Skip to content

RELEASES.md

Latest commit

 

History

History
96 lines (73 loc) · 4.96 KB

RELEASES.md

File metadata and controls

96 lines (73 loc) · 4.96 KB
Raw

Releases

v2.1.0 — Evidence-first Idea Lab

Digital Unconscious now closes the missing reverse loop: papers and local dataset structure can generate traceable, testable research ideas.

New

  • du ideate / du explore for PDF, text, Markdown, JSON, JSONL, CSV, and TSV inputs.
  • A local source ledger with stable IDs and SHA-256 content identity.
  • Exact-anchor evidence cards; invented model quotes fail deterministic validation.
  • Privacy-preserving dataset profiles that exclude raw rows and example values from model prompts.
  • Named opportunity operators for contradictions, limitations, boundary extensions, measurement gaps, method transfer, replication, robustness, and dataset reuse.
  • Falsifiable study cards with hypothesis, null, design, required variables, confounders, smallest useful test, and falsifier.
  • DeepSeek extraction, Codex synthesis, Claude methods review, deterministic final ranking, and normal provider fallback.
  • An Idea Lab dashboard screen and portable workspace/ideation/session_* reports.
  • Optional papers extra for PDF text extraction.

Reconstructed

  • One IdeaBacklog repository now owns identity, deduplication, ranking, and retention across the engine, CLI, tray, and Idea Lab, with locked atomic writes for concurrent processes.
  • Response-local IDs such as idea_001 are replaced by content-derived app IDs and reconciled without cross-window collisions.
  • The idea budget is enforced once per daily cycle, not once per compressed time window.
  • Full structured idea cards survive promotion into research through research_seed.json and the run manifest.
  • Real resume reloads completed artifacts; a new path or changed file content invalidates analysis, drafting, and review while preserving upstream work.
  • Model routing diagnostics include the three Idea Lab workloads.

Safer defaults

  • Daily scans no longer launch research, browser automation, or autostart by default.
  • Research no longer evolves prompts from its own review scores by default; learning is explicit.
  • Setup saves preferences without unexpectedly starting a background service.
  • Automatic browser acquisition was removed from normal literature and dataset stages.
  • Direct PDF downloads require a PDF signature and record their checksum.
  • Acquisition guidance permits lawful open access only and stops at paywalls, login, CAPTCHA, MFA, terms, and payment.
  • Dataset prompts omit local paths, filenames, raw rows, and example values.
  • Dashboard-rendered model content is HTML-escaped and state changes require an exact same-origin local host and port.

Verification

  • Evidence graph, source parser, privacy, backlog identity, resume, and acquisition regression tests.
  • Full unit and integration suite on the core install.
  • Clean package build and wheel smoke installation.
  • Browser inspection of the dashboard and Idea Lab at desktop and mobile widths.

v2.0.0 — Subscription-first reconstruction

Digital Unconscious is now a smaller, safer, genuinely zero-key application.

New

  • Real local Codex support through isolated codex exec calls.
  • Rebuilt Claude Code runner using safe mode, subscription auth, native effort, and structured output.
  • Workload-first routing for DeepSeek V4, GLM-5.1, Codex, Claude Code, GPT-5.6, Claude API, and Kimi.
  • du doctor for secret-free runtime readiness diagnostics.
  • Zero-key vision through temporary Codex image attachments with Claude fallback.
  • Modular backend kernel: base contract, local runners, hosted adapters, and router.
  • New visual identity, diagrams, README, architecture, model-routing, security, and reconstruction docs.

Fixed

  • codex no longer incorrectly aliases the OpenAI API.
  • Local model calls no longer inherit repository or home-directory access.
  • Setup no longer stores an unclassified API key in plaintext JSON.
  • Privacy language now states when text and screenshots cross the local boundary.
  • Automatic startup no longer mutates ~/.claude/agents.
  • Auto mode uses both signed-in local CLIs and falls back across available providers.

Removed

  • Stale binary PRD exports and overlapping planning/status documents. Git history remains the archive.
  • Startup-time Claude agent installation.
  • The claim that screenshots never leave the device.

Verification

  • Unit and integration suite on the lightweight install.
  • Real Codex subscription smoke call.
  • Real Claude Code subscription smoke call.
  • Package build and clean-wheel installation.
  • Browser inspection of the local dashboard and setup flow.

Earlier line

Version Theme
1.7 Dashboard and product interaction redesign
1.6 Vision observation and web search
1.5 Bounded long-term memory
1.4 Failover, reasoning controls, and usage tracking
1.3 Initial multi-provider API layer
1.2 Claude research-agent experiments
1.0 One-command desktop workflow
0.9 and earlier Research pipeline and daily loop foundations

The complete historical notes remain available in Git before tag v2.0.0.