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.

dsh analysis, Rho results and a selected-row question

The current 0.1.4 local bundle matches tools and the launcher to each conversation's project; see project matching. Version 0.1.2 added the visible entries; see Desktop installation. The earlier November 5 episode records 0.1.0. The first episode below records the retired 0.0.1 embedded-view prototype, whose source remains in Git.

2026-10-31 local / 2026-11-01 UTC. This episode implements the requested dsh-rho prototype: analyse in dsh, show the outputs with existing Rho components, select a result and continue the original conversation. It follows the native plugin example in dsh-context and the installed dsh 0.2.1-rc.5 public APIs. The result is a local package, not a published npm or signed distribution.

What was built

The application adapter lives in the 0.0.1 adapter. Its Cordis host registers rho_present and an authenticated result-reading route; its browser entry contributes Rho to dsh's conversation.view slot. Native session identity supplies both the tool's caller and the view's result scope. dsh owns analysis, permissions, tools, model and history. It is not started as a second Agent inside Rho.

The selected owners are Files 0.5.0, Tables 0.5.0, Media 0.4.0 and Contracts 0.4.0 from Rho-plugins commit 1173789010e4679180804aa309e661b218198f70. Their public APIs, TableView and MediaView are consumed unchanged. The adapter supplies composition, session archives and a transient selection port; it reuses Rho's existing POSIX writer lease. No Core or Rho-plugins source changed, and neither the Core Host nor the Rho backend is started.

rho_present captures completed project files. CSV columns and types are explicit; row selection needs a unique string key. Questions resolve those keys against the captured version, then use dsh's native revision-checked draft insertion and normal Send action. A question can be read or revised before sending. Closing the view, reading a result and reconnecting do not execute analysis.

Real episode and observations

The owned fixture used a disposable project and isolated DSH_HOME, dsh 0.2.1-rc.5, Node 22.22.3, Python 3 and headless Google Chrome on macOS. It used the existing configured gateway with westlakehpc/deepseek-v4.1-flash; credentials were read in memory and not put in source, patches or evidence. The person's running Desktop Harness and projects were not restarted.

The six input values were alpha: 1, 2, 8 and beta: -3, 4, 7. A single native prompt asked dsh to write and run a standard-library Python analysis once, append one run marker, emit a CSV and SVG, and present both with rho_present. The model wrote analysis.py, ran it and presented both outputs. The run took 3 minutes 32 seconds.

Group	Observations	Mean in retained CSV
alpha	3	3.6666666666666665
beta	3	2.6666666666666665

The browser showed both rows and the labelled two-bar figure. Selecting alpha changed the question label to Ask about 1 selected row. A Chinese question asked for its group, count and mean and whether this was the overall mean. The pre-existing draft 请用中文回答。 survived insertion. The added context contained only alpha, its exact numeric values, original path, snapshot digest, capture time and the same session ID. Sent once through the native composer, the model answered in 13 seconds: alpha, n = 3, mean approximately 3.67; this was alpha's mean and did not establish the combined mean. It ran no analysis for the question.

Retained identities:

• Session: session-38a2a4c4-1e62-46f3-8ff6-234d7d098ca7.
• CSV tool call: chatcmpl-tool-a6473a9aeace26b0; SHA-256 e2bba75703875eb8484e302157f29d28a42705fa3b21332897bf86f87ba31017.
• SVG tool call: chatcmpl-tool-a9ee726ca0c532ff; SHA-256 1cb9430aed9b827880f63b866e539215b2daeab70926119ff4dfb22f9a027cb0.

Failures that changed the implementation

1. Loading an absolute package directory as a host module failed with an ESM directory import error. The source fixture now resolves the package by name; delivery uses an ordinary tarball installed by dsh plugin.
2. The first model session used dsh's retained default workspace instead of the disposable folder. It found no input and did no analysis. The fixture now explicitly creates/selects its workspace with dsh's public controller. The installation instructions require choosing the matching folder in dsh.
3. Comparing the browser Origin with the handler's internal request URL rejected valid result reads. dsh uses a carrier-neutral internal URL. The adapter now relies on the connection service's authenticated /api boundary and retains the cross-site refusal; it still validates original session and project scope.
4. Files' captured resources name archive paths. Routing those paths back into the analysis project made four initial store checks fail. The adapter's path provider now resolves captured internal paths in its archive and ordinary paths through Files' project containment. All five store checks then passed.
5. An exclusive lock file survived the first owned-host shutdown. Reusing Rho's kernel-held writer lease removed stale-unlink/PID-reuse recovery. A second writer is refused, and the installed plugin reopened the same archive after both normal shutdown and an abrupt kill of its owned dsh process.

