omnius 1.0.807 → 1.0.809

package/README.md

@@ -133,6 +133,11 @@ the local discover → exact expansion → explicit web-tool workflow.
 
 ## Common Workflows
 
+For large PDF collections, Omnius retains originals in a durable scoped index
+and processes ordering, manipulation, summary and extraction in resumable steps.
+See the [PDF collections guide](docs/guides/pdf-collections.md) for coverage,
+source provenance, numeric checks and complete delivery of large results.
+
 ```bash
 omnius "inspect this repo and summarize the main entrypoints"
 omnius serve

package/dist/api/py-embed.js

@@ -294593,9 +294593,127 @@ init_vision();
 init_system_deps();
 init_vision();
 
+// packages/execution/dist/tools/pdf-corpus-archive.js
+init_process_async();
+
 // packages/execution/dist/tools/ocr-pdf.js
 init_process_async();
 
+// packages/execution/dist/tools/pdf-transform.js
+init_process_async();
+var BUFFER_BYTES = 128 * 1024;
+var MAX_RECORD_BYTES = 64 * 1024;
+
+// packages/execution/dist/tools/source-corpus.js
+var SEGMENT_CHARS = 64 * 1024;
+var RECORD_BYTES = 64 * 1024;
+var ACTIONS = /* @__PURE__ */ new Set(["ingest", "inventory", "search", "read", "begin", "process", "status"]);
+var SOURCE_CORPUS_TOOL_DEFINITION = {
+  name: "source_corpus",
+  aliases: ["pdf_corpus"],
+  description: "Acquire and process a scoped PDF corpus without a total document/page-count cap. ingest uses an immutable JSONL manifest, one {ref,sha256} per source, and returns a durable import ID; repeat it to resume bounded batches. Local directory acquisition requires an explicit host policy and resumes bounded traversal/hash/manifest work with ingest_id. It captures an observed, version-fenced filesystem selection, not an atomic live-tree snapshot, and excludes only this workspace's owned .omnius state tree. inventory/search are cursor-paged; read returns exact channel-labelled segments with character offsets. begin preserves the full processing request and selected source versions; process checkpoints complete native pages and resumes after cancellation/restart. Source bytes, full native text and bbox are archived before bounded views. Explicit filters.channel='ocr_text' schedules resumable CPU OCR; rendered pages, engine/language, full OCR text and TSV are preserved separately. For OCR summaries/extractions use corpus_work filters.channel='ocr_text', complete anchored work, export, and corpus_export_review before delivery; source_answer_review only accepts native facts. Native or OCR engine completion does not prove visual understanding, summaries or final delivery.",
+  parameters: { type: "object", properties: {
+    action: { type: "string", enum: [...ACTIONS] },
+    corpus_id: { type: "string" },
+    manifest: { type: "string", description: "Host-authorized JSONL file or alias; one {ref,sha256} record per PDF." },
+    manifest_sha256: { type: "string" },
+    directory: { type: "string", description: "Explicit local-admin directory enumeration only." },
+    ingest_id: { type: "string", description: "Resume directory acquisition or the immutable import without resupplying its directory/manifest. Use a new ID for a deliberately changed directory selection." },
+    source_budget: { type: "integer", minimum: 1, maximum: 64 },
+    operation_id: { type: "string" },
+    request: { type: "string", description: "Complete extraction processing request; not authority to replace the user's goal." },
+    page_budget: { type: "integer", minimum: 1, maximum: 64 },
+    page_size: { type: "integer", minimum: 1, maximum: 128 },
+    cursor: { type: "string" },
+    query: { type: "string" },
+    filters: { type: "object" },
+    direction: { type: "string", enum: ["asc", "desc"] },
+    page_id: { type: "string" },
+    ocr_language: { type: "string", description: "Explicit OCR operation language, e.g. eng or eng+deu. Immutable after begin; uses installed Tesseract languages." },
+    char_offset: { type: "integer", minimum: 0 },
+    max_chars: { type: "integer", minimum: 2, maximum: SEGMENT_CHARS }
+  }, required: ["action", "corpus_id"], additionalProperties: false }
+};
+
+// packages/execution/dist/tools/corpus-work.js
+var ACTIONS2 = ["begin", "next", "read", "commit", "status", "complete", "cancel", "export", "aggregate", "order"];
+var CORPUS_WORK_TOOL_DEFINITION = {
+  name: "corpus_work",
+  aliases: ["pdf_corpus_work"],
+  description: "Durable model-driven summary or extraction over a scoped corpus. Native text is default; explicitly select filters.channel='ocr_text' for previously processed OCR evidence. begin preserves the FULL request and immutable selection. next gives bounded leased claims; read exposes each complete input through resumable exact character ranges. commit requires every input fully delivered through a successful host model request and exact quote anchors for every input. A read result prepares an exact range; same-response read+commit cannot establish model exposure. Full derived content is archived, never substituted for originals. Summary parents use host-issued bounded child membership and require complete child reads. Repeat next until leaves and reduction finish. Processing coverage is actor-attested, not proof of factual correctness, visual understanding or final delivery. Optional anchored decimal/count records support exact deterministic aggregates; record selection remains actor-attested. Exact anchored typed fields support SQL ordering before pagination through order, after processing completes. Text uses explicit binary collation; decimal is exact, dates calendar, clocks require declared fixed/source offsets. Missing and multiple values require explicit policies; field meaning and selection remain actor-attested. export resumably streams all derived leaf records and the summary root to JSONL. Repeat the exact export while completeExport=false; partial work has no finished artifact. byte_budget bounds each call, not the final output. Public OCR results require corpus_export_review prepare/process/finalize on the returned retained artifactPath; source_answer_review does not ground OCR inline prose. There is no total corpus or output count cap.",
