Merge PR #7881: docs(agent-hub): document subagent observability (@roboomp)
This commit is contained in:
can1357
@@ -143,6 +143,8 @@ Split a job across workers and get typed results back. task fans out into isolat
|
||||
|
||||
_[Watch the capture ↗](https://omp.sh/clips/irc.mp4)_
|
||||
|
||||
Watch the fan-out while it runs: `Alt+A` opens [Agent Hub](docs/agent-hub.md), where the roster shows current activity and usage for every subagent. Open one to read its live transcript, type a steering message, revive a parked worker, or kill a stuck one without aborting the parent session.
|
||||
|
||||
### 06 · A second model, watching every turn.
|
||||
|
||||
Pair a reviewer model to the 'advisor' role and it reads every turn the main agent takes, injecting notes inline — a quiet aside, a concern, or a hard blocker. It runs on its own context and its own model, so it catches what the doer rushed past. The main agent sees the note and course-corrects, or tells you why it won't.
|
||||
|
||||
@@ -326,8 +326,8 @@ Paths derive from the owning session file (not the shared artifacts root), so ea
|
||||
Why a file:
|
||||
|
||||
- **Usage attribution.** `omp stats` scans each session folder recursively, so advisor assistant turns (with their usage/cost) are attributed to the same project/session like any other subagent. Advisor "session update" prompts are persisted as `synthetic`, agent-attributed user messages so they never inflate user-message metrics.
|
||||
- **Observability.** The Agent Hub discovers legacy and named `__advisor*.jsonl` files on open and shows each as a read-only `advisor`-kind transcript under its owning session.
|
||||
- **Observability.** [Agent Hub](./agent-hub.md) discovers legacy and named `__advisor*.jsonl` files on open and shows each as a read-only `advisor`-kind transcript under its owning session.
|
||||
|
||||
The file follows session switches: on `/new`, resume/switch, and branch the recorder reopens at the new session's path on the next advisor turn; before a `/drop` deletes the old artifacts dir the recorder feed is detached and drained so a queued write cannot recreate the deleted file. The on-disk log is append-only and independent of the in-memory context — re-primes and compaction never truncate it.
|
||||
|
||||
The advisor is never a peer. The `advisor`-kind registry ref is excluded from every agent-facing surface — the `hub` peer roster and broadcast targets, the subagent peer prompt, and the `history://` index/lookup/completions — and cannot be messaged (`hub` send and collab chat refuse it) or revived/killed from the Agent Hub or collab. It is not addressable as a peer, regardless of what tools it has been granted.
|
||||
The advisor is never a peer. The `advisor`-kind registry ref is excluded from every agent-facing surface — the `hub` peer roster and broadcast targets, the subagent peer prompt, and the `history://` index/lookup/completions — and cannot be messaged (`hub` send and collab chat refuse it) or [revived or killed from Agent Hub](./agent-hub.md#persisted-agents-and-advisors) or collab. It is not addressable as a peer, regardless of what tools it has been granted.
|
||||
|
||||
@@ -0,0 +1,95 @@
|
||||
# Agent Hub
|
||||
|
||||
Agent Hub is the interactive TUI for watching and controlling subagents associated with the current session. It combines a live roster, per-agent activity and usage, transcript access, steering, revive, and kill controls. The main agent is not listed because its conversation is the ambient session view.
|
||||
|
||||
The Hub also discovers parked subagents from the current session's persisted artifacts when a session is resumed. Advisor transcript files appear as read-only rows.
|
||||
|
||||
## Open the Hub
|
||||
|
||||
| Input | Behavior |
|
||||
| -------------- | ---------------------------------------------------------------------------------------------- |
|
||||
| `Alt+A` | Open or close Agent Hub through `app.agents.hub`. This opens the roster even when it is empty. |
|
||||
| `Ctrl+S` | Open or close the same Hub through the legacy `app.session.observe` action. |
|
||||
| Double-tap `←` | Open the Hub from an empty main-session editor when the current session has an agent to show. |
|
||||
|
||||
Run `/hotkeys` to see the active chords. Remap either action in `~/.omp/agent/keybindings.yml`:
|
||||
|
||||
```yaml
|
||||
app.agents.hub: Alt+A
|
||||
app.session.observe: Ctrl+S
|
||||
```
|
||||
|
||||
The double-`←` gesture is not a keybinding action. While focused on a subagent, double-`←` returns to the main session instead of opening the Hub.
|
||||
|
||||
## Roster and inspector
|
||||
|
||||
The roster updates from the session's agent registry and progress events. Its responsive rows show:
|
||||
|
||||
- status (`running`, `idle`, `parked`, or `aborted`), agent identity, parent, and unread IRC count;
|
||||
- model role, resolved model, and age since last activity;
|
||||
- assigned task or current activity;
|
||||
- cost, active time or elapsed span, request count, tool-call count, and tokens.