A browser assertion initially read an empty draft before asynchronous insertion completed. Waiting for the native Chat view switch resolved the observation; the same captured values were then verified. No analysis prompt was resent. Package assembly also caught nested AG Grid package metadata without adjacent licenses; notices now come from each actual bundled dependency root, including ag-stack. An installation-guide dry run also showed that npm --prefix ... pack selected the root application. The verified command now names npm pack ./adapters/dsh-rho.

Verification and installation

npm run build, five store tests and a TypeScript check over the adapter's three source files passed. The tests cover snapshot overwrite/reopen, original-call retries, scope/path/key rejection, budgets and competing writers. They do not claim full Rho application acceptance; its unrelated suite was not run.

In the real browser:

• Table and SVG views, row selection, the existing Chinese draft and same-session model continuation all worked. Screens were inspected at 1440, 768 and 390 px; narrow results use scrolling. No browser page errors were recorded.
• Offline result reads showed an error and Try again while keeping displayed results. Restoring the connection recovered those same results.
• Editing the chat draft while a delayed question request was pending refused insertion and kept Newer draft ✓ intact.
• Refresh and a normal owned-host restart restored the original conversation and both results. The test then installed the tarball through dsh plugin --profile web add <tarball> --ignore-scripts --offline, which added and enabled its bundle. Its archive and draft insertion worked after launch.
• Overwriting the original CSV with alpha mean 999 did not alter the displayed snapshot or the selected-row question, which still carried 3.6666666666666665. The test restored the original CSV afterward.
• After an abrupt owned-host process kill, the installed package reopened the archive and original session without removing the lock file. At the end, run-count.txt still had exactly one line and the archive still had two results.

Local evidence is under work/dsh-rho/prototype-03/: evidence.json, chat.txt, the native question drafts, original CSV, screenshots of empty/results/selection/ answer/failure/recovery/narrow states, and dsh-rho-0.0.1.tgz. The private launch address stays outside the repository. The fixture and native session history are retained in the disposable /tmp/rho-dsh-prototype-03 directory. The package's composition manifest records exact bundle inputs and selected local overrides; it is not an official release attestation.

The delivered package is work/dsh-rho/package/dsh-rho-0.0.1.tgz, SHA-256 4ffe8b2798d2d0ead6e558f9d26599bd63b1d5d0b7eecd54efa85684e5396266. Final review keyed result choices and questions by original tool call as well as snapshot, so two presentations of identical bytes retain their own titles. The five tests and strict type check passed; the final tarball was installed again, and its table, figure, native draft insertion and original call ID were checked in the same session without resending analysis. The run count remained one.

What this establishes and what remains

This observed episode establishes the requested local loop with reused components and one native dsh conversation. The adapter is deliberately one-project and one selected dsh release candidate. It has no general Studio, notebook, live variable inspection or automatic output discovery. Row selections are transient, image questions carry references rather than pixels, and a question's structured context is visible in the native draft. React comes from Harness; broader compatibility with the components' declared React peer version is not established by this run.

The 0.0.1 package README names the budgets, configuration and refusal behavior. Invalid table content can leave a bounded captured file unlisted. There is no silent deletion of earlier results or automatic replay of unknown work. Independent researcher use, Desktop/TUI entry points, Windows locking and public package distribution remain untested.

Reuse the whole Rho application

2026-11-05. The user asked to reuse the existing Rho at the lowest integration cost. Inspection of Harness 0.2.1-rc.5's installed MCP client and bundle reference, and its public bundle documentation, showed that a bundle can contain configuration without Host or Client entry code. Rho already supplies the project grant, stdio MCP bridge, full WebUI and frozen review-context handoff. These existing boundaries were sufficient for the flow.