+  parameters: { type: "object", properties: {
+    action: { type: "string", enum: [...ACTIONS2] },
+    corpus_id: { type: "string" },
+    operation_id: { type: "string" },
+    mode: { type: "string", enum: ["summary", "extraction"] },
+    request: { type: "string" },
+    filters: { type: "object" },
+    query: { type: "string" },
+    direction: { type: "string", enum: ["asc", "desc"] },
+    fan_in: { type: "integer", minimum: 2, maximum: 64, description: "Per-parent child allocation, not total corpus limit." },
+    byte_budget: { type: "integer", minimum: 1, maximum: 67108864, description: "Per-call export write/fixity byte budget; repeat export to finish arbitrarily large output." },
+    page_size: { type: "integer", minimum: 1, maximum: 32 },
+    claim_id: { type: "string" },
+    input_id: { type: "string" },
+    char_offset: { type: "integer", minimum: 0 },
+    max_chars: { type: "integer", minimum: 2, maximum: 65536 },
+    content: { type: "string", description: "Complete nonempty model-derived output for this claim." },
+    anchors: { type: "array", minItems: 1, items: { type: "object", properties: {
+      input_id: { type: "string" },
+      char_start: { type: "integer", minimum: 0 },
+      char_end: { type: "integer", minimum: 0 },
+      quote: { type: "string" }
+    }, required: ["input_id", "char_start", "char_end", "quote"], additionalProperties: false } },
+    fields: { type: "array", items: { type: "object", properties: {
+      field: { type: "string" },
+      kind: { type: "string", enum: ["text", "decimal", "date", "time", "datetime"] },
+      anchor: { type: "integer", minimum: 0 },
+      literal: { type: "string", description: "Entire exact literal in the indicated source anchor." },
+      timezone: { type: "string", description: "Dates: calendar. Clocks: source_offset when the literal has Z/offset; otherwise UTC or a fixed +/-HH:MM offset. No inferred timezone." }
+    }, required: ["field", "kind", "anchor", "literal"], additionalProperties: false } },
+    order_spec: { type: "object", properties: {
+      field: { type: "string" },
+      kind: { type: "string", enum: ["text", "decimal", "date", "time", "datetime"] },
+      unit: { type: "string", enum: ["source", "page"] },
+      direction: { type: "string", enum: ["asc", "desc"] },
+      missing: { type: "string", enum: ["first", "last", "exclude", "error"] },
+      multiple: { type: "string", enum: ["min", "max", "error"] },
+      collation: { type: "string", enum: ["binary"] },
+      timezone: { type: "string", enum: ["calendar", "recorded_offset"] },
+      timeSemantics: { type: "string", enum: ["utc_clock"] },
+      where: { type: "object", properties: { equals: { type: "string" }, from: { type: "string" }, through: { type: "string" } }, additionalProperties: false }
+    }, required: ["field", "kind", "unit", "direction", "missing", "multiple"], additionalProperties: false },
+    cursor: { type: "string" },
+    numbers: { type: "array", items: { type: "object", properties: {
+      field: { type: "string" },
+      kind: { type: "string", enum: ["decimal", "count"] },
+      anchor: { type: "integer", minimum: 0 },
+      literal: { type: "string" }
+    }, required: ["field", "kind", "anchor"], additionalProperties: false } },
+    field: { type: "string" }
+  }, required: ["action", "corpus_id", "operation_id"], additionalProperties: false }
+};
+
+// packages/execution/dist/tools/pdf-corpus-transform.js
+init_process_async();
+var LINE_BYTES = 64 * 1024;
+var ACTIONS3 = /* @__PURE__ */ new Set(["begin", "process", "status"]);
+var PDF_CORPUS_TRANSFORM_TOOL_DEFINITION = {
+  name: "pdf_corpus_transform",
+  aliases: ["pdf_corpus_compose"],
+  description: "Resumable original-page PDF merge, split and reorder over an acquired source_corpus. begin accepts an immutable JSONL recipe with one {sourceId,sourceSha256,page} per output page; order and repetitions are exact. process commits bounded partial PDFs and a durable merge tree; repeat the same operation_id after cancellation/update/restart. Source references must already belong to this host-scoped corpus. No text/OCR extraction is required for page copying. Returns a canonical PDF and streamed flattened original-page lineage with exact hashes; document metadata/forms/signatures are not preserved. No total PDF/page-count cap; each physical merge is atomic and may require a longer host work deadline for very large containers.",
+  parameters: { type: "object", properties: {
+    action: { type: "string", enum: [...ACTIONS3] },
+    corpus_id: { type: "string" },
+    operation_id: { type: "string" },
+    operation: { type: "string", enum: ["merge", "split", "reorder"] },
+    request: { type: "string", description: "Complete caller-declared transformation request; does not replace host goal authority." },
+    manifest: { type: "string" },
+    manifest_sha256: { type: "string" },
+    work_budget: { type: "integer", minimum: 1, maximum: 64, description: "Bounded recipe import or leaf/merge work units for this invocation." }
+  }, required: ["action", "corpus_id", "operation_id"], additionalProperties: false }
+};
+
 // packages/execution/dist/tools/pdf-visual-markdown.js
 var MAX_IMAGE_BYTES = 8 * 1024 * 1024;
 

package/dist/index.js

@@ -1,4 +1,4 @@
 [diffend] Oversized file quarantined before diffing.
