fix: macOS arm64 libchdb load failure and recoverable agent tool failures as error issues

apps/ai/src/chat/turn-runner.ts

@@ -26,6 +26,7 @@ import { workerEnvLayer } from "@maple/infra/worker-runtime"
 import { workerTelemetryConfig } from "@maple/infra/worker-telemetry"
 import { Cause, Effect, Exit, Layer, ManagedRuntime, Match, Option } from "effect"
 import type { WorkersAiBinding } from "../platform/WorkersAiHttpClient"
+import { ReturnedToolFailuresOkLayer } from "../platform/genai-spans"
 import type { ChatSession } from "./ChatSession"
 import type { ChatTurnEvent } from "./events"
 import { withToolTranscript } from "./close-out"
@@ -252,7 +253,7 @@ export const runChatSessionTurn = async (input: RunChatSessionTurnInput): Promis
 			Layer.provideMerge(layerPg),
 			Layer.provideMerge(mapleDbConnectionLayer(input.env)),
 			Layer.provideMerge(workerEnvLayer(input.env)),
-			Layer.provideMerge(telemetry.layer),
+			Layer.provideMerge(ReturnedToolFailuresOkLayer.pipe(Layer.provideMerge(telemetry.layer))),
 		),
 	)
 

apps/ai/src/mcp/tools/run-sql.ts

@@ -109,7 +109,8 @@ const runSqlDescription =
 	`at most ${MAX_RAW_SQL_RESULT_ROWS} are fetched and the first ${MAX_RENDERED_ROWS} rendered, with the true count reported. ` +
 	"Use it to answer what query_data cannot express, to spot-check data, or to test SQL before saving it " +
 	"as a raw_sql widget with add_dashboard_widget. For trends and top-N prefer query_data. " +
-	"describe_warehouse_tables gives table and column names."
+	"Spans are in `traces` (there is no otel_spans, otel_traces or spans table) and log lines in `logs`; " +
+	"describe_warehouse_tables gives every table's columns."
 
 /** A warehouse cell as JSON: 64-bit ints can arrive as bigint, and nothing else is JSON-unsafe. */
 const toJson = (value: unknown): Schema.Json => {

apps/ai/src/platform/genai-spans.test.ts

@@ -3,10 +3,16 @@
  * and tool results run to 50k, so these paths run in production far more often than the happy one.
  */
 import { assert, describe, it } from "@effect/vitest"
-import { Effect, Schema } from "effect"
+import { Effect, Exit, Schema } from "effect"
 import { Tool } from "effect/ai"
 import { makeRecordingTracer } from "@maple/backend/testing/recording-tracer"
-import { invokeAgentAttributes, messagesJson, toolCallJson, withToolCallContent } from "./genai-spans"
+import {
+	invokeAgentAttributes,
+	messagesJson,
+	toolCallJson,
+	withReturnedToolFailuresOk,
+	withToolCallContent,
+} from "./genai-spans"
 
 const TRUNCATION_MARKER = "…[truncated]"
 
