bug(orchestration): a child whose provider starts in >30 s is adopted without its bootstrap prompt #1370

Problem

orchestration_create_agent waits at most 30 s for the renderer's create-agent answer (OrchestrationBridge.dispatchRendererRequest, TIMEOUT_MS = 30_000). The renderer answers only after the child's provider has started. createOrchestrationAgent awaits spawn, then window.api.spawnSession, then main's session.start().

A provider start that is slow but healthy therefore makes the create outcome unknown. The caller gets OrchestrationOutcomeUnknownError ("do not repeat"), and the MCP handler never delivers the bootstrap prompt. When the renderer answers later, adoptLateResponse files the child with bootstrapPromptDelivered: false. The parent ends up with an idle child that never received its brief.

Evidence

• Found by review a of fix(opencode): give opencode serve a startup budget that survives a loaded machine #1367 (temp/review-1367/report-a.md).
• Reproduced timing: under CPU contention, the real bundled OpenCode 1.18.31 took 16.7, 18.3 and 43.1 s to report its URL. fix(opencode): give opencode serve a startup budget that survives a loaded machine #1367 raises the serve wait to 120 s, so the 43.1 s start now succeeds in the pane but still exceeds this 30 s bridge deadline.
• Tests: src/main/orchestration/timeoutOutcome.test.ts pins both the 30 s unknown outcome and the late adoption.

Not decided here

Whether to:

• (a) deliver the bootstrap prompt when a timed-out create is adopted late (main holds the request, and armPromptWhenReady already waits for readiness); or
• (b) give create-agent a deadline derived from the provider's start budget.

Option (b) conflicts with MCP clients' own tool-call timeouts: a Codex parent's defaults are shorter than 120 s.

🤖 Generated with Claude Code