外部 ACP 智能体

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。

This document owns the product contract for local external ACP Agent onboarding, authentication, model selection, input, disconnect, and endpoint compatibility. For collaboration setup and workspace controls, see Participants. Layer ownership lives in Architecture. For Host settings clients and model-provider OAuth, see Model Connections.

Protocol versions

The default public ACP endpoint and unselected third-party processes use stable ACP v1. Stable initialization negotiates v1 when a client requests version 1 or newer; version 0 is invalid. A stable outbound client requires the negotiated response to be exactly v1 before Session work. The AppServer user API and Application integrations keep their existing stable types independently of the ACP wire profile.

Repo-owned subprocess children use the explicitly selected schema-v2.0.0-alpha.8 profile. The hosted Codex adapter keeps stable v1 by default. Direct native main/child delegation stays inside Control and Runtime and does not acquire an ACP transport. User-provided process overrides remain stable unless their link explicitly selects a draft. The Host-owned adapter descriptor, one-use channel grant and helper invocation carry the same exact schema. An unsupported schema fails before granting or serving the channel; the selected draft never falls back to v1.

caelis acp --acp-draft=schema-v2.0.0-alpha.8 serves product Sessions through the existing principal-bound Control clients. Initialization requires protocol version 2, the v2 info shape and an exact _meta["caelis/acp_schema"] tag. Responses advertise the implemented Session capabilities and the conditional-prompt profile described below. Initialization is pinned once per connection. Unknown methods and unimplemented optional features fail explicitly. Only actual image/MCP/configuration support is advertised; native ACP Sessions do not accept lifecycle MCP server declarations or additional directories that their existing owner cannot apply.

The Codex draft channel requires native app-server 0.160.0 or newer with clientUserMessageId correlation. The adapter assigns a private submission token and acknowledges only the corresponding actual native user item. Older or unverifiable native versions fail this selected channel without a stable fallback. Codex authentication remains Host-owned; neither product draft endpoint advertises remote authentication or elicitation without an implementing owner. It serves standard v2 lifecycle and active-only insertion as an explicit experimental channel, but advertises idleOnly: false: native multi-agent input can start an owned Thread independently, and native turn/start lacks atomic start-only-if-idle admission. The controlled Caelis client rejects this missing capability during initialization. Safe default migration needs that native admission contract; it does not disable native multi-agent behavior or silently fall back after selecting a draft.

Draft session/prompt acknowledges insertion with a real messageId, echoed in a live user-message update and kept unchanged when retained and replayed. This acknowledgement is distinct from queue acceptance and foreground completion. state_update reports running, requires-action and idle; only the matching idle state settles the internal synchronous producer. session/cancel requests cancellation and still waits for the actual state. Each Session has independent foreground state, receipt and permission ownership. Reconnect uses standard session/resume, with replayFrom: {type: "start"} when replay is needed, and keeps the existing ordered replay barrier. Transport loss before settlement remains unknown_outcome; there is no automatic resubmission.

Native permission requests keep the same original options and resolution owner. Where alpha.8 cannot express completeness, sensitivity or per-option scope, the bounded public approval subject travels in _meta["caelis/approval_subject"]. The receiver stores it as a reported display hint, joins only the original option identities and marks remote claims in presentation. Guardian continues using the canonical standard subject and private native evidence; the hint does not grant authority or change resolution.

Full message/tool/terminal updates retain absent versus explicit null/empty patch semantics. Message replacements reuse their message identity, terminal chunks append only their decoded bytes, and terminal snapshots replace existing output, including an empty snapshot. These are normalized into the same live, canonical and recovery projections used by stable clients.

Exact retained narrative chunks remain deltas during session/resume catch-up. Each replay epoch resets an identified narrative once before its first retained delta; it does not replace the message with each chunk. Repeating a resume on the same client therefore reconstructs the same text without duplicate appends. Session message ownership survives Activity boundaries: repairing a message from an earlier Activity updates its display without taking the new Activity’s result.

The explicit Codex draft channel preserves native skill and mention references by their supplied name and path without reading those paths. Native negative command exit sentinels produce an exit status without an unsigned exit code. If a route fails before native completion is confirmed, alpha.8 cannot express an Agent-side Session closure. The affected ACP connection closes observably, pending native requests are cancelled and the owner attempts interruption of the actual native Turn. Sessions on that connection become unknown outcomes; other ACP connections and the shared native backend remain independent. Cleanup does not manufacture a successful idle state.