The 0.1.0 adapter now contains two configuration files, README and license. Its patch mounts Harness's existing MCP client and runs rho --mcp with an existing agent connection file. It adds no tool implementation, browser entry, service, result archive, Skill copy or model configuration. The previous 297 lines of runtime TypeScript, build script, lockfile and obsolete store tests were removed. No Core or component source changed. Rho's WebUI stays in its own window; its iframe policy was not relaxed. Unlike 0.0.1, returning a question requires one explicit copy/paste into the original Harness conversation.

Installed application episode

The disposable project was /private/tmp/rho-dsh-reuse-20261105/projects/analysis, with separate service state and DSH_HOME. The service used an already installed @rho-workspace/rho@0.5.3 from /tmp/rho-npm-053, with its packaged components; no application build was needed. Source main was 0.5.4, but the registry returned 404 for that version at the time of this check. The person's other Rho service, Harness Desktop and projects were left running unchanged.

An owned Rho service granted this project to session dsh. The 0.1.0 tarball was installed through dsh plugin --profile web add … --ignore-scripts --offline. The unchanged skills/rho directory was linked into this trial's Harness skill folder. The model actually read it, resolved Rho operations, wrote a Python standard-library script, ran it through kernel.run and registered a keyed CSV and SVG through resource.register. Model credentials stayed in memory, outside the bundle and evidence.

The first configured model route, westlakehpc/deepseek-v4.1-flash, failed with an upstream duplicate-item-ID error after its initial tool calls. Both Chat Completions and Messages attempts failed, including a clean Messages session. The project still contained only the input file, and no Notebook cell had run. The trial then selected the already configured default route fusecode-lab/claude-opus-5-5[1m] through Messages. This succeeded. These are the configured route names, not independent verification of the model providers. The bundle makes no model selection and did not work around the failed gateway.

The successful analysis took 59 seconds in Harness. The six input values were again alpha: 1, 2, 8 and beta: -3, 4, 7. The script wrote:

Group	n	CSV mean
alpha	3	3.666667
beta	3	2.666667

Notebook cell k-441c12dcff1d7d0b, change ID analysis-run-20261105-1, completed in the shared Python 3.13.1 session. Its output included both means, the CSV and run-count lines: 1. In the real browser, the Notebook showed the code, output and live ns value; existing table and media views displayed the two rows and the 400 × 300 SVG. Existing project notes were readable alongside the Studio table.

Selecting alpha and asking in Studio created question q-99d55f47-8bcb-4f1b-bc57-b9bd9bcb7cc7, with frozen context 1f4d589da930eaf21214d9024bfabedc7c5b76e1c7f655c674c1fd613269636d. The retained CSV digest was 014d7d55f837f35a400dededec75999c5becd392a5355461be2795862509a4ca. Through Review's External Agent · copy request, that context request was pasted and sent in the same Harness session. In 46 seconds the model read the context and wrote annotation.reply, retaining its original message identity and exact result versions. Rho displayed the Chinese answer: alpha, n = 3, mean 3.666667, the mean of that group rather than all six observations. The answer created a reply document; it did not run the analysis again.

The actual route needed Open in Studio: the ordinary file preview selected rows but did not show the question control. On this installed version, Review this question left the Studio visible, so the trial used the Review sidebar. Reading the clipboard before the asynchronous copied confirmation initially returned empty text; the disabled Harness Send button prevented an empty turn. Waiting for Rho's copied confirmation supplied the exact context request. The guide now follows these observed controls. No claim of one-click delivery is made.

Packaging and recovery checks

The final package contains exactly package.json, cordis.patch.yml, README.md and LICENSE: 15,053 bytes compressed, 41,643 bytes unpacked. No dependencies or install/build scripts run. It was installed again with the ordinary offline bundle installer. A missing connection file then produced ENOENT and Harness's rho-mcp … initial connection or tool synchronization failed warning; Harness opened with the entry unavailable. The Plugins page showed installation, which by itself did not establish MCP readiness. Rho was not started or restarted by this failure.

While the owned Harness was stopped, the Rho Notebook remained idle with its original live value and one cell. Restoring the correct connection and restarting only Harness reopened the same conversation and allowed a passive read of the original cell. The run counter stayed one. The bundle's five-minute MCP reply wait is configurable; this episode did not exercise long-running cells, all four tools' large-result paths, other operating systems, Desktop or TUI.

