Add daemon-backed HarnessClient receive path

[chughtapan]

Summary

• add raw same-conversation turn delivery to MoltZapChannelCore without daemon-side presentation enrichment or checkpoint writes;
• make moltzapd own the active MoltZapService, channel core, WebSocket connection, loopback MCP listener, and turn publisher;
• expose an Effect-based HarnessClient whose MCP-backed turns carry their originating conversation and a payload-only bound reply closure;
• own the narrow MoltZap subscriptions/listen extension adapter while delegating discovery, tools, standard subscriptions, and unrelated lifecycle behavior to the official MCP SDK;
• bound graceful MCP response draining so a stalled loopback reader cannot retain daemon shutdown; and
• remove local duplicate-reply suppression so every bound reply(payload) invocation sends.

Why

This is the first operational production slice of #926. It establishes the daemon-owned receive/reply boundary after #941 removed production dispatch leases, without recreating leases, reply tokens, action IDs, turn IDs, or generic authority fields.

The result is a real loopback MCP path from a peer message through moltzapd to HarnessClient, with a reply returning through the originating conversation and no Unix-socket dependency.

SDK extension boundary

The package-local adapter intercepts only the exact MoltZap turn-ready filter. Standard, mixed, malformed, inherited-property, and unrelated requests remain SDK-owned. An independent landability audit found mixed-filter and inherited-property isolation gaps; both are fixed and regression-tested.

The upstream official SDK proposal remains a possible future replacement, not a prerequisite: modelcontextprotocol/typescript-sdk#2593

Review fix

Pre-landing lifecycle review reproduced a stalled SSE reader holding scope closure behind Node response backpressure. Shutdown now gives handler closure and response draining one second, then destroys remaining scope-owned connections and joins cleanup. The regression test pauses a reader behind an 8 MiB event and proves scope closure completes without client cooperation while the existing small slow-reader test still receives terminal completion.

Validation

• pnpm nx run workspace:precommit — 61/61 targets
• pnpm nx run @moltzap/client:test — 30 files, 221 tests
• focused real daemon integration — 3 PGlite/WebSocket/MCP tests
• OpenClaw unit suite — 66 tests
• NanoClaw unit suite — 24 tests
• independent contract, lifecycle, and test/maintenance reviews — no remaining landability blocker

Deliberate follow-up scope

This PR does not yet cut OpenClaw or NanoClaw over to HarnessClient, define the deferred management/search/context Schemas, or remove the existing CLI, Unix RPC socket, and generic send. Those remain ordered follow-up slices in #926 and are retained until their MCP replacements are operational.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
moltzap	🟢 Ready	View Preview	Aug 4, 2026, 4:11 AM

💡 Tip: Enable Workflows to automatically generate PRs for you.