The selected draft retains message identities for later repairs and settled Activity identities to reject reuse. Its private display cache has a 32 MiB serialized budget per Session and 128 MiB per connection, with at most 4,096 message, terminal and settled Activity records of each kind per Session and 1,024 Sessions per connection. Exhaustion is explicit and closes the channel; it does not evict old identities and reinterpret a repair as new work. Accepted work then stays unknown_outcome. Already projected terminal bytes remain with the canonical terminal owner; the private cache buffers only output awaiting its first tool association, then releases those bytes.

Connect

The /connect overlay, including authentication method choices, supports keyboard and mouse navigation. Hovering a catalog row selects it without activating it; hovering a form field highlights it without moving the editing cursor. Buttons have brackets and highlight on hover. A click activates only when the left mouse button is released on the same control. Read-only text supports drag-to-copy.

Run /connect and choose an ACP agent. The guided flow:

1. prepares the selected local command and discovers its capabilities;
2. completes one declared authentication method when required;
3. selects the Agent’s default or an advertised remote model;
4. commits the Agent and resulting ModelProfile against the current Host configuration revision.

Preparation, authentication, and connect are Host-scoped, principal-bound, idempotent commands. Durable intent is recorded before process, authentication, Session, or configuration effects. A committed configuration write is not rolled back because later refresh or durability observation reports a warning.

Preparation records are secret-free, time-bounded recovery evidence, not a second operation ledger. Ambiguous process or protocol cleanup remains unknown_outcome and is never retried blindly.

Terminal authentication is a caller capability rather than wire data. An embedded interactive client may run a declared terminal login. HTTP clients cannot carry that executable callback and fail before starting the login; Agent-managed ACP authentication works through either transport.

Authentication recovery

Caelis detects ACP auth_required by its structured JSON-RPC code, not message text. Authentication methods come from initialize.authMethods.

For an authenticated operation, the bridge:

1. calls the operation;
2. on auth_required, selects one declared method;
3. performs agent-managed authentication in-band or directs terminal login back through interactive /connect;
4. retries the original operation once.

A second auth_required is returned without another side effect. Session open, resume, prompt, and negotiated steering use this path. If authentication succeeds but resume proves the remote Session unavailable, the controller may create a new Session only while it can still prove that the current prompt was not submitted.

See the upstream ACP authentication methods RFD for standard wire behavior.

Speed and service tiers

ModelProfiles expose a typed speed capability independently of reasoning effort. Each choice maps a product value to the exact backend wire value; ACP profiles also retain the advertised config ID. /model and /team show Fast only when both standard and Fast are available. Provider capabilities retain the official endpoint checks. A binding’s empty speed preserves an explicitly configured profile default in SessionDefaults, or inherits the backend default when none is configured; standard and fast are explicit selections. Binding sets retain that choice. Host and Session model commands retain their explicit Fast boolean: false selects standard on a capable profile. Guardian remains provider-only.

The Codex adapter reads model/list.serviceTiers and defaultServiceTier and publishes service_tier. Every catalog-known model exposes the protocol baseline default, including models without additional tiers. Additional choices use catalog IDs (currently priority for Fast). Missing or empty catalogs do not advertise Fast; there is no additionalSpeedTiers or model-name fallback. Thread start/resume responses supply the effective tier. Missing/null tier stays inherited internally, even when the selector displays its discovered default. Control sends explicitly configured values even when they equal that display; only an omitted selection inherits. When a combined model/tier selection requests Standard, Control applies it before switching models if the current model advertises it, then confirms the requested defaults against the destination configuration. Changing a model validates an explicit tier against the destination capability. Explicit standard (default) remains valid without additional tiers; an incompatible Fast selection fails without silently downgrading it.

Set-config validates the current catalog and stages the next Turn selection. The next turn/start.serviceTier applies it to that and subsequent Turns; serviceTierForTurn is not used. Configuration cannot change during an active Turn. A rejected start retains the complete staged model, effort and tier so a retry sends the same selection; it does not combine a new model with an old tier. An ambiguous start closes the route rather than claiming a successful change. A newly opened Codex thread cannot be resumed until it has persisted its first Turn, so set-config does not fabricate a resume as an acknowledgement. Resume reads actual backend state again. See the Codex app-server protocol.

