Long-Document Analysis

CoWork OS has a bounded, read-only review path for directly referenced long documents. It is intended for requests such as manuscript review, continuity checking, contradiction detection, chapter-transition review, and full-document summarization where one model prompt would not reliably cover the source.

This path currently supports .docx, .pdf, .md, and .txt sources. It is separate from document editing and from visual PDF inspection.

When the bounded path is used

The deterministic pipeline is selected when all of the following are true:

• the request directly identifies a supported source file or a strongly named document in the workspace;
• the request asks to read, review, summarize, inspect, or analyze the document;
• the request does not ask to write, edit, convert, replace, or save the source.

Requests that include document mutation continue through the normal execution path and its write/approval controls. Ordinary targeted reads can still use parse_document; PDF layout, formatting, charts, diagrams, and scanned-page appearance remain read_pdf_visual work.

Source discovery and safety

If the request does not resolve to one exact path, CoWork ranks supported files by filename and task-text matches. Discovery is bounded to four directory levels and skips:

• symbolic links and Office lock files such as ~$draft.docx;
• hidden directories;
• .cowork, .git, node_modules, dist, build, and release.

The selected file is resolved with realpath and must remain inside the active workspace. DOCX text is extracted with Mammoth, PDF text uses the OCR-aware PDF extractor, and Markdown/plain-text sources are read as UTF-8. Empty or unsupported sources fail explicitly.

Coverage model

Extracted text is split into stable overlapping windows. The default maximum is 24,000 characters with 400 characters of overlap; paragraph and line boundaries are preferred, and every source character remains covered.

The runtime then:

1. records the selected path, source-text SHA-256 digest, character count, and chunk count;
2. analyzes every chunk with its character range and the original user request;
3. splits an oversized failed chunk once more before giving up;
4. reduces very large evidence sets in bounded groups; and
5. produces one evidence-grounded synthesis plus a coverage-verification line.

If final synthesis times out after all chunks were analyzed, CoWork preserves the range-labelled evidence digest and reports partial success instead of discarding the completed review work. Cancellation before complete coverage is reported as an interruption.

Local models

Ollama uses smaller bounded output budgets for chunk findings and synthesis. The planner can recover numbered or bullet steps from plain-text local-model responses when valid JSON planning output is absent. Empty responses and token-limit truncation remain explicit failures; they are not presented as a completed review.

Maintainer validation

Run the focused regression coverage with:

npx vitest run \
  src/electron/agent/__tests__/document-analysis-pipeline.test.ts \
  src/electron/agent/__tests__/executor-plan-parsing.test.ts \
  src/electron/agent/__tests__/executor-tool-allowlist.test.ts \
  src/electron/agent/llm/__tests__/ollama-provider.test.ts \
  src/electron/agent/tools/__tests__/document-parser-tools.test.ts

See Execution Runtime Model for the wider planning and tool-policy contract and Document Artifacts for created or edited document outputs.