feat(tui): add side questions for running flows

docs/.vitepress/config.mts

@@ -184,6 +184,7 @@ export default defineConfig({
           collapsed: false,
           items: [
             { text: 'Talking to a running turn', link: '/guide/steering' },
+            { text: 'Side questions (/btw)', link: '/guide/btw' },
             { text: 'Many conversations at once', link: '/guide/conversations' },
             { text: 'Showing the working (/details)', link: '/guide/details' },
             { text: 'The shape of a run (/status)', link: '/guide/status' },

docs/guide/btw.md

@@ -0,0 +1,32 @@
+# Side questions — `/btw`
+
+Ask about a long-running flow while it keeps working, without sending another message to the
+flow's agent or waiting for its current turn.
+
+## Try it
+
+```
+/btw what is the reviewer waiting for?
+```
+
+## What It Reads
+
+The command takes a snapshot of the active flow: its name and task, each agent's current state
+and turn count, observed handovers, spending, and the latest agent events. A separate short-lived
+session answers from that snapshot. It is given read-only permissions and no flow skills, and it
+is not registered with the run's monitor or cycle.
+
+The answer appears in the transcript when it is ready. The original flow keeps its sessions,
+queued messages and context untouched, so `/btw` is safe to use while an agent is thinking or
+while several agents are working at once.
+
+`/btw` needs an active flow and a question. It reports an error rather than starting a new flow
+or falling back to a write-enabled agent when no read-only backend is available. Observations are
+bounded and treated as untrusted data; the side agent is told not to follow instructions found in
+the flow's output.
+
+## See also
+
+- [The shape of a run](/guide/status)
+- [Talking to a running turn](/guide/steering)
+- [Permissions](/guide/permissions)

docs/guide/index.md

@@ -21,6 +21,7 @@ These are for looking things up. If you have not used humanize before, the
 | | |
 | --- | --- |
 | [Talking to a running turn](/guide/steering) | A line typed mid-turn goes *into* it, not after it |
+| [Side questions](/guide/btw) | `/btw`: ask about progress without steering the flow |
 | [Many conversations at once](/guide/conversations) | One transcript, **tab** between the agents that are working |
 | [Showing the working](/guide/details) | `/details`: tool calls and thinking, or only what the agent says |
 | [The shape of a run](/guide/status) | `/status`: who is working, who handed to whom, what it cost |

docs/guide/questions.md

@@ -203,6 +203,7 @@ agent.hooks.on(Moment.NOTIFICATION, lambda occasion: ring_a_bell(occasion.said))
 
 ## See also
 
+- [Side questions](/guide/btw)
 - [Being away](/guide/afk)
 - [Answers in a shape](/guide/shapes)
 - [The person as an agent](/guide/human-agent)

docs/guide/status.md

@@ -23,6 +23,10 @@ so there is nothing to ask. What `/status` draws is kept from **the turns going
 same `begins`/`ends` events any [watcher](/reference/agents#watching-a-turn-as-it-happens)
 sees.
 
+`/btw` uses this same live observation, together with the task, agent turn counts and handovers,
+to answer a quick question. Its read-only side session receives a frozen snapshot, so asking it
+does not add a message to, pause, or otherwise steer the flow.
+
 That is also why the person, driven as [an agent](/guide/human-agent), is not in the graph.
 Their turns are not bracketed by those events. Counting them would put a human in the handover
 graph and spin a clock at them while they thought.
@@ -87,6 +91,7 @@ sessions, one slice per thing the agent did.
 
 ## See also
 
+- [Side questions](/guide/btw)
 - [Cost and rate](/guide/tally)
 - [Many conversations at once](/guide/conversations)
 - [Tracing](/guide/tracing)

docs/reference/tui.md

@@ -149,6 +149,7 @@ list appears under the editor with a line about each.
 | `/providers` | | [The accounts](#the-accounts-themselves) an agent may be run as: what there is, and what can happen to one — made, taken away, and, on enter, corrected, signed in again, pointed at what it falls back to or told how it is tried again. |
 | `/settings` | | [What humanize remembers](#what-humanize-remembers): two pages, one for what is true of this machine and one for what is remembered about this directory. |
 | `/status` | | [How the run is going](#how-the-run-is-going), and the shape of it: a box per agent, marked as it works, with the handovers between them drawn as the arrows joining them. Enter reads an agent. **esc** opens it. |
+| `/btw` | `<question>` | Asks a side question about the running flow from a read-only snapshot of its progress. It runs in a separate session and never steers the flow. |
 | `/details` | `[on\|off]` | Shows or hides everything a turn did on the way to its answer: tool calls, thinking, and whatever a backend printed on its way past. One question — how much of the working to show — so one switch. **Off** to begin with. |
 | `/afk` | `[on\|off]` | Whether an agent may stop and ask you something. See [below](#questions-and-being-away). |
 | `/fallback` | | Where a turn goes when what was taking it cannot: an agent that has nowhere left to run, and an account that has gone down. See [below](#where-a-turn-goes-when-it-cannot-be-taken). |