参与者参考

本页为已锁定版本的英文上游参考，尚未翻译。中文入门指南可从左侧导航进入。

A participant is an addressable Agent conversation within a Caelis work Session. The main controller, native collaborators running Caelis’s built-in runtime, and external ACP agents share one collaboration network. A participant keeps its own conversation and can exchange messages with the controller and peers; finishing a Turn does not prevent later follow-up work.

This guide covers configuration and the terminal workspace. The external ACP contract owns mailbox delivery, authorization, and endpoint capability requirements.

Configure collaborators

Use /connect to add a provider model or external ACP agent, then /model to choose the main Session controller. External executables must already be on the Host’s PATH. The ACP catalog includes a Custom command for other stdio agents; see connection and endpoint setup.

For Jev, choose Connect a judgment model in /connect, then typesafe, the API key, and jev-1.13.0. This flow omits conversation-only controls. The key uses the same managed credential store as other providers. Judgment profiles appear in /team and /disconnect; they cannot become the main Session model, a participant, or Reviewer.

Open /team to configure the profiles available for collaboration:

• self uses the current Session controller’s model, reasoning effort, Fast selection, and approval mode. With auto-review, the child’s own Guardian handles its approvals. Full access is inherited from the shared Host. Manual approval continues through the ACP permission bridge.
• breeze, orbit, and zenith are named profiles that you bind to a provider model or ACP agent. Their descriptions help the controller select a fit for the work; the name itself does not select a model.
• Custom roles give another profile a stable handle and capability description.
• Binding sets save named snapshots of explicit profile bindings.

In the model picker, Tab and Shift+Tab cycle through the selected model’s Model, Effort, and Fast cells. Up/down selects a model and retains the focused field when supported. Left/right adjusts effort (or speed while Fast is focused); F toggles Fast. Fixed effort and unsupported Fast controls are skipped. Click a model once or press Enter to apply; clicking effort or Fast edits the draft. Escape discards the draft. Fast off is explicit standard speed. Binding sets preserve independent profile Fast choices.

On the main list, Tab switches between a role’s model and its auxiliary field. A single click opens either selector. Left/right and F also update effort and Fast directly on an explicitly bound model. Only the focused cell is highlighted; the filled dot in the picker marks the currently applied model.

The same overlay includes fixed system roles; configuring a binding does not start a conversation:

Role	Binding behavior
ToolSearch	A restricted selector inherits the Main Agent’s resolved provider model when unbound and can consume model tokens when candidates are ready. Bind a separate provider model or Jev judgment to override it. An explicit provider binding retains its effort and Fast setting. Jev scores the scoped name/description catalog; search failures are reported without keyword fallback.
Guardian	Uses a provider model, or the Main Agent model when unbound. An auxiliary classifier selector, labeled Classifier, appears on the same row when a judgment model is connected; it is off until selected.
Reviewer	Uses a provider model or ACP agent for the fixed review scene.

ToolSearch’s private model requests use their own context capacity and do not trigger the main Session’s compaction watermark. Their provider attempts and reported usage enter the Session journal without adding selector messages to the main model conversation.

On Guardian’s row, use Tab or click the main model or auxiliary classifier to configure it. A selected classifier runs first, choosing only among the original approval options. Direct approval requires a decisive approve distribution and low probabilities of both missing decision-relevant facts and visible violations. Direct refusal requires a decisive deny distribution and a concrete visible violation; it has no generated explanation and can settle despite other missing facts. Otherwise, missing facts, conflicting or unclear judgments, malformed answers, an oversized input, or a provider failure continue to Agent review within the same approval deadline. Protocol-valid options whose kinds are not supported by screening go directly to the Agent; option names and IDs are never guessed. The Agent supplies a reason when its review rejects the action. Screening uses only user messages and the current approval ticket with its exact action. It receives no historical tool observations and never waits for Tasks. Diagnostics distinguish input-budget, option-eligibility, provider, and inconclusive-judgment fallbacks without recording request content. Changing or disabling screening preserves the Agent model selection. See Guardian’s evidence contract.

/team takes no arguments and opens the configuration overlay. The TUI accepts /subagent as an alias; completion shows one /team (subagent) entry. Likewise, /quit accepts /exit and appears as /quit (exit). Built-in commands take precedence over same-name Skills and custom roles. A conflicting role stays in configuration with a name-conflict warning; rename it to use its slash command. The role name experimental is reserved for new role creation. Existing roles, bindings and sets with that name remain valid; /experimental opens the feature editor and the role remains addressable through participant APIs.

The product term participant describes the collaboration role; subagent still names native execution contracts, SDK packages, and existing wire or storage fields. Those technical names do not imply a separate messaging network.

Coordinate work

Ask the main controller for a concrete division of work and shared outcome:

Use two participants to investigate this failure: one should trace the code
path, the other should reproduce it with a focused test. Have them exchange
findings before you propose a fix. Do not edit production code yet.

The controller starts conversations with StartThread and reads their public results with ReadThread or WaitThread. Each participant can use ListThreads to discover the Session roster and SendMessage to send mail to another handle, including the reserved controller address parent. Follow-up messages reuse the same conversation rather than starting a new isolated task.

