Skip to content

2.0.7

Choose a tag to compare

@karilaa-dev karilaa-dev released this 05 Sep 04:42
c5c8bed

This release includes every merged change since v2.0.3: model defaults, runtime and file-delivery improvements, Office authoring and validation, reusable image assets, presentation design guidance, and the review fixes. It contains PRs #38, #39, and #40.

Office documents and presentations

  • Replace OfficeCLI with pinned docx-cli 0.25.0 for Word, PptxGenJS 4.0.1 for new presentations, python-pptx 1.0.2 for existing presentations, and openpyxl 3.1.5 for spreadsheets. Focused skills cover creation, targeted edits, tracked changes, comments, tables, charts, and image placement.
  • Render saved Office files through LibreOffice and Poppler, with up to four preview pages per call. Office previews work without Browser Use credentials.
  • Require structural checks and explicit visual review of every page before delivery. Approval is bound to the exact file contents and validated format. Edits invalidate approval; failed drafts remain in the workspace. Technical validation captions are no longer appended to attachments.
  • Add image search and stronger presentation guidance for relevant photographs, sourced illustrations, generated concepts, typography, composition, and review of rendered slides.
  • Accept spreadsheet formulas that legitimately return blank values. Normalize equivalent internal relationship paths in private rendering copies while preserving original files and all other package parts.
  • Block external Office resource relationships before rendering while preserving ordinary web and email hyperlinks. Reject mislabeled Office exports and retain failed cleanup paths for retry.

Reusable image generation

  • generate_image saves one original workspace asset per call and returns its path, dimensions, provider metadata, and an image preview to the model. Generation continues the conversation without automatically attaching the image.
  • The bot can inspect, edit, and embed generated assets, then deliberately deliver an image or only the finished document through finish_response or create_file.
  • Support references from current-thread chat files and workspace assets, up to five combined, with scoped writes, cancellation handling, provider fallback, and persistence when assets are selected for delivery.

Agent runtime and file delivery

  • Use GPT-6 Astra with low reasoning by default for the primary model and OpenRouter fallback. Preserve provisional streamed text across subsequent tool calls and supply the selected model identity without accumulating it in conversation history.
  • Add finish_response and combined Bash execution/image inspection to reduce model round trips. Keep the core prompt compact and put detailed workflows in approved skills.
  • Prepare files with two workers sharing a bounded 40 MiB delivery buffer. Preserve attachment order, overlap preparation with delivery, group compatible adjacent files, and retain successful files when another attachment needs repair.
  • Improve cancellation, failure budgets, delivery recovery, fork-aware visibility queries, snapshot pruning, and runtime telemetry. Consolidate sandbox, attachment, and response ownership and add reproducible benchmarks and regression coverage.

Upgrade and verification

  • The published sandbox image is ai-tg-bot-tools:v2.0.7, still using 2 vCPU and 2 GiB RAM. Existing sandboxes upgrade through verified staging directories and one remote installer lock, preserving workspaces and saved sources and removing obsolete bundle files.
  • Deploy the application at this release. Leave E2B_TEMPLATE unset to use the version-derived image, or update an explicit override to ai-tg-bot-tools:v2.0.7. Keep E2B_DEPLOYMENT_ID unchanged during the upgrade.
  • Typecheck, all 472 tests, and build passed on the release commit in GitHub CI, including PostgreSQL integration coverage. Local verification passed 470 tests, with two PostgreSQL tests skipped locally.
  • The E2B image release and live upgrade from v2.0.6 passed, including all Office backends, actual-file rendering, workspace preservation, concurrent installer locking, and pause/resume. Live model checks also covered reusable image generation, deck embedding, and image acquisition from an ordinary presentation request.

Included pull requests

Full changelog: v2.0.3...v2.0.7