ACP and this platform: fit and roadmap
Running record of the design discussion that followed the protocol and ecosystem study in synthesis.md. These are directional decisions, not final commitments. This record is frozen as decision-time input: where a conclusion here differs from an accepted record in the ADR index or from the current spec position in ACP Conformance, the ADR or the conformance document is authoritative.
What this platform already has
The exact spec position, SDK pin, schema cap, and per-method conformance status drift too often to repeat numbers here; see ACP Conformance for the current version table and matrix. At the time of this study, the relevant shape was: wire v1, the official HTTP/WebSocket transport adopted for remote traffic, a hand-rolled NATS leg by design (ADR#0020) over a shared JSON-RPC-over-NATS codec (ADR#0011), typed decode rather than passthrough forwarding at the bridge (ADR#0021), and the ACP and A2A crate families kept in separate directory groups (ADR#0034; that grouping is navigational, not itself a dependency-boundary rule). A known defect recorded in the conformance matrix: the ext/session/prompt_response completion path has no production waiter-registration site, so the bridge-extension's receive half is routed and tested but the completion half is dead code in a release build. No service crate wires the ACP-over-NATS bridge to a process-spawning path yet.
The gap
Every major coding agent CLI ships an ACP agent mode (gemini --acp, codex-acp, claude-agent-acp, goose acp, opencode acp, cline --acp, cursor-agent agent acp, grok agent stdio, devin acp), all wire v1 over stdio. This platform cannot call any of them yet. The ACP crate family speaks the agent role and bridges NATS to HTTP/WebSocket, but nothing in it spawns a child process and speaks the client role against that child's stdio. The missing component is one ACP client host: something that spawns a CLI as a stdio child, serves fs/terminal RPCs against sandboxed workspaces, routes session/request_permission through a policy decision point instead of a blind passthrough, injects provider credentials as environment variables at spawn, and bridges the resulting session onto the existing NATS subjects. The current pinned SDK already ships the subprocess helper needed for the spawn side; buzz-acp and Devin Desktop are working reference designs for the host role as a whole (see Host Role and Invocation Mechanics).
Recommended build sequencing
- Build the ACP client host crate (working name
acp-host): spawn a CLI as a stdio child, speak the client role, serve fs/terminal RPCs against sandboxed workspaces, forwardsession/request_permissioninto a policy decision point instead of a blind passthrough, inject provider credentials as environment variables at spawn time, and bridge sessions onto the existing NATS subjects. Validate against the community conformance tools (a CLI smoke-test harness, a trace viewer for debugging). - Fix the
ext/session/prompt_responsedead-code defect: the waiter registration that would complete the bridge extension's send half is currently absent from production builds, so out-of-turn responses are logged and dropped instead of delivered. Land this before any host work depends on out-of-turn responses. - Bump the schema pin per the upgrade ritual in ACP Conformance; this is what unblocks the tool-call
nameconformance gap recorded there. - Wire the first three agents in this order: Gemini CLI first (native agent, cleanest headless auth via an API key or Vertex ADC), then
codex-acp, thenclaude-agent-acp. Goose, OpenCode, Cline, Cursor, Grok, Hermes, and Buzz follow as demand dictates; see the product dossiers for per-product invocation and auth notes. - Channel adapters: adopt the OpenClaw-validated pattern from Channel Mapping and Channel Bridge Mechanics. Channels (console, chat surfaces, A2A callers) act as ACP clients in front of the gateway, sessions map one-per-conversation, and replay goes through
session/loadwhile still on wire v1. Keep ACP as the boundary protocol rather than merging channel-specific semantics into ACP payloads, consistent with the protocol/transport layering rules in ADR#0003 and ADR#0004. - Watch-only, unchanged: ACP wire v2 (adopt once upstream marks it preview, per the conformance policy), MCP-over-ACP routing (the three methods stay unrouted until runner demand appears), and the
-polyfill/-rmcp/-conductorcompanion crates (adoption triggers are in the Rust Crate Inventory).
Design dossiers for the acp-host build
Four component-level designs support the acp-host build:
- Sandboxed Per-Session Workspaces: a tiered design combining always-on RPC-boundary validation, an OS-sandbox layer around a per-session git worktree with default-deny egress, and a policy-triggered escalation tier for higher-risk sessions.
- Permission Decision Point: a policy authorizer sitting behind
session/request_permissionthat combines verified identity, delegation chain, session key, and tool risk class into an issue/require-interactive/deny decision, and fails closed. - Credential Injection at Spawn: resolving provider credentials per spawn and injecting them as environment variables, built on the OpenBao-backed secret custody direction in ADR#0023. Note that the specific guarantee that spawned agent implementations never receive upstream provider credentials directly, and that ACP messages carry no upstream credentials, authorization headers, or grant tokens, is recorded in the draft ADR#0032 rather than in ADR#0023 itself. The secret custody direction fixes the secrets-service architecture those guarantees are built on, but had not reached accepted status as of this writing.
- Media Store Decision: S3/MinIO object storage with lifecycle TTLs, a
media://resource-link scheme resolved to fresh presigned URLs, and an inline-base64 guardrail for small payloads.
Full supporting evidence, including the original callability matrix and later product additions, is in synthesis.md and the product dossiers.
