Add report_dreaming_artifact tool to engine-tools MCP server

[dario-nunez]

What

Adds a report_dreaming_artifact tool to the engine-tools MCP server so Dreaming jobs can report a generated artifact back to the platform.

Implements the SDK piece of github/agents issue 1476 (Dreaming M0: add Dreaming artifact reporting tool).

Why

Dream jobs (e.g. memory-voting) produce an artifact that needs to reach the platform. MCP tools do not have the engine job-progress token, so the tool hands the artifact to PlatformClient, which already holds the token and nonce and posts to the job-progress endpoint. The token never reaches the model.

Changes

• client.ts
  • DreamingArtifact type (title, summary, optional data).
  • DREAMING_PROGRESS_NAMESPACE = "dreaming", DREAMING_ARTIFACT_KIND = "artifact", DREAMING_ARTIFACT_VERSION = 0.
  • PlatformClient.sendDreamingArtifact() posts a progress payload on the first-party dreaming namespace (deliberately not the default sessions-v2), so downstream consumers can filter on it.
• mcp-server.ts
  • Registers report_dreaming_artifact, mirroring the existing reply_to_comment tool.
  • Guards on config.platformClient and enforces MAX_DREAMING_ARTIFACT_CONTENT_BYTES (256 KiB) before sending, so oversized artifacts are rejected up front rather than failing on the downstream hydro topic.
• index.ts
  • Exports the new public symbols.

Provisional decisions (M0)

These three items from the issue are resolved here with reasonable defaults and may change:

• Tool name: report_dreaming_artifact
• Payload schema: { title, summary, data? }
• Size limit: 256 KiB of serialized artifact content

Verification

• npm run build (typecheck + esbuild bundle) passes.
• Local end-to-end smoke test (not committed) against the built dist, using the MCP in-memory transport and a stubbed fetch, confirms:
  • the tool is registered and exposes the title + summary input schema;
  • a small artifact triggers exactly one POST to jobs/{jobId}/progress with namespace: "dreaming", kind: "artifact", version: 0, the auth bearer + job-nonce headers, and content round-tripping the artifact;
  • an oversized artifact returns isError and issues no HTTP call.

This repo has no test runner, so no automated tests were added.

Follow-ups (out of scope here)

• sweagentd must add the dreaming progress namespace to its emit list so the payload is forwarded to hydro.
• The copilot-dreams memory-voting task must list report_dreaming_artifact in its tools and prompt for it.
• The hydro consumer + storage that persists the artifact is separate future work.