Skip to content

armgit / macro

Public Imported from GitHub
0

Macro is a unified workspace for teams: email, chat, docs, tasks, agents, calls, and CRM — @-linked together with shared AI memory.

fix: use proper separate worker retry machines (#4762)

Wolf Mermelstein69c5ebeff2d4
diff --git a/crates/documents/src/outbound/editing_worker_client.rs b/crates/documents/src/outbound/editing_worker_client.rs
index c20d80e5f6..36f3490e73 100644
--- a/crates/documents/src/outbound/editing_worker_client.rs
+++ b/crates/documents/src/outbound/editing_worker_client.rs
@@ -47,9 +47,6 @@ impl EditingWorkerService for ReqwestEditingWorkerClient {
"documentToken": document_token.as_str(),
"documentId": document_id,
"prompt": instructions,
- // Each role lists models tried in order: Cerebras primary, with an
- // Anthropic fallback so edits keep working if Cerebras is rate-limited
- // or down.
"models": {
"supervisor": [
{ "provider": "anthropic", "model": "claude-opus-4-8" },
@@ -61,20 +58,9 @@ impl EditingWorkerService for ReqwestEditingWorkerClient {
{ "provider": "anthropic", "model": "claude-sonnet-4-6" },
],
"coding": [
- { "provider": "anthropic", "model": "claude-haiku-4-5" },
- { "provider": "cerebras", "model": "gpt-oss-120b" },
- ],
- // Snippet composition: fast/cheap by default, escalating to a
- // strong prose model when the supervisor marks a spec
- // `effort: "high"`.
- "snippet": [
{ "provider": "cerebras", "model": "gpt-oss-120b" },
{ "provider": "anthropic", "model": "claude-haiku-4-5" },
],
- "snippetHigh": [
- { "provider": "anthropic", "model": "claude-sonnet-4-6" },
- { "provider": "cerebras", "model": "gpt-oss-120b" },
- ],
},
"interpret": false,
});
diff --git a/services/ai-editing-worker/src/ai-editing/agents/supervisor.test.ts b/services/ai-editing-worker/src/ai-editing/agents/supervisor.test.ts
index 9fa8543bb2..89a33178df 100644
--- a/services/ai-editing-worker/src/ai-editing/agents/supervisor.test.ts
+++ b/services/ai-editing-worker/src/ai-editing/agents/supervisor.test.ts
@@ -136,9 +136,7 @@ describe('supervisor — streamed dispatch', () => {
const models = {
supervisor: supervisorModel,
interpret: codingModel,
- coding: codingModel,
- snippet: codingModel,
- snippetHigh: codingModel,
+ coding: () => codingModel,
} as unknown as ResolvedModels;
const session = createEditingSession();
diff --git a/services/ai-editing-worker/src/ai-editing/agents/supervisor.ts b/services/ai-editing-worker/src/ai-editing/agents/supervisor.ts
index 5cb78eed2e..b90950bfe7 100644
--- a/services/ai-editing-worker/src/ai-editing/agents/supervisor.ts
+++ b/services/ai-editing-worker/src/ai-editing/agents/supervisor.ts
@@ -54,7 +54,7 @@ export async function supervisor(
const dispatch = createDispatchTool({
session,
- childModel: models.coding,
+ makeChildModel: models.coding,
tracker,
request: intent ? `${request}\n\n<intent>\n${intent}\n</intent>` : request,
params: opts.params,
diff --git a/services/ai-editing-worker/src/ai-editing/tools/dispatch.test.ts b/services/ai-editing-worker/src/ai-editing/tools/dispatch.test.ts
index cb2c602a8d..b37d04a67a 100644
--- a/services/ai-editing-worker/src/ai-editing/tools/dispatch.test.ts
+++ b/services/ai-editing-worker/src/ai-editing/tools/dispatch.test.ts
@@ -140,7 +140,7 @@ function setup() {
const requests: Array<string | undefined> = [];
const dispatch = createDispatchTool({
session,
- childModel,
+ makeChildModel: () => childModel,
tracker,
request: 'make it formal',
runner: () => [],
diff --git a/services/ai-editing-worker/src/ai-editing/tools/dispatch.ts b/services/ai-editing-worker/src/ai-editing/tools/dispatch.ts
