Skip to content

CHANGELOG.md

Latest commit

 

History

History
76 lines (62 loc) · 3.55 KB

CHANGELOG.md

File metadata and controls

76 lines (62 loc) · 3.55 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

Format follows Keep a Changelog. This project adheres to a pre-1.0 cadence: minor bumps are breaking/notable, patch bumps are additive or fixes. Dates are in YYYY-MM-DD.

[Unreleased]

Nothing yet.

[0.2.0] — 2026-04-21

First tagged release post-MVP. The pipeline is end-to-end production-quality on the default MiniMax M2.7 model, with real image attachment gated by CLIP + alt-quality + vision-LLM rerank.

Added

  • PDF export via scripts/mkpdf.mjs (Chromium headless print-to-PDF) — weasyprint's CJK output was garbled on macOS Preview, Chromium's Type 3 font embedding is cross-reader stable. Remote images are downloaded into out/_pdf_assets/ first so the PDF is offline-reproducible.
  • CLIP warm-up (#26) — warmClipModel fire-and-forget from runTopic2md / regenSection entry; cuts cold-start from ~114s to ~1.4s when the Replicate container is still warm from a prior run.
  • Persistent image embeddings (#27) — image_embeddings(url, model_version, embedding BLOB) table in SQLite caches jina-clip-v2 embeddings across runs. 100% cache-hit on regen of prior runs; ~$0.17/run → near-zero for recurring URLs.
  • Alt-quality penalty (#28) — soft multiplier on raw CLIP cosine for empty/generic alt, short alt, and CMS-dump upload-path URLs.
  • CLIP relevance gate — zsxkib/jina-clip-v2 on Replicate, threshold 0.30 calibrated on 68 labeled candidates (AUC 0.85, recall 0.97, kills ~60% irrelevant). REPLICATE_API_TOKEN optional; absent ⇒ bypass.
  • Candidate pool + vision rerank — all image plugins contribute candidates; qwen3-vl-32b picks one per section. pickIndex = -1 is respected — no keyword fallback when vision deliberately rejects.
  • Background context field (#24) — WorkflowInput.background (freeform) threads through research → outline → sections. Persisted in runs.background. CLI --background / --background-file; Web UI via collapsible textarea. regen reuses the source run's background unless overridden.
  • SQLite persistence — runs + run_stages tables record every runTopic2md invocation; topic2md list / show / regen --section N.
  • LLM model fallback chain with per-generation Langfuse telemetry (generation.* events).
  • Plugin dispose?() lifecycle hook — runner calls in finally to release Playwright browsers and DB handles.

Changed

  • generateObject forced to mode: 'json' to avoid the malformed tool-call JSON path on OpenRouter-routed providers (MiniMax/GLM). Cut sections step from ~500s → ~60s.
  • Open-source hygiene overhaul: CONTRIBUTING.md, SECURITY.md, GitHub issue / PR templates, full English README.en.md.

Fixed

  • Outline retries once on first-attempt failure before giving up.
  • Section two-retry failure now renders outline bullets as a bullet-list fallback rather than leaving the section empty.
  • regen runId parsing accepts nanoids that start with - / _ (previously filtered by !arg.startsWith('-')).
  • Docker image copies the full workspace + installs build-essential so better-sqlite3 can build from source on unusual glibc/node combos.
  • DATABASE_URL anchored via import.meta.url so Web UI and CLI share one run history regardless of cwd.

[0.1.0] — ~2026-03

Initial MVP. Mastra workflow skeleton, Tavily source, file publish, Playwright screenshot, Next.js trigger UI. Single model, no persistence, no regeneration. Historical record only; not tagged.