feat(acp): per-chat ACP runtime — hot-swap a coding agent from the model picker (ADR 0082 S2+S3)

[mabry1985]

Draft — for local testing before merge (as requested). This is a runtime behavior change CI can't fully exercise (it needs a real ACP agent driving a chat), so it's held for extensive local testing.

Implements the routing + picker for ADR 0082 (Accepted, #1994). Builds on the agent-registration surface #1993 (S1, merged).

What it does

Pick an ACP coding agent (Claude Code / Codex / proto / …) from a chat tab's model dropdown and that chat runs on it — hot-swapping the runtime the way a gateway model already hot-swaps, per tab.

S2 — routing (backend)

• runtime/acp_runtime.py: resolve_turn_runtime(model, config) — a per-tab model=acp:<agent> selects the ACP runtime FOR THIS TURN, overriding the global agent_runtime (even when it's native); anything else falls back to the global. AcpRuntime(agent=…) can now run on a thread whose global runtime is native/a different agent.
• server/chat.py: the ACP session registry is keyed by (thread, agent) — a mid-chat swap builds a fresh session while the previous agent's survives for a resumable swap-back. Both dispatch sites (streaming + non-streaming) + _acp_turn_collected thread the agent through.
• graph/middleware/model_override.py: guard so an acp:* value can never build create_llm's text-only aux relay as the lead model (Trap A). acp:* stays valid for the explicit aux slots.

S3 — picker + boundary (console)

• ComposerModelSelect.tsx: lists ACP agents alongside gateway models; shows the effective agent + a "coding agent" badge; stops replacing the menu with a static label under a global ACP runtime. Under a global ACP runtime it offers coding agents only (a gateway pick would be inert per the routing). chat-store/api needed no change — session.model already carries an arbitrary string end to end.
• Option-C boundary note (D4): on a runtime crossing (native↔acp, acp:A↔acp:B — never gateway↔gateway) it drops a system note via noteToThread ("earlier context isn't carried across runtimes"). Native history (checkpointer) and ACP history (the agent's own stateful session) don't sync, so this is the honest boundary rather than faked continuity.

Tests / gates (local)

• Backend: ruff ✓, lint-imports (3 kept) ✓, full pytest tests/ — 3695 passed, 5 skipped. New tests: the resolver, explicit-agent guard, (thread,agent) keying + swap-back-resume, the middleware guard, per-tab-over-native routing.
• Frontend: tsc (both configs) ✓, vitest — 548 passed ✓, vite build ✓.

Please test locally (needs a real ACP agent CLI installed + authed)

• Global native runtime: open a chat, pick an ACP agent from the model dropdown → the turn runs on it; a boundary note appears; other tabs stay on the gateway.
• Swap agent mid-chat (acp:A → acp:B → back to A) → each crossing notes; swapping back to A resumes A's session.
• Pick a gateway model again → back to native; no note between two gateway models.
• Global acp:X runtime: the picker offers coding agents only, X badged "default".
• Non-streaming surface (/v1, desktop /api/chat) with model=acp:<agent> routes to ACP.

Out of scope (per ADR 0082)

Cross-runtime history bridging (Option B); /effort//compact/HITL/skill parity under ACP; a UI to author a custom acp.agents.<id> (registration stays config/YAML).

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro

Run ID: db3b6fcd-93c8-4fa2-89e2-df66d773b830

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/acp-per-chat-routing

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[mabry1985]

Triage: keep held. S2/S3 genuinely aren't on main (runtime/acp_runtime.py has resolve_runtime but no resolve_turn_runtime; server/chat.py keys the ACP registry by thread_id only, not (thread, agent) — so no per-tab swap/resume). It recomputes CLEAN/MERGEABLE and #2145 didn't touch these files, so it flips to a clean merge the moment you run local QA with a real ACP CLI. Parked deliberately, not stale — no action needed until that QA.