index 2a1ea7e7a0..4d0dd475b5 100644
--- a/services/ai-editing-worker/src/ai-editing/tools/dispatch.ts
+++ b/services/ai-editing-worker/src/ai-editing/tools/dispatch.ts
@@ -208,7 +208,9 @@ export type CoderRunCode = { code: string; snippets?: Record<string, string> };
export type DispatchToolOptions = {
session: LexicalSession;
- childModel: LanguageModel;
+ /** Called once per dispatched coder so each gets its own model with private
+ * fallback state, instead of racing a shared instance. */
+ makeChildModel: () => LanguageModel;
tracker: TokenTracker;
/** The user's original edit request, shown to every writer for tone/intent. */
request?: string;
@@ -240,7 +242,7 @@ export type DispatchEdit = z.infer<typeof DispatchEditSchema>;
export function createDispatchTool(opts: DispatchToolOptions) {
const {
session,
- childModel,
+ makeChildModel,
tracker,
request,
params,
@@ -257,13 +259,19 @@ export function createDispatchTool(opts: DispatchToolOptions) {
const serialize = opts.serialize ?? serializeWithXml;
type CoderRun = Awaited<ReturnType<typeof runTask>>;
- type EditRun = { run: Promise<CoderRun>; trace: DispatchEditTrace };
+ type EditRun = {
+ run: Promise<CoderRun>;
+ trace: DispatchEditTrace;
+ /** This coder's own model, kept for usage attribution after it finishes. */
+ model: LanguageModel;
+ };
// Coders already started from onInputAvailable, keyed by tool call id.
const calls = new Map<string, EditRun>();
async function runEdit(
edit: DispatchEdit,
- trace: DispatchEditTrace
+ trace: DispatchEditTrace,
+ childModel: LanguageModel
): Promise<CoderRun> {
// Serialized at launch time so an edit launched early sees whatever earlier
// coders have already applied.
@@ -308,7 +316,8 @@ export function createDispatchTool(opts: DispatchToolOptions) {
coderFinishedAt: 0,
runCodeAt: [],
};
- const entry: EditRun = { run: runEdit(edit, trace), trace };
+ const model = makeChildModel();
+ const entry: EditRun = { run: runEdit(edit, trace, model), trace, model };
// A streamed run can reject before `execute` awaits it; keep that rejection
// for the join instead of letting it escape as an unhandled rejection.
entry.run.catch(() => {});
@@ -329,7 +338,7 @@ export function createDispatchTool(opts: DispatchToolOptions) {
calls.delete(toolCallId);
const result = await entry.run;
onEditTrace?.(entry.trace);
- tracker.add(childModel as { modelId: string }, result.totalUsage);
+ tracker.add(entry.model as { modelId: string }, result.totalUsage);
const blocked = findBlocked(result);
const summary = blocked
? `⚠ BLOCKED -- ${blocked.message}.`
diff --git a/services/ai-editing-worker/src/endpoints/edit.ts b/services/ai-editing-worker/src/endpoints/edit.ts
index 0edec6fb9a..b5e1925129 100644
--- a/services/ai-editing-worker/src/endpoints/edit.ts
+++ b/services/ai-editing-worker/src/endpoints/edit.ts
@@ -117,7 +117,8 @@ edit.post('/', zValidator('json', EditBody), async (c) => {
models: {
supervisor: resolveModel(models.supervisor),
interpret: resolveModel(models.interpret),
- coding: resolveModel(models.coding),
+ // Fresh fallback per coder — see ResolvedModels.coding.
+ coding: () => resolveModel(models.coding),
},
typingAnimations,
sleep,
diff --git a/services/ai-editing-worker/src/run-edit.ts b/services/ai-editing-worker/src/run-edit.ts
index 2383d9e69c..b31d9c2bf7 100644
--- a/services/ai-editing-worker/src/run-edit.ts
+++ b/services/ai-editing-worker/src/run-edit.ts
@@ -30,7 +30,7 @@ export type Models = {
export type ResolvedModels = {
supervisor: LanguageModel;
interpret: LanguageModel;
- coding: LanguageModel;
+ coding: () => LanguageModel;
};
export type RunEditArgs = {