Explicit speed is sealed in the existing Placement: provider service_tier or ACP session_config_values. It contributes to the placement fingerprint. Rebinding changes future work; active participants keep their frozen selection. Session overrides do not mutate Codex’s global configuration or another thread.

Input and collaborator Sessions

An idle external collaborator receives session/prompt on its existing ACP Session. On the controlled draft links, running input also uses standard session/prompt. Initialization negotiates _meta["caelis/conditional_prompt"] = {version: 1, activeOnly: true, idleOnly: true}. State updates carry the owner-assigned Activity identity in _meta["caelis/activity_id"]; running-only input sends _meta["caelis/active_only"] = {activityId: "observed identity"}. The peer atomically checks the actual active native Turn before insertion. Initial foreground calls send _meta["caelis/active_only"] = {idleOnly: true} instead; the actual owner admits them only while idle, so another caller starting work cannot turn the initial prompt into mid-turn injection. Both conditions in one guard are invalid. An unmet condition returns the profile’s structured promptRequired no-effect error. Callers do not silently retry it as a different operation. Ordinary unguarded external v2 prompts keep standard insertion behavior. ACP alpha.8 has no equivalent running-only or idle-only admission/CAS field, so this minimal negotiated metadata remains. Pure commands may use their durable interaction identity as an opaque Activity ID; that ID never authorizes steering into a native Turn.

Each successful controlled-draft PromptResponse also carries _meta["caelis/activity_id"]: the actual admitted native Turn, or the durable interaction identity for a command that creates no Turn. Its standard messageId identifies the canonical inserted input and the mandatory same-ID live user echo. Callers bind completion to this receipt’s Activity and its true terminal state, including when state notifications arrive before the response. Another caller’s running/idle pair cannot complete the request. Missing receipt identity is an unknown outcome; callers do not infer it from ambient Session state or retry the input.

Stable v1 collaborators receive _session/steering only when they advertised _meta.steering.supported. That maintained v1 extension remains available for third-party peers; direct running input stays unsupported without it. StartThread reports the effective link capability as supports_steering. Product mailbox, input queue, receipt and Task lifecycles remain Control-owned: a v2 inserted message receipt does not replace their separate queued/sent/applied or durable Task states.

Control appends a tagged collaboration setup block to the child’s initial session/prompt, after the task body and sender. It names the assigned handle, the reserved parent address parent, and the participant role. It supplies the stable MCP discovery key caelis-collaboration and asks children to report meaningful progress through SendMessage, process returned mail, and end the current Turn when finished or blocked. New messages resume the same Session. The setup is not appended to later idle prompts, steering, or reconnects to the same Session. Its text is English; task and message bodies retain the sender’s language. Credentials, mailbox contents, and Session or Task identifiers stay outside it. Built-in spawned Sessions receive their identity through system-prompt assembly and do not inherit controller-only tool guidance. Transferred parent context follows the initial task body.

Product Agents share one Control-owned mailbox service within their owning work Session. Children expose ReadMessages, ListThreads, and SendMessage; the controller also has StartThread, ReadThread, and WaitThread. Built-in children and child MCP servers register only their tool set. ListThreads discovers all collaborators.

SendMessage {to, message, reply_to?} atomically commits a message to the shared log and recipient’s persistent mailbox, then independently takes up to 32 messages from the caller’s own mailbox. The result contains {id, status:"queued"} and optional messages; it never echoes the outgoing body. Empty or failed inbox checks omit messages and do not change the successful send acknowledgement. A failed send does not take inbox messages. Incoming batches are bounded by encoded JSON size within the shared 4 MiB response limit; remaining mail stays queued. Automatic delivery, SendMessage replies, and controller waits share the atomic read-and-delete path: a message is consumed by only one path. There are no automatic retries if a response is lost or a dispatch outcome is unknown. Message text is limited to 65,536 bytes; reply_to, when present, is a canonical message UUID.

ReadMessages {limit?, cursor?} reads shared explicit collaboration messages. The default page contains at most 32 messages; the limit can be 1–128 and the encoded response is bounded by 4 MiB. Each entry includes sequence, message ID, sender, directed recipient and optional reply ID. @handle and handle address the same recipient. Public results, private reasoning, tool traces, original controller conversation and direct user input are not copied into this log. New members may read messages sent before they joined, within retention.

