LumenBox 0.3.4

LumenBox 0.3.4 makes connected services usable by the default team and makes the record honest about what delegated engines, retries and metering actually did.

• Connected services reach the default team. Tool lists can name an MCP service (notion__*) or every service the box carries (mcp:*). The starter team and the catalog experts get mcp:*, except the two reviewers. Existing agents are not widened automatically; add mcp:* to one in its settings.
• A delegation is done only when the engine says so. Claude Code, pi and opencode now run in their machine-readable modes, and the outcome comes from the engine's own completion report: done, failed, stopped, or unknown. Measured on the pinned versions, pi exits 0 on a provider error and Claude reports a 400 as "success" with is_error; neither counts as done any more. A thread Claude cannot resume is started fresh, with the reason.
• One retry layer. The model SDK no longer retries on its own underneath the host, so every retry is visible and counted once.
• Metering does not record nothing. The relay records a conservative estimate when a provider reports no usage or a stream is cut, counts it against the ceiling, and reports it apart from measured usage.
• Approvals stay on the record. The policy log archives instead of dropping old approvals, and the audit export now carries who asked, who approved or denied, and what delegated engines called.
• Two timing-dependent tests now assert order instead of milliseconds.

Validation: 2241 tests passed; typecheck and lint clean; npm run release:check passed against these artifacts and the published image; npm run smoke passed 42/42 against a fresh box from the 0.3.4 image. The engine outcomes were measured against a stub model server on claude 2.1.250, pi 0.85.1 and opencode 1.18.25; long real-model runs with tool calls were not measured.

Downloads: signed macOS ARM64 and x64 DMG/ZIP (not notarized). Docker: fakechris/lumenbox:0.3.4 and latest, linux/arm64.