@@ -175,3 +181,44 @@ describe("withToolCallContent", () => {
 		}),
 	)
 })
+
+describe("withReturnedToolFailuresOk", () => {
+	// The engine writes `failure_handling` on the span before it ends it with a failed exit.
+	const endedSpan = (name: string, failureHandling: string) =>
+		Effect.gen(function* () {
+			const { spans, tracer } = makeRecordingTracer()
+			yield* Effect.annotateCurrentSpan({ "effect_agent.tool.failure_handling": failureHandling }).pipe(
+				Effect.andThen(Effect.fail("bad sql")),
+				Effect.withSpan(name),
+				Effect.withTracer(withReturnedToolFailuresOk(tracer)),
+				Effect.exit,
+			)
+			const span = spans[0]
+			assert.isDefined(span)
+			assert.strictEqual(span.status._tag, "Ended")
+			return { span, failed: span.status._tag === "Ended" && Exit.isFailure(span.status.exit) }
+		})
+
+	it.effect("ends a failure returned to the model as Ok, with error.type", () =>
+		Effect.gen(function* () {
+			const { span, failed } = yield* endedSpan("execute_tool run_sql", "returned-to-model")
+			assert.isFalse(failed)
+			assert.strictEqual(span.attributes.get("error.type"), "ToolCallFailed")
+		}),
+	)
+
+	it.effect("keeps a propagated failure as Error", () =>
+		Effect.gen(function* () {
+			const { span, failed } = yield* endedSpan("execute_tool run_sql", "propagated")
+			assert.isTrue(failed)
+			assert.isUndefined(span.attributes.get("error.type"))
+		}),
+	)
+
+	it.effect("leaves spans other than execute_tool alone", () =>
+		Effect.gen(function* () {
+			const { failed } = yield* endedSpan("chat.turn", "returned-to-model")
+			assert.isTrue(failed)
+		}),
+	)
+})

apps/ai/src/platform/genai-spans.ts

@@ -18,8 +18,7 @@ import {
 	MAPLE_GENAI_INPUT_MESSAGES_DROPPED_ATTR,
 	MAPLE_GENAI_MODEL_DURATION_MS_ATTR,
 } from "@maple/domain/gen-ai"
-import { Effect, Option, Predicate, Stream } from "effect"
-import type { Tracer } from "effect"
+import { Effect, Exit, Layer, Option, Predicate, Stream, Tracer } from "effect"
 import * as AiError from "effect/ai/AiError"
 import type * as LanguageModel from "effect/ai/LanguageModel"
 import type * as Prompt from "effect/ai/Prompt"
@@ -525,6 +524,41 @@ const executeToolSpan = (span: Tracer.Span): Option.Option<Tracer.Span> =>
 		? Option.some(span)
 		: Option.filter(span.parent, isExecuteToolSpan)
 
+/** A tool failure handed back to the model as the call's result: it retries, the run carries on. */
+const isReturnedToolFailure = (span: Tracer.Span, exit: Exit.Exit<unknown, unknown>): boolean =>
+	Exit.isFailure(exit) &&
+	span.name.startsWith(EXECUTE_TOOL_SPAN_PREFIX) &&
+	span.attributes.get("effect_agent.tool.failure_handling") === "returned-to-model"
+
+/**
+ * Export a returned tool failure as `Ok` + `error.type` rather than `Error`, so it stops opening error
+ * issues while Agent Sessions still counts it failed (the gateway stamps `maple_ai.error` from
+ * `error.type`). Propagated failures keep their Error span.
+ */
+export const withReturnedToolFailuresOk = (tracer: Tracer.Tracer): Tracer.Tracer =>
+	Tracer.make({
+		...(tracer.context === undefined ? undefined : { context: tracer.context }),
+		span(options) {
+			const span = tracer.span(options)
+			if (!options.name.startsWith(EXECUTE_TOOL_SPAN_PREFIX)) return span
+			const end = span.end.bind(span)
+			return Object.assign(span, {
+				end: (endTime: bigint, exit: Exit.Exit<unknown, unknown>) => {
+					if (!isReturnedToolFailure(span, exit)) return end(endTime, exit)
+					if (span.attributes.get("error.type") === undefined)
+						span.attribute("error.type", "ToolCallFailed")
+					end(endTime, Exit.void)
+				},
+			})
+		},
+	})
+
+/** Wraps the current tracer; put it over the telemetry layer with `Layer.provideMerge`. */
+export const ReturnedToolFailuresOkLayer = Layer.effect(
+	Tracer.Tracer,
+	Effect.map(Effect.tracer, withReturnedToolFailuresOk),
+)
+
 /**
  * Untraced on purpose: a span of its own would become the current span, one level further from the
  * `execute_tool` span this looks for.