 name: package/dist/index.js
-size: 39849803 bytes
-sha256: 1e4dc06c6231b25e6176245614d6b3b3fe397fcc40fad72792c34bbd6b9ac772
+size: 41434427 bytes
+sha256: 19cfa3387615fd7148ce7841a0c4aed39284bab0e7018dfeacdf47de97e59a16

package/dist/library.js

@@ -47751,6 +47751,20 @@ var init_immutableEvidenceStore = __esm({
   }
 });
 
+// packages/memory/dist/artifactReviewSourceIndex.js
+var init_artifactReviewSourceIndex = __esm({
+  "packages/memory/dist/artifactReviewSourceIndex.js"() {
+    "use strict";
+  }
+});
+
+// packages/memory/dist/sourceCorpusStore.js
+var init_sourceCorpusStore = __esm({
+  "packages/memory/dist/sourceCorpusStore.js"() {
+    "use strict";
+  }
+});
+
 // packages/memory/dist/structureAwareChunker.js
 var init_structureAwareChunker = __esm({
   "packages/memory/dist/structureAwareChunker.js"() {
@@ -49093,6 +49107,8 @@ var init_dist = __esm({
     init_db();
     init_contextMemoryLedger();
     init_immutableEvidenceStore();
+    init_artifactReviewSourceIndex();
+    init_sourceCorpusStore();
     init_structureAwareChunker();
     init_evidenceIngestion();
     init_structuredMemoryStore();
@@ -51865,10 +51881,19 @@ var init_vision_action_loop = __esm({
   }
 });
 
+// packages/execution/dist/tools/pdf-corpus-archive.js
+var init_pdf_corpus_archive = __esm({
+  "packages/execution/dist/tools/pdf-corpus-archive.js"() {
+    "use strict";
+    init_process_async();
+  }
+});
+
 // packages/execution/dist/tools/ocr-pdf.js
 var init_ocr_pdf = __esm({
   "packages/execution/dist/tools/ocr-pdf.js"() {
     "use strict";
+    init_pdf_corpus_archive();
     init_process_async();
     init_pdf_document();
   }
@@ -51882,6 +51907,243 @@ var init_pdf_to_text = __esm({
   }
 });
 
+// packages/execution/dist/tools/pdf-transform.js
+var BUFFER_BYTES, MAX_RECORD_BYTES;
+var init_pdf_transform = __esm({
+  "packages/execution/dist/tools/pdf-transform.js"() {
+    "use strict";
+    init_process_async();
+    BUFFER_BYTES = 128 * 1024;
+    MAX_RECORD_BYTES = 64 * 1024;
+  }
+});
+
+// packages/execution/dist/tools/source-corpus-ocr.js
+var init_source_corpus_ocr = __esm({
+  "packages/execution/dist/tools/source-corpus-ocr.js"() {
+    "use strict";
+    init_dist();
+    init_pdf_corpus_archive();
+  }
+});
+
+// packages/execution/dist/tools/corpus-directory-manifest.js
+var init_corpus_directory_manifest = __esm({
+  "packages/execution/dist/tools/corpus-directory-manifest.js"() {
+    "use strict";
+    init_pdf_corpus_archive();
+  }
+});
+
+// packages/execution/dist/tools/source-corpus.js
+var SEGMENT_CHARS, RECORD_BYTES, ACTIONS, SOURCE_CORPUS_TOOL_DEFINITION;
+var init_source_corpus = __esm({
+  "packages/execution/dist/tools/source-corpus.js"() {
+    "use strict";
+    init_dist();
+    init_pdf_corpus_archive();
+    init_source_corpus_ocr();
+    init_corpus_directory_manifest();
+    SEGMENT_CHARS = 64 * 1024;
+    RECORD_BYTES = 64 * 1024;
+    ACTIONS = /* @__PURE__ */ new Set(["ingest", "inventory", "search", "read", "begin", "process", "status"]);
+    SOURCE_CORPUS_TOOL_DEFINITION = {
+      name: "source_corpus",
+      aliases: ["pdf_corpus"],
+      description: "Acquire and process a scoped PDF corpus without a total document/page-count cap. ingest uses an immutable JSONL manifest, one {ref,sha256} per source, and returns a durable import ID; repeat it to resume bounded batches. Local directory acquisition requires an explicit host policy and resumes bounded traversal/hash/manifest work with ingest_id. It captures an observed, version-fenced filesystem selection, not an atomic live-tree snapshot, and excludes only this workspace's owned .omnius state tree. inventory/search are cursor-paged; read returns exact channel-labelled segments with character offsets. begin preserves the full processing request and selected source versions; process checkpoints complete native pages and resumes after cancellation/restart. Source bytes, full native text and bbox are archived before bounded views. Explicit filters.channel='ocr_text' schedules resumable CPU OCR; rendered pages, engine/language, full OCR text and TSV are preserved separately. For OCR summaries/extractions use corpus_work filters.channel='ocr_text', complete anchored work, export, and corpus_export_review before delivery; source_answer_review only accepts native facts. Native or OCR engine completion does not prove visual understanding, summaries or final delivery.",
+      parameters: { type: "object", properties: {
+        action: { type: "string", enum: [...ACTIONS] },
+        corpus_id: { type: "string" },
+        manifest: { type: "string", description: "Host-authorized JSONL file or alias; one {ref,sha256} record per PDF." },