|
||||
|
||||
The header aggregates status and usage across measured agents. Press `t` to switch between the stable flat roster and a parent/child tree.
|
||||
|
||||
On a wide terminal, the selected agent's inspector appears beside the roster. On a narrow terminal, press `Tab` to replace the roster with it. The inspector adds:
|
||||
|
||||
- the current tool and arguments, last intent, and retry state;
|
||||
- context-window use when available;
|
||||
- parent and child lineage;
|
||||
- output and patch paths, plus isolated-worktree branch metadata when present.
|
||||
|
||||
Metrics depend on the progress or persisted usage data available for that agent. Missing data appears as `usage —` rather than an estimate.
|
||||
|
||||
### Roster controls
|
||||
|
||||
| Key or input | Action |
|
||||
| --------------------------- | ---------------------------------------------------------------------------- |
|
||||
| `j` / `k`, `↑` / `↓`, wheel | Select an agent. |
|
||||
| `Enter` or click | Open the selected agent. |
|
||||
| `t` | Toggle flat and parent/child views. |
|
||||
| `Tab` | Toggle the inspector on narrow terminals. |
|
||||
| `PageUp` / `PageDown` | Scroll an open inspector. |
|
||||
| `r` | Revive the selected parked agent. |
|
||||
| `x` | Abort a running turn if necessary, then kill and release the selected agent. |
|
||||
| `Esc` | Close the inspector first on narrow terminals, then close the Hub. |
|
||||
|
||||
Only `parked` agents can be revived. `x` is immediate; use it only when you intend to discard that agent instance.
|
||||
|
||||
## Read and steer a subagent
|
||||
|
||||
For a normal local subagent, `Enter` or click focuses the main TUI on that agent's session and closes the Hub. Focusing a parked agent revives it. The transcript, status line, and editor then belong to that subagent:
|
||||
|
||||
1. Read its live transcript and tool activity.
|
||||
2. Type a message and press `Enter` to steer a running turn or prompt an idle agent.
|
||||
3. Press `Esc` with an empty editor, or double-tap `←`, to return to the main session.
|
||||
|
||||
Steering uses the normal prompt path, so the message and response are written to the subagent's persisted session history. While a subagent is focused, `Esc` returns to the main session; it does not interrupt the subagent.
|
||||
|
||||
Contexts without a local focusable session use the Hub's full-screen transcript viewer instead. This includes collab guests and advisor rows. The viewer incrementally tails the file-backed transcript and provides an input line only when the selected agent can be messaged. Sending there has the same semantics: revive if parked, steer if running, and prompt if idle.
|
||||
|
||||
## Persisted agents and advisors
|
||||
|
||||
Opening the Hub for a persisted session scans that session's artifact tree. Historical subagent JSONL files become parked rows; a killed agent's tombstone keeps it aborted. Nested subagents retain their parent/child lineage. Output and patch artifacts are attached to the corresponding inspector row.
|
||||
|
||||
Advisor transcript files (`__advisor*.jsonl`) appear as `advisor`-kind rows under their owning session. They are observability records, not peers:
|
||||
|
||||
- their transcripts can be opened and followed;
|
||||
- they cannot be messaged;
|
||||
- they cannot be revived;
|
||||
- they cannot be killed.
|
||||
|
||||
These restrictions also apply to collab guests controlling the host's Hub.
|
||||
|
||||
## Related surfaces
|
||||
|
||||
Agent Hub is the human-facing live session view. Adjacent commands and internal URLs serve narrower purposes:
|
||||
|
||||
- `/jobs` prints a snapshot of running and recently settled asynchronous tool jobs. It does not replace the per-agent transcript or control view.
|
||||
- `history://<id>` gives the coding agent a concise transcript for a live or parked subagent.
|
||||
- `agent://<id>` resolves a subagent's saved final output artifact; it is not the live transcript.
|
||||
- `hub` `list` exposes the peer roster to the coding agent, and `hub` `send` steers or follows up with a normal subagent programmatically. Messaging a parked subagent revives it.
|
||||
|
||||
Advisor rows are intentionally excluded from the agent-facing `hub`, `history://`, and `agent://` peer workflows.
|
||||
|
||||
See also [Task Agent Discovery and Selection](./task-agent-discovery.md), [Collaboration](./collab.md), and [Advisor, WATCHDOG.md, and WATCHDOG.yml](./advisor-watchdog.md).
