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.

Agent panel live activity

The user asked on 2026-10-27 to make the Agent UI show the Agent's work as it happens: streaming replies, every action visible while it runs, and a distinct treatment for the actions that need a person. This record keeps the setup, the observations before and after, the owner changes and the checks. It is evidence that these paths work, not evidence that another person can follow them.

Setup

• One controlled ACP peer wrote text chunks every 80 ms and reported a read, a command, a write, a plan, usage and a native permission request. It ran through the real local service and Chrome 155.0.8059.84 on 2026-10-27, with no model credentials.
• One real Claude ACP turn ran through Westlake HPC (westlakehpc/deepseek-v4.1-flash, Claude Code 2.1.312) in a disposable project. The prompt asked for ls, Read input.txt and a short sum.
• The recorded conversations from the person's own 2026-10-27 runs were inspected, including the 24-step ggtree turn that the Overview and Agent panel showed as three collapsed "Worked in the project · 1 step" lines.

What the panel did before

• Reply text arrived in the recorded 250 ms batches, but the panel had no caret, no elapsed time and no reading position: every update pulled the pane to the end. Markdown was printed literally (**Files in …**, fenced code and lists).
• Consecutive tool steps folded into one line. While a turn ran, the newest step was rendered as Used <tool>…; a real Bash step therefore said "Used Bash…" and never showed the command. File reads and writes did not name their file.
• A permission request was a plain bordered card; the person's choice left no trace in the turn. The Agent's plan sat in a closed section below the composer.
• The Agent plugin preferred a tool's programmatic name over its human title and discarded the ACP category and input, so the durable record could not support a better presentation later.

Changes

Agent 0.4.11 (Rho-plugins/plugins/agent) keeps three bounded additions on each tool step: the Agent's human title, its ACP tool category (read, edit, execute, search, …) and one identifying native input (command, file path, pattern or query). They are cut to one 512-character line; raw tool output, reasoning and history replay remain excluded. The README and fixture tests state and check the new fields.

Rho now derives presentation from the retained record instead of rendering protocol events as bubbles:

• src/conversation-presentation.ts turns steps into text, one line per action, or a folded run. Each action names its verb and target (Read data.csv, Ran Rscript analysis.R, Wrote results/summary.csv, Called data.page), keeps the Agent's title and reported location behind the row's disclosure, and distinguishes running, completed, refused and unknown outcomes. Short runs stay visible; more than three consecutive actions fold into what they accomplished.
• src/agent-markdown.ts and web/markdown.tsx render the Agent's Markdown as paragraphs, headings, lists, quotes, inline code, fenced code and http(s) links. No raw HTML is produced; unsafe link targets remain labels. An unclosed fence renders as code while it streams.
• The panel adds a caret on the live text and an elapsed time in the ending line. It follows new activity only while the person is at the end; scrolling up keeps the position and offers "New activity · Jump to latest". The composer's status line says new activity is below.
• A permission request is an amber "Needs your choice" card with the Agent's title, the reported command/path and only its offered choices. After the choice the turn keeps a line naming what the person chose.
• The plan's current entry (or Plan · complete) is visible above the composer; the full plan and usage remain one disclosure away. A failed turn offers "Ask again", which only returns the original words to the composer; it sends nothing.
• The conversation list names the Agent backend. Tool status is the Agent's own report and is labelled "Tool status"; the ACP path is labelled "Reported path", not Rho's independent file observation. Work progress keeps that separate observation and its hashes.

The public conversation.read schema now describes the bounded fields; all existing records remain readable because the fields are optional. No new operation, runtime, read path or automatic resend was added.

Checks

• npm run build passed; npm run check:contracts passed (97 published operations; 96 typed).
• Rho node --test test/*.test.mjs: 202/202 passed, including the new presentation and Markdown checks. The first full run cancelled one test at its existing 30 s limit; two following full runs passed 202/202. Rho-plugins: 89/89 passed.
• npm run acceptance:agent-ui passed seven checks in Chrome 155.0.8059.84: text grew in more than one visible piece before the turn ended; the read, command and write rows named their file/command; the plan followed the in-progress entry; the permission was shown, answered and retained; Markdown rendered; reading earlier lines was not interrupted and Jump to latest returned to the stream; no page errors. Evidence, peer log and screenshots are in work/agent-ui-live-20261027/final-ui/.
• npm run acceptance:frontend passed its 11 checks again, including the Agent panel draft and the link handoff. Evidence: work/agent-ui-live-20261027/ final-frontend/.
• The real Claude turn completed. The retained record kept Bash with title and input ls -la /private/tmp/…/project, Read input.txt, and a Markdown answer with bold, a list, a fenced code block and inline code. The rendered panel at 1440 and 390 CSS pixels is in real-rendered-1440.png and real-rendered-390.png; the record is real-report.json. No page errors.
• The controlled probe showed text appearing about 250–500 ms after a chunk was emitted, matching the service's 250 ms durable-write schedule. The follow check held a scrolled-up position while the newest activity continued and returned to the end only after the explicit choice.

Limits and open questions

• Streaming is still a sequence of durable observations (up to about four per second), not a per-token channel. Reducing the write interval would multiply fsyncs of the conversation record; no separate live transport was added. Long replies can therefore look stepwise.
• Markdown support is a selected subset: no tables, images, nested lists or raw HTML. Links open external pages; the panel does not fetch them.
• Action titles, inputs and locations come from the Agent's ACP report. Rho's independent file observation, integrity and managed execution facts remain separate surfaces. The panel does not claim that an Agent action changed the bytes it names. A bounded command or path is now kept in the local conversation record as the Agent reported it, not only when a permission asked; it is not a diagnostics log and may contain whatever the Agent typed.
• The panel is still the 340 px inspector; the conversation header scrolls with its content, and there is no expanded conversation view. Long commands clamp to two lines and the full input is in the row's disclosure.
• These checks show the paths work. They do not establish that a person can scan a long trace, that the chosen action vocabulary is clear, or that four updates per second is enough for a long answer. Those need use with real work.