refactor(client): move the status contract to the harness plane

[chughtapan]

Plan step 6. Stacked on #951. Mechanical, and the single dependency edge that would otherwise make the CLI/socket deletion break the MCP status tool.

Problem

local-daemon-rpc.ts → statusCommandRpc was simultaneously the Unix-socket RPC and the schema source for two consumers that survive the socket's deletion:

• harness-mcp-wire.ts → registerStatusTool (input/output schemas)
• harness-client.ts → readActiveAgentId (result decoder)

So the plane that outlives the socket depended on the plane that does not. Deleting local-daemon-rpc.ts later would have broken the MCP status tool and the client identity read.

Change

The harness plane now states the status contract itself in harness/runtime.ts, alongside the reply and start-conversation contracts it already owned, and exposes it the same way those are exposed — JSON Schemas, derived types, and a strict decoder. No new export style.

The socket surface states the same shape independently rather than importing it. That is not an oversight: agent-code-guard/no-exported-brand-constructor rejects exporting a schema constructor, so the constructor genuinely cannot cross the boundary. Drift is still a compile error in both directions, because service-local-daemon.ts and moltzapd.ts each implement a handler against its own side — I verified this by injecting a field into each schema in turn and confirming each produced a type error.

I drafted a *.types-check.ts drift canary for the duplication and then deleted it: the probe showed it never fires on its own, since the two handlers already catch every case. Shipping it would have been a vacuous gate.

Dropping the RPC handler context ((payload, context) → (payload)) also let the status tool registration lose its Rpc.isWrapper unwrap and the cast justified by "HandlersFrom widens that known-empty environment to any". Net: one fewer cast.

statusCommandRpc is now module-private, matching sendCommandRpc in the same file — knip flagging it as an unused export was the confirmation the coupling was gone.

Gate

git grep -n 'local-daemon-rpc' packages/client/src \
  | grep -v '/cli/' | grep -v 'service-local-daemon\|local-socket-server\|service.ts'

returns nothing — the harness plane no longer reaches into the socket plane. This also required decoupling harness-mcp-server.test.ts, which was building an entire socket handler table via makeLocalDaemonHandlers just to extract its .status field.

• pnpm typecheck --skip-nx-cache — green
• pnpm lint — 0 warnings, 0 errors (incl. knip, arch:check, arch:config:check)
• pnpm nx run @moltzap/client:test — 233 passed / 32 files

No behavior change: the same shape is advertised and decoded, only its owner moved.

ADR conformance (R-pass)

Governing outcome	Record	Owner	Binds this diff?	Verdict	Evidence
Tool named status on the local MCP surface	20260728-endpoint-daemon-speaks-modern-mcp	v2	no	CONFORMS	Governs v2/* per the step-0 sweep (52 records classified, posted on #926). Tool name and shape unchanged regardless.
Search/history/status are MCP management operations, not adapter-facing service methods	20260801-harness-client-owns-runtime-context	v2 (contested)	no	CONFORMS	status stays a management tool; it is not added to HarnessClientService. readActiveAgentId remains private.
—	—	—	—	—	No main-governing record covers client module layout. The 5 main-governing records are agents-md-single-source, v2-lives-top-level, code-first-simulator-kernel, effect-native-evaluation-results, principal-io-uses-runtime-gateways — none touches this diff.

No new divergence introduced; no ledger row changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01P76aaa1STr3WPZ3nDascta

[chughtapan]

Superseded by #974, which collapses this stack into one PR against main. This branch's commits are all contained in it — nothing is lost, and this thread stays readable. Closing to leave one place to review.