+        manifest_sha256: { type: "string" },
+        directory: { type: "string", description: "Explicit local-admin directory enumeration only." },
+        ingest_id: { type: "string", description: "Resume directory acquisition or the immutable import without resupplying its directory/manifest. Use a new ID for a deliberately changed directory selection." },
+        source_budget: { type: "integer", minimum: 1, maximum: 64 },
+        operation_id: { type: "string" },
+        request: { type: "string", description: "Complete extraction processing request; not authority to replace the user's goal." },
+        page_budget: { type: "integer", minimum: 1, maximum: 64 },
+        page_size: { type: "integer", minimum: 1, maximum: 128 },
+        cursor: { type: "string" },
+        query: { type: "string" },
+        filters: { type: "object" },
+        direction: { type: "string", enum: ["asc", "desc"] },
+        page_id: { type: "string" },
+        ocr_language: { type: "string", description: "Explicit OCR operation language, e.g. eng or eng+deu. Immutable after begin; uses installed Tesseract languages." },
+        char_offset: { type: "integer", minimum: 0 },
+        max_chars: { type: "integer", minimum: 2, maximum: SEGMENT_CHARS }
+      }, required: ["action", "corpus_id"], additionalProperties: false }
+    };
+  }
+});
+
+// packages/execution/dist/tools/corpus-population.js
+var init_corpus_population = __esm({
+  "packages/execution/dist/tools/corpus-population.js"() {
+    "use strict";
+  }
+});
+
+// packages/execution/dist/tools/source-corpus-workspace.js
+var init_source_corpus_workspace = __esm({
+  "packages/execution/dist/tools/source-corpus-workspace.js"() {
+    "use strict";
+    init_dist();
+    init_source_corpus();
+    init_corpus_population();
+  }
+});
+
+// packages/execution/dist/tools/corpus-derived-order.js
+var init_corpus_derived_order = __esm({
+  "packages/execution/dist/tools/corpus-derived-order.js"() {
+    "use strict";
+  }
+});
+
+// packages/execution/dist/tools/corpus-work.js
+var ACTIONS2, CORPUS_WORK_TOOL_DEFINITION;
+var init_corpus_work = __esm({
+  "packages/execution/dist/tools/corpus-work.js"() {
+    "use strict";
+    init_dist();
+    init_pdf_corpus_archive();
+    init_corpus_derived_order();
+    ACTIONS2 = ["begin", "next", "read", "commit", "status", "complete", "cancel", "export", "aggregate", "order"];
+    CORPUS_WORK_TOOL_DEFINITION = {
+      name: "corpus_work",
+      aliases: ["pdf_corpus_work"],
+      description: "Durable model-driven summary or extraction over a scoped corpus. Native text is default; explicitly select filters.channel='ocr_text' for previously processed OCR evidence. begin preserves the FULL request and immutable selection. next gives bounded leased claims; read exposes each complete input through resumable exact character ranges. commit requires every input fully delivered through a successful host model request and exact quote anchors for every input. A read result prepares an exact range; same-response read+commit cannot establish model exposure. Full derived content is archived, never substituted for originals. Summary parents use host-issued bounded child membership and require complete child reads. Repeat next until leaves and reduction finish. Processing coverage is actor-attested, not proof of factual correctness, visual understanding or final delivery. Optional anchored decimal/count records support exact deterministic aggregates; record selection remains actor-attested. Exact anchored typed fields support SQL ordering before pagination through order, after processing completes. Text uses explicit binary collation; decimal is exact, dates calendar, clocks require declared fixed/source offsets. Missing and multiple values require explicit policies; field meaning and selection remain actor-attested. export resumably streams all derived leaf records and the summary root to JSONL. Repeat the exact export while completeExport=false; partial work has no finished artifact. byte_budget bounds each call, not the final output. Public OCR results require corpus_export_review prepare/process/finalize on the returned retained artifactPath; source_answer_review does not ground OCR inline prose. There is no total corpus or output count cap.",
+      parameters: { type: "object", properties: {
+        action: { type: "string", enum: [...ACTIONS2] },
+        corpus_id: { type: "string" },
+        operation_id: { type: "string" },
+        mode: { type: "string", enum: ["summary", "extraction"] },
+        request: { type: "string" },
+        filters: { type: "object" },
+        query: { type: "string" },
+        direction: { type: "string", enum: ["asc", "desc"] },
