neuroflow

---

What's new in 0.2.13

• /autoresearch — infinite improvement loop for any research artifact: point it at any file(s), and a worker-evaluator cycle runs indefinitely (one focused change per iteration, keep or revert, loop never stops until interrupted); live progress dashboard at localhost:8765; per-phase criteria auto-loaded from existing project memory; triggers via /autoresearch or any phase command with the word autoresearch in the prompt
• Agent cleanup — removed 16 unused phase agent files that were never spawned by commands; commands follow phase skills directly; the 8 agents that are actually spawned (paper-writer, paper-critic, poster-critic, literature-review, scholar, sentinel, sentinel-dev, flowie) remain unchanged

What's new in 0.2.12

• Notes → flowie sync (/notes) — after every notes session, Claude offers to copy the formatted note to .neuroflow/flowie/notes/ for GitHub sync (default: yes); a local config.json stores per-project defaults for type, speaker, and project relation
• Daily wellbeing tracking (/flowie --assess) — opt-in daily self-assessment for anxiety, energy, and happiness on a 1–10 scale (5=neutral); stored in flowie/wellbeing/; Claude prompts on any sync operation if today's entry is missing; enabled via /flowie --init or /flowie --assess

What's new in 0.2.11

• Removed broken pubmed-mcp-server — replaced by paper-search-mcp-nodejs (the biorxiv server), which already includes search_pubmed and requires no credentials; PUBMED_EMAIL is no longer needed

What's new in 0.2.10

• Global device config (/setup) — credentials can now be saved to ~/.neuroflow/integrations.json (global, shared by all projects on the machine) instead of per-project; per-project still takes precedence and overrides global; Step 0 of the wizard asks which scope to use
• Windows support in setup — /setup, neuroflow:setup, and the e-INFRA reference now include Windows-specific paths and PowerShell env var syntax throughout
• Proxy model-name fix (proxy.mjs) — the proxy now restores the original claude-* model name in every response chunk, preventing Claude Code's "unexpected model" error when using custom LLM providers; flowie now enforces that integrations.json is gitignored in the flowie sync repo

What's new in 0.2.8

• Session logging overhaul — removed the noisy [tool] PostToolUse hook; Claude now owns all session logging and writes entries broadly (most actions, not just milestones); neuroflow-core logging rules are now marked MUST and non-negotiable
• flow.md purity rule — flow.md is now explicitly a pure index table; narrative content, figure maps, and cross-references must go in dedicated .md files in the phase subfolder
• mind.js consistency check — sentinel-dev Check 11 audits that every skill, command, and agent has a node in docs/javascripts/mind.js; missing humanizer node added; neuroflow-develop release workflow now flags this as a blocking step
• scholar sequential search + batch downloads — searches now run PubMed → bioRxiv → fallbacks one at a time (was simultaneous); paper downloads processed in batches of 2 to reduce concurrency pressure and make failures easier to diagnose

What's new in 0.2.6

• Scholar agent: download reporting fixes — .pdf/.txt files now correctly marked ⏭️ already downloaded; .md-only stubs re-attempt download unless reason: unavailable; ✅ downloaded is gated on a confirmed .pdf or .txt write; download summary counter now labelled ✅ [n] downloaded (PDF/text) with a new ⏭️ [n] unavailable (metadata cached) bucket
• Scholar agent: search coverage fixes — Semantic Scholar 429 rate-limit triggers a 3 s wait + retry then falls back to CrossRef/arXiv with a visible warning; PubMed query-overlap detection auto-generates 2–3 diversified queries when < 15 unique results or > 80% overlap; arXiv keyword fallback added when bioRxiv returns 0 results; mandatory coverage summary table printed before results, with any ⚠️/❌ row also surfaced as an inline warning block

What's new in 0.2.5