shortlist with category, archived-health, interaction, and risk metadata; no cataloged service is contacted. |
 | `tool.osint-show` | OSINT Show | Expand one exact local OSINT resource ID into its URL, aliases, original category placements, archived health, and safe next-tool route; no cataloged service is contacted. |
 | `tool.pdf-to-text` | Pdf To Text | pdf_to_text is a directly callable Omnius tool. |
+| `tool.pdf-transform` | Pdf Transform | pdf_transform is a directly callable Omnius tool. |
 | `tool.pdf-visual-markdown` | Pdf Visual Markdown | pdf_visual_markdown is a directly callable Omnius tool. |
 | `tool.phase-recall` | Phase Recall | phase_recall is a directly callable Omnius tool. |
 | `tool.play-sound` | Play Sound | play_sound is a directly callable Omnius tool. |

package/docs/corpus-ocr.md

@@ -0,0 +1,57 @@
+# Explicit scanned-PDF corpus OCR
+
+The public `source_corpus` tool can process an acquired PDF population with
+`begin` using `filters: {channel: "ocr_text"}` and optional `ocr_language`
+(default `eng`). Repeated `process` calls resume a durable per-source page
+cursor. Installed CPU Poppler and Tesseract are required. The default channel
+remains `native_text`; completing native extraction of an empty scanned page
+does not complete its OCR extraction.
+
+Each OCR page retains the original PDF identity, rendered PNG, full Tesseract
+text and TSV geometry in durable content-addressed files, with exact hashes,
+page lineage, engine version and language. Immutable segmented evidence powers
+bounded search and read operations. OCR receipts preserve the original issuing
+owner; subsequent reads bind their exact exposed range to the current owner.
+OCR TSV geometry is separate from native PDF geometry.
+
+`corpus_work` can explicitly select the complete `ocr_text` population for
+anchored extraction or summary. A read checkpoint is prepared when text is
+returned, then confirmed only after the production runner successfully sends
+the exact protected OCR text carrier to its configured backend. Failed
+transport cannot confirm that read or authorize a derived commit. Derived
+content remains `actor_attested`, requiring independent factual review.
+
+OCR completion means the engine processed every selected page. It does not
+prove transcription accuracy or visual comprehension: `visualCoverage`
+remains `not_assessed`, and OCR does not receive native evidence authority.
+Complete OCR-derived text exports can pass the separate export reviewer after
+it reopens every original PDF, rendered page, OCR text, TSV and immutable
+source/producer/exposure record. Fresh bounded factual review sees explicit
+`ocr_text` transcription provenance and the unassessed visual limitation.
+The complete source and output unions remain required before retention and
+provider delivery. Exact numeric source-line transcription can be verified
+against OCR engine output; this does not verify that output against the image.
+
+Native table-cell, count, sum and aggregate quantity gates remain strict: OCR
+text and TSV do not confer native field or arithmetic authority. OCR-derived
+arithmetic needs a separately reviewed transcription/operand population before
+approval; matching digits or a prose verdict cannot provide it. The inline
+Telegram source-answer review path still has no OCR-specific source-index
+adapter; the verified delivery route here is the full derived text export.
+Both OCR-only and mixed native/OCR factual inline candidates are refused after
+an actual successful OCR model handoff. The barrier is durable for the admitted
+session/run/epoch; ingestion, metadata and failed transport do not create it.
+The refusal preserves existing OCR evidence and gives the exact work/export/
+review steps instead of requesting another native-empty read.
+
+After fresh `corpus_export_review` completes under the current request, use
+`task_complete` with `deliver_artifacts: true` and the exact neutral notice
+`The reviewed OCR results are attached.` The host reopens the complete export,
+its original OCR source proof and current signed semantic receipt. Completion
+still requires actual provider acknowledgement for every selected output.
+That file review does not approve unrelated inline factual prose, and the
+neutral notice alone cannot waive missing review or delivery receipts.
+
+Population traversal and page work are resumable and have no total source or
+page count ceiling. A single-page rasterization still needs physical CPU/RAM
+resources; this implementation does not tile unusually large page images.