Local evidence is in work/dsh-rho/reuse-01/: chat and handoff text, the missing-file log, package inventory, screenshots of Notebook, table, figure, question and reply, and the observation summary. The private launch addresses and credentials remain in the disposable trial directory. The final tarball is work/dsh-rho/package/dsh-rho-0.1.0.tgz, SHA-256 8b3b56f9d1c3370e8256ce167cece8753acd5a569e822e3036f4baadca64fd1e.

The minimum entry route is an ordinary Harness bundle. Its market reader uses user-configured indexes and has no default index in the tested release. Git subdirectory packages are supported by that installer, so a listing can point to an available source commit without packaging Rho again. Public Git installation, npm publication and market inclusion were not performed. The local package stays private. The old 0.0.1 source history and archive bytes remain; no migration or archive deletion was attempted. This is one observed integration episode, not full application acceptance or independent usability evidence.

Desktop installation and visible entry points

2026-11-05. The user installed the connection-only bundle and found no Rho tab or button in Harness. The missing entry was visible in their screenshot. Version 0.1.2 adds Rho to conversation.view and sidebar.panellist, both opening a small launcher. Open Rho workspace uses the existing sidebarRight.openTab Browser to display the complete application beside the conversation. Open in system browser remains available. The Rho application and its iframe policy are unchanged; Desktop uses Harness's native webview.

A small Host route reads at most 64 KiB from the chosen agent connection file, extracts its local service address and project ID, and returns only the project URL. Credential fields never leave the Host. It refuses remote or credential-bearing addresses, missing project scope and oversized files. The launcher neither starts a service nor submits analysis. The existing MCP entry and four tool names are unchanged. React comes from Harness; only the launcher is built, with esbuild 0.25.12. No Rho component code is bundled or modified.

The real installation used the person's already running Harness Desktop 0.2.1-rc.5 on macOS and the existing Rho 0.5.3 trial project analysis. Its original service, Notebook, one cell and retained results remained in place. Desktop refused dsh plugin --profile desktop because the Electron application exclusively owns that profile. Its Plugins → Add plugin → Enable now path installed the local tarball successfully. The prior profile files were backed up outside the repository. The unchanged Rho Skill was linked into Harness's skills directory.

An initial partial MCP config override failed schema validation: a profile config replaces the complete object, so command and arguments alone dropped serverName and transport. A complete override and a bundle off/on recovered it. Desktop's Finder launch did not use the shell exports, so its profile now explicitly names the same agent connection for the MCP and launcher entries. Neither the other configured MCP servers nor their grants were changed.

In the actual Desktop window, both Rho entries appeared and the final bundle reported 2 running components. The conversation tab's open button and the sidebar's open button both displayed the full Rho project inside Harness's Browser. Switching back to Chat kept Rho beside it. The existing summary.csv displayed alpha (3, 3.667) and beta (3, 2.667), with alpha's prior selection retained. The final screen was visually inspected and left open for the person. This check opened existing work; it did not run a new model analysis or establish another selected-row model continuation beyond the earlier episode.

An isolated Web Harness exercised a missing launcher connection: the entry stayed visible, Open Rho workspace was disabled, and Try again was available. Restoring the file and retrying recovered the same launcher without restarting Rho. Another run found that a missing Browser provider switched away from the launcher before displaying its error. The final client checks the tab registry before navigation and restores the prior panel on a failed open. Repeating the same Web action then kept the launcher, its explanation and its system-browser link visible. Reusing the same local tarball version initially retained an older pnpm asset; checking actual installed bytes caught this, and the final 0.1.2 package uses a distinct version and tarball identity.

The adapter build and both credential-projection/error-route tests passed. The final installed Desktop files were compared to the source package. The tarball has six files and is 19,011 bytes compressed (51,727 unpacked), at work/dsh-rho/package/dsh-rho-0.1.2.tgz, SHA-256 60dbeb0a5417d92682fa169f1fcb6d33e9efd6418bb4704b71bfe7a68c0d11cd. Local evidence is under work/dsh-rho/launcher-01/, including the Desktop window, missing-connection and recovery screenshots and the installed-package summary. The owned Web check is disposable; the person's Desktop and existing Rho services stay available for their trial. Public distribution, TUI, other operating systems and independent usability evidence remain outside this check.