Control commits default read progress before returning. Concurrent reads from one member serialize; a cancelled transaction does not advance progress. A lost response may therefore lose a page from the unread view. Supplying an earlier cursor explicitly replays retained bodies without changing saved progress. Use the returned cursor to continue; has_more reports remaining entries at the read snapshot. Messages appended later remain available on subsequent reads. These are observation positions, separate from ReadThread cursors and delivery. Reading never drains a mailbox or starts execution. A message read from the group log can still arrive later through its pending directed delivery.

Successful sends, mailbox takes and successful automatic input admissions mark individual IDs known to that reader. Default group reads skip those IDs while scanning intervening unseen records in order; they never jump over unread gaps. Admission and server-side response completion do not prove model consumption. An ambiguous or failed dispatch leaves its body readable in the group log, even though its directed queue entry was consumed. Explicit replay can include known bodies. None of these paths promises exactly-once delivery.

Reader identity includes the immutable Task/participant/attachment generation, or the parent’s controller epoch. Credential renewal and reconnect within that instance retain progress. Handle reuse, participant replacement and controller handoff start a new reader. Progress survives Host restart for unchanged instances; a newly activated controller epoch starts afresh. Old credentials cannot read as a replacement. Control retains the newest 4,096 messages per work Session; truncated_before reports when an older cursor crosses discarded history. Session closure discards the log, read positions and setup markers along with pending mail. Existing mailbox rows are not backfilled into shared history.

ReadThread returns a participant’s latest public result and observation cursor, not its reasoning or complete conversation. Supplying after suppresses already observed output. WaitThread waits up to 60 seconds for incoming mail or new terminal/attention states among at most eight selected threads. It returns when any target needs attention or mail arrives, rather than waiting for all targets. The result contains the wake reason, consumed messages and thread observations; timeout does not cancel work. Neither tool exposes the parent transcript or cross-Session routing. Thread observations use one cursor with handle, state, and optional name or output, without internal Task IDs or a duplicate revision. Received mail retains id, from, message, and optional reply_to; its implicit recipient is omitted. Wait results omit empty message and thread lists. TUI keeps display-only decoding of retained ReceiveMessages results for historical transcripts; that compatibility can be removed when those records are no longer supported. The tool itself is not exposed or callable.

Only the controller receives StartThread. It creates a persistent participant conversation and starts its initial prompt, returning its handle, status and steering capability without folding its result into the creation result. Follow-up messages reuse that conversation. Participant removal is an internal Control operation; it is not exposed as a model tool and preserves Session history. Running or unresolved participants must settle before removal.

Taking a message or claiming it for automatic delivery removes it from the mailbox. A failed dispatch or lost tool response can lose the message; there is no acknowledgement, retry, or redelivery protocol. Repeating SendMessage creates another message. Pull and automatic delivery share the same atomic removal. Automatic delivery is serial per recipient and independent across recipients, with a ten-second deadline per attempt. Each attempt atomically claims the pending messages that fit the encoded batch budget and submits them as one input admission, preserving their order and individual source identities. The same rule applies to the parent: a running local controller accepts the batch at its next safe model boundary, and a running ACP controller receives one negotiated running-input request. Local safe boundaries follow a completed model response or tool step; mail does not cancel an in-flight tool. Accepted local batches are committed atomically before entering model context. A final empty drain closes input admission so late mail can select the next Turn without being acknowledged into a completed Run. The 32-message pull limit does not split automatic delivery; messages beyond the encoded budget remain queued. A deadline does not establish whether the peer executed the input and does not trigger a retry. Agents without steering can take mail through MCP during their current turn; otherwise the Host waits for a known terminal activity before submitting the queued batch as one prompt on the existing Session. An unresolved execution does not qualify as idle. Confirmed Session closure discards its pending mail, including mail recovered after Host restart; transient discovery failures do not authorize cleanup.

Native tools and the caelis collaboration mcp --stdio bridge call the same AppServer service. The bridge uses Host-issued CAELIS_COLLABORATION_URL and CAELIS_COLLABORATION_TOKEN environment values, never general Host credentials. External controller and child creation/load/resume inject this stdio server through ACP mcpServers when the Host has a child Control endpoint. The built-in Codex adapter translates stdio MCP declarations into per-thread app-server overrides; HTTP and SSE injection are not supported by that adapter. Bridge exit does not cancel queued messages.