+        fan_in: { type: "integer", minimum: 2, maximum: 64, description: "Per-parent child allocation, not total corpus limit." },
+        byte_budget: { type: "integer", minimum: 1, maximum: 67108864, description: "Per-call export write/fixity byte budget; repeat export to finish arbitrarily large output." },
+        page_size: { type: "integer", minimum: 1, maximum: 32 },
+        claim_id: { type: "string" },
+        input_id: { type: "string" },
+        char_offset: { type: "integer", minimum: 0 },
+        max_chars: { type: "integer", minimum: 2, maximum: 65536 },
+        content: { type: "string", description: "Complete nonempty model-derived output for this claim." },
+        anchors: { type: "array", minItems: 1, items: { type: "object", properties: {
+          input_id: { type: "string" },
+          char_start: { type: "integer", minimum: 0 },
+          char_end: { type: "integer", minimum: 0 },
+          quote: { type: "string" }
+        }, required: ["input_id", "char_start", "char_end", "quote"], additionalProperties: false } },
+        fields: { type: "array", items: { type: "object", properties: {
+          field: { type: "string" },
+          kind: { type: "string", enum: ["text", "decimal", "date", "time", "datetime"] },
+          anchor: { type: "integer", minimum: 0 },
+          literal: { type: "string", description: "Entire exact literal in the indicated source anchor." },
+          timezone: { type: "string", description: "Dates: calendar. Clocks: source_offset when the literal has Z/offset; otherwise UTC or a fixed +/-HH:MM offset. No inferred timezone." }
+        }, required: ["field", "kind", "anchor", "literal"], additionalProperties: false } },
+        order_spec: { type: "object", properties: {
+          field: { type: "string" },
+          kind: { type: "string", enum: ["text", "decimal", "date", "time", "datetime"] },
+          unit: { type: "string", enum: ["source", "page"] },
+          direction: { type: "string", enum: ["asc", "desc"] },
+          missing: { type: "string", enum: ["first", "last", "exclude", "error"] },
+          multiple: { type: "string", enum: ["min", "max", "error"] },
+          collation: { type: "string", enum: ["binary"] },
+          timezone: { type: "string", enum: ["calendar", "recorded_offset"] },
+          timeSemantics: { type: "string", enum: ["utc_clock"] },
+          where: { type: "object", properties: { equals: { type: "string" }, from: { type: "string" }, through: { type: "string" } }, additionalProperties: false }
+        }, required: ["field", "kind", "unit", "direction", "missing", "multiple"], additionalProperties: false },
+        cursor: { type: "string" },
+        numbers: { type: "array", items: { type: "object", properties: {
+          field: { type: "string" },
+          kind: { type: "string", enum: ["decimal", "count"] },
+          anchor: { type: "integer", minimum: 0 },
+          literal: { type: "string" }
+        }, required: ["field", "kind", "anchor"], additionalProperties: false } },
+        field: { type: "string" }
+      }, required: ["action", "corpus_id", "operation_id"], additionalProperties: false }
+    };
+  }
+});
+
+// packages/execution/dist/tools/corpus-work-workspace.js
+var init_corpus_work_workspace = __esm({
+  "packages/execution/dist/tools/corpus-work-workspace.js"() {
+    "use strict";
+    init_corpus_work();
+    init_source_corpus();
+    init_source_corpus_workspace();
+  }
+});
+
+// packages/execution/dist/tools/corpus-host-input.js
+var init_corpus_host_input = __esm({
+  "packages/execution/dist/tools/corpus-host-input.js"() {
+    "use strict";
+    init_pdf_corpus_archive();
+  }
+});
+
+// packages/execution/dist/tools/pdf-corpus-transform.js
+var LINE_BYTES, ACTIONS3, PDF_CORPUS_TRANSFORM_TOOL_DEFINITION;
+var init_pdf_corpus_transform = __esm({
+  "packages/execution/dist/tools/pdf-corpus-transform.js"() {
+    "use strict";
+    init_process_async();
+    init_pdf_corpus_archive();
+    init_pdf_transform();
+    LINE_BYTES = 64 * 1024;
+    ACTIONS3 = /* @__PURE__ */ new Set(["begin", "process", "status"]);
+    PDF_CORPUS_TRANSFORM_TOOL_DEFINITION = {
+      name: "pdf_corpus_transform",
+      aliases: ["pdf_corpus_compose"],
+      description: "Resumable original-page PDF merge, split and reorder over an acquired source_corpus. begin accepts an immutable JSONL recipe with one {sourceId,sourceSha256,page} per output page; order and repetitions are exact. process commits bounded partial PDFs and a durable merge tree; repeat the same operation_id after cancellation/update/restart. Source references must already belong to this host-scoped corpus. No text/OCR extraction is required for page copying. Returns a canonical PDF and streamed flattened original-page lineage with exact hashes; document metadata/forms/signatures are not preserved. No total PDF/page-count cap; each physical merge is atomic and may require a longer host work deadline for very large containers.",
