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.

Code, Markdown and runs in a project

Updated 2026-10-29.

Question

The user asked to put Python scripts and a Markdown report into the test project and see how Rho shows code, runs it and renders Markdown, then chose to work in three steps: code highlighting, Markdown rendering, and legible Agent runs. A Run button for the person is left out for now.

Setup

test/ozone-report/ in the dev service: analysis.py (pandas, matplotlib), model.R (ggplot2), data/airquality.csv, outputs in figures/ and results/ from real runs, and report.md with two tables, three figures, inline and display math, Python and R code blocks, links to the scripts, a quote and a task list. Observation scripts and screenshots are in work/code-md-episode/ (ignored). Later checks use a copy of the project in a disposable service.

Observed before

Rho 65dcc8d3 with Documents 0.4.5 (before/, run-2026-10-29T06-50-54-612Z/):

1. .py and .R showed as plain monospace text, in place and in Studio. The editor had line numbers, wrapping and Agent marks, no language.
2. report.md showed as raw Markdown: tables as pipes, broken across lines in a narrow Studio column, images as ![…](https://github.com/YuLab-SMU/Rho/blob/main/docs/…), math as $…$. Rho's Markdown renderer exists only for Agent replies and has no images, math, task lists or relative links.
3. The person had no way to run a script. Asked to run both and check the report, the Agent (workbuddy-ai/deepseek-v4.1-flash, the saved Westlake settings untouched) ran python3 analysis.py and Rscript model.R in its own shell and compared each number. The Graph showed question → runs → the figures and table that changed during each run, but a run node said only "Returned": no exit code, duration or output. Run titles carried rtk, a wrapper from the person's own Claude Code hook.

Step 1: highlighting

Changed

• Documents 0.4.6: the editor takes extensionsFor(path), CodeMirror extensions the application adds for a document's path. It is asked once the buffer's path is known and applied in place, so the text shows at once and the buffer is untouched; a failure leaves the text plain. Exact-version views take the same. Annotations 0.4.7 widens its peer.
• Rho chooses the language by file name (.py, .R, .Rprofile, .md, .Rmd, .qmd) and loads it the first time a file of that kind opens: lang-python, CodeMirror's legacy R mode, lang-markdown with fenced Python and R, including R Markdown's {r label, …} chunks. One highlight style uses Rho's colour tokens (--code-*). Each language is its own chunk; the main bundle grew by about 3.6 KB.

The first version derived the path from the project snapshot. The sources acceptance caught that a quiet project's snapshot does not list a document opened since, so the text stayed plain; the editor now asks with the buffer's own path.

Observed

highlight/ and after-hl/: Python keywords, strings, numbers and called functions are coloured; R comments, strings and numbers (the legacy mode does not mark calls); Markdown headings bold, markers dimmed, links tinted, and fenced code coloured by its language. Typing and undo work as before.

Checks

Rho-plugins 97/97. Rho 218/218 serially; acceptance:sources now checks that an R script's comment and string take their colours; frontend and person (9/9 at 1440, 768 and 390) passed.

Step 2: Markdown read as formatted text

Changed

• A new package in Rho-plugins, @rho-workspace/plugin-markdown 0.1.0, is a UI extension over the Documents capability (component markdown.reader), so Documents stays format-agnostic. It reads the buffer every view of the document shares and renders it with markdown-it 14.3.2 (tables, task lists, autolinks) and KaTeX 0.16.47 for math; the document's own HTML stays text. Read and Source sit in the view's header; Source is the Documents editor on the same buffer. The renderer and math fonts are their own chunk (about 375 KB), loaded when a Markdown view first shows. Above 512 KiB a document opens in Source with a note.
• Documents 0.4.6 exports useDocumentBuffer, the hook its cards already used, so another view reads the same buffer.
• Rho selects first-party extensions in ui-extensions.json, which builds the reader into every application; RHO_UI_EXTENSIONS adds to it. A .md or .markdown file opens in the reader in place and in Studio.
• Views receive shell, the shell's services: a link to a project file opens it in place, or beside in Studio; an image is read as its current bytes (at most 64 per page, 8 MiB each); Rho's languages serve the editor and code blocks with one set of token classes coloured in theme.css. Images on the web are not loaded (the page policy allows only its own, data: and blob: images); they show their description.