|
||||
+1
-1
@@ -85,7 +85,7 @@ Guests with a full link can:
|
||||
- read the entire session (including the back-transcript at join time),
|
||||
- prompt the agent (rendered with their name badge on every participant's transcript; the LLM sees the prompt text verbatim — names are display-only),
|
||||
- interrupt the agent (Esc),
|
||||
- use the Agent Hub against the host's subagents: live table and progress, chat (steers the host's subagent), kill, revive, and transcript viewing (fetched from the host on demand).
|
||||
- use [Agent Hub](./agent-hub.md) against the host's subagents: live table and progress, chat (steers the host's subagent), kill, revive, and transcript viewing (fetched from the host on demand).
|
||||
- answer host interactive `select` and `editor` requests. The host broadcasts each pending request only to writable guests; the first submitted or cancelled response settles it and dismisses the other presentations.
|
||||
|
||||
Guests with a view-only link can read everything live — back-transcript, streaming text, tool cards, subagent transcripts — but the host rejects prompting, interrupting, and agent control from them.
|
||||
|
||||
+1
-1
@@ -47,7 +47,7 @@ app.history.search: []
|
||||
| `app.clipboard.pasteImage` | Linux: `Ctrl+V`; macOS: `Ctrl+V`, `Cmd+V`; Windows: `Ctrl+V`, `Alt+V` | Paste from the clipboard (image preferred, text fallback) |
|
||||
| `app.stt.toggle` | Unbound (hold `Space`) | Toggle speech-to-text. By default there is no key chord — hold the space bar to record (push-to-talk) and release to transcribe; bind a chord here for a press-to-toggle alternative |
|
||||
| `app.live.toggle` | `Ctrl+L` | Start or stop live voice mode (same as `/live`) |
|
||||
| `app.agents.hub` | `Alt+A` | Open the agent hub |
|
||||
| `app.agents.hub` | `Alt+A` | [Open the Agent Hub](./agent-hub.md) |
|
||||
|
||||
On Windows Terminal, `Ctrl+V` may be handled by the terminal paste command before `omp` sees it; use the `Alt+V` fallback when clipboard image paste appears to do nothing. When the clipboard holds no image, `app.clipboard.pasteImage` pastes the clipboard text instead, so hosts that deliver only this chord (VS Code's integrated terminal when configured to forward `Ctrl+V`, Windows clipboard history via `Win+V`) work for both payload kinds. Windows Terminal also swallows `Ctrl+Enter`, so the `app.message.followUp` chord also binds `Ctrl+Q` — the same chord GitHub Copilot CLI uses — and the same chord submits the agent dashboard's new-agent description and hook-editor prompts. If your existing `keybindings.yml` already assigns `Ctrl+Q` to another action, that user remap wins and follow-up keeps `Ctrl+Enter` unless you explicitly bind `app.message.followUp`.
|
||||
|
||||
|
||||
@@ -85,6 +85,10 @@ For a dispatch, set the agent name and task:
|
||||
|
||||
`/model`'s Roles view can assign and persist custom role mappings such as `review`, `fast`, and `good`. Changing only the active or default session selection does not remap those roles.
|
||||
|
||||
## Watch running agents
|
||||
|
||||
After dispatch, press `Alt+A` to open [Agent Hub](./agent-hub.md). Its live roster shows each task agent's status, current activity, model, age, and usage. Select an agent to read its transcript and steer it directly; parked agents can be revived from the same view.
|
||||
|
||||
### `vibe_spawn` tier routing
|
||||
|
||||
`vibe_spawn` maps `fast` to bundled `sonic` and `good` to bundled `task`. Both resolve through `task.agentModelOverrides` before their bundled agent model defaults (`src/vibe/runtime.ts`, `src/task/agents.ts`).
|
||||
|
||||
@@ -99,6 +99,10 @@ Then use `isKeyRelease()` / `isKeyRepeat()` if needed.
|
||||
- Overlay APIs exist in `TUI` (`showOverlay`, `OverlayHandle`). In interactive extension/custom UI, `custom(..., { overlay: true })` mounts your component through `TUI.showOverlay(...)`; without `overlay`, it replaces the editor component area directly.
|
||||
- Overlay custom UI is anchored at `bottom-center` with full terminal width/max height and is removed through the returned overlay handle when `done(...)` closes the flow.
|
||||
|
||||
### Built-in full-screen surfaces
|
||||
|
||||
The coding-agent integration also mounts built-in full-screen surfaces outside `ctx.ui.custom(...)`. [Agent Hub](./agent-hub.md) is the live roster and control surface for subagents. Its file-backed transcript viewer borrows the alternate screen while it is open, then restores the Hub beneath it on close.
|
||||
|
||||
## Mount points and return contracts
|
||||
|
||||
## 1) Extension UI (`ExtensionUIContext`)
|
||||
|
||||
Reference in New Issue
Block a user