Historical record

This page records what happened on a date, with its observed failures and limits. It is not a requirement for the current implementation. Current behaviour is in Status.

macOS arm64 preview 0.5.1

This is candidate 2 of the unsigned local npm tarball and WebUI. It supersedes the retained first candidate after fixing the newly created review conversation target, so later feedback returns to that conversation. The selected artifact is rho-workspace-local-0.5.1.tgz, 274921791 bytes, with SHA-256:

33f4def5b10475836e11d49cfdecb5f36ccb51b99081f62cd6ace67b1617c32a

The local deliverables are under work/review-evidence-implementation/delivery-051-r2/: tarball, SHA256SUMS, composition.json, ACCEPTANCE.json, INSTALL.md, selected component tarballs and example-evidence.zip. Registry publication, GitHub binary releases, signing and other platforms are outside this preview.

Fixed source and component selection

Owner	Commit
Application	c90742dbd08e69c3f943ea8fc58c7c49a80cd32b
Plugins	72553eebba77a091e75cbc91cbff0354f89d31de
Core	4f5bed8631e15acad44cc86b63ea4779ab8af342

Both source selections were clean. The Core is the earlier accepted binary with SHA-256 434a92fced13595d281b1f9e2d6d49cf9fa7c8a451da08aa3c1ba86cec85eab3; it was not rebuilt. The composition also records Git trees, build-input/lock/browser-asset digests and component tarball identities. Markdown 0.1.0 is a selected browser extension, compiled into the WebUI; its source is included in the Plugins input selection.

Runtime component	Version
contracts	0.4.0
documents	0.4.7
files	0.5.0
r	0.4.1
requirements	0.4.0
tables	0.5.0
media	0.4.0
charts	0.4.0
agent	0.4.18
annotations	0.5.1

Install and start

Use Node 22.22.x, Python 3, and a configured Claude Code or Codex client for Agent conversations. R is required only for R analyses. The checks used Node 22.22.3, Python 3.13.1 and R 4.5.2 on macOS arm64. The tarball includes its selected Core; it downloads or replaces no runtime on startup.

shasum -a 256 -c SHA256SUMS
mkdir -p "$HOME/rho-preview-0.5.1" "$HOME/rho-projects"
npm install --prefix "$HOME/rho-preview-0.5.1" --omit=dev --ignore-scripts /absolute/path/rho-workspace-local-0.5.1.tgz
"$HOME/rho-preview-0.5.1/node_modules/.bin/rho" --root "$HOME/rho-projects" --state-dir "$HOME/.rho-preview-0.5.1" --open

Stop the owned launcher normally with Ctrl+C. Keep the printed private connection files private. An unavailable R installation does not prevent native editing, review or Agent work. Missing prerequisites are reported for the operation that needs them. A wrong Core platform or digest is refused.

Acceptance of this artifact

The same tarball installed outside all repositories without development source links or Core/runtime overrides. The installed launcher used its bundled Core, served the WebUI and exited with no owned Host left. Installation took 8.2 s, ready 0.76 s, and the first browser render 1.32 s on this machine; these are single observations rather than a performance benchmark.

The frozen source passed Rho 230/230 and Plugins 110/110 checks. Contract checks passed. Frontend's 13 and Sources' eight browser checks passed. Installed review checks cover actual R output, keyed multi-field differences, captured code, whole-resource questions, Unicode reasons, lost question/decision responses and original-identity retry, modal Escape/focus return, refresh, Graph focus/backlink, service restart with the same review/result/conversation, and two direct Review sends that keep the same newly created Agent conversation. Screenshots were read at 1440, 1024, 768 and 390 widths.

Python and R each ran once and agreed with their Markdown report; overwriting the scripts/report kept old captured bytes and Graph refs readable. The installed service also retained an acknowledged dirty Unicode document across shutdown and restart. Candidate 1’s actual macOS folder chooser returned and registered an existing directory; its content was not edited. Candidate 2 reuses that observation for its byte-identical picker implementation. An expected cancellation was instead a selection, so native cancellation is unobserved. Wrong-platform and digest fault cases were refused; the original installed composition bytes were restored. The existing no-Core runtime path was separately checked with Core and R absent.

The candidate 2 installation completed the two-turn question, result and human feedback loop using configured real Claude and Codex clients. Claude retained three permission records with its default 32-choice budget; Codex retained 41 with an explicitly selected 64-choice per-turn budget. An earlier installed Codex attempt exhausted the default 32; its Rho grant scope did not change. ACCEPTANCE.json records the selected budgets and results. Earlier failed attempts remain evidence. The client models use the selected Westlake HPC route; these are client integration checks, not a comparison of model scientific reasoning.

Independent offline reader milestone

The installed artifact separately produced a read-only evidence package. Its manifest and reader distinguish complete package generation from unestablished analysis evidence completeness. The source service was closed, the project moved and browser networking disabled; example-evidence.zip is a verified, independently readable example. index.html still showed the selected question, replies, decisions, table and captured files at four widths. Changed bytes failed integrity verification. Use rho --verify-evidence /path/to/package for inventory checks without starting a service. The package neither restores a workspace nor reruns analysis.

Limits and follow-up

No independent researcher trial or physical IME observation establishes usability. The episode contains a researcher trial guide. Default native permission retention is 32 choices per turn; a model that exceeds it needs an explicitly selected larger conversations.maxPermissions budget. Captures and output already obtained remain available. Native permissions retain the Agent's own policy and do not authorize broader Rho access.

Windows/Linux folder choice, POSIX locks, executable discovery, packaging and installation remain portability work. No account/cloud/team/billing, desktop shell, public npm publication or signed release is included. Human acceptance records a version-specific work decision; it does not establish scientific correctness.