[Yash-arch-ui]

Closes #118

What was wrong

With an mcp.json present, processLine sent every line to completeWithMcp, which completed with { mcp } alone. The nine built-in actions the system prompt advertises were never offered to the model, so a wallet request was a no-op — and a scripted run still exited 0. A built-in call the model did make was answered by the server's tool catalog instead of our safety boundary (Refused: no handler is registered for MCP tool "get_address").

What changed

• src/agent.mjs — completeWithMcp now takes the agent's own tools alongside mcp (the SDK merges both lists into one request; an empty list keeps the request byte-identical to the MCP-only shape it replaces) and routes by name: built-in names go to invokeNativeToolCall, server names to the existing invokeToolCall gate.
• src/nativeToolLoop.mjs — dispatchNativeCall is the seam both turn drivers share. Writes (send_mon, send_token, transfer_nft, swap) and account-with-index go through the same handleAction a slash command uses (resolve → policy → preview → mainnet ack → y/N); reads go through dispatchToolCall. boundary says who printed, so results are shown once, not twice or never. Refusals, tool errors and a hit round limit all fold into hadFailure.
• src/nativeToolLoop.mjs — runMcpTurn is the new MCP turn driver, living beside runNativeToolLoop; src/cli.mjs only passes in the handles it needs, same as the native loop. invokeMcpToolCall is exported next to handleAction so the gate is testable.

The MCP confirmation gate is untouched: every server tool still asks, and a declined call never reaches the server.

Tests

Coverage of the combined CLI dispatch offline (only the model is faked — no live model, no GPU, no funded wallet):

• test/mcp-native-dispatch.test.mjs (13) — completion receives tools and mcp on every round; a built-in write reaches the real handleAction (refused, exit 1) and never the gate; a read goes through dispatchToolCall, prints once, and a refusal still fails the run; MCP calls route to the gate and nowhere else; the gate refuses an unapproved call; toolError and round limit fail the run; routing unit tests for dispatchNativeCall.
• test/mcp-cli.test.mjs (5) — dist/cli.mjs end to end with a real stdio MCP server (test/helpers/mcp-stdio-server.mjs): server connects, tools=9 mcp=1 in the actual request, built-in read/write answered by our boundary (exit 1 on refusal, server untouched), a declined server call never reaches the server (exit 1), and the same call approved by a scripted y really runs on it (exit 0).
• test/agent-mcp.test.mjs (+4) — request shape (tools absent when empty), name-based routing, and the missing-handler error.

Verified the new tests fail against the current code: tools=0 mcp=1 and the built-in read/write misroute (3 of 5 subprocess tests fail on main).

npm run build   ✓
npm test        ✓ 573 pass, 0 fail
npm run smoke   ✓ SMOKE_OK

Scope: no new surface — this restores the built-ins the prompt already promised alongside MCP, on the existing safety boundary.

[portdeveloper]

The combined native/MCP routing and policy coverage look good. One error path still needs a fix in runMcpTurn (src/nativeToolLoop.mjs:301-313). If dispatchNativeCall throws, boundary stays "handleAction", so the catch creates an error result but the following print branch skips it.

I reproduced this with get_balance throwing Error("RPC unavailable"): the model receives Error: RPC unavailable, hadFailure is true, and the operator output is only newlines. A scripted run can therefore fail without showing the cause; an interactive run depends on the model choosing to repeat it.

Please print caught exceptions directly and keep the no-duplicate behavior for results already printed by handleAction. Add a regression for a throwing read (and a throwing write boundary) asserting the error appears once, reaches the model, and still fails scripted runs. The existing 573 tests pass locally.

[Yash-arch-ui]

Thanks — fixed in 3d9eabc.

The bug: in runMcpTurn's catch, boundary was still "handleAction", so the read-only print branch skipped the Error: … result it had just built. The read boundary that would have printed it is the one that threw, so nothing did: blank lines for a scripted run, and an interactive run relying on the model to repeat the cause.

The fix (src/nativeToolLoop.mjs): the catch prints the exception itself, in red, and returns — nobody else has printed it, so there is no duplicate. Everything else is unchanged: a write result from handleAction still prints only from handleAction, a read still prints only from dispatchToolCall, and the tool result reaches the model either way. Refusal detection, hadFailure and the round/error paths are untouched.

Regressions added (test/mcp-native-dispatch.test.mjs, counting both print channels):

• throwing read (get_balance → RPC unavailable): printed exactly once, Error: RPC unavailable is the tool result in history, handleAction never called, hadFailure → exit 1;
• throwing write boundary (send_mon → signer unavailable): same four assertions, and dispatchToolCall never called;
• a successful read is still printed exactly once, so the no-duplicate contract is pinned too.

Both throwing tests fail on the reviewed commit (12 pass / 2 fail) and pass with the fix.

npm run build   ✓
npm test        ✓ 576 pass, 0 fail
npm run smoke   ✓ SMOKE_OK

[portdeveloper]

lgtm, thanks for fixing the missing error output