Controller MCP declarations include the activation’s configured delegation names and role descriptions from the same catalog as native StartThread. The Host passes these bounded definitions in CAELIS_COLLABORATION_TOOLS, so tool search works during ACP setup before the grant is bound. Definitions describe tools; each invocation still requires the current authenticated controller grant.

The controller’s first work-bearing prompt includes a lightweight maintainer instruction with the stable caelis-collaboration search key. A durable marker belongs to the work Session and exact remote Session, so ordinary turns, credential renewal, same-remote resume and Host restart do not repeat it. A new remote Session gets its own instruction. A proven non-submission releases the marker; an ambiguous submission retains it. The prompt grants no authority.

Controller grants bind the current controller epoch and exact remote Session. Before prompting a replacement remote Session, Runtime commits its binding under the admitted Turn’s fence. A fresh remote starts at checkpoint zero and receives full context; only successful prompt completion acknowledges context delivery. Restoring a controller can replace its epoch within an admitted Turn. Bound Turn approval, steering and cancellation use the exact Handle/Run/Turn target; new commands still validate the controller epoch observed before admission. StartThread reuses Runtime’s native Spawn, permission, journal, Task and sealed placement owners under the live controller Turn’s execution fence. Participants cannot obtain this authority through tool arguments or an MCP role flag. Hosts embedded without a child Control endpoint cannot inject collaboration MCP; endpoint and transport failures are reported by Session setup or tool invocation. The Codex adapter forwards explicit empty-form MCP tool approvals through ACP permission requests. It does not answer arbitrary elicitation forms or URLs.

The Host keeps random bearer grants in memory. Each grant binds one work Session, immutable participant instance, and exact ACP Session. A pending grant must bind after a successful handshake within two minutes. An active grant has a fixed 24-hour deadline; tool calls and repeated binding cannot extend it. Before another idle turn, a connection with less than one hour remaining is replaced and resumes the same ACP Session with a new grant. Expired credentials are rejected even during a long turn. A peer that cannot resume reports a reconnection error rather than silently starting another Session.

Every call checks the current participant instance and work-Session lifecycle; mailbox waits recheck authorization as they wait. The first call may wait up to two seconds for the initial participant commit. Connection close or failure, participant removal, replacement activation and Host shutdown invalidate the affected grant. A closed work Session cannot use existing grants. Host restart requires new grants. A bearer copied elsewhere still represents its original identity; it never grants access to a caller-selected Session.

Credential values remain outside model prompts, tool schemas, results and canonical history. Tool schemas are stable across credential changes. Pending mail survives connection replacement, but consumed mail is never retried.

Delivered input follows the ordinary Agent-communication context path. The model receives the task body followed by mail reference lines (Message-ID and optional In-Reply-To) and a single From footer, without an embedded mailbox JSON object or repeated recipient identity. Typed source identity and the original display body remain separate from this text. Accepted input is projected as ACP session/update with user_message_chunk on v1 and the corresponding canonical user_message on the controlled draft; display-only source metadata may use _meta.caelis.agent_communication, while typed event identity remains authoritative. Caelis mailbox IDs are separate from peer-owned ACP message IDs. Delivered context remains in canonical Session history after its mailbox entry is removed. The external history reader removes collaboration setup and mail footers from child display and attributes preceding content blocks to the footer’s sender. Its legacy header reader is display-only; remove that reader once supported external histories no longer contain header-format prompts.

Direct user input quotes each text block as a JSON string inside a <caelis_user_input version="1"> prompt envelope. ACP peers need not retain custom content metadata, so this text encoding distinguishes user content from mail syntax across retained history (session/load on v1 and session/resume on v2). The child model receives the envelope; the Task transcript decodes it once and displays the original text with a user source. Images remain ordinary ACP content blocks. The envelope carries no principal identifier and grants no authority. Replay joins fragmented envelope text before decoding and never interprets its decoded body as mail or setup. Unmarked retained history still uses the legacy display parser; an already stored ambiguous message has no reliable evidence for retroactive attribution.