+      parameters: { type: "object", properties: {
+        action: { type: "string", enum: [...ACTIONS3] },
+        corpus_id: { type: "string" },
+        operation_id: { type: "string" },
+        operation: { type: "string", enum: ["merge", "split", "reorder"] },
+        request: { type: "string", description: "Complete caller-declared transformation request; does not replace host goal authority." },
+        manifest: { type: "string" },
+        manifest_sha256: { type: "string" },
+        work_budget: { type: "integer", minimum: 1, maximum: 64, description: "Bounded recipe import or leaf/merge work units for this invocation." }
+      }, required: ["action", "corpus_id", "operation_id"], additionalProperties: false }
+    };
+  }
+});
+
+// packages/execution/dist/tools/pdf-corpus-recipe.js
+var init_pdf_corpus_recipe = __esm({
+  "packages/execution/dist/tools/pdf-corpus-recipe.js"() {
+    "use strict";
+    init_corpus_derived_order();
+    init_pdf_corpus_archive();
+  }
+});
+
+// packages/execution/dist/tools/pdf-corpus-transform-workspace.js
+var init_pdf_corpus_transform_workspace = __esm({
+  "packages/execution/dist/tools/pdf-corpus-transform-workspace.js"() {
+    "use strict";
+    init_corpus_work();
+    init_pdf_corpus_transform();
+    init_pdf_corpus_recipe();
+    init_source_corpus_workspace();
+  }
+});
+
+// packages/execution/dist/tools/pdf-corpus-recipe-workspace.js
+var init_pdf_corpus_recipe_workspace = __esm({
+  "packages/execution/dist/tools/pdf-corpus-recipe-workspace.js"() {
+    "use strict";
+    init_corpus_work();
+    init_pdf_corpus_recipe();
+    init_source_corpus_workspace();
+  }
+});
+
 // packages/execution/dist/tools/pdf-visual-markdown.js
 var MAX_IMAGE_BYTES;
 var init_pdf_visual_markdown = __esm({
@@ -297303,6 +297565,18 @@ var init_dist3 = __esm({
     init_vision_action_loop();
     init_ocr_pdf();
     init_pdf_to_text();
+    init_pdf_corpus_archive();
+    init_pdf_transform();
+    init_source_corpus();
+    init_source_corpus_workspace();
+    init_corpus_work();
+    init_corpus_work_workspace();
+    init_corpus_population();
+    init_corpus_host_input();
+    init_pdf_corpus_transform();
+    init_pdf_corpus_transform_workspace();
+    init_pdf_corpus_recipe();
+    init_pdf_corpus_recipe_workspace();
     init_pdf_native_tables();
     init_pdf_visual_markdown();
     init_pdf_document();
@@ -297390,6 +297664,8 @@ var init_dist3 = __esm({
     init_transcribe_python_runtime();
     init_verificationCommand();
     init_verificationInvocation();
+    init_corpus_derived_order();
+    init_ocr_pdf();
   }
 });
 
@@ -301202,6 +301478,12 @@ init_dist();
 // packages/orchestrator/dist/contextSNR.js
 init_dist();
 
+// packages/orchestrator/dist/corpus-native-source-proof-memo.js
+init_dist();
+
+// packages/orchestrator/dist/corpus-page-review-source.js
+init_dist();
+
 // packages/orchestrator/dist/contextEngine.js
 init_dist();
 

package/dist/update-worker.js

@@ -295027,9 +295027,127 @@ init_vision();
 init_system_deps();
 init_vision();
 
+// packages/execution/dist/tools/pdf-corpus-archive.js
+init_process_async();
+
 // packages/execution/dist/tools/ocr-pdf.js
 init_process_async();
 
+// packages/execution/dist/tools/pdf-transform.js
+init_process_async();
+var BUFFER_BYTES = 128 * 1024;
+var MAX_RECORD_BYTES = 64 * 1024;
+
+// packages/execution/dist/tools/source-corpus.js
+var SEGMENT_CHARS = 64 * 1024;
+var RECORD_BYTES = 64 * 1024;
+var ACTIONS = /* @__PURE__ */ new Set(["ingest", "inventory", "search", "read", "begin", "process", "status"]);
+var SOURCE_CORPUS_TOOL_DEFINITION = {
+  name: "source_corpus",
+  aliases: ["pdf_corpus"],
+  description: "Acquire and process a scoped PDF corpus without a total document/page-count cap. ingest uses an immutable JSONL manifest, one {ref,sha256} per source, and returns a durable import ID; repeat it to resume bounded batches. Local directory acquisition requires an explicit host policy and resumes bounded traversal/hash/manifest work with ingest_id. It captures an observed, version-fenced filesystem selection, not an atomic live-tree snapshot, and excludes only this workspace's owned .omnius state tree. inventory/search are cursor-paged; read returns exact channel-labelled segments with character offsets. begin preserves the full processing request and selected source versions; process checkpoints complete native pages and resumes after cancellation/restart. Source bytes, full native text and bbox are archived before bounded views. Explicit filters.channel='ocr_text' schedules resumable CPU OCR; rendered pages, engine/language, full OCR text and TSV are preserved separately. For OCR summaries/extractions use corpus_work filters.channel='ocr_text', complete anchored work, export, and corpus_export_review before delivery; source_answer_review only accepts native facts. Native or OCR engine completion does not prove visual understanding, summaries or final delivery.",