Public Git distribution

2026-11-05. The user asked to complete the publisher's work for market discovery, leaving user configuration alone. The npm account check returned HTTP 401, so this release uses Harness's documented commit-pinned Git package source. No npm publication is claimed. The existing source package omitted lib/, which makes a source-only Git install incomplete without running a build. Version 0.1.3 tracks the prebuilt client with its source and adds an exported English display locale, the existing Rho SVG icon, author, homepage and search keywords. It has no install or prepare lifecycle script, and keeps the checked engines.dsh: 0.2.1-rc.5 range.

The Host, built client and MCP patch match the installed 0.1.2 package byte for byte. No Rho component or runtime behavior changed. The launcher build and both adapter tests passed. The packed candidate has eight files, is 19,642 bytes compressed (53,288 unpacked), and has SHA-256 34f0ac6e29b562a331ea2536c55d2f85d50b31f439d410499174311fbf2f9c2f. The inventory is under work/dsh-rho/publish-01/package/.

adapters/dsh-rho/market.mjs reads one explicit, full source commit, including its built client, locale and icon, and generates the index from that version's manifest and commit timestamp. Public Git installation, official index validation, search and public index delivery are the next observations; the person's existing Harness configuration and Rho sessions are not part of those checks.

Public installation

Source f3248a8a was published before the index reference. Harness 0.2.1-rc.5 installed its exact Git subdirectory target in two new profiles, using separate empty pnpm stores and no Git credential helper. The first disabled install scripts; the second used the normal defaults. Both completed successfully, downloading one package, in about seven seconds. The scripts-disabled run printed pnpm's generic Git build warning, but its eight distribution files were present and matched the source. The normal installation needed no build approval or additional dependencies.

Harness's own dsh-plugin-market-index 0.2.1-rc.5 accepted the generated index and produced the same commit-pinned install target. Its Rho title, icon, author, homepage, keywords, version, source and compatibility are read from that fixed package. The index's timestamp is the source commit's committer time. The publisher's checks use disposable Harness homes; the person's installed Desktop bundle and existing Rho/Notebook sessions are unchanged.

In an isolated Web Harness, the optional Discover bundle read a local preview of that same index. Searching Rho displayed one card with the original rho icon, Prototype label, version 0.1.3 and YuLab-SMU author. An unmatched query showed No plugins match; restoring Rho recovered the card. Its detail page displayed the exact commit and package directory and marked 0.1.3 Recommended Installed. The public package's sidebar launcher loaded too: with no Rho connection supplied, it kept the disabled open button, explanation and Try again. The fixture disabled rho-mcp, so this was a package, metadata and discovery check, without a Rho service or model analysis. Screenshots are in work/dsh-rho/publish-01/.

The page assembler now carries site/market/ into the Pages output alongside the existing documentation. The documentation build and coverage check pass with 54 documents; the rendered site's local links and anchors resolve, and the 123 existing files outside the front page and documentation remain byte-identical. The generated index and public guide are the publisher's deliverables. Enabling a market source in a person's Harness is outside this change.

Conversation project matching

2026-11-05. A person's Harness conversation had opened their real project in a second Rho service, while the plugin's global grant and launcher still selected an earlier disposable analysis project. The conversation and Open Rho workspace therefore referred to different folders and services. The launcher did not show either identity. Changing only the URL would leave the tools on the wrong project.

Version 0.1.4 selects an existing grant by the conversation's canonical working folder. It verifies that folder against the project's read-only Rho observation, then gives the same selected descriptor to the conversation's own official Harness MCP client and launcher. The global rho-mcp component is removed. Both launcher entries show the project name, full folder and service address. No match or multiple matches refuse both the link and tool access; the panel names the conversation folder and reachable configured projects. Subfolders do not broaden a grant, and a symlink to the same canonical folder does match.

The first real Harness check exposed a lifecycle mistake: agent/created delivers { agent }, not the Agent itself. After correcting that owner code, two same-named Unicode projects on separate disposable Rho 0.5.4 services each returned their own project ID from a real mcp__rho__rho_index call. Each ID matched its launcher's project URL. A third, unconnected conversation had no Rho tools, returned a launcher refusal and refused a forced tool call. Switching projects in Web Harness visibly changed the full folder and service address. The unconnected card kept Open Rho workspace disabled. The Web fallback opened the selected project's actual Rho page in a separate browser tab.

