0.6.11

What's new

• Each study is one typed graph. The study program and its effects (effects.json) are read together as a single graph of conditions, steps, variables, analyses, results, effects and the constructs the paper names. Program nodes can now link to each other directly, not only through an effect. Agents read it with study_graph (MCP) or study_core.py graph (CLI), and can filter by node or edge type.
• A fixed vocabulary that can grow. The graph uses a core set of node and edge types (graph-types.json, vocabulary 2.0.0). Effects schema 2 no longer accepts other: new kinds of contrast (linear, ratio, model_comparison) and metrics (cohens_q, auc, ratio, percent_change) cover cases that used to fall into it. Anything else is declared as an x: extension with a definition and a core parent, or reported as a gap. Nothing is forced into a type that does not fit.
• The paper's own words. Constructs and phenomena keep the paper's wording exactly. Nothing is merged across papers at this stage.
• Schema 1 files stay valid.

0.6.10

What's new

• Studies say what they can do. Each study has a capabilities file covering read, run, human data, the authors' code, and whether the paper's results are recovered from the human data. Each answer comes with its reason. Agents refuse work a study cannot do, and say why, instead of giving an approximation.
• Human data from their source. The human-data tool downloads the original data and checks every file's fingerprint (sha256). The data are never copied into the study. reproduce recomputes the paper's key results from those data.
• Smaller packages. Packages no longer store copies of the study program.
• Agent skills. Exported skills gain capabilities, get_human_data and reproduce_paper (MCP and CLI).

HumanStudy Hub 0.6.9

New

• Export a study as an agent skill. Ask the agent to export your study and you get one folder any AI agent can use: SKILL.md, a readable reference for each study with page numbers, the program and replay package, and scripts to read, search, run and score it. Add --mcp for an MCP server. Works with Claude Code, Codex and OpenCode. The paper PDF and sources/ are left out unless you choose to include them.
• Standalone skill (humanstudy-skill.zip). Build studies in Codex or Claude Code without the app.
• Quote check. Builds check that verbatim fields really appear in the paper.

Notes

• Windows: the installer is still unsigned. On first run, SmartScreen shows a warning: choose "More info", then "Run anyway".

HumanStudy Hub 0.6.8

What's changed

• Studies shows only extracted papers. The library now lists papers that already have a Human Program or study materials, so opening one gives you a project with content. Indexed papers that haven't been extracted yet are in their own view, linked under the list.

Notes

• Windows: the installer is still unsigned. On first run, SmartScreen shows a warning: choose "More info", then "Run anyway".

HumanStudy Hub 0.6.7

What's fixed

• Importing a paper starts the build. You no longer get an empty chat and "No study yet". Older projects with a paper but no study get a link to build it from the paper.
• Chat and simulation stay separate. Asking the agent to edit something no longer starts a simulation. A run's status comes only from its own files. Stop always returns to a state you can retry. Empty run folders are no longer counted as results.
• Edits aren't lost. Unsaved drafts are kept, and the app asks before you leave. Saving one Study field changes only that field.
• Source lookup is more precise. Results show which file a quote comes from. Numbers and signs are compared exactly, so p<.001 vs p>.001 is flagged. A replaced PDF is reloaded.
• Run and the cost estimate wait for a complete package. If something is missing, the app lists it.
• Large papers stay under the request limit. Older page images are kept locally instead of being resent on every request, so long builds no longer hit "request too large".
• Windows reliability. Fixes for first-time setup on machines with Git installed, "Access is denied" during saves, folder names that differ only in capitalisation, antivirus locking progress files, and line endings.

New

• The agent can search the Hub's library of 302 indexed papers and prepare community contributions.

Notes

• Windows: the installer is still unsigned. On first run, SmartScreen shows a warning: choose "More info", then "Run anyway".
• Research data: collection follows your research agreement. Nothing new is collected from the website: collection there is still switched off.

HumanStudy Hub 0.6.6

Papers and sources

• Open long PDFs without extracting the entire paper up front. Offscreen pages release their display resources; zoom and cross-page highlights are preserved.
• View all supporting quotations for an item. Citation checks require the complete quotation, including every segment around an ellipsis.
• New build instructions retain complete supporting sentences and necessary context. Existing saved short quotations need agent re-extraction.

Updates

• See the installed version and downloaded version together.
• “Later” is remembered for the same version across launches; stale events for the installed version are ignored.

Mac installers for Apple Silicon and Intel are signed and Apple notarized. Windows remains an unsigned preview.

HumanStudy Hub 0.6.5

OpenCode tools are now available directly in the study workspace. Type / or press Cmd/Ctrl+Shift+P for commands: compact context, choose an agent/model, branch conversations, undo/restore turns, review history and changes, connect MCP tools/providers, or open an interactive terminal.

Full conversations remain saved. Reopening conversations preserves their agent/model choices; undo keeps the original local history. Chat rendering and scrolling have been improved, and the fixed starter suggestions have been removed.

Mac Apple Silicon and Intel installers are signed and notarized. Windows remains a preview without code signing or real-machine acceptance. External provider sign-in and optional workspace services depend on the account/service being connected.

HumanStudy Hub 0.6.4

Build recovery now resumes automatically when a model response reaches its generation limit or a malformed tool call prevents a file from being written. It checks saved work and asks the agent to split the remaining task into smaller complete writes while preserving source details.

• Handles truncated/invalid tool arguments and late output-limit signals.
• Continues in the same conversation, with at most two adjusted recovery attempts.
• Honors Stop, unanswered questions and permission decisions; unrelated errors remain visible.

Verified against the real OpenCode v1.19.0 server with local deterministic model fixtures, including an actual recovered file write. Mac installers are signed and notarized. Windows remains an unsigned preview; real Windows/Intel hardware testing is pending.

0.6.3

v0.6.3

Merge pull request #5 from HumanStudy-Hub/codex/recover-cancelled-shu…

0.6.2 — Safe saves and update recovery

HumanStudy Hub 0.6.2

• Save the complete conversation before quitting or restarting for an update. A save failure keeps the app open so it can be retried.
• Ask before interrupting a study agent that is still working. Keep working is the default.
• Recover interrupted update downloads with up to three automatic retries, followed by manual retry if needed.
• Show missing update information and verification failures as errors instead of claiming the app is current.
• Wait for macOS update verification before preparing a restart.

macOS Apple Silicon and Intel apps are Developer ID signed, notarized and stapled. Windows remains an unsigned preview, pending real-device acceptance.

Validation: regression scripts, main/renderer typechecks and production build passed. An isolated Electron test confirmed busy-quit cancellation, real save-failure blocking, successful retry and complete conversation persistence before exit. No research content is shortened by these changes.