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.

Base-R collaboration episode

This episode exercises an ordinary R project: select a result row, ask about its method, let a real external Agent edit and execute the analysis, compare both outputs, and return after connection loss. It uses the public owner contracts and installed Chrome. The person is automated; this is not a human usability study or scientific acceptance of the method. Each run ends with a review that decides what the next run should show.

Question and objects

The project fixture uses datasets::airquality. The initial script selects complete observations across all variables and calculates monthly sample counts and mean ozone. It produces monthly-ozone.csv, with stable keys month-05 through month-09, and a base-R PNG. The only runtime is R and its included packages; the tested R is 4.5.2.

In the browser, the person adds a Unicode code comment, saves it, leaves a separate note unsaved, sorts the table, filters May and selects its row. The Chinese question asks why May has only 24 observations and asks the Agent to drop only missing ozone values, preserve the comment, recalculate every month and link both new outputs.

The question freezes the table identity, version, month-05 key and confirmed draft. Copying its request is an explicit handoff to an external Agent. The runner passes that actual clipboard text to Claude Code with a project-scoped MCP session; it does not generate an answer or edits in place of the model. The model has no shell/filesystem tools or layout-write grant. It reads the frozen context and shared document, saves against observed conditions, submits one new managed R operation, checks artifacts, reads rows and records an explicit linked reply.

Comparison and ownership

The expected descriptive results are independently calculated by a separate base-R expression, not by the Agent's modified script or explanation:

Month	Complete-case n	Complete-case mean, ppb	Ozone-only n	Ozone-only mean, ppb
May	24	24.125000	26	23.615385
June	9	29.444444	9	29.444444
July	26	59.115385	26	59.115385
August	23	60.000000	26	59.961538
September	29	31.448276	29	31.448276

The original table remains bound to its original version. Its May selection does not migrate to the new table. The person opens the exact table linked in the reply and chooses Arrange views → Split right to inspect both versions, each with its own producing-run line. Both original and linked PNGs remain accessible. Reload restores the saved comparison; below 768 CSS pixels the two views remain reachable as separate tabs. At narrower widths a real horizontal wheel exposes the mean column, and the full numeric value is checked against the oracle. Acknowledged unsaved notes remain in the document owner.

This exposed an application gap: an opened linked result was a local view outside the saved composition tree, so its Arrange views menu was absent. Rho now includes only that exact view in the person's conditional arrangement write. It neither rebinds an existing same-id node to different bytes nor persists unrelated previews. Existing composition freshness, reference, availability and protection checks still apply. Core and plugin semantics did not change.

The source line distinguishes completed execution, verified bytes and the absence of a domain check. The external Agent's reply is its interpretation. The separate oracle verifies these five aggregates; it does not establish missingness assumptions, uncertainty, significance or causal conclusions. The person's note leaves adoption pending. View placement and execution success do not record scientific approval.

Failure and return

Two failures belong to this episode:

1. The question is accepted, but its browser response is deliberately dropped. Reload and Retry original question send the original body and change identity. There is one immutable question, with the original text, row key and version.
2. The browser goes offline while the Agent's R operation is active. The script has an explicit six-second test window; normal fixture execution does not wait. The Agent continues under its existing grant. The person restores the connection and clicks Reconnect. Rho observes the same request digest, Core caller/request identity and native operation identity. There are exactly two native submissions: the baseline and the requested change.

The browser also reloads after saving the comparison. The old selection and dirty notes survive. Sort/filter controls return to their existing defaults on reload; the selected stable entity and frozen question remain exact. This record does not claim Host crash recovery, restoration of every local presentation setting or automatic resumption of an external model session. Those require separate episodes.

Review of the 2026-10-25 run

The run asked whether a real external Agent can answer an anchored question through Rho's public owners, and whether the person keeps their place and work across the change and a disconnection. It takes this episode to the runs stage (stages). The observations come from the retained screenshots at 1440 × 900 (01-point-ask, 02-question-ack-lost, 03-offline-during-run, 05-compare-tables, 07-linked-figure) and at 390 px (comparison-390).

What works. The Agent read the frozen context, changed only the inclusion rule and its comment, kept the person's Unicode note in the script, ran R once and linked both outputs. The selection, both versions and the unsaved notes survived; Reconnect observed the same request, and there were exactly two native submissions. The offline notice says what the person is looking at and offers Reconnect without covering the table. The one application gap the run exposed, a linked result that could not be arranged, is fixed.

What does not work yet.

