Macro is a unified workspace for teams: email, chat, docs, tasks, agents, calls, and CRM — @-linked together with shared AI memory.
TypeScript agentaiai-agentsall-in-onecrmcrm-systememaillinearmcpmessagingnotesnotionofficerustslackslack-alternativestartupstartup-toolstypescriptworkspace
chg(import): fallback from oss->nano (#5231)
teo57a4992d1f26
diff --git a/apps/web/src/features/setup/flow/BuildingStep.tsx b/apps/web/src/features/setup/flow/BuildingStep.tsxindex 509e816b23..52aa521609 100644--- a/apps/web/src/features/setup/flow/BuildingStep.tsx+++ b/apps/web/src/features/setup/flow/BuildingStep.tsx@@ -2,7 +2,7 @@ import { LoadingSpinner } from '@core/component/LoadingSpinner';import { createSignal, onCleanup, onMount, Show } from 'solid-js';/** The breather before the summary: brand animation + cycling phrases for- * ~15s, then auto-advance. Pure theater — nothing waits on real work. */+ * ~22s, then auto-advance. Pure theater — nothing waits on real work. */const BUILDING_PHRASES = ['Building your unified workspace…',@@ -13,7 +13,7 @@ const BUILDING_PHRASES = ['Waking up your Macro…',];-const BUILDING_HOLD_MS = 15_000;+const BUILDING_HOLD_MS = 22_000;const PHRASE_MS = BUILDING_HOLD_MS / BUILDING_PHRASES.length;export function BuildingStep(props: { onDone: () => void }) {diff --git a/crates/import/src/domain/service.rs b/crates/import/src/domain/service.rsindex 8e7a516286..9aaa382907 100644--- a/crates/import/src/domain/service.rs+++ b/crates/import/src/domain/service.rs@@ -44,6 +44,11 @@ mod test;/// A raw provider-prefixed id (not [`PredefinedModel`]) because the registry/// has no Cerebras tier — the model router resolves the `cerebras/` prefix.const GATHER_MODEL: &str = "cerebras/gpt-oss-120b";+/// Fallback when the primary gather session fails: Cerebras enforces tight+/// org-wide rate limits, so a burst of concurrent onboardings can 429 every+/// gather at once. gpt-5.4-nano is slower per token but rides OpenAI's much+/// higher limits, and a slower gather beats a failed section.+const GATHER_FALLBACK_MODEL: &str = "openai/gpt-5.4-nano";/// Turn cap for gather sessions: a couple of searches plus one staging tool/// call per candidate (providers may batch several per turn).const GATHER_MAX_TURNS: usize = 24;@@ -464,6 +469,31 @@ where}}+ // Staging is idempotent (the ledger dedups already-staged rows), so+ // rerunning the whole session on the fallback model is safe even+ // when the primary died mid-way through staging.+ match self+ .gather_agent_session(user, source, GATHER_MODEL, mcp_tools.clone())+ .await+ {+ Ok(()) => Ok(()),+ Err(e) => {+ tracing::warn!(model = GATHER_MODEL, error = ?e, "gather session failed; retrying on the fallback model");+ self.gather_agent_session(user, source, GATHER_FALLBACK_MODEL, mcp_tools)+ .await+ }+ }+ }++ /// One agent gather session on a specific model.+ #[tracing::instrument(skip(self, user, mcp_tools), err)]+ async fn gather_agent_session(+ &self,+ user: &MacroUserIdStr<'static>,+ source: ImportSource,+ model: &str,+ mcp_tools: Arc<McpToolSet>,+ ) -> anyhow::Result<()> {let native = gather_toolset::<Self>();let toolset = NativePlusMcp::new(native, mcp_tools);let context = ImportToolContext {@@ -473,7 +503,7 @@ whereself.drive_session(user,- GATHER_MODEL,+ model,GATHER_MAX_TURNS,toolset,context,