Skip to content

claude code: /plugin install agentmon-live --marketplace codenotary/agentmon-live

Blog 7 min read

AgentMon Live 0.48: turn steps, feedback from the pane and sessions from everywhere

By Dennis

AgentMon Live 0.48: turn steps, feedback from the pane and sessions from everywhere

AgentMon Live 0.48 is out. Since 0.36 it gained a step-by-step view of every turn, a feedback button in the pane, sessions from Codex, Kiro Crew and containers in one web view, and a done gate that is harder to talk past. Most of it started as an issue or a message from someone who used it.

If you haven’t met it yet: AgentMon Live is an open source Claude Code plugin. It puts a live pane beside your chat and the same session in a browser tab, and shows where the time and the context go: every tool call, subagent, file, test run and token. A done gate checks for unfinished work before Claude stops and sends it back once if a test failed or a todo is still open. Across sessions it learns what keeps going wrong, and a self benchmark tells you whether a change to your setup helped. The first post walks through the pane, and the last one covers the web view.

Built from your feedback#

A big thank you to everyone who installed AgentMon Live, filed an issue, sent us a message or told us what got in the way. This release is mostly your list:

  • The status line is gone. It repeated what the pane already said and looked like a warning when nothing was wrong.
  • Focus puts your work first. Findings about this session sort to the top. Findings that are the same in every session (MCP schema size, quota pace, memory files) move under an Environment label.
  • The memory check finds more. It checks paths relative to the project, a memory file’s frontmatter and each file on its own, and skips /tmp.
  • The same-error item names the error. A failed shell call is keyed as exit N: <first output line>, not one bucket called “Exit code”.
  • The agent flow no longer looks empty when the agents finished before the window you look at: it says how many are outside it.

Keep it coming: the issue tracker is open, and from 0.47 on you can send feedback without leaving the session (see below).

Turn steps: what happened in this turn, in order#

The pane told you what was running and what failed. It did not answer the plain question “what has Claude done since I pressed Enter?” Turn steps does. A new pane section and web panel list the current turn newest first: who acted, what it was, and how long it took.

AgentMon Live web view replaying a Claude Code turn on a demo shop-api project: subagents fan out in the agent flow, turn steps fill in, the steps list is filtered to Memory, Hooks and Agents, and the session list shows a container session and a Codex chat beside it

Each row has a category:

  • Model, Shell, Files, Agents, Network, MCP: the usual work, with short labels. A shell call shows its description (“Run the coupon tests”), a file its last two path parts, a web call its host. The raw command stays one step away: in the tooltip, behind d in the pane, and in the call drawer.
  • Memory (new in 0.48): a read or an edit of CLAUDE.md, MEMORY.md or a memory file, and every instruction file Claude Code loads during the turn, with why it loaded (session start, folder entered, path rule, after compaction).
  • Hooks (new in 0.48): context a hook added to your prompt, a call a hook blocked, and a hook that sent the turn back at its end. The send-back shows in the turn that goes on.
  • Skills: a slash command you typed.

Filter by category with /agentmon steps memory in the pane (/agentmon steps shows all again), or with the menu above the list in the web view, which counts the rows per category. With details on, each pane row starts with its clock time. Quiet repeats fold into one row, so twenty reads of the same file don’t push everything else out.

Feedback from the pane#

Press b in the pane. Claude drafts the feedback from what it saw in the session, and the pane shows the draft. Press y to send it or n to drop it. Nothing leaves your machine before y.

  • Anonymous by default. Press w to put your name on it: that adds an install ID the plugin made itself and, if you set feedbackIdentity, your email or handle.
  • What you read is what is sent. Secrets are masked and control and invisible characters are removed before the draft is shown.
  • GitHub instead, if you prefer. A problem or an idea can go out as a GitHub issue from a link that opens it filled in.
  • Sent feedback cannot be deleted, and the draft says so. feedback: off removes the feature.