1. The answer is not on any screen. May goes from n 24 to 26 and mean 24.125 to 23.615, August from 23 to 26 and 60.000 to 59.962, and the other months do not change. To learn this, the person compares two tables cell by cell. The Agent's explanation sits behind a button labelled "Explicit Agent response"; in place it is one truncated line of raw Markdown (Agent: "**为什么五月原来只有 24 条：** 原 a...).
2. The comparison is two unrelated tables. Their headers and first rows are 48 px apart. The left is named by its rule ("Original · complete cases"), the right by its file (monthly-ozone.csv). Nothing marks the changed cells, and one column shows 60 beside 59.12.
3. Implementation facts lead. "Open linked result · b517bba0" and "Open linked result · fcde896f" leave the person to guess which is the table and which the figure. Each result view is subtitled "completed · bytes verified · no domain check". After the lost acknowledgement, nine lines describe one question, the word "retained" appears five times, and the first line is the raw browser error "Failed to fetch".
4. Everything has the same weight. Three stacked bars, tabs, four bordered buttons and a status line come before the data. The first data row starts nearly 400 px down a 900 px window, while about 300 px stay empty below five rows. "Step away" does not say what it changes.
5. At 390 px the versions cannot be seen together. They become separate tabs, so comparing means remembering values across a switch: recall instead of recognition.

Analysis. Problems 1–4 share one cause at the concept level. The workspace is assembled from the views that exist, one tab per file and one button per link, rather than from the episode: a question, what the Agent changed, what changed in the results and the evidence behind it. Restyling would not fix that. Problem 2 also has a capability part: the Tables plugin offers a keyed data.compare, but no view presents its differences. The mixed precision in problem 2 belongs to the table view. The labels in problem 3 are each true, but they describe the implementation rather than the work. The question's labels and linked-result buttons come from the Annotations plugin's table-question view in Rho-plugins; the run subtitles come from Rho's shell. Problem 5 is composition: below 768 px the shell shows one view at a time.

Next run. Present the episode around the question: what was asked, the Agent's answer, the change between the two results by month key with the changed values marked, and the code change, with tables, figure and provenance one step away. Keep the project, prompt and failure injections so that the runs can be compared. The run succeeds if the change in May and August and its reason can be read from one screen at 1440 px and reached at 390 px; then a person who did not build Rho tries to say what changed and why. If the person still has to compare tables to find the change, the presentation is wrong, not merely unpolished.

Reproduce

From a configured source checkout, after building the selected plugin packages:

npm run build
RHO_CORE_BINARY=/absolute/path/to/rho-core npm run acceptance:base-r

The command needs the installed Chrome executable, Rscript and an already configured Claude Code client. It uses the existing local model configuration without printing credentials. RHO_UI_CHROME selects Chrome; RHO_RSCRIPT selects Rscript. RHO_EPISODE_EVIDENCE selects the evidence directory, and RHO_EPISODE_HEADED=1 shows the browser. The runner creates its own temporary project and owned service, closes its browser/service on completion and retains the project and evidence. It does not restart a user's Host or R session.

The model is bounded to 48 turns and a ten-minute process deadline. On exhaustion, failure or incomplete evidence, the command fails and retains obtained output; there is no deterministic fallback and no automatic new execution identity. Tables have five rows, each read is bounded, and the ordinary artifact/output and view budgets still apply. No add-on R packages are installed by this episode.

For debugging later browser steps, RHO_EPISODE_RESUME can point to this runner's saved evidence directory. That mode reuses recorded outputs and performs no model run or native submission. Its report explicitly names the prior evidence; it must not be counted as a fresh end-to-end pass.

Evidence

The selected run and its remaining limits are recorded in Status. Local evidence is under work/base-r-episode-20261025/: the report, model activity receipts, native identity before/after connection loss, exact result references, all compared rows, independent oracle output and browser screenshots. Private model transcripts and connection files are not source artifacts.

This is a source-checkout composition. It uses the accepted Core binary from ebe351043e74529ccf49960c0b07c805d1e50c07, with SHA-256 recorded in the report, and Plugins 3588175e25371ae1fa953d92bc24e5897489d7f5 with the pre-existing local table-style override. This work leaves that override untouched. The report records every selected UI entry and the browser asset digest; implementation file hashes identify the application/runner/fixture bytes. No component API, package selection, binary distribution or signed-release claim is introduced.

The first runner attempt corrected document-field/DOM readiness assumptions. A later attempt preserved a successful real Agent run after a runner response-shape error. before-fix/ reproduced the absent arrangement control on its actual new table. after-fix-2/ verifies the repaired browser path against those retained outputs. full-run/ retained a second real Agent's successful output when the runner incorrectly expected network restoration to replace the explicit Reconnect action. scroll-check-2/ checks column access against those outputs after correcting a grid DOM selector. These are diagnostic evidence, separate from the final fresh run under accepted/.

The layout regression checks cover two explicit bindings, id collision refusal, unchanged sibling weights, movement and removal. The existing frontend acceptance also exercises saved views, keyboard navigation, Unicode/undo, handoff and reload. Human observation, a DESeq2 episode, external-write conflicts and broader unknown outcomes remain separate work in the roadmap.