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.

Timeline events and the Notebook's name

Updated 2026-11-01. The user asked for two things. Plugins should hand the timeline "who did what, when" as events, so the person's Notebook cells and Console commands appear on it. And they asked whether the Notebook should be renamed. In a Rho conversation they had asked "可以在notebook 中演示一下ggplot2吗？", and the Agent made a Quarto document instead of using Rho's Notebook.

The Notebook's name

Observed. That conversation began on 2026-10-27, before Rho told its Agents about the Notebook (2026-10-30). In the turn, Claude Code with DeepSeek V4.1 Flash checked R, ggplot2, IRkernel and Quarto with Bash. It found Jupyter with only a Python kernel, wrote ggplot2-demo.qmd and rendered it. It never looked up kernel.run. The conversation's grant held kernel.run. The ACP adapter (claude-agent-acp 0.97.0) passes Rho's appended instructions when it loads a session as well as when it creates one, so the Notebook part was present.

Reproduced. The question was asked with the person's exact words, the same model, Rho's appended instructions, Claude Code's native tools and Rho's MCP tools under the conversation profile (work/notebook-name/headless.mjs). Nothing asked for permission: Rho's tools, reads and listings ran, and other calls were refused and recorded.

Run	First Rho step	Cells run in the R session
New session	rho_index for kernel	8
First message answered without the Notebook part, then resumed	rho_index for kernel	4
As above, with the person's own Claude Code settings	rho_index for kernel	3
New session with the sentence below	rho_index for kernel	10

Every run used the shared R session; none looked for Jupyter or Quarto. Replaying the failed conversation's own history meant writing a copy of its transcript into ~/.claude/projects. Claude Code's permission check refused that as transcript tampering, so the failure itself was not reproduced. A session whose earlier messages worked in other ways is the remaining difference.

Decided. The Notebook keeps its name. The person says "notebook" because the page is called that, and Agents in new sessions took the word as Rho's sessions in every run. A rename would change the words on every screen without touching the case that failed. Agents started in Rho now read one more sentence: something asked for "in the notebook", or in the R or Python session, means these sessions on Rho's Notebook page, not a Jupyter, Quarto or R Markdown notebook, unless a file is named. Whether it changes the outcome for a long, older conversation is not established.

Who did what, when

Before. The timelines drew only what conversations record. The person's cells and commands were absent. An Agent's own cells were absent too, because a rho_call to kernel.run is recorded as a project call, not a run.

Changed.

• Plugin contracts carry one event shape: who (by), what (label), where, when it began and ended, and how it ended (running, done, failed, stopped, unknown). A plugin names its events read in timeline. Notebook 0.5.0 reads kernel.events and Console 0.3.0 reads console.events, newest first, without code or output.
• Rho lists each plugin's timeline in the catalog, reads every one the tab may call, and places each event by principal. The person at this machine has a lane named You, above the conversations. An Agent started in Rho is known by its conversation's principal, so its cells join that conversation's lane. Anyone else shares a lane named Outside Rho. Pointing at a block says "You ran summary(fit) · R", with Failed, Stopped or Outcome unknown when it applies.
• The Overview's strip appears when only the person has worked, as What you did, and otherwise as What you and the Agent did.
• A conversation's Studio draws its own Agent's events and the person's since it began, the same share its Notebook view shows. A replayed moment at one of them brings the Notebook or Commands view forward when the Studio has one.

Runs

work/timeline/episode.mjs copied the person's Rho-Playground to /tmp, with its conversation and its Notebook cell. In the browser the person ran a failing cell, a working one and two commands. A caller holding an outside grant ran a cell, as an Agent outside Rho would. The Overview, the Graph and the conversation's Studio were then shown at 1440 px.

Pass	Observed	Change
1	Lanes You, the conversation and Outside Rho with the right events, and a tooltip naming the command. The 20 seconds of new work had about 10 px after a 4.7 h pause, beside a stretch of days	Each stretch first gets room for its steps (stretchSizes), then shares the rest by duration
1	The conversation's brand mark stood above its name: the global .conversation style matched the lane's class	The lane marks who it is with data-who
2	Four separate blocks in You with the failed cell outlined, the outside cell below, and the person's cells after the pause in Studio's side timeline. The outside cell was left out there	—
Person only	A new project with two cells and a command shows What you did with one lane, with no page errors and no sideways scroll at 390 px	—

Checks

Rho 267/269 in the ledger. A rerun showed its two misses: the MCP v2 test at its 30 s limit and the .rho cleanup race after a Console test. Each passes alone, MCP v2 in 23.9 s twice. Plugins 163/163, with new tests for both events reads. The pinned contracts export list gained the timeline names.

The graph, frontend and pages acceptances passed. agent-canvas failed where it clicked 10 px into Studio's side timeline. The click had depended on a cramped first stretch to land after the first read; with the new sizing it landed on the question, and an earlier step's check had been passing on an empty Studio. The acceptance now clicks the first read mark, passes 10/10, and its replay shows the data and script read then.

Left open

• No Agent started in Rho ran cells in these browser runs, because its permission prompts are the person's to answer. Its cells in its conversation's lane are covered by unit tests only.
• A replayed moment at the person's cell bringing the Notebook forward was not exercised.
• Lane names are hidden at phone width, for conversations as before.
• The Graph page still shows "Nothing to draw yet" without conversations, even when the person has worked.