+  parameters: { type: "object", properties: {
+    action: { type: "string", enum: [...ACTIONS] },
+    corpus_id: { type: "string" },
+    manifest: { type: "string", description: "Host-authorized JSONL file or alias; one {ref,sha256} record per PDF." },
+    manifest_sha256: { type: "string" },
+    directory: { type: "string", description: "Explicit local-admin directory enumeration only." },
+    ingest_id: { type: "string", description: "Resume directory acquisition or the immutable import without resupplying its directory/manifest. Use a new ID for a deliberately changed directory selection." },
+    source_budget: { type: "integer", minimum: 1, maximum: 64 },
+    operation_id: { type: "string" },
+    request: { type: "string", description: "Complete extraction processing request; not authority to replace the user's goal." },
+    page_budget: { type: "integer", minimum: 1, maximum: 64 },
+    page_size: { type: "integer", minimum: 1, maximum: 128 },
+    cursor: { type: "string" },
+    query: { type: "string" },
+    filters: { type: "object" },
+    direction: { type: "string", enum: ["asc", "desc"] },
+    page_id: { type: "string" },
+    ocr_language: { type: "string", description: "Explicit OCR operation language, e.g. eng or eng+deu. Immutable after begin; uses installed Tesseract languages." },
+    char_offset: { type: "integer", minimum: 0 },
+    max_chars: { type: "integer", minimum: 2, maximum: SEGMENT_CHARS }
+  }, required: ["action", "corpus_id"], additionalProperties: false }
+};
+
+// packages/execution/dist/tools/corpus-work.js
+var ACTIONS2 = ["begin", "next", "read", "commit", "status", "complete", "cancel", "export", "aggregate", "order"];
+var CORPUS_WORK_TOOL_DEFINITION = {
+  name: "corpus_work",
+  aliases: ["pdf_corpus_work"],
+  description: "Durable model-driven summary or extraction over a scoped corpus. Native text is default; explicitly select filters.channel='ocr_text' for previously processed OCR evidence. begin preserves the FULL request and immutable selection. next gives bounded leased claims; read exposes each complete input through resumable exact character ranges. commit requires every input fully delivered through a successful host model request and exact quote anchors for every input. A read result prepares an exact range; same-response read+commit cannot establish model exposure. Full derived content is archived, never substituted for originals. Summary parents use host-issued bounded child membership and require complete child reads. Repeat next until leaves and reduction finish. Processing coverage is actor-attested, not proof of factual correctness, visual understanding or final delivery. Optional anchored decimal/count records support exact deterministic aggregates; record selection remains actor-attested. Exact anchored typed fields support SQL ordering before pagination through order, after processing completes. Text uses explicit binary collation; decimal is exact, dates calendar, clocks require declared fixed/source offsets. Missing and multiple values require explicit policies; field meaning and selection remain actor-attested. export resumably streams all derived leaf records and the summary root to JSONL. Repeat the exact export while completeExport=false; partial work has no finished artifact. byte_budget bounds each call, not the final output. Public OCR results require corpus_export_review prepare/process/finalize on the returned retained artifactPath; source_answer_review does not ground OCR inline prose. There is no total corpus or output count cap.",
+  parameters: { type: "object", properties: {
+    action: { type: "string", enum: [...ACTIONS2] },
+    corpus_id: { type: "string" },
+    operation_id: { type: "string" },
+    mode: { type: "string", enum: ["summary", "extraction"] },
+    request: { type: "string" },
+    filters: { type: "object" },
+    query: { type: "string" },
+    direction: { type: "string", enum: ["asc", "desc"] },
+    fan_in: { type: "integer", minimum: 2, maximum: 64, description: "Per-parent child allocation, not total corpus limit." },
