@oh-my-pi/pi-agent-core 18.3.5 → 18.4.1

package/CHANGELOG.md

@@ -2,6 +2,20 @@
 
 ## [Unreleased]
 
+## [18.4.1] - 2026-09-28
+
+### Fixed
+
+- Fixed fitted output caps overshooting the context window by a few tokens on strict Chat Completions hosts (e.g. llama.cpp), causing 400s.
+- Fixed native remote compaction sending requests already estimated past the model's context window (e.g. after re-expanding history behind another provider's native boundary); it now fails fast so the next configured compaction method runs ([#13502](https://github.com/can1357/oh-my-pi/issues/13502))
+- Fixed V2 remote compaction retrying a standalone stream `error` event three times and reporting it as `stream closed before response.completed`; the upstream status, code, and message (e.g. `context_too_large`) are now surfaced ([#13502](https://github.com/can1357/oh-my-pi/issues/13502))
+
+## [18.4.0] - 2026-09-28
+
+### Changed
+
+- Updated telemetry attribute names from the `pi.*` namespace to the `omp.*` namespace.
+
 ## [18.3.3] - 2026-09-27
 
 ### Added

package/README.md

@@ -462,7 +462,7 @@ const runCoverage = aggregateAgentRunCoverage(coverages);
 
 ### Tool status reporting
 
-`execute_tool` spans carry `pi.gen_ai.tool.status` ∈
+`execute_tool` spans carry `omp.gen_ai.tool.status` ∈
 `"ok" | "error" | "skipped" | "blocked" | "timeout" | "aborted"`.
 `beforeToolCall` blocks throw a distinguishable `ToolCallBlockedError`
 internally; the catch path reports `status: "blocked"` instead of conflating

package/dist/types/compaction/branch-summarization.d.ts

@@ -55,7 +55,7 @@ export interface GenerateBranchSummaryOptions {
     convertToLlm?: ConvertToLlm;
     /**
      * Optional telemetry handle. When provided, the branch summary LLM call is
-     * wrapped in an OTEL chat span tagged with `pi.gen_ai.oneshot.kind = "branch_summary"`.
+     * wrapped in an OTEL chat span tagged with `omp.gen_ai.oneshot.kind = "branch_summary"`.
      */
     telemetry?: AgentTelemetry;
     /**

package/dist/types/compaction/compaction.d.ts

@@ -180,7 +180,7 @@ export interface SummaryOptions {
     /**
      * Optional telemetry handle. When provided, every LLM call emitted during
      * compaction is wrapped in an OTEL chat span tagged with
-     * `pi.gen_ai.oneshot.kind` (`compaction_summary`, `compaction_short_summary`,
+     * `omp.gen_ai.oneshot.kind` (`compaction_summary`, `compaction_short_summary`,
      * or `compaction_turn_prefix`). `undefined` keeps the call paths zero-cost.
      */
     telemetry?: AgentTelemetry;
@@ -243,7 +243,7 @@ export interface HandoffOptions {
     metadata?: Record<string, unknown>;
     /**
      * Optional telemetry handle. When provided, the handoff LLM call is
-     * wrapped in an OTEL chat span tagged with `pi.gen_ai.oneshot.kind = "handoff"`.
+     * wrapped in an OTEL chat span tagged with `omp.gen_ai.oneshot.kind = "handoff"`.
      */
     telemetry?: AgentTelemetry;
     /**

package/dist/types/compaction/openai.d.ts

@@ -33,6 +33,8 @@ export interface TrimRemoteCompactionInputResult {
     rewrittenOutputs: number;
     estimatedTokensBefore: number;
     estimatedTokensAfter: number;
+    /** Whether `input` fits the model window; false means it must not be sent. */
+    fits: boolean;
 }
 /**
  * Preserve the full native transcript unless trailing tool outputs alone push a
@@ -41,6 +43,15 @@ export interface TrimRemoteCompactionInputResult {
  * matching Codex's recovery path for oversized tool turns.
  */
 export declare function trimRemoteCompactionInputToContextWindow(input: Array<Record<string, unknown>>, tokenizer: Tokenizer, contextWindow: number | null | undefined, instructions: string, tools?: unknown[]): TrimRemoteCompactionInputResult;
+/**
+ * Refuse a native compaction request whose prepared input cannot fit the model
+ * window, before any network I/O. Re-expanded history behind an unreadable
+ * native boundary can exceed the window even when live context does not.
+ *
+ * @throws Error flagged `ContextOverflow` when `trimmed.fits` is false, so
+ *   compaction callers skip retries and advance to the next method.
+ */
+export declare function assertRemoteCompactionInputFits(trimmed: TrimRemoteCompactionInputResult, model: Model): void;
 export type OpenAiRemoteCompactionItem = {
     type: "compaction" | "compaction_summary";
     encrypted_content?: string;

