[germago119]

Summary

The installed GitHub Copilot CLI streams live tool activity (text chunks, tool start/progress/result, permission decisions, completion) over the ACP protocol (copilot --acp --stdio, JSON-RPC protocol version 1). This adds an opt-in CopilotAcpRuntime alongside the existing copilot backend/runtime, preserving the original copilot -p runtime unchanged as the default.

• No new --runtime backend enum: ACP is a transport option on the existing copilot backend, selected via config (see feat(config): add opt-in Copilot ACP transport).
• Text/tool/permission/completion events are normalized into the existing runtime event projection and persistence — no parallel/UI-specific event system was introduced.
• src/ouroboros/copilot/acp_client.py-adjacent protocol boundary owns a fresh child process and one prompt per invocation, validates envelopes/response IDs/protocol version/session IDs/stop reasons, bounds frames and startup/idle waits, drains stderr concurrently, and sends cancel before closing/reaping owned processes.
• Fallback to the legacy transport is intentionally not attempted mid-prompt (fallback is impossible once a prompt may have been submitted); auth/authorization errors are classified and surfaced rather than silently replayed through another transport.

Testing

350 targeted unit/integration tests added (see diffstat), covering the ACP client/events/permissions, the new runtime, config wiring, and a standalone read-only streaming probe fixture (tests/fixtures/fake_copilot_acp.py). Existing legacy Copilot runtime tests were re-verified to still pass unmodified.

Full implementation notes, including the initial ACP diagnostic (protocol handshake, observed update types, subagent-ancestry caveat) and documentation cross-link updates (CONTRIBUTING's Orchestrator → Doc Mapping, runtime guide, capability matrix, CLI/config reference, getting-started, architecture, README): .acp-artifacts/IMPLEMENTATION-REPORT.md (not intended to ship, happy to drop it from this PR if preferred — flagging here so reviewers have the trail).

Why draft

Opening as draft to get early feedback on the transport-vs-backend design decision and event-projection approach before final polish. No credentials, config, or unrelated projects were touched — this branch only affects src/ouroboros/copilot/, src/ouroboros/orchestrator/, and their tests/docs.

🤖 Generated with GitHub Copilot CLI

[Q00]

Reviewed the ACP client/runtime boundary and focused tests. The opt-in transport design is coherent and preserves the legacy path. Before marking ready, please add an issue reference to satisfy the PR hygiene check and resolve the current CI failure; this PR is intentionally still draft, so I am leaving a comment review rather than approval.

[lee3Q]

APPROVE for the reviewed code at this head. I reviewed the opt-in configuration/factory path, ACP client framing and response/session validation, process lifecycle and cancellation, permission decisions, event normalization, and persistence integration. The legacy CLI transport remains the default; the change does not introduce a second runtime backend or replay a submitted prompt through fallback. I found no demonstrated contract violation in the changed scope.

I ran the ACP client/events/permissions/runtime/config/integration selection: 128 passed. These were local checks using existing cached packages rather than the exact lock-file environment. An authenticated live Copilot ACP run was not performed, and repository-required CI remains a separate merge gate. This review is not a merge or deployment authorization.

Reviewed head: 758af3bae87edb744e9e636ef17b7a4e5c24e3c3.