Release Notes

This page summarizes what changed in recent Hypatia releases. GitHub releases use the same version-specific notes from the repository RELEASES.md file.

Unreleased

  • Native ideas and experiment setup in chat: /idea <goal> or /experiments generate <goal> uses Hypatia models/auth with built-in QML, retrieval/ranking, RAG/agent evaluation, statistics, ML/LLM and general research guidance, or custom project knowledge. No AutoResearch clone or Python backend is needed for new idea runs. /experiments create <goal> and proposal selection prepare contracts through chat, with inferred defaults, grouped prerequisite questions and one final approval per generation/benchmark. Request caps, caches, provenance and review gates persist; old Python-backed runs remain compatible.

  • Legacy Idea Forge provider configuration: optional providerConfig selects a private JSON for generation and persists across worker launches/resume. Relative AUTORESEARCH_CONFIG paths resolve against the backend clone rather than the experiment workspace.

  • Experiment environment fingerprints: new /autoresearch contracts bind the environment label, host platform/architecture/Node runtime, benchmark executable and hashes of declared frozen inputs. Include dependency lockfiles in inputFiles; inherited environment variables and installed-package inventories are deliberately not captured.

  • Legacy Python-backed Idea Forge generation: /experiments generate combines one signal and one knowledge direction with an explicitly approved 3–6-model backend panel. API attempt/time/token caps, frozen config/source/knowledge identity, response checkpoints and pause/resume persist. Generated proposals can prepare an experiment in chat with verified artifact provenance; proposal/model-review validity remains unverified. Proposal preview is readable and scrollable; View status shows request/cache/receipt integrity. Uses a local AutoResearch Python backend with its own credentials; collectors/publishers are not invoked.

  • Experiment setup and independent reviews: /experiments create <goal> prepares benchmark scope, metric, seeds, environment and budgets in chat, with one contract approval. create-json preserves JSON import. New contracts automatically run a fresh read-only critic before benchmarks and an auditor before closure. Review verdicts bind exact evidence hashes; findings, model identity, usage and capped attempts persist across sessions. Blocked, malformed or stale reviews cannot pass a required gate.

  • Persistent benchmark experiments: approved contracts freeze benchmark scope and budgets; observed receipts, code snapshots, cached baselines and pause/resume survive chat sessions. /experiments manages runs and validates AutoResearch idea imports. Reports and provenance appear in /outputs; the footer follows durable state. Execution supports macOS/Linux initially; Windows supports management/import.

  • Renamed bundled research agents: the five roles are now evidence-scout, scribe, auditor, critic, and reviser. Existing model-route and Pi override keys migrate to the new names. The old writer, verifier, and patcher selectors remain aliases; use the new names for researcher and reviewer, which Pi reserves for built-in agents. Startup settings normalization is serialized; command-specific model and service-tier writes remain outside this lock. A timeout names the lock file, which can be removed only after confirming no Hypatia startup is active.

  • Centered new-session logo: an empty fullscreen chat renders the supplied Hypatia PNG directly when the terminal supports inline images, and uses an automatic truecolor rendition otherwise. Slash commands and keyboard shortcuts dismiss it immediately; it adds no transcript or model content. Regular mode uses a compact scale.

  • Collapsible warning inbox: TUI extension warnings appear as a footer count; F2 opens a list, and selecting an item shows its full text. When a user skill shadows a bundled Hypatia skill, Hypatia filters the skipped package copy before Pi loads resources, keeping the winning user skill active without reporting a collision. Press d to dismiss warnings. Other unresolved Pi skill diagnostics remain available through F2.

  • Inline image attachments in chat: pasting a clipboard image shows an accent-blue [Image #1] marker in the composer instead of a temporary file path. On send, the image is attached as image content for the model.

  • Coordinated composer borders: when Pi shows an active Working status in the upper composer border, the lower border now uses the same Pi border color; the idle composer keeps its muted lower border.

  • Grouped chat activity and turn receipts: a live line groups observed operations, safe file basenames, and issue counts. Settled turns leave a compact outcome/time receipt; Ctrl+O expands operations and available tool error details.

  • TUI mode application notices: /settings reports when a TUI mode is applied immediately, and a reloaded preference that differs from the active layout reports that restart is needed.

  • Contextual English chat tips: tips appear when the editor is empty and the main agent, tools, subagents, and blocking input are idle. Successful visible tool results suggest the configured expand shortcut; new or updated reports in outputs/ or papers/ suggest /outputs. Report detection reads metadata and excludes hidden files, drafts, provenance, and links; it makes no verification claim.

  • Multiline composer keys: new Hypatia installations use Enter to send, Shift+Enter or Ctrl+J for a newline, and Alt+Enter to queue a follow-up. Existing installations retain user mappings; edit ~/.hypatia/agent/keybindings.json and run /reload to apply this mapping.

  • Hypatia chat presentation: You/Hypatia labels clarify the transcript; answers retain an open layout and an accent marker. The composer uses a thin › Message border and a research placeholder. Contextual tips live in the chat area above the composer when idle and empty, and add no model context. Multiline input, autocomplete, shell mode, and Pi’s working indicator are retained. Run /reload while idle to apply; an existing custom editor takes precedence.

  • Citation-aware answer rendering: arXiv, OpenReview, and DOI URLs use compact labels, clickable through OSC 8 where supported. Bibliography numbers and titles stand out while venue/year details recede; deeper Markdown headings no longer expose ### prefixes; explicit verification and completeness notes use amber callouts. The transformation affects TUI rendering only, preserving transcript text, model context, and exports.