Benchside 0.4.1

Benchside 0.4.1

This release improves faithful scientific recording, fixes the reviewed journal
and sharing issues, and publishes the Python package for pip/uv installation.

Available on PyPI and
GitHub Releases.
Local agents can run uv tool install benchside, or
python -m pip install benchside inside a persistent virtual environment, then
the appropriate bench setup command in SETUP.md.

Update every app sharing a notebook

All apps sharing a private notebook must use Benchside 0.4.1 or newer. Once
0.4.1 saves a notebook using the hypothesis-reviewer role or multiple disciplines,
older apps—including 0.4.0—refuse to open it and request an update. This prevents
older readers from treating newer records as damaged. Keep the existing notebook
and update each app; do not recreate or overwrite it to bypass that check.

Changes

• Research-role guidance better distinguishes supplied evidence, researcher
  decisions and suggestions. The supervisor supports exploratory discussions
  without requiring a hypothesis; focused hypothesis review remains separate.
• ODT/Markdown journal exports work for either format or both. Reading copies
  refresh once per command rather than after every event in that command.
• Survey answers that are blank or match sensitive-text checks are refused before
  sending, including older pending drafts. Ordinary answers still use the study
  participation agreement without a second approval.
• Full selected-conversation history is retained. At its notebook allowance,
  analysis pauses and explains alternatives; there is no automatic summarisation.
• Migration detects conflicting writes; malformed contributor identity files are
  preserved with a repair message. New notebook formats give older readers a
  clear update-required response.
• Local Claude Code setup checks linked settings before creating setup files or
  changing the storage preference. It preserves shared settings and explains
  manual options. Normal setup leaves unrelated linked settings alone.
• Python distributions include licence, Science Works maintainer attribution,
  source/support links and standalone setup instructions, retaining Rory's author
  credit. GitHub Trusted Publishing releases the Python package after clean wheel
  and source installation checks.

Known limits

The survey scanner currently also rejects full URLs and some local paths, even
when harmless. It cannot identify all confidential content; contextual review
remains necessary. Public-resource link handling is a separate proposed refinement.
Blank/invalid survey answers can still surface with a generic setup error in some
app paths. Receiver-side blank-answer enforcement is a separate deployment.

The core regression suite has 582 passing tests on Python 3.11 and 3.14. Package
tests exercise saved records and fresh-process restoration; they do not establish
live account-level persistence in every ChatGPT or Claude account. Installed-app
acceptance for this final build remains a separate check.

On 6 November 2026, fresh public PyPI installations through pip (Python 3.11)
and uv tool install (Python 3.14) passed the CLI, eight-role setup, private
save/restore, ODT/Markdown and sharing-off checks. No submissions were sent.
The public wheel and source archive checksums match the GitHub release assets.

See the review follow-up,
Python publishing procedure, and
successful publication run.

Downloads

• Claude: benchside-claude-plugin-0.4.1.zip (plugin) or benchside-claude-0.4.1.zip (custom skill).
• ChatGPT / OpenAI-compatible hosts: benchside-plugin-0.4.1.zip. Installation and storage tools must be available in the account.
• Local Python agents: uv tool install benchside or python -m pip install benchside in a persistent virtual environment, followed by the appropriate bench setup command. The wheel and source archive are also attached below.

Keep plugin ZIPs intact; GitHub’s automatically generated source-code archives are not plugin packages. Follow SETUP.md for setup and a save/restore check.

Benchside 0.4.0

Benchside 0.4.0

Eight research roles, simpler setup and readable private records.

• Broad supervisor: pi supports exploratory, qualitative and planning work without demanding a hypothesis. pi-hypothesis retains focused hypothesis critique.
• Private continuity: save research questions and answers alongside decisions; preserve existing sitting history and decision wording.
• Readable local records: agent/ holds structured state; researcher/projects/<project name>/ holds automatically refreshed copies of the complete saved history. Setup asks once for ODT (default), Markdown or both.
• Lighter setup: keep five independent participation choices together, defer analysis-source/activity-label questions until needed, allow multiple disciplines, and suggest relevant roles sparingly.
• Community direction: contribute scientific role profiles through PRs. Shared workaround discovery and researcher connections are planned, not built.
• Agent-led installation: clearer installer discovery and honest capability checks across hosted and local agents.

Download

Installation route	Asset
Claude plugin	benchside-claude-plugin-0.4.0.zip
OpenAI / portable plugin	benchside-plugin-0.4.0.zip
Claude standalone skill	benchside-claude-0.4.0.zip
Python runtime	benchside-0.4.0-py3-none-any.whl

Give a capable agent the repository and ask it to set up Benchside using SETUP.md. It should install through available tools and ask only for choices or approvals it needs. Hosted accounts without agent-accessible installation tools may still require manual ZIP upload. Installation does not itself grant execution, durable storage or network access.

This is a downloadable GitHub release, not a public plugin-directory listing or PyPI publication. Do not use GitHub's generated source ZIP as a plugin.

Updating

Update every agent sharing a store before writing new research-question or multiple-discipline records. Existing open sittings retain their original role rules; begin a new sitting to use the broader supervisor.

Existing flat local stores remain supported. Organise them into the split layout only after every sharing agent has been updated and stopped. Migration verifies a copy, retains a private backup and preserves project identities, preferences and receipts. Generated ODT/Markdown edits do not update the journal; ask Benchside to record corrections. Native cloud notebooks keep their existing storage format.

