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.

Agents and conversations

Updated 2026-10-28.

Question

The user asked to improve the Agent settings and the conversation management, and gave four screens from another application as references: one row per Agent with its mark, version, connection, model and a switch, opening to model tiers, a model list grouped by provider and settings; and a list of sessions with each one's model, effort, message count, tokens and cost, continue and delete, opening to its time, folder, session identity, routing and an export. They also gave the marks of Claude Code, Codex and DeepSeek.

The references set the kind of page, not values to copy. The questions:

1. Can a person see which Agents Rho can start, what each offers and what new conversations will use, and change it, without reading a row of chips?
2. Can a person find a conversation, see what it used and cost, continue it, keep a copy, and clear it away?

Observed before

Looked at with the real Claude Code's settings (101 models) and records of real conversations (work/agent-settings-20261028/before/):

• The Agent page opened on a card: "New conversations use" with a Claude Code / Codex switch and five chips in one line (model, effort, mode, fast mode, ultracode). No Agent had a mark, a version or a state; Codex showed nothing until chosen.
• The model chip opened one list of 101 models. Claude Code's own names came first, read as "Opus 5.5 (sonnet)", with no word on what they are.
• Each conversation was a large card with every recent message's work. Finding one, or seeing what it used, meant reading the cards. Nothing showed context use or cost, though each record kept them, and nothing searched.
• A project keeps at most 64 conversations started in Rho; past that, starting one is refused, and nothing removed one.
• An answer's Markdown table showed as raw | Month | … | lines.

Changed

Rho only:

• The record lists itself (src/conversations.ts, src/protocol.ts). Each conversation's summary carries the Agent's own title for the session, the Agent and its adapter, the last model, the person's choices, the context use and cost the session reported, its session identity and whether it is in the trash. Summaries written before are rebuilt from their records once.
• A trash (conversation.list, conversation.trash, conversation.restore, conversation.delete, person only). Moving a conversation to the trash keeps it whole and frees its place among the 64; it cannot be continued until restored, and an idle connection to its Agent ends. It is refused while the Agent answers. The trash holds as many again. Deleting works only from the trash and removes Rho's record; the Agent's own session files stay. The request that started a deleted conversation is remembered (the last 256), so repeating it is refused rather than starting it again.
• The Agent page has two tabs (web/agent-manage.tsx, web/agent-manage.css, web/brand-marks.tsx):
  • Conversations: a search, an Agent filter when both Agents have conversations, and the Trash; rows by day, each with its Agent's mark, title, when, model, effort, message count and state, context use with a meter and cost; Continue (Follow while it answers), export as Markdown and Move to the trash. A row opens to its facts (the first message when the title differs, time, Agent and adapter, model and its value, other settings, context, cost, session identity and record path, each copyable) beside what its latest messages did. Moving to the trash offers Undo; the Trash restores, or deletes after a confirmation. Runs and sessions started outside Rho stay below.
  • Agents: one row per Agent with its mark, adapter version, Default or Make default, whether and when it was asked and how many models from how many providers, and its model. Details holds its settings; Claude Code's own names (default, opus, sonnet…) with what each runs and Choose; the providers with their counts, each opening its models; and who offered these, when, with Check again. Asking is still only the person's click.
• Marks: Claude Code, Codex and DeepSeek in their own marks; any other provider as its initial on one of the theme's hues. The message box's chip and its model list use them too.
• Markdown tables (src/agent-markdown.ts, web/markdown.tsx): a table after prose, alignment, pipes in code spans and short rows; it scrolls on its own.
• A conversation's names are read again while its Agent has none here, so a Codex row named its model once Codex's first session told Rho what it offers.

Runs

Evidence: work/agent-settings-20261028/ and work/agent-page-acceptance/ (ignored by Git).

Run	What was looked at	Observed	Change
look.mjs	The page before, real settings and records	As above	The two tabs
look2.mjs 1	Both tabs, real data	The marks sat on green squares: .brand-mark was already Rho's logo; search icon at the top; Settings repeated the model	Own class; centred icon; model only on its line
trash.mjs	A copy of the local service's project: three real records written before	Summaries rebuilt; DeepSeek V4.1 Flash with Max effort, a failed message and a Chinese title shown; trash, Undo, restore and delete worked; no page errors	—
look2.mjs 390	Phone width	The cost ran under Continue	The row wraps; the meter hides
acceptance	A Codex conversation	Its row showed gpt-5.6: the page read the names before Codex's first session gave them	Names read again while missing
composer.mjs	The message box's chip	Marks on the Agents and providers; a search for DeepSeek lists it under four providers	—
ask.mjs	The real Claude Code and Codex asked from the tab	Claude Code 3.2 s, claude-agent-acp 0.97.0, 101 models from 10 providers, its five own names with what each runs; Codex 0.8 s, codex-acp 2.3.0, 103 models from 11 providers, effort, mode and collaboration mode. "Check again" fell to its own line	Kept in its sentence

Checks

Run serially on 2026-10-28 (work/agent-settings-20261028/final/):

• build:service, build:web, the web type check and check:contracts (104 operations) passed.
• Rho 214/214 and Plugins 92/92 (ledger), including "a conversation in the trash is kept whole, frees its place, cannot be continued, and is deleted only from there" and the Markdown table test.
• acceptance:agent-page 5/5 with controlled Claude Code and Codex peers: the Agents tab asks only when asked and saves the model, effort and default Agent; rows are found by search and by Agent; a row opens to its facts with its table rendered and exports as Markdown; the trash keeps a conversation until restored or deleted, with Undo, and never takes one the Agent is answering; no horizontal overflow at 1024 and 390 and no page errors.
• graph 8/8, agent-canvas 10/10, frontend 13/13, agent-ui 7/7, continuity 7/7, sources 7/7, objects 5/5, actions 6/6, and shell, pages, person and plugin UI passed.

Limits

• Context use and cost are what the Agent reported for its session; a conversation whose Agent reports neither shows neither. They are not Rho's measurement.
• Provider groups are the Agent's own names: Codex offers both "fusecode-lab" and "fusecode Lab", and they are shown as two.
• Search reads titles, models and Agents, not the messages.
• Claude Code's own names follow Claude Code's settings; Rho shows what they run and does not change them.
• Deleting does not remove the Agent's own copy of the session.
• No person other than the builder has used the page.