Skip to content

CHANGELOG.md

Latest commit

 

History

History
226 lines (204 loc) · 12.6 KB

CHANGELOG.md

File metadata and controls

226 lines (204 loc) · 12.6 KB
Raw

Changelog

All notable changes to tofu-open are documented in this file.

[Unreleased]

[0.16.0] - 2026-07-31

Versions 0.11.0 – 0.15.2 were never released. They exist as VERSION bumps (and, for 0.15.0 / 0.15.2, as orphan git tags) but no GitHub Release was ever published behind them — the last published release was v0.14.2, and the macOS build leg starved on a retired runner label before the release job could run. Rather than reconstruct changelog entries for releases that never shipped, their content is folded into this entry.

This is a minor bump, not a patch: VERSION sat at 0.15.2 from 2026-07-23 while ~1250 further commits landed, adding six new top-level capability packages — four of which expose their own HTTP surface (routes/api_v1/research.py, motion.py, skills.py, private_hosts.py).

Added

  • Auto-research pipeline (lib/research/, routes/api_v1/research.py). Give it a research direction and it harvests recent literature into a local paper corpus (parsed once, then reused), surveys it to map what has already been done, and proposes scored ideas screened against that corpus so they are genuinely new rather than A+B recombinations. Rejected ideas are reported with the reason. Exposed as the produce_research tool.
  • Long-form research reports (lib/longform/). research → outline → sections(×N) → assemble, published as a cited markdown artifact. The stage list is data-dependent (one stage per outline section), which the static video stage list never exercised.
  • Motion-graphics video pipeline (lib/motion_video/, routes/api_v1/motion.py). Topic → researched script → real-TTS-timed storyboard → per-scene composed MP4 → concat → narration mux. Every fact card carries a real source URL and is credited on an end card. Per-scene authoring degrades to a template floor so a single bad scene can never fail the film.
  • Production Substrate (lib/production/). The horizontal layer under every "one sentence → finished product" capability: a checkpointed stage graph where a stage's artifact is committed as soon as its gate passes, so a killed process resumes at the first unfinished stage. Crash-resume is a correctness contract here, not an optimisation.
  • Text-to-speech / narration (lib/tts/, routes/api_v1/audio.py) and voice input (speech-to-text) with a mic button in the composer.
  • Skills as a first-class noun (lib/skills/, routes/api_v1/skills.py). User-installed skill packages are now decoupled from model-authored memories: an always-visible <available_skills> index plus an load_skill progressive-disclosure loader. The model channel is read-only; install/uninstall/toggle are user-only.
  • Project Brain — cross-conversation coordination (lib/conversations/). Charter with human-reviewed decisions, an epic board with claims and leases, an activity feed, direct peer messaging, path leases, and a "Needs you" attention surface that aggregates everything genuinely waiting on a human.
  • Request Inspector / debug panel. Per-request snapshot store with server-side folding, a </> affordance on bubbles and tool rows, and incremental retention — so "what exactly went on the wire?" is answerable without a debugger.
  • Manual compaction (/compact). Context compaction is no longer only automatic: an explicit command with a REST endpoint, a frontend card, and a live streaming summary.
  • Air / Pro / Studio capability dial. One toolbar control replaces the separate Enhance/Tools/Mode toggles, selecting a coherent capability profile per turn.
  • Remote Worktree / Desktop Agent. Run a project on a remote machine from the desktop app — per-user bridge tokens, a "Remote devices" picker group, tray connect, and run_command parity with the local path.
  • Auto-escaping safeHtml tagged template (static/js/core/safe_html.js). The chat-render path built HTML by string concatenation, hand-wrapping every interpolation in escapeHtml() — correct but fragile (one forgotten wrap is an XSS hole). safeHtml\...`escapes EVERY interpolation by default, with an explicitraw(x)opt-out for trusted HTML. A lint rule intests/test_frontend_safe_html.py` blocks new bare template-string HTML sinks in adopted render files.
  • Frontend type-check harness (tsc --checkJs, no build step). Root tsconfig.json + static/js/globals.d.ts catch cross-file global misuse — typos, stale renames, dead typeof guards — that the shared-window-scope design otherwise fails silently at runtime. Wired as make typecheck and enforced by a monotonically-decreasing error-budget ratchet.
  • Release gates. scripts/release_assets.py (is the release complete? — per-platform assets plus a size floor that catches a hollow build) and scripts/changelog_gate.py (is this VERSION documented? — this file is now a build gate, which is why the nine-version gap above can never recur).