Observed

markdown-1/ and markdown-2/ in a disposable service with copies of the project: report.md opens as a page: headings, the two tables with numbers right-aligned, the three figures from figures/, inline and display math, highlighted Python, the quote and four task boxes. analysis.py in the text opens the script in place. Source shows the editor with Save and the change highlight; Read returns. In Studio the tab carries Read and Source; at 390 px the page reflows without horizontal scrolling.

A closing parenthesis after inline math wrapped onto its own line; punctuation against a formula now stays with it (markdown-3/). A Studio tab placed for report.md before this change stays the editor, because a tab keeps the component it was placed with; opening the file in a workspace without that tab places the reader.

Checks

Rho-plugins 103/103 (new: paths, links, images and their budget, escaped HTML, tables, task lists, math, fenced code, the manifest). Rho 218/218 serially; acceptance:sources 8/8 now reads a Markdown note with its image, math, table and coloured code, follows an edit by another writer, switches to Source and opens a linked script; frontend, person (9/9) and plugin-dx passed.

Step 3: what an Agent's run did

Changed

• Agent 0.4.17 asks Claude's ACP adapter for its terminal_output extension, so a command's printed output and exit code arrive as data instead of a console text block the record dropped. A command step keeps its exit code (never invented when the Agent sends none) and the last lines it printed, at most 20 lines and 2 KiB with terminal codes removed and carriage-return rewrites resolved. Other tool output is still not kept.
• Rho records when each tool step started: its last description before any result, because the adapter reports a call while the model is still writing its input. A message at its size limit gives up a command's output before its outcome.
• The panel row names the command without a leading rtk (the person's own output wrapper; the details keep what ran), with the time it took and, when it failed, its exit code. Its details show the exit code, the time and what it printed, opened at the last line. The Graph's run card reads "Exit 0 · 4.3 s · 04:06" and its details show the output. A command that ran and failed is "failed" and stays on the Graph; before, the adapter's "failed" status made it "refused" and the Graph left it out.

Observed

Four real runs of the same request in a disposable service (run-2026-10-29T08-…, runs-after-*, panel-rows-*), workbuddy-ai/deepseek-v4.1-flash, mode auto:

1. Every command carried exit 0 and 6–30 printed lines; R's ANOVA table reads in the panel and the Graph with its columns intact (output is no longer wrapped).
2. Times of 2–8 s for ls and find: between the call's last description and its result Claude Code also decides permission (auto mode) and runs its hooks, which ACP does not report. The time is shown as how long the call "took", not as run time.
3. Asked to run a failing script, the Agent first appended ; echo "EXIT: $?", so the shell exited 0 and the output's last line reads EXIT: 1; asked for the bare command, the adapter reported exit 1 and the record kept the traceback's end, FileNotFoundError: … 'data/june.csv'. The panel said "refused" until the "failed" outcome was added.
4. One command's output carried rtk's own lines ([+1 hidden: rtk recall …]): the person's hook rewrites commands after the Agent names them, so the record shows the Agent's command and the rewritten command's output.

Checks

Rho-plugins 104/104 (new: exit codes, the bounded tail without terminal codes, an unreported exit). Rho 223/223 in a direct serial run (new: start time, the output given up first, command names, durations and failed versus refused, a failed run on the Graph). In three runs of npm run verify the real stdio MCP v2 test, which this step does not touch, hit its 30 s limit; alone it passes in 21 s. The machine's load was about 6.5 from other applications, and the ledger keeps the failed run. sources 8/8, agent-ui 7/7, agent-page 7/7, agent-canvas 10/10, frontend, person 9/9 and plugin-dx passed. acceptance:graph fails at "both outputs changed while their command ran": after Fit the Graph sits at 60 % zoom, below the 62 % at which edge labels are drawn. It fails the same way with the Graph view at HEAD; its last recorded pass was at 0d98b571, before the timeline lanes and Overview commits.

Limits and next

• The reader opens .md and .markdown; R Markdown and Quarto files stay in the editor (their front matter and chunk options are not rendered). Front matter in a .md shows as text.
• A Studio tab placed before the reader keeps the editor.
• The Documents card's outline still splits Markdown headings inside Documents.
• Command times include Claude Code's permission decision and hooks. Codex and the person's own Claude Code outside Rho are not covered by the terminal extension yet.
• No Run button for a person, as chosen.