A stable v1 session/prompt remains open until its execution reaches a Turn terminal. A draft prompt acknowledges insertion earlier; the internal producer still observes its matching foreground settlement separately. If ACP forwarding fails, the bridge can no longer reliably service permissions: it requests Control cancellation and waits for the actual terminal, rather than silently draining a Run that may be waiting for approval. Cancellation requests do not prove cancellation completion. If observation is lost, or cancellation cannot be settled within 30 seconds, the prompt reports an error, not a successful end_turn or cancelled response. A completed Turn wins a race with observation cancellation and returns end_turn; unrelated forwarding failures remain errors, even when their cleanup successfully cancels the Turn.

A standard RPC internal error, request-cancellation error, or unrecognized peer error does not prove that remote execution stopped. The child Task records unknown_outcome, retaining the response phase and numeric RPC code when available without exposing peer text or data. Such an activity rejects follow-up input, including after Runtime reload: session/resume restores access to a Session, not proof that its old execution is idle. Automatic execution reconciliation is not available through standard ACP resume; unresolved Tasks remain isolated rather than retrying a prompt blindly. Proven admission rejections and normal completed Turns retain ordinary follow-up behavior.

The Host records ACP prompt-response, settlement and cleanup failures in <Store>/logs/runtime.jsonl, with Task, activity, parent-call and Session identities, RPC code, submission classification when available, and a bounded original error chain. This private sink is independent from model context and Task output. It uses owner-only file access and the existing 2 MiB rotation with one .1 backup. Error details may contain sensitive peer data or paths; review logs before sharing. Prompts, launch environments and child stderr are not attached to these records.

ReadThread and WaitThread observe subsequent public results. Thread identity and ACP Session placement survive Runtime or Host restart; later input resumes that exact Session rather than substituting session/new.

Task addresses individual asynchronous Jobs. Its model-facing operations reject participant handles. Job input and cancellation depend on the producer’s capabilities; RunCommand is the current built-in Job producer. The standalone SDK direct-input SendMessage tool is available only when an embedder explicitly assembles it. Product assembly always uses the Control mailbox tools; Runtime never injects a legacy SendMessage tool implicitly.

A nested Spawn performed inside a third-party participant stays behind that participant boundary. Caelis may render its final standard tool result, but it does not create another parent Task workspace or flatten the nested transcript.

Main Session controller

The product ModelProfile catalog is the single /model surface for provider and ACP backends. Selecting an ACP profile transfers the selected Session controller from the SDK Kernel to that Agent. Caelis continues to own durable Session state, feed/replay, permissions, and handoff.

The durable controller binding freezes the Agent, remote model, configuration, effort, remote Session ID, controller epoch, and context-sync position. Runtime reattachment uses that binding rather than resolving the current profile again. When a remote Session is gone, the bridge may create a replacement and transfer canonical context only before the new prompt has been submitted.

An ACP-backed Host default is stored as a dormant binding for new Sessions and starts no Agent process until work activates it. ACP main Turns do not require a local provider. Local-only capabilities such as Runtime compaction are omitted or rejected while the external Agent controls the Session.

The latest standard ACP usage_update is retained as the main context gauge. Collaborating participants’ gauges remain on their Tasks and contribute once to Session totals.

Disconnect

/disconnect acp supports selecting multiple Agents. Each Agent, its ACP profiles, and bindings are removed in one revision-aware Host command. Enter submits the selected targets without a second confirmation. Commands run in selection order and stop at the first error; confirmed removals remain applied.

Control immediately revokes the removed placement from live Runtime catalogs, detaches matching participants, and repairs affected main-controller bindings. An accepted in-flight operation may finish with already resolved values, but later work cannot select or display the deleted profile. A post-commit repair warning never restores disconnected configuration.

/disconnect provider lists configured provider models grouped by Provider and supports the same multi-selection flow. It never removes an ACP Agent connection.

Models

ACP model catalogs are optional. If an Agent advertises none, Caelis creates one product-only Agent default profile and sends no synthetic model ID to the Agent.

Guided onboarding selects the remote model but does not impose a reasoning effort. The Agent-advertised choices become profile capabilities; fixed Agent bindings and participant attachment choose an explicit effort later.

Provider selectors exposed by Caelis

Caelis publishes configured provider models as provider[@endpoint]/model in AppServer presentation, ACP model configuration options, and /model completion. Only the literal default endpoint is omitted: deepseek/deepseek-flash, xiaomi@api-cn/mimo-v2.5-pro, and xiaomi@token-plan-cn/mimo-v2.5-pro name distinct routes. Adding another endpoint never changes a route’s selector. Custom model aliases retain their fully qualified configuration ID so multiple configurations of one upstream model remain distinct.

