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.

Connecting your own Agent with one prompt

Date: 2026-11-06. Source changes after Rho 0.5.5; package publication and installation are separate. This episode changes Rho only. Core and Plugins are unchanged.

Question and observed problem

The person could not tell whether Agents or Connections was where to connect their own Agent. Agents configured conversations started inside Rho; Connections offered brand-specific Claude Code and Codex buttons beside an ACP relay option. Those choices made a generic MCP capability look like a two-client restriction.

The question was whether the person's own Agent could configure itself from one copied prompt, without another brand-specific form in Rho.

What changed

The Agent page has Conversations and Connect your agent. The default Agent and model choices now sit under New conversation settings in Conversations and explain that they apply inside Rho. Connect your agent leads with a copyable request and a selectable full prompt. Existing grants appear under their actual session names, with waiting, verified last call, refusal or ended access states.

The local-browser/operator-only session.instructions reads the running launcher's program and argument array and exact project identity. It includes no credential contents, starts nothing and grants no access. The Agent runs the existing grant CLI, chooses its own stable session name and registers the returned stdio MCP entry using its own client. The added --project-id selects a registered project without reopening a guessed path, including a folder selected outside the service root. The old named CLI shortcuts remain compatible.

Repeated setup with the same scope keeps its valid credential. Changed methods, expired access or a stale address require a fresh one. Granting now emits a browser invalidation: the first browser run exposed a grant that existed but was invisible until another event or reload. Refused calls remain visible and a successful retry uses the same connection.

Real runs and corrections

node scripts/agent-connect-episode.mjs creates a disposable Unicode/spaced project, an isolated Claude configuration directory and a real service without Core. Chrome copies the prompt through the real clipboard. The Agent receives that exact prompt, plus the boundary that its configuration belongs in the disposable environment. Existing model credentials travel only through the process environment.

Two initial harness failures were corrected: reload returned to the project's Overview rather than its Agent page, and a variadic CLI option consumed the prompt. Neither was evidence of an Agent configuring itself.

The next run configured a shared project MCP entry. Claude's own approval and the harness's disabled settings sources kept the fresh client from loading it. The Agent had separately made a direct stdio probe, and the original assertion wrongly counted that earlier receipt as the new client's success. Reading both replies invalidated that run (work/agent-connect/1793992731271/report.json, review field). A direct server probe proves the server answers, not that the client's tools work.

The prompt now requests private local configuration outside version-controlled files and checks scopeHint.projectId, which rho_index actually returns. The harness uses the isolated client's normal settings sources. Verification requires both a fresh native mcp__…__rho_index tool-use event and a new successful server receipt after the second session starts; an earlier probe cannot satisfy it. A further run loaded the tools but the harness’s wildcard permission did not authorize them in Claude’s dontAsk mode. The harness now names the exact rho_index tool it is testing and requires zero permission denials; it changes no user configuration or native permission policy.

The final run (work/agent-connect/1793993054868) passed: Claude Code 2.1.319 configured its own private local entry in 27.1 seconds, then a fresh client called mcp__rho-study__rho_index in 8.5 seconds. The returned project ID matched, both sessions had zero permission denials, and the first session honestly reported that its own tools needed a new session. No project configuration file was written. The harness asks only for the ID that rho_index supplies; an earlier request for a folder path led to additional, ungranted diagnostic calls and was narrowed.

Checks and screen review

• npm run build: service and WebUI compiled. The existing large-bundle and mixed static/dynamic import warnings remain.
• npm run check:contracts and npm run test:contracts: generated schemas/types match the implementation; 11 contract checks pass.
• The Agent connection, CLI refusals, browser, MCP and progressive-disclosure tests: 30 pass with the existing Core debug binary explicitly selected. The first run lacked that binary for its managed MCP case and omitted Origin in a new pairing test; both failures are retained in the local logs.
• npm run acceptance:agent-page: settings, conversations, export, trash, generic grants, refused calls and recovery pass in Chrome. Its conversation peers are controlled fixtures, not model trials.
• The connection episode checks a real clipboard, a failed instructions request and retry, denied clipboard access with selectable text, Unicode/spaced paths, and the setup screens at 1440, 1024, 768 and 390 px. The screenshots show the prompt before connection details, no brand picker and no horizontal overflow; the narrow layout wraps the copy action and keeps the full prompt accessible.
• npm run docs:assemble, npm run docs:check and npm run docs:build: 58 pages and 324 rewritten links; changed pages rendered and read, with their new anchors resolving. The initial check correctly required regenerating navigation after the new episode was added.

The private run directories and raw model outputs stay in ignored work/. No credentials or private service addresses are source evidence. Setup with another client, remote/HTTP-only MCP clients, Windows and independent usability remain unverified by this episode. This work adds no new Agent execution adapter and does not move an external conversation's input into Rho.