package/docs/guides/pdf-collections.md

@@ -0,0 +1,76 @@
+# Large PDF collections
+
+Ask Omnius to process the collection in ordinary language, specifying the result
+you need: search, ordering, original-page manipulation, summary, or extraction.
+For example: “Review every supplied PDF, extract each recorded duration with its
+source page, order the documents by the invoice date, and deliver the complete
+results.” A field's meaning, units, missing-value handling, and ordering direction
+remain part of the request; a date appearing somewhere in a PDF does not by
+itself establish that it is the requested invoice date.
+
+The collection is stored outside the model context. Original PDFs, complete
+native page text, page geometry, immutable source identities and hashes remain
+available in the private source archive. SQL applies scope, version and request
+filters before sorting and paging. Retrieval can reopen exact source ranges;
+condensing a model prompt never replaces the original evidence.
+
+Search uses cursor-paged lexical matches in the authorized collection. A search
+result, including a result with no matches, does not establish that every source
+was extracted or reviewed. Exhaustive requests retain the complete selected
+population and track source coverage separately from relevance ranking.
+
+There is no application limit on the total number of PDFs, pages, selected
+records, or output files. Per-call work widths keep memory and prompts finite;
+the model's actual context window, available storage, compute and each provider's
+per-file upload limit still apply. Processing the entire collection takes work
+proportional to that collection. Targeted search and exhaustive extraction use
+different coverage requirements.
+
+| Task | Runtime tools and retained result |
+| --- | --- |
+| Ingest and search | `source_corpus` retains originals, indexes native pages, and returns scoped inventories and exact ranges. Telegram's `manifest="current_task"` selects the host-admitted attachment population. |
+| Scanned pages | Explicit `source_corpus` OCR work retains page images, full engine text, word positions and engine receipts in the separate `ocr_text` channel. Interrupted page work resumes without replacing native text. |
+| Summary and extraction | `corpus_work` preserves the complete request and selection, issues bounded input claims, records ranges that actually reached a successful model request, and commits derived records with source anchors. |
+| Field ordering | `corpus_work` stores source-anchored typed fields and orders through SQL. Exact decimal and date handling, ties, missing values and multiple values have explicit policies. Model-selected field meaning remains subject to source review. |
+| Original-page manipulation | `pdf_corpus_recipe` and `pdf_corpus_transform` plan and execute original-page ordering, merge and split operations without reconstructing PDFs from summaries. |
+| Factual acceptance | `source_answer_review` and `corpus_export_review` independently review complete selected output and source coverage. Numeric aggregates reopen every selected native input and verify exact operands and arithmetic. |
+| Large result delivery | `corpus_export_review` partition actions retain upload-sized data and reconstruction files. `partition_finalize` inherits a complete current parent review and verifies every child; actual provider receipts still establish delivery. |
+
+Processing and delivery checkpoints survive restart. Committed work and accepted
+uploads are reused under the current task; stale owner or request authority must
+be renewed. A Telegram upload quota wait retains the exact reviewed candidate
+and resumes at the provider retry time. Completion requires receipts for every
+selected output, including all reconstruction files. An extracted index, a model
+claim of completion, or an intermediate progress message does not establish that
+the final answer was delivered.
+
+Directory acquisition preserves enumeration, original-file acquisition and
+manifest-writing checkpoints. Restart verifies the previously observed directory
+and file identities before continuing; a changing directory cannot silently
+become a different collection. This is an observed, verified manifest, not an
+atomic filesystem snapshot. Export writing and fixity checks also resume within
+an individual JSONL record, so output size does not impose a total export cap.
+
+Verified durable work can continue beyond the default emergency turn budget.
+An explicit finite budget still applies. Repeating a call without advancing its
+current source-bound checkpoint earns no continuation credit; progress does not
+approve facts or fulfill the final-output and delivery requirements.
+
+After restart, the host reopens the current pending SQL job before applying its
+retry guard. An authentic job can receive one admission attempt at its existing
+producer clock, including an initialized job that has not advanced yet. Reopening
+the job creates no progress event. The consumed attempt survives restart and
+cannot be renewed by switching job IDs; actual verified work must advance the
+clock before another attempt becomes available.
+
+Native text and geometry can be incomplete or ambiguous. Unresolved extraction,
+field association or arithmetic stays explicit; visual review can resolve the
+relevant page on the configured capable backend. The research basis and
+stage-by-stage evaluation criteria are in the
+[PDF distillation synthesis](../../.aiwg/research/pdf-knowledge-distillation/synthesis.md).
+
+Numeric review distinguishes populated cells from recorded events. A nonempty
+cell can contain zero or an incomplete time placeholder; it does not necessarily
+record an event. Bounded examples expose exact distinct native spellings without
+frequencies or expected answers. The reviewer selects the meaning from the full
+request, and the host verifies arithmetic against the complete original field.

package/docs/guides/telegram.md

@@ -145,6 +145,12 @@ a lease, bounded backoff, and the returned Telegram message ID.
 | Audio | `audio_edit` | inspect, trim, concatenate, typed splice and silence, mix, normalize, fade, convert, resample | `audio_analyze` for sound/music or `transcribe_file` for speech |
 | PDF or DOCX | `document` | create from heading/paragraph/bullet/page-break blocks, inspect, hash-guarded revise | `document action='inspect'` on the exact output |
 