Changed

  • SQLAlchemy Core table-definition layer (lib/database/_core_schema.py). Tables are defined ONCE as Core Table objects and compiled to correct DDL and DML for BOTH backends (PG JSONB/IDENTITY ↔ SQLite JSON/autoinc, paramstyle, dialect-correct upserts), retiring the hand-maintained twin-DDL path. Compile-only: no SQLAlchemy Engine is opened; execution stays on the existing connection. Generated DDL is byte-equivalent to the legacy hand-DDL (tests/test_core_schema_parity.py). Adds sqlalchemy>=2.0.
  • Unified the LLM SSE streaming core. lib/llm/stream.py (sync) and lib/llm/astream.py (async) each carried a ~480-line copy of the identical SSE parsing loop, so every fix had to land twice and the copies drifted. That logic now lives once in lib/llm/_sse_core.py; the two modules are thin transport shells keeping only retry/backoff and transport-native handling. Pure code-motion — anomaly fields are emitted byte-for-byte as before, locked by tests/test_sse_core_parity.py. Net −432 lines.
  • Account ↔ wire-face separation in provider config, so one account can serve both OpenAI- and Anthropic-shaped endpoints without duplicate entries.
  • Web search and fetch extracted to the standalone tofu_search package; the app seams via lib/search_bridge.py. lib/fetch/ and lib/search/ no longer exist in-tree.
  • Desktop release workflow is VERSION-driven, not tag-driven. A tag is a product of releasing, not evidence of it; the gate now asks the Releases API whether a complete asset set exists.

Fixed

  • Weak image-caption escaping in renderMessage. The image tile's title tooltip escaped only double-quotes, leaving </>/& unescaped. Now uses the full escapeHtml().
  • Chat didn't re-render on language switch / debug-mode toggle. i18n.js::_onLanguageChange and settings/save_export.js::saveSettings called renderMessages() behind a typeof … === 'function' guard, but that function never existed (the real repaint is renderChat(conv)), so the guard silently swallowed the no-op. Caught by the new tsc --checkJs harness.
  • Duplicate common.close i18n key in static/js/i18n.js removed.
  • Desktop downloads 404'd during a release window. URLs were built as /releases/latest/download/<cached filename>, whose two halves have different lifetimes. They now come from one API payload with the tag pinned.
  • Intel Macs could not install Tofu. Only the arm64 DMG shipped; the macOS build is now a per-architecture matrix on live runner labels, and the release refuses to publish a partial asset set.
  • Numerous fixes across tool lifecycle (per-tool completion events rather than round-barrier), streaming transport, cache accounting, MCP launching, scheduler, and the paper Reading Mode pipeline.