Six adapter regression tests pass, including two-service matching, Unicode, symlinks, ambiguous grants, missing connections, restored-file recovery, per-Agent attachment and replacement of a live descriptor. A read never resumes a dormant Agent or starts a bridge. Try again can retry an initially failed connection for a live Agent; changing an already bound descriptor refuses new calls until plugin reload, without replaying work. The plugin checks the current folder again before each Rho call. It permits eight active conversation bridges by default (maxBridges); exceeding the limit refuses the new bridge and keeps existing ones. Probe and response bounds are recorded in the package README.

The person's Desktop 0.2.1-rc.5 profile was backed up before upgrading the local bundle to 0.1.4 and replacing the two old overrides with one configuration of already-existing project grants. Their original conversation's Rho tab showed the correct full folder, and its button opened that same real project in the Harness Browser. The running Rho services, scientific sessions, conversation history and background task were retained. No new model turn or scientific execution was requested for this check. Local evidence and disposable fixtures are under work/dsh-rho/workspace-01/; the tarball is under work/dsh-rho/package/. All nine installed files match the source; the 23,231-byte final tarball has SHA-256 f27327bfacc067049f4ef541a6eb7891ffc3c879e3669d43e09a45f7be4ebe8a. That local check left the public market at 0.1.3; the publication below updates it. TUI, other operating systems and independent usability remain unobserved.

Public workspace fix

2026-11-05. The user asked to update the public index and online installation guide after the 0.1.4 workspace fix. The selected package source is 4bc6f6bb, published before the index reference. Its runtime files and prebuilt launcher match the already checked workspace fix in b646ed51; only its package README now describes the public 0.1.4 route. The index retains the original icon, display locale and checked Harness 0.2.1-rc.5 compatibility, and pins the full source commit and its timestamp. The adapter build and all six tests pass.

Two new Harness homes installed that exact Git subdirectory, with empty npm caches and pnpm stores and no Git credential helper. One disabled install scripts and one used normal defaults. Both completed, in about 66 and 49 seconds. All nine distribution files, including the manifest, matched the selected source. The scripts-disabled install printed pnpm's generic Git build warning; the prebuilt client was present. Both printed a missing MCP-client peer warning: Harness supplies that peer when it loads the plugin. The earlier real Harness/MCP check covers the unchanged runtime files; this publication does not repeat the model or scientific analysis. The official market-index parser accepted the generated 0.1.4 index and returned the same fixed install target.

The online guide, start page, front-page entry, package README and release summary now name 0.1.4. The guide replaces the old two-entry configuration with one dsh-rho entry and describes removal of the old rho-mcp override, exact conversation-folder matching, refused or ambiguous connections and recovery. Rho’s index is a manually configured source in Harness’s optional Discover feature; it is not an official default source. No npm publication or installed user-profile change is part of this update. Local installation and publication evidence is under work/dsh-rho/publish-02/.

The public index returned HTTP 200 with version 0.1.4 and bytes identical to the checked index. An owned, freshly installed Web Harness read that public URL directly: searching Rho displayed the original icon, 0.1.4, Prototype and YuLab-SMU. A nonmatching query showed No plugins match; restoring Rho recovered the card. Its detail page showed the full selected commit and package directory, with Recommended Installed. This also loaded the installed Host and prebuilt launcher despite the installer's peer warning. The public guide, start page and front page returned HTTP 200 and matched their deployed bytes.

The first local documentation preview used the development /docs/ prefix and showed missing styles. Rebuilding with RHO_DOCS_BASE=/Rho/docs/ corrected it before publication. The rendered changed sections were inspected, and the full Pages tree passed 3,971 local link/asset and 765 anchor checks. All 123 existing files outside the front page, documentation and market remained byte-identical, including the demo, update feeds, setup page and Skill. The initial Pages deployment is 244f0167 from source 43e23fd5; the observed public verification is recorded here in the following documentation commit. Public screenshots and HTTP receipts are in work/dsh-rho/publish-02/.