Control’s modelconfig package owns selector generation and resolution. Model options and their current value use the same public selector; selection resolves to the existing internal configuration and ModelProfile identities before any Session write. Credentials, durable bindings, and upstream model IDs are not renamed. Remote model IDs owned by external ACP Agents are unaffected.

Selection first matches an exact internal ID, then a public selector, and only then an unambiguous historical alias. A historical alias cannot shadow a public selector: deepseek/model names the default route when that route exists, while deepseek@office/model names the office route even if both configurations share an alias. Without a matching public selector, ambiguous aliases such as xiaomi/model are rejected rather than resolved through the Host default.

Compatibility aliases are input-only, not additional options. Historical alias resolution is not a stable route binding when the catalog changes; clients that need one must use an advertised selector or an exact internal ID. Alias support remains while historical client references are supported and can be removed only through an explicit breaking migration. A public selector colliding with another configuration’s public selector or exact internal ID is still a catalog error; shared historical aliases alone do not invalidate the catalog.

Endpoint catalog

The built-in catalog contains stable commands for official ACP stdio modes plus a Custom command:

Catalog ID	Command
antigravity	agy_acp_server.par on macOS; agy_acp_server.par --uid= on Linux; agy_acp_server.exe on Windows
grok	grok agent stdio
kimi	kimi acp
opencode	opencode acp
copilot	copilot --acp
qoder	qoder --acp, falling back to qodercli --acp
gemini	gemini --acp
qwen-code	qwen --acp
auggie	auggie --acp
cline	cline --acp
factory-droid	droid exec --output-format acp-daemon
goose	goose acp
kilo	kilo acp

Installed commands are discovered on the Host process PATH. Antigravity also supports the explicit installation flow below and discovery in its default installation directory. Caelis persists ordinary commands and arguments; it does not pin runtime versions or run background updates. Use Custom for any other ACP stdio command.

Executable discovery does not make an Agent available as a collaborator. Open /team to choose its participant profile and binding in the Team overlay.

Google Antigravity installation and updates

/connect lists Google Antigravity immediately after Codex. It uses Google’s standalone ACP runtime; installing the agy CLI alone does not supply this endpoint. When the runtime is missing, the wizard offers two actions:

• Install agy_acp_server.par (.exe on Windows) checks the official ACP Registry entry for a download matching the Host’s operating system and CPU architecture. The confirmation page links to the archive and lets you edit the destination. Install and continue downloads the complete bundle from Google, extracts it, sets executable permissions where needed, and continues to authentication and model selection. The runtime archive is downloaded only after this confirmation.
• Manual setup shows Caelis instructions inside the wizard: a direct ZIP download link for the Host platform, the extraction directory, and executable permission commands where needed. It does not open an editor-specific tutorial or run those commands. After installing, choose Check installation to continue. Drag over the download URL or instructions to select text; release to copy it. Copied commands retain their original line breaks even when the terminal wraps them. The overlay expands with the terminal; use the arrow keys or mouse wheel to scroll on a small terminal. Dragging at the top or bottom edge scrolls the selection.

When the current Turn is idle and a model is configured, the manual page also offers Send to agent. It sends an installation task to the current Session’s Agent, including the Host OS, CPU architecture, resolved destination, official archive, and launch arguments. The task asks the Agent to download and extract the complete bundle, apply platform-specific permissions, and verify an ACP initialize handshake. This uses ordinary prompt submission; it preserves the composer draft and attachments. The button is hidden while a Turn is running or no model is configured. The current Agent needs access to the Host and installation tools. Browser sign-in remains an interactive user step. Use Tab to switch between the footer actions, or click either button.

The default destination is ~/.local/share/antigravity-acp on macOS and Linux, and %LOCALAPPDATA%\antigravity-acp on Windows. The Host resolves these paths for its current user; the displayed directory is not a fixed username or a path derived from the TUI client’s machine. The installer accepts a new or empty directory and preserves existing files by refusing to overwrite them. Failed downloads or extraction leave no published runtime. The connection uses an absolute executable path, reused when reconnecting, so this flow needs no PATH change. The directory is user-owned; Caelis has no adapter cache, version lock, or automatic update policy.

