Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
56 lines (46 loc) · 2.93 KB

ARCHITECTURE.md

File metadata and controls

56 lines (46 loc) · 2.93 KB

Bauer architecture (current)

Architecture (code organization)

At a high level, Bauer is a CLI with three core subsystems: Google Docs extraction, prompt generation, and Copilot execution.

  • cmd/bauer/main.go: CLI entry point and the end-to-end orchestrator.
  • internal/config: flag parsing + validation for --doc-id, --credentials, --chunk-size, --output-dir, --dry-run, --model, --summary-model, --page-refresh.
  • internal/gdocs: Google Docs/Drive client + extraction pipeline.
    • service.go: auth + API clients.
    • extraction.go: fetch doc, walk document tree, build structure/anchors.
    • grouping.go: group and merge suggestions by location and ID.
    • types.go: data contracts for suggestions, metadata, and structure.
    • comments.go: Drive comments fetch (currently not wired into the main flow).
  • internal/prompt: prompt engine + embedded instruction templates.
    • templates/: copy-docs + page-refresh instructions, plus Vanilla pattern references.
  • internal/copilotcli: Copilot SDK wrapper to run each chunk and (optionally) generate a summary.

How the flow operates (implementation details)

flowchart TD
    A[CLI start] --> B[Load config + validate]
    B --> C[Init GDocs client]
    C --> D[Fetch doc with SUGGESTIONS_INLINE]
    D --> E[Extract raw suggestions]
    E --> F[Build doc structure + metadata]
    F --> G[Build actionable suggestions]
    G --> H[Group by location + merge by ID]
    H --> I[Filter metadata table suggestions]
    I --> J[Render chunk prompts]
    J --> K{Dry run?}
    K -- Yes --> L[Write outputs + stop]
    K -- No --> M[Start Copilot CLI]
    M --> N[Execute chunks sequentially]
    N --> O{>1 chunk?}
    O -- Yes --> P[Generate summary session]
    O -- No --> Q[Done]
Loading

A few specifics worth knowing:

  • Google Docs is fetched in SUGGESTIONS_INLINE mode so insertions/deletions are embedded in the document tree.
  • Style suggestions are currently skipped entirely (only insert/delete operations become actionable suggestions).
  • Anchor text is generated from the surrounding document text to help Copilot find exact matches.
  • Suggestions are grouped first by logical location (section + heading + table), then merged by suggestion ID to form a single change.
  • Chunking is by number of location groups (not “suggestions per chunk”).
  • Each chunk file includes: instructions → Vanilla pattern references → JSON suggestions.
  • Copilot runs one session per chunk, streaming output to the terminal; summary is a separate session only when there are multiple chunks.

Understanding the output files

  • bauer-log.json: JSON logs (debug-level) for the entire run.
  • bauer-doc-suggestions.json: Full ProcessingResult (document metadata + actionable + grouped suggestions). Useful for debugging or re-running prompt generation.
  • bauer-output/chunk-X-of-Y.md: One prompt per chunk. Each file embeds the instruction template, Vanilla patterns reference, and the JSON suggestions for that chunk.
⚡