Notes, footers and replies

The transcript says what happened to a session and how each turn ended, and you can reply to any row of it.

The rest of the transcript is in The transcript.

Quiet rows

Quiet rows say what happened to the session, most of them a thin rule across the transcript with their words in the middle:

• a change into or out of waiting, such as Running → Waiting for you;
• Stopped, with why when it wasn’t you: “Stopped: the agent crashed”, “Stopped: the daemon stopped mid-turn” or “Stopped: signed out”;
• a change of model or effort, and of permission mode;
• “Context cleared”, a rewind, “Compacting the conversation…” then “Conversation compacted”, or “Conversation compacted automatically”, with the tokens before and after where the agent says them, and Summary to open where it sent one (or “Compaction didn’t finish”);
• a goal set, changed or cleared, and a turn a goal or a loop starts by itself (see Conversation commands);
• errors, each led by Error, Moontors’ words for it with the agent’s own reason in full beneath, some with Sign in (see Signing in);
• notes from the agent or from Moontors, each a light row with an info icon, such as “The agent stopped: its reply reached the length limit.”, “The agent stopped: it reached its step limit for one turn.”, “The agent declined to continue.”, “The agent’s saved conversation was gone, so it started a new one.”, which Grok Build’s reopen also gets where it restored none of the conversation, that the agent couldn’t save the session’s conversation, or its own settings, as the disk is full or your disk quota is used up, which Claude Code says only when its run ends by itself, never when you stop the session, sign out or Moontors starts the agent again, that Claude Code’s own settings, yours or your organization’s, refuse every action that would ask for your approval, so the session never waits for you, that the instruction files the agent reads, as CLAUDE.md, are longer than recommended for the model, that the agent’s moderation blocked your message, its reply, a tool call or a tool’s result, as Codex’s can, or “The agent warns:” followed by the agent’s own words, as Claude Code’s and Codex’s warnings, suggestions and notes are, such as what a hook says, or each of your permission rules Grok Build couldn’t load, as a session opens, once a session;
• a turn the agent’s safety check stopped, with the ways it offers to continue (see When the agent’s safety check stops a turn);
• while Codex or Grok Build tries again by itself, one note each try updates, saying why: “The agent lost its connection — reconnecting…”, “The agent couldn’t connect — retrying…”, or “The agent’s service is overloaded — retrying…”, as when the service holds back capacity, each followed by the agent’s own reason where it gives one, or “The agent is retrying:” then the agent’s own words, with its tries counted where the agent counts them, as Grok Build’s at its rate limit; then a note once it goes on, as “The agent lost its connection and reconnected.”

When the agent’s safety check stops a turn

Codex’s safety check can stop a turn it judges could lead to harm in biology or cybersecurity, as Codex’s own terminal does.

• While it looks for safer ways to continue, the turn’s last row says “Finding safer ways to continue…”.
• Once the turn ends, after the error Codex gives, a row says “The agent’s safety check stopped this turn.”, then what the turn touched on, in Moontors’ own words, such as “It touches on access to accounts or systems beyond what’s allowed.”
• Under it, Ways to continue offers the safer requests Codex suggests, up to three, each by its own name, with the words it sends shown when you point at it. Choosing one starts a new session from before the stopped turn, as /branch does, and sends those words there, as written; this session stays as it is.
• The ways are offered while the stopped turn is still the session’s latest: a later message, /clear, a rewind, a compaction or a fork takes them away, and the row keeps its words alone.
• The phone shows the row, and its ways, as the window does.

What Moontors leaves out is under Known limits.

A turn’s footer

A turn that finishes ends with a footer; a stopped or failed one ends with its Stopped or error row instead. The footer shows:

• how many files the agent and its subagents edited in the turn, with the lines added and removed, a file both edited counted once: click it for each file, shown from the session’s folder as a tool’s row shows it, its lines added and removed at the right;
• how long the turn took and when it ended, shown when you hover over the turn’s last reply or its footer, or move the focus there. Times follow your system’s 12- or 24-hour clock. A turn that ended on an earlier day says the day too, counted in your own calendar days and always in days: yesterday at 1:21 AM, or 46 days ago at 1:21 AM. It moves on at midnight with the transcript open, or as your computer wakes;
• Copy reply, the copy icon beside that time on a turn with a reply: it copies the turn’s reply as the agent wrote it, without its tools, thinking or questions. Its icon turns to a check for a moment. Words you select in the transcript copy with a right-click’s Copy.

Replying to a row

To answer one thing in the transcript, a reply of the agent’s, your own message, a tool’s row or a note, hover over its row, or move the keyboard’s focus to it, and choose Reply: at the row’s top right, or beside your own message with its other actions. A click in the row leaves it hidden once the pointer goes. Or select words in a row and choose Reply under them. Either way, the focus goes to the message box, after any words you’ve typed. A permission, a question or a plan still waiting for you is answered rather than replied to, so it offers Reply only once it’s answered. Files sent alone have no words to reply to, and a turn’s footer isn’t replied to either.

• The row’s words, a command named in the agent’s own words with what it was given, the command among it, or the words you selected, show at the top of the message box as a quote, above any files and your words, two lines of it at most, with an X to take it off. Escape takes it off too. There’s one quote at a time: another Reply takes its place.
• What you send then replies to it. The agent reads the quote first, as a Markdown quote, then your words, so it knows what you mean even once the conversation has been compacted. A reply needs words or files of its own, and a command can’t be one.
• A quote keeps up to 2,000 characters of a row, cut there with an ellipsis.
• Your message shows its quote above its words, dimmed, with no Markdown marks. Click the quote, or press Enter on it, to scroll to the row it quotes, or to the subagent’s row that holds it, which is marked for a moment, inside its own bubble where it has one, and takes the focus, so a screen reader reads it. It keeps the focus once the mark is gone: the arrow keys scroll on from there, and Tab goes on from it. Followed by a click, it shows its focus ring at your next key. Once that row is gone, as after a rewind, the quote leads nowhere.
• A reply sent while the agent works goes as any message does then (see Sending while the agent works). Rewinding to before it brings back its quote with its words.
• The exported transcript shows a reply’s quote above its words.