For manual installation without another editor:

1. Open Manual setup and download the ZIP from its Google download link. The URL comes from the official ACP Registry entry and matches the Host platform. If downloading from the registry yourself, select the matching distribution.binary entry and open its archive URL.

2. Create the default directory above and extract the entire ZIP into it. Keep companion files such as localharness_external beside the runtime. On Windows, use Extract All; on macOS or Linux, an archive utility or unzip works.

3. On macOS or Linux, run the commands displayed in the wizard on the Host. For the default directory, they are:

   cd "$HOME/.local/share/antigravity-acp"
   chmod +x agy_acp_server.par
   [ ! -f localharness_external ] || chmod +x localharness_external

4. Return to Caelis and choose Check installation. No PATH change is needed for the default directory. For another directory, use Custom command below, or add it to the Host’s PATH and restart the Host from the updated environment with caelis service restart before checking again.

To use another existing directory without changing PATH, choose Custom command and enter the absolute executable path, quoted if it contains spaces. On Linux, append --uid=, for example: "/path/to/antigravity/agy_acp_server.par" --uid=. Custom connections retain their own generated identity.

For manual updates, use a fresh ZIP link from Manual setup or the registry. Stop active Antigravity sessions, download the desired official archive, and replace the full bundle together. Keep the configured directory and command stable, or reconnect with the new path. New sessions use the replacement runtime. Authentication belongs to the official runtime; personal Google accounts use its Google sign-in option. Select the connected model in /model to use Antigravity as the main controller, or bind it in /team as a participant.

Compatibility

Compatibility stays inside the Host-private ACP bridge and is selected from observed message shape or advertised capabilities, never a guessed peer version.

Path	Enabled condition	Removal event
Flat Session configuration options	Standard options fail normalization; standard shapes always win	Supported peers and upgrade fixtures no longer emit the flat shape
Legacy models and session/set_model	No standard model option is advertised and the requested model exists in the legacy catalog	Every supported selectable peer uses standard model configuration and no fixture needs the legacy channel
Legacy tool names	Standard name is absent or null; projection reads retained tool-name metadata, and permission conversion falls back to metadata, title, then kind. The main-controller bridge uses raw output/input names, then kind, when only a title/kind fallback is available	Supported peers and retained histories all supply standard tool names
Prompt image name	Standard image content is valid and a non-empty top-level name supplies display metadata only	Supported peers use standard image URI/reference metadata
Codex MCP display identity	_meta["codex/mcp_tool"] names server caelis-collaboration and tool SendMessage, with no conflicting standard kind or existing exact display name	ACP supplies a standard structured MCP identity that replaces the provider hint
Antigravity command output	A completed or failed command has no standard content or terminal stream metadata, and rawOutput contains typed combinedOutput, commandLine, workingDir, and exitCode fields; explicit content, including an empty collection, wins	Supported runtimes and retained traces supply standard content for these results

Standard ACP v1 Session notices use session/update and are sent only when the client advertises clientCapabilities.session.notices. The SDK validates the wire shape. Unknown severity strings remain presentation hints, and malformed optional fields are ignored. The bridge projects notices through the existing transient Notice event, never reasoning, model input, public child results, or approval authority.

Codex MCP compatibility belongs to the built-in adapter: arguments becomes standard rawInput, and supported result.content blocks become standard ACP tool content. Complete provider results, including structuredContent and unknown content blocks, remain in rawOutput. The Host-private bridge consumes the structured MCP display hint; it never derives a tool name from a title. Generic tools render standard input even without a recognized display profile. Antigravity command output compatibility preserves rawOutput and projects the complete text as a replaceable standard content collection. It never emits the complete result as a terminal delta or infers execution or permission authority from provider result fields.

Older persisted connections may still use package_exec or managed launchers. Runtime keeps them read-compatible, but new onboarding cannot create or repair them. A Codex connection that points into the retired Store-owned acp-agents cache is migrated to the built-in hosted_adapter only when the Host confirms that codex is available on PATH; its stable connection, Agent, profile, and binding identities are preserved while stale discovery is dropped. The cache is reclaimed only after neither external connections nor live ACP preparations still reference it. Remove the general legacy launcher reader only after every supported upgrade source can be migrated to a user-owned executable or a built-in hosted adapter.