ReadMessages lets each Agent observe explicit group messages, including mail addressed to other participants. The TUI shows only messages addressed to the Agent in that pane: parent in the main transcript, or the selected participant in a child pane. Successful reads with no matching messages leave no visible tool row; failed reads remain visible. This display filter does not change model context, read progress, or mail delivery. Outgoing SendMessage rows remain visible.

Only the controller can start participants or observe their public results through thread tools. Participants receive discovery and messaging tools, not creation or handoff authority. Their private reasoning and full transcripts do not become the controller’s model context merely because the TUI displays them.

Native tools and the external MCP bridge call the same mailbox service. A send acknowledges queueing, not delivery or execution. Mail is consumed once, without automatic retry or redelivery if dispatch or a tool response is lost. External agents need to accept injected MCP tools to discover peers and send mail; running input additionally depends on negotiated steering. Without steering, queued mail can be collected through MCP during a Turn or submitted when the Agent is known to be idle. See the delivery contract for the full guarantees and limits.

Participant workspace

Enter /orbit <task> (or another configured role) to start a background participant and open its pane automatically. The main controller remains free to work. Use /{handle} <message>, /orbit(handle) <message>, or the pane’s input to continue that same Task through its user-input mailbox. Initial input and follow-ups retain their user role and attachments. Child approvals use the Session approval queue, including while the main controller is idle. Follow-ups retain an input receipt: queued means waiting for admission, while sent means accepted by the child endpoint, not necessarily applied by its model. The participant pane and ACP output report delivery failures and uncertain outcomes; uncertain input is not automatically resent.

/review [instructions] starts a new background Reviewer and opens its pane. The fixed review scene and configured Reviewer model or ACP agent apply to every review. The child remains available after its first result: use its handle (for example /lina recheck the fix) or /reviewer(lina) recheck the fix to request another review with the same context. The main controller can discover these handles through ListThreads and contact them through SendMessage. Core commands and configured role commands take precedence over bare handles; qualified names such as /reviewer(lina) select the participant explicitly.

Existing foreground ACP attachments remain addressable until detached; new direct starts, including reviews, use background Tasks.

The TUI discovers child Tasks from the selected Session’s Control directory, including children started by an external ACP controller through the collaboration MCP bridge. Discovery does not depend on a visible StartThread row or load child transcripts. Opening a participant pane subscribes to that child’s content; hiding the pane releases its content subscription while status remains observed.

Click the running/done count in the footer or a participant’s label in the transcript to open one participant pane. Message rows split that click: the peer label (handle[agent], or @handle[agent] when you send) opens the peer’s pane, while the rest of the row expands or collapses the message in place. Long incoming messages start collapsed in both the main transcript and participant panes. The name dropdown switches participants; the layout dropdown chooses Overlay, Split left/right, or Split up/down. The dropdown lists Agents, not individual command Jobs.

Drag the divider to resize within 30–70%. It previews the new position; the transcripts reflow once on release. Caelis remembers the layout and separate horizontal/vertical ratios in the Host’s UI preferences. Small terminals use an overlay temporarily and restore the preferred split when space permits.

Focus and layout

Action	Control
Focus a composer	Click its pane
Switch composer focus	F6; Shift+F6 in reverse
Show or hide the participant pane	F7
Open the pane and participant dropdown	Ctrl+G from either composer
Open the layout dropdown	Ctrl+L while the pane is open
Resize with the keyboard	Choose Resize split, then use arrows along the divider axis

F6 also opens the selected participant when hidden. Both composers use the same focus colors: brighter in dark themes, deeper in light themes. The participant title also marks focus; both transcripts and progress hints remain readable. Main-composer Tab completion and Shift+Tab mode switching keep their usual behavior.

Keyboard resizing previews five-percentage-point adjustments. Enter applies and saves; Esc cancels. Terminal resizing cancels an unconfirmed resize preview. F7, the footer’s clickable F7 Hide, or the title’s close button hides the pane while the participant keeps working. The footer shows the short model ID on the left and the latest ACP-reported context usage at the far right, with F7 Hide immediately to its left. Agents that do not report context usage leave the gauge empty.

Send input

Enter sends a prompt to the selected participant. Shift+Enter or Ctrl+J adds a line. Up recalls the last submitted prompt when that composer’s draft is empty. Drafts and scroll positions survive switching participants during the Session.

Click within the participant composer to position the cursor; drag text to copy on release. Image paste uses the main composer’s platform shortcut: Ctrl+V on macOS/Linux, Ctrl+Alt+V on Windows/WSL. Images stay with the selected participant’s draft and are included when recalling its last prompt.

Esc dismisses a participant menu or selection; it does not hide the pane or interrupt an Agent. Main-composer Esc retains its interruption behavior. Model and context usage are read-only; the main controller retains orchestration. Direct user input is separate from Agent mail and does not copy the prompt into the main conversation. A queued input receipt is not proof that the model has applied it; uncertain delivery outcomes are not automatically resent.