Routes cannot raise the 60 s render session limit (maxElapsedMs); long-polling tools must clamp their own waits

[ScriptedAlchemy]

Summary

Every rendered route (CLI route or MCP tool) runs inside a render session bounded by DEFAULT_AGENT_RENDER_LIMITS.maxElapsedMs = 60_000 (packages/rsc-runtime/src/agent-document.ts). The CLI runtime and the MCP projector construct the dispatcher with no limit overrides, and neither CliRouteConfig nor ToolConfig has a field to raise it. A route whose legitimate work exceeds a minute — a long-poll such as cargo-hauler's hauler_await — throws elapsed-time-exceeded at emit time, after the wait has already succeeded, and the caller loses the result.

Observed downstream as ScriptedAlchemy/cargo-hauler#32: hauler await cc-N --max-wait-ms 90000 waited 91 s, then printed only Agent render elapsed time exceeds 60000ms and exited 1. The daemon had the result; the render session refused to deliver it.

Workaround in place

cargo-hauler now clamps one await call to 55 s (schema ceiling, guidance text, README) and tells the agent to call again. That is a caller-visible regression in ergonomics purely to stay under a framework constant, and every long-polling plugin will have to rediscover it.

Request

Let a route declare its render budget, e.g.

export const config = {
  description: '…',
  render: { maxElapsedMs: 7_200_000 },   // or `limits: Partial<AgentRenderLimits>`
} satisfies ToolConfig;

• Validated at build time (positive safe integer; perhaps an upper bound per host, since Claude/Codex/Cursor tool timeouts are the real ceiling and AB… could warn when the route's budget exceeds the host's tool timeout).
• Applied by both the CLI runtime and the MCP projector when they create the render dispatcher for that route.
• Surfaced in agent-bundle inspect output.

Alternatively (smaller): treat a streamed Agent.Progress update as keep-alive and only count elapsed time since the last emitted event, so a route that is visibly making progress is not killed by a wall-clock cap.

Version

agent-bundle preview 4edbd493b (main); also present at 9bb5d0dc (DEFAULT_AGENT_RENDER_LIMITS unchanged, no config surface for limits — rg "AgentRenderLimits" packages/agent-bundle/src matches only test/render.ts).

[ScriptedAlchemy]

Additional evidence from the downstream workaround: ScriptedAlchemy/cargo-hauler#47. With hauler await clamped to 55 s per call, an agent that follows the tool's own guidance (hauler result → "call hauler await … up to two hours") hits the schema ceiling; waiting on a 10-minute build becomes ~11 sequential tool calls, which is exactly the polling the plugin's skill tells agents to avoid. Still present on main 10a98a0fb (packages/rsc-runtime/src/agent-document.ts:143, no limits field on CliRouteConfig/ToolConfig). Related: the validator's raw ZodError output that the same report hit is #465.