Sessions from everywhere, in one web view#

The web view started as a view of your Claude Code sessions on one machine. It now lists more:

  • Codex. Codex chats show as sessions, read from Codex’s own session files, and the repository installs as a Codex plugin too.
  • Kiro Crew. Every open Kiro Crew chat shows beside your Claude Code sessions, on any harness. AgentMon Live also installs as a Kiro Crew app and shows in a tab beside the chat.
  • VMs and containers. A session that cannot reach the host writes its snapshots to a folder. Mount the host’s ~/.claude/agentmon-live/spool at /var/lib/agentmon-live/spool and the host’s web view lists it, marked sandbox. AGENTMON_SPOOL changes the path. The server reads at most 64 files of 4 MiB, follows no links and drops a session that went quiet for 90 seconds.
  • Headless runs. claude -p loads the plugin and shows in the web view with no terminal open.

A done gate that is harder to talk past#

The done gate is the part that sends Claude back when it says “done” too early. In this release it got stricter:

  • A claim has to match a check of the same kind: “tests pass” needs a test run, “typecheck clean” needs a typecheck.
  • A failed test, or an objection from the reviewing model, no longer passes.
  • The reviewer also flags features nobody asked for.
  • Edits made through the shell count as edits.
  • Fewer false send-backs: in a chain of checks only the last one reads the shared output, and edits outside the project folder don’t count.

Smaller things worth knowing#

  • Web view stays readable when busy. The Events panel filters by kind with counts, subagent runs fold in the session log, long lists fold behind “Show all”, and the agent flow draws only the lanes in view.
  • Old data never looks current. When the connection drops, the page keeps the last state with a “no signal” label and the time of the last update.
  • Only your account gets in. The browser pass is bound to the server run. Servers of different versions no longer replace each other.
  • Focus warns about a plain-text secret in the project’s Claude settings.
  • PowerShell calls count as shell calls.
  • Everything ships in English, German, French, Spanish, Chinese and Japanese.

Changelog since 0.36#

0.48.0

  • Turn steps: Memory category, hook rows (context added, send-back), Skills rows for typed commands.
  • /agentmon steps <category> filter in the pane, category menu with counts in the web view, clock times with details on.
  • Fixes: a subagent’s instruction load is no main row; hook moments no longer hide a running request.

0.47.0

  • Feedback button in the pane (b, y, n, w), anonymous by default, GitHub issue link as an alternative.
  • Friendly call labels; new Turn steps section and panel.
  • Codex sessions in the web view; the repository as a Codex plugin.
  • Same-error item keyed by exit code and first output line; status line removed; PowerShell counts as shell.
  • Focus: session findings first, machine-wide ones under Environment; memory check covers project paths, frontmatter and a per-file cap.

0.39.0

  • Sessions in a VM or a container through a shared snapshot folder, marked sandbox.

0.38.1

  • Kiro Crew chats in the web view on any harness; AgentMon Live as a Kiro Crew app.
  • Browser pass bound to the server run; servers of different versions stop replacing each other.
  • Focus warns about a plain-text secret in project settings; checks read more exactly.
  • Headless claude -p runs show in the web view.

0.37.0

  • Events filter by kind; “no signal” label when data is stale; folding for busy sessions.
  • Done gate: claims must match checks, failed tests and reviewer objections block, unrequested features flagged, shell edits count, fewer false send-backs.

The full notes are on the releases page.

Try it#

AgentMon Live is open source under Apache 2.0. Install it from a Claude Code session:

/plugin install agentmon-live --marketplace codenotary/agentmon-live

Already installed? Update the marketplace and restart the session. Press o in the pane, or run /agentmon open, to open the web view. It needs node on your PATH, listens on 127.0.0.1 only, and masks secrets before anything reaches it.

The source and the full guide are on GitHub. Thanks again to everyone who told us what to fix: keep the issues and the feedback coming, they shape what we build next.

Keep reading