Full release notes · Readable-record guide

Validation

536 core tests, 72 receiver tests and 64 reporting tests pass. GitHub CI passed for the updated build on Python 3.11 and 3.14. All three archives pass offline cold-install checks across all eight roles, local and native restore, ODT/Markdown export and damaged-dependency refusal. Independent review findings were fixed and documented on PR #10. The compatible receiver is deployed. A registered synthetic usage report verified the new role and multiple disciplines through the production endpoint, and Benchside’s client verified its receipt and duplicate retry. The reporting import confirmed the synthetic object was excluded and existing research observations and aggregates stayed unchanged. No real research data was submitted for these release checks.

These checks cover code and packaging, not advice quality for every model or live storage capability in every cloud account. Verify saving and fresh-chat restoration in the intended account before relying on it for research.

Benchside 0.3.1

Benchside 0.3.1

Benchside brings seven scientific roles, private project memory and optional research-workflow reporting to capable AI assistants. This is the first downloadable GitHub release, intended for research pilots and feedback.

Choose your download

Where you will use it	Download
Claude web or desktop with custom plugins enabled	Claude plugin ZIP
ChatGPT/Codex with a supported custom-plugin installation flow	OpenAI / portable plugin ZIP
Claude custom-skill installation	Standalone Claude skill ZIP
Python installation or development	Python wheel

Download the appropriate ZIP without extracting it, then install it through your app's plugin or skill controls. In Claude, open Plugins (or Customize → Plugins on older interfaces) and upload the Claude plugin ZIP. Start a fresh conversation, select Benchside and say:

Set up Benchside for my research project. Ask where I want my private records stored, explain the optional sharing choices, and verify that you can save and restore a synthetic decision.

This GitHub release provides installation packages; it does not create a public ChatGPT or Claude directory listing. Custom-plugin availability and permissions depend on the account. Do not use GitHub's automatically generated "Source code" ZIP as an app plugin.

Let an agent handle setup

The repository now starts with a short setup prompt and an agent installation guide. Hand the agent the repository URL; it should handle supported installation steps and ask only for the storage, project and participation choices it needs. Hosted apps without installation tools still require their supported package-upload step.

In Codex, Claude Code or another agent with filesystem and command tools, use:

Set up Science Works' Benchside v0.3.1 from https://github.com/science-works/benchside/tree/v0.3.1. Follow SETUP.md and the guided onboarding instructions. Perform setup for this app, ask where I want my private records stored and whether I want optional research sharing. Verify saving and restoring a synthetic decision. I should not need to type terminal commands.

Private records must live outside the source checkout. The runtime supports Python 3.11–3.14 in macOS/Linux environments; the app ZIPs bundle dependencies for supported Intel/Apple Silicon and Linux x86-64/ARM64 environments. A browser chat needs code execution and suitable durable storage tools. Reading the repository alone does not install Benchside.

Included

• Seven roles: research supervisor, study-design consultant, methods troubleshooter, results discussant, reproducibility auditor, journal-club discussant and scientific peer reviewer.
• Private project records with decisions, unresolved questions, role-specific outputs and resumption across conversations.
• Session tracking and protections against mixing concurrent conversations or overwriting newer notebook state.
• Optional, reviewed workflow-problem reports, usability feedback and usage reporting; separate choices for surveys and selected-conversation analysis.
• Optional problem context including public tool/dataset names, versions, frequency, impact and workarounds.
• Claude Docs file-based notebook transfer, its smaller document limit and clearer guidance when a network policy blocks sending.

Storage and optional sharing

Setup checks the storage tools actually available to the researcher. Local records are separate from the installed package; native or connected cloud storage needs complete reads, guarded updates and readback. A cloud execution folder alone is not durable storage.

Optional participation starts off. Gap, feedback, usage and activity submissions require review and approval of the exact content. Survey answers follow the study agreement accepted before answering. New opted-in contributions include a random store-level contributor ID, so contributions can be grouped over time; this is pseudonymous, not anonymous. Scientific findings and identifying research details are excluded from model-prepared workflow reports.

The Science Works receiving endpoint is included; users do not need Cloudflare access or API credentials. In Claude cloud execution, approved sending may require allowing ingest.benchside.science.works in the account's network settings. Installation does not change those settings.

Validation and limits

• Published packages come directly from the successful CI run for the released commit e81b5f25953b3c0bf893406740835ebfa2a82c14, including the updated installation documentation.
• 486 tests passed on Python 3.11 and 3.14, alongside lint, persona compilation and installation checks for all three archives.
• ChatGPT web 0.3.1 passed a synthetic private save/readback, independent fresh-conversation restore and read-only endpoint check on 29 October.
• The Claude transfer changes were tested with a 52 KB synthetic cloud notebook before merge. The rebuilt 0.3.1 Claude upload has not had a separate installed-account acceptance pass. Its blob replacement operation is not documented in Anthropic's Docs guide; unsupported or failed saves must be reported as unsaved.
• Host and account capabilities vary. Verify persistence in the intended account before relying on it for research, and assess scientific advice independently.
• The time-limited 29 October demo watcher has expired. Future live demonstrations need a new reporting window or an on-demand import; ordinary research reporting is separate.

Updating later

Future changes will be published as new versioned releases. Use the host's complete plugin update flow, preserve the chosen private storage location and reopen the existing project. For source-installed skills, have the agent update the checkout and rerun setup. Do not overlay a packaged ZIP onto a source-installed skill directory; follow the onboarding update instructions.

Getting started · Report an issue · Checksums