[0.10.0] - 2026-05-09

Added

  • Daily Optimizer (self-tuning loop). New lib/optimizer/ package mines the prior day's logs, audit events, and daily reports, asks an LLM for optimisation proposals, and either auto-applies whitelisted low-risk actions (currently block_search_domain, with TTL-based auto-revert) or stages everything else as pending_review for human approval. Runs nightly at 03:30 via the scheduler (Daily Optimizer task, auto-registered on boot). REST API in routes/optimizer.py; review UI in static/js/optimizer.js. Gated by OPTIMIZER_ENABLED setting.
  • Skills Store (curated catalogue + drag-and-drop installer). Settings → Skills tab now has an App-Store-style layout (search + Catalogue/Installed scope tabs + category pills) backed by lib/memory/catalog.py. One-click install downloads a .zip over HTTPS (≤ 50 MB) and unpacks it via lib/memory/installer.py. Anthropic / OpenClaw / AgentSkills .zip packages can also be drag-dropped onto the tab; bundled install.sh scripts are surfaced as hints, never auto-executed.
  • Pluggable token counter. New lib/token_counter/ package routes token counting through provider-specific backends (Anthropic / Gemini / DeepSeek / HuggingFace / tiktoken / heuristic) with a usage cache, replacing the scattered ad-hoc estimators.
  • File-history store. New lib/file_history/ records per-file edit history so write tools and the diff viewer can show a coherent timeline of changes across a session.
  • Memory prefetch. lib/memory/prefetch.py surfaces likely-relevant memories at turn start via the <relevant_memories> block, so the model doesn't have to call search_memories as a generic discovery step.
  • Compaction archive viewer. New routes/conversations_compaction.py
    • static/js/compaction-viewer.js let you inspect the archived layers produced by 3-layer context compaction.
  • Conversation full-text search endpoints moved into a dedicated routes/conversations_search.py Blueprint (extracted from routes/conversations.py).
  • Provider templates. Added Meituan and Tencent provider one-click templates in Settings → Providers.

Improved

  • routes/chat.py decomposition. Extracted chat_human_io.py, chat_queue.py, and chat_tool_state.py so the chat blueprint is smaller and individual concerns (stdin/human-guidance responses, server-side message queue, tool-toggle PATCH) live in their own modules.
  • PDF parsing. Added lib/pdf_parser/docling.py as an additional backend alongside the existing text/VLM/math paths.
  • Project tools. New lib/project_mod/gitignore_suggest.py proposes .gitignore entries for files the indexer keeps re-scanning.
  • Multi-root workspace robustness. Extra roots now persist across conversation switches (frontend sends projectPaths; backend ensure_project_state() accepts extra_paths). The system prompt's multi-root section explicitly warns about new-file creation in non-primary roots, since there is no auto-detection until the file exists.
  • requirements.txt. Pin lxml_html_clean>=0.4 so trafilatura keeps working on lxml 5.2+ where lxml.html.clean was extracted.

Fixed

  • Numerous small fixes in browser dispatch, conv_ref handling, image generation, LLM sanitisation, scheduler timer/manager, and trading decision routes (see file-level diffs).

[0.9.3] - 2026-04-22

Fixed

  • MCP launcher pre-flight check. When an MCP server is configured with a command that is not on PATH (e.g. uvx without uv installed, npx without Node), we now emit a clear, actionable install hint instead of a cryptic FileNotFoundError. Covers uvx / npx / pipx / node / python3.

Improved

  • Overleaf MCP auto-install resilience. The catalog entry and migration rules now pin overleaf-mcp-plus[compile]>=0.1.3, the slimmer release that drops the unused playwright dependency (~100 MB faster first-run install).
  • Auto-migration upgraded. Stale server entries from prior versions are rewritten on load even when only the args list differs — user-supplied env vars and credentials are always preserved.

[0.9.2] - 2026-04-20

Fixed

  • Fixed Overleaf MCP server failing to launch with FileNotFoundError: 'overleaf-mcp' on machines where the package was not pre-installed. The curated registry entry now uses uvx --from overleaf-mcp-plus[compile] so the server is auto-fetched from PyPI on first launch, matching the behavior of the other MCP cards.

[0.9.1] - 2026-04-20

Improved

  • Further optimized support for Claude Opus 4.7.

Added

  • Added support for the Overleaf MCP server in the curated registry (edit/read/compile/history on Overleaf LaTeX projects).

Fixed

  • Fixed incorrect retry behavior of the model when invoked by tools.

[0.9.0]

  • Previous release.