apps/cli/test/native-archive-crash-recovery-probe.sh

@@ -25,13 +25,13 @@ PORT="${2:-45291}"
 REPO="$(cd "$(dirname "$0")/../../.." && pwd)"
 WORKER="$REPO/apps/cli/test/probes/archive-crash-worker.ts"
 RECONCILE="$REPO/apps/cli/test/probes/archive-reconcile-worker.ts"
-# The bundled `maple` binary bakes __CHDB_VERSION__=v26.1.0 via bun --define at
+# The bundled `maple` binary bakes __CHDB_VERSION__=v26.7.3 via bun --define at
 # compile time; running the workers from SOURCE defaults CHDB_VERSION to "dev",
 # which makes the restore version-check reject a checkpoint made by the bundle.
 # Define it to match so the source-tree workers are version-consistent with the
 # bundle's checkpoints. (The deployed binary is consistent by construction.)
 CHDB_VER="$("$MAPLE" --version 2>/dev/null | grep -oE 'chdb v[^ ]+' | sed 's/chdb //')"
-[ -z "$CHDB_VER" ] && CHDB_VER="v26.1.0"
+[ -z "$CHDB_VER" ] && CHDB_VER="v26.7.3"
 BUN=(bun --define "__CHDB_VERSION__=\"${CHDB_VER}\"")
 
 command -v duckdb >/dev/null 2>&1 || { echo "FAIL: duckdb required" >&2; exit 1; }

apps/cli/test/native-archive-gc-probe.sh

@@ -39,7 +39,7 @@ command -v duckdb >/dev/null 2>&1 || { echo "FAIL: duckdb required" >&2; exit 1;
 [ -f "$LIBCHDB" ] || { echo "FAIL: libchdb not found at $LIBCHDB" >&2; exit 1; }
 
 CHDB_VER="$("$MAPLE" --version 2>/dev/null | grep -oE 'chdb v[^ ]+' | sed 's/chdb //')"
-[ -z "$CHDB_VER" ] && CHDB_VER="v26.1.0"
+[ -z "$CHDB_VER" ] && CHDB_VER="v26.7.3"
 BUN=(bun --define "__CHDB_VERSION__=\"${CHDB_VER}\"")
 
 pass=0

apps/cli/test/native-local-store-migration.sh

@@ -123,7 +123,7 @@ bounded 60 "legacy source fixture" \
 # compatibility evidence the v0 resolver is allowed to use.
 step "installing legacy marker"
 CHDB_VERSION="$(bounded 60 "maple --version" env MAPLE_LIBCHDB="$LIBCHDB" "$MAPLE" --version 2>/dev/null | sed -n 's/.*chdb \([^ ]*\).*/\1/p')"
-[[ -n "$CHDB_VERSION" ]] || CHDB_VERSION="v26.1.0"
+[[ -n "$CHDB_VERSION" ]] || CHDB_VERSION="v26.7.3"
 printf '%s\n' "{\"chdb\":\"$CHDB_VERSION\",\"maple\":\"native-probe\",\"createdAt\":\"unknown\",\"schema\":\"428701854f9fd30e\"}" >"$ROOT/maple-store-version.json"
 chmod 600 "$ROOT/maple-store-version.json"
 

scripts/build-local-binary.sh

@@ -10,7 +10,7 @@
 #   4. Compile apps/cli (the CLI + the OTLP-ingest/query server) into a single
 #      executable with `bun build --compile`. The schema artifacts and SPA are
 #      embedded; the OTLP encoders run in-process; chDB is reached via bun:ffi.
-#   5. Download libchdb (v26.1.0, matching what we test against) for the host
+#   5. Download libchdb (v26.7.3, matching what we test against) for the host
 #      platform and place it beside the binary. At runtime `maple` dlopens the
 #      sibling libchdb (resolved relative to its own path) — no rpath tricks.
 #   6. Restore the committed ui-embed.gen.ts stub so the tree stays clean.
@@ -25,7 +25,7 @@ set -euo pipefail
 
 REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 OUT_DIR="${OUT_DIR:-$REPO_ROOT/dist}"
-LIBCHDB_VERSION="${LIBCHDB_VERSION:-v26.1.0}"
+LIBCHDB_VERSION="${LIBCHDB_VERSION:-v26.7.3}"
 UI_EMBED="$REPO_ROOT/apps/cli/src/server/ui-embed.gen.ts"
 
 # Version baked into the binary via `bun build --define`. The release workflow
@@ -62,16 +62,19 @@ case "$(uname -s)-$(uname -m)" in
 	Darwin-arm64)        ASSET="macos-arm64-libchdb.tar.gz" ;;
 	*) echo "ERROR: unsupported platform $(uname -s)-$(uname -m)" >&2; exit 1 ;;
 esac
-# Expected sha256 per platform asset, pinned to LIBCHDB_VERSION: release assets
-# are mutable even for a fixed tag, so this is verified before extraction.
-# Bumping LIBCHDB_VERSION means re-hashing each asset here. A `case` rather than
-# an associative array because macOS still ships bash 3.2.
-case "$ASSET" in
-	linux-x86_64-libchdb.tar.gz)  EXPECTED_SHA256="3f192cb85cbdb6153622db8ad6f6a25d7163bbda57f283e5ba8fd22cadd0da81" ;;
-	linux-aarch64-libchdb.tar.gz) EXPECTED_SHA256="08ee80488822dfa04776841bb5a4b7ddce35e161ef579b5615b80a88b177428d" ;;
-	macos-x86_64-libchdb.tar.gz)  EXPECTED_SHA256="425db8a197288e1826b60958b3e233d0a7a5b4b46b8065c8bb7489366dbaf47f" ;;
-	macos-arm64-libchdb.tar.gz)   EXPECTED_SHA256="78cfd42203cd33cb0304952f4a6c39921fd9a583c614c2eef96ef6060e9677eb" ;;
-	*) echo "ERROR: no pinned sha256 for $ASSET (LIBCHDB_VERSION=$LIBCHDB_VERSION)" >&2; exit 1 ;;
+# Expected sha256 per (LIBCHDB_VERSION, platform asset): release assets are
+# mutable even for a fixed tag, so this is verified before extraction. Bumping
+# LIBCHDB_VERSION means adding each asset's hash here. A `case` rather than an
+# associative array because macOS still ships bash 3.2.
+# v26.1.0's macos-arm64 dylib put its LINKEDIT string pool at a non-8-byte-aligned
+# offset, which newer dyld rejects ("mis-aligned LINKEDIT string pool"). Check
+# `otool -l libchdb.so` (LC_SYMTAB stroff % 8 == 0) before pinning a new release.
+case "$LIBCHDB_VERSION/$ASSET" in
+	v26.7.3/linux-x86_64-libchdb.tar.gz)  EXPECTED_SHA256="bc33260c32acf78eade2ac41a9115f38e00404651fa42e3bb4c419e4f011c031" ;;
+	v26.7.3/linux-aarch64-libchdb.tar.gz) EXPECTED_SHA256="d153adad1ff39b2e3caf0417f09d8bd9edd41939c7c67a3c4978a61e73fb9227" ;;
+	v26.7.3/macos-x86_64-libchdb.tar.gz)  EXPECTED_SHA256="af5ded3ed3e84c31af1cd198dcf459f11d2b6aad4f6ddeccc04b8a519b0300fc" ;;
+	v26.7.3/macos-arm64-libchdb.tar.gz)   EXPECTED_SHA256="5640e50dccf711bf3dd5551333d08e43f433edf7bd94b2289f36c2539e627762" ;;
+	*) echo "ERROR: no pinned sha256 for $ASSET at LIBCHDB_VERSION=$LIBCHDB_VERSION" >&2; exit 1 ;;
 esac
 
 URL="https://github.com/chdb-io/chdb-core/releases/download/$LIBCHDB_VERSION/$ASSET"