+Photo intake and `telegram_image_analyze` use the configured endpoint and model
+with the original image. The backend must declare image support. If that
+capability is unavailable, Omnius retains the image and reports it as unassessed.
+These paths do not automatically select another vision model. An uncaptioned
+photo retains its native media identity without inventing a caption or request.
+
 These tools do not accept shell commands, raw FFmpeg filters, ImageMagick
 expressions, raw PDF/OOXML bytes, HTML, macros, or paths outside the chat
 workspace. Image and audio edits never overwrite an artifact. Document

package/docs/rest/endpoints/chat.md

@@ -43,11 +43,11 @@ Important body fields:
 | `tools` | array | OpenAI-shape tools array |
 | `tool_choice` | any | OpenAI-shape tool choice |
 | `parallel_tool_calls` | boolean | Forwarded to backend |
-| `timeout_s` | number | Per-request timeout |
+| `timeout_s` | number | Per-backend-request liveness timeout |
 | `agent_loop` | boolean | Run server-side tool loop |
 | `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_timeout_s` | number | Deprecated compatibility field; ignored |
 | `agent_max_tool_rounds` | integer | Daemon tool rounds before forced final synthesis; default 1 |
 | `agent_prefetch_web_search` | boolean | Explicit one-generation web-search prefetch; factual-first enables it automatically |
 | `max_turns` | integer | Server-side agent loop turn cap |
@@ -71,7 +71,7 @@ Realtime example:
 
 ## `/v1/chat`
 
-This is the Omnius stateful chat endpoint. It uses direct backend inference by default. Set `tools: true` or `use_tools: true` to explicitly launch the full Omnius agent stack with tools, memory, skills, and multi-agent context. `realtime: true` always uses the direct backend path. Explicit agent runs default to a 180-second daemon-owned wall-clock deadline and reject equivalent concurrent work even when callers omit `session_id`.
+This is the Omnius stateful chat endpoint. It uses direct backend inference by default. Set `tools: true` or `use_tools: true` to explicitly launch the full Omnius agent stack with tools, memory, skills, and multi-agent context. `realtime: true` always uses the direct backend path. Explicit agent runs reject equivalent concurrent work even when callers omit `session_id` and continue until completion, failure, explicit Stop, or daemon shutdown. Agent runs have no whole-run wall-clock deadline; legacy `agent_timeout_s` and `OMNIUS_CHAT_AGENT_TIMEOUT_S` are ignored.
 
 Body fields:
 
@@ -84,7 +84,7 @@ Body fields:
 | `tools` | boolean | Set true to explicitly launch the full agent stack |
 | `use_tools` | boolean | Omnius-native alias for `tools: true` |
 | `timeout_s` | number | Per-backend-request liveness timeout |
-| `agent_timeout_s` | number | Whole-agent daemon deadline, default 180 seconds and maximum 3600 |
+| `agent_timeout_s` | number | Deprecated compatibility field; ignored |
 | `realtime` | boolean | Short ASR/TTS conversation mode |
 | `realtime_options` | object | Realtime settings |
 
@@ -144,4 +144,4 @@ For ASR/TTS systems that only need the text brain, use `/realtime` or `/v1/realt
 
 `/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. For a normal loop, the planning turn is capped at 96 output tokens; one daemon-tool round is the default, after which Omnius removes daemon schemas for final synthesis. `prompt_template: "factual-first"` skips that model-planning round entirely: Omnius executes the already-mandated, authorized `web_search` using the latest user turn, inserts the protocol-correct tool evidence, and performs one grounded synthesis generation. `agent_prefetch_web_search: true` opts into the same path directly. Generic/deeper tool workflows remain available through `agent_max_tool_rounds`. Omnius returns typed HTTP 504/508 failures, caps tool evidence at 6,000 characters, and limits the implicit catalog unless `daemon_tool_names` requests exact additions.
+Ollama-backed loops use its native `/api/chat` tool protocol. `timeout_s` applies to each backend round. There is no whole-loop wall-clock deadline; legacy `agent_timeout_s` and `OMNIUS_AGENT_LOOP_TOTAL_TIMEOUT_MS` are ignored. Each daemon tool retains its own execution timeout and cancellation handling, and the loop still respects client disconnects, `max_turns`, and repeated-call guards. For a normal loop, the planning turn is capped at 96 output tokens; one daemon-tool round is the default, after which Omnius removes daemon schemas for final synthesis. `prompt_template: "factual-first"` skips that model-planning round entirely: Omnius executes the already-mandated, authorized `web_search` using the latest user turn, inserts the protocol-correct tool evidence, and performs one grounded synthesis generation. `agent_prefetch_web_search: true` opts into the same path directly. Generic/deeper tool workflows remain available through `agent_max_tool_rounds`. Backend liveness timeouts return typed HTTP 504 failures, and identical repeated tool calls return HTTP 508. Omnius caps tool evidence at 6,000 characters and limits the implicit catalog unless `daemon_tool_names` requests exact additions.

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.807",
+  "version": "1.0.809",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.807",
+      "version": "1.0.809",
       "bundleDependencies": [
         "image-to-ascii"
       ],