+    byte_budget: { type: "integer", minimum: 1, maximum: 67108864, description: "Per-call export write/fixity byte budget; repeat export to finish arbitrarily large output." },
+    page_size: { type: "integer", minimum: 1, maximum: 32 },
+    claim_id: { type: "string" },
+    input_id: { type: "string" },
+    char_offset: { type: "integer", minimum: 0 },
+    max_chars: { type: "integer", minimum: 2, maximum: 65536 },
+    content: { type: "string", description: "Complete nonempty model-derived output for this claim." },
+    anchors: { type: "array", minItems: 1, items: { type: "object", properties: {
+      input_id: { type: "string" },
+      char_start: { type: "integer", minimum: 0 },
+      char_end: { type: "integer", minimum: 0 },
+      quote: { type: "string" }
+    }, required: ["input_id", "char_start", "char_end", "quote"], additionalProperties: false } },
+    fields: { type: "array", items: { type: "object", properties: {
+      field: { type: "string" },
+      kind: { type: "string", enum: ["text", "decimal", "date", "time", "datetime"] },
+      anchor: { type: "integer", minimum: 0 },
+      literal: { type: "string", description: "Entire exact literal in the indicated source anchor." },
+      timezone: { type: "string", description: "Dates: calendar. Clocks: source_offset when the literal has Z/offset; otherwise UTC or a fixed +/-HH:MM offset. No inferred timezone." }
+    }, required: ["field", "kind", "anchor", "literal"], additionalProperties: false } },
+    order_spec: { type: "object", properties: {
+      field: { type: "string" },
+      kind: { type: "string", enum: ["text", "decimal", "date", "time", "datetime"] },
+      unit: { type: "string", enum: ["source", "page"] },
+      direction: { type: "string", enum: ["asc", "desc"] },
+      missing: { type: "string", enum: ["first", "last", "exclude", "error"] },
+      multiple: { type: "string", enum: ["min", "max", "error"] },
+      collation: { type: "string", enum: ["binary"] },
+      timezone: { type: "string", enum: ["calendar", "recorded_offset"] },
+      timeSemantics: { type: "string", enum: ["utc_clock"] },
+      where: { type: "object", properties: { equals: { type: "string" }, from: { type: "string" }, through: { type: "string" } }, additionalProperties: false }
+    }, required: ["field", "kind", "unit", "direction", "missing", "multiple"], additionalProperties: false },
+    cursor: { type: "string" },
+    numbers: { type: "array", items: { type: "object", properties: {
+      field: { type: "string" },
+      kind: { type: "string", enum: ["decimal", "count"] },
+      anchor: { type: "integer", minimum: 0 },
+      literal: { type: "string" }
+    }, required: ["field", "kind", "anchor"], additionalProperties: false } },
+    field: { type: "string" }
+  }, required: ["action", "corpus_id", "operation_id"], additionalProperties: false }
+};
+
+// packages/execution/dist/tools/pdf-corpus-transform.js
+init_process_async();
+var LINE_BYTES = 64 * 1024;
+var ACTIONS3 = /* @__PURE__ */ new Set(["begin", "process", "status"]);
+var PDF_CORPUS_TRANSFORM_TOOL_DEFINITION = {
+  name: "pdf_corpus_transform",
+  aliases: ["pdf_corpus_compose"],
+  description: "Resumable original-page PDF merge, split and reorder over an acquired source_corpus. begin accepts an immutable JSONL recipe with one {sourceId,sourceSha256,page} per output page; order and repetitions are exact. process commits bounded partial PDFs and a durable merge tree; repeat the same operation_id after cancellation/update/restart. Source references must already belong to this host-scoped corpus. No text/OCR extraction is required for page copying. Returns a canonical PDF and streamed flattened original-page lineage with exact hashes; document metadata/forms/signatures are not preserved. No total PDF/page-count cap; each physical merge is atomic and may require a longer host work deadline for very large containers.",
+  parameters: { type: "object", properties: {
+    action: { type: "string", enum: [...ACTIONS3] },
+    corpus_id: { type: "string" },
+    operation_id: { type: "string" },
+    operation: { type: "string", enum: ["merge", "split", "reorder"] },
+    request: { type: "string", description: "Complete caller-declared transformation request; does not replace host goal authority." },
+    manifest: { type: "string" },
+    manifest_sha256: { type: "string" },
+    work_budget: { type: "integer", minimum: 1, maximum: 64, description: "Bounded recipe import or leaf/merge work units for this invocation." }
+  }, required: ["action", "corpus_id", "operation_id"], additionalProperties: false }
+};
+
 // packages/execution/dist/tools/pdf-visual-markdown.js
 var MAX_IMAGE_BYTES = 8 * 1024 * 1024;