package/dist/types/output-budget.d.ts

@@ -2,6 +2,16 @@ import type { Context, Model } from "@oh-my-pi/pi-ai";
 import type { Tokenizer } from "./tokenizer.js";
 /** Smallest output cap {@link fitOutputTokensToContextWindow} will request. */
 export declare const MIN_FITTED_OUTPUT_TOKENS = 1024;
+/**
+ * Absolute headway subtracted from the remaining room (unconditionally, on anchored and
+ * fully-local counts alike). Proportional padding covers
+ * tokenizer drift that scales with the prompt, but a host can still count a few tokens more
+ * than any local estimate can see (chat-template framing, reasoning wrappers). Measured
+ * against a 262,144-token host: fitted caps landed +11 to +42 tokens over and were rejected
+ * with 400s. 64 tokens covers the observed drift with margin to spare; the
+ * {@link MIN_FITTED_OUTPUT_TOKENS} floor still applies.
+ */
+export declare const OUTPUT_FIT_HEADWAY_TOKENS = 64;
 /**
  * Output cap for a request, so prompt plus output stays inside the model's
  * context window.
@@ -26,7 +36,8 @@ export declare const MIN_FITTED_OUTPUT_TOKENS = 1024;
  * OpenRouter-hosted model with no caller cap: the transport omits the catalog
  * default there so each upstream self-caps, and a fitted value would turn into
  * an explicit cap that filters upstreams). Otherwise returns
- * the remaining room (never below {@link MIN_FITTED_OUTPUT_TOKENS}); a
+ * the remaining room minus {@link OUTPUT_FIT_HEADWAY_TOKENS} (never below
+ * {@link MIN_FITTED_OUTPUT_TOKENS}); a
  * prompt that fills the whole window still overflows and is left to the
  * caller's compaction. Near a full window the floor means a turn can stop on
  * `length` instead of failing with a 400.

package/dist/types/telemetry.d.ts

@@ -2,7 +2,7 @@
  * OpenTelemetry instrumentation for the agent loop.
  *
  * Implements the OpenTelemetry GenAI semantic conventions
- * (https://opentelemetry.io/docs/specs/semconv/gen-ai/) plus `pi.gen_ai.*`
+ * (https://opentelemetry.io/docs/specs/semconv/gen-ai/) plus `omp.gen_ai.*`
  * extension attributes for run summaries, dashboard summaries, and cost hints
  * that are useful to downstream observability UIs.
  *
@@ -78,35 +78,41 @@ export declare const enum OpenAIAttr {
     ResponseServiceTier = "openai.response.service_tier"
 }
 /** Project extension attributes. Kept out of the reserved `gen_ai.*` namespace. */
-export declare const enum PiGenAIAttr {
-    AgentStepNumber = "pi.gen_ai.agent.step.number",
-    AgentStepCount = "pi.gen_ai.agent.step.count",
-    RequestReasoningEffort = "pi.gen_ai.request.reasoning.effort",
-    RequestToolChoice = "pi.gen_ai.request.tool.choice",
-    RequestAvailableTools = "pi.gen_ai.request.available_tools",
-    RequestMessages = "pi.gen_ai.request.messages",
-    ResponseText = "pi.gen_ai.response.text",
-    ResponseToolCalls = "pi.gen_ai.response.tool_calls",
-    ResponseUpstreamProvider = "pi.gen_ai.response.upstream_provider",
-    UsageTotalTokens = "pi.gen_ai.usage.total_tokens",
-    UsageServerSideTools = "pi.gen_ai.usage.server_tool_requests",
-    CostEstimatedUsd = "pi.gen_ai.cost.estimated_usd",
-    CostInputUsd = "pi.gen_ai.cost.input_usd",
-    CostOutputUsd = "pi.gen_ai.cost.output_usd",
-    CostUnavailableReason = "pi.gen_ai.cost.unavailable_reason",
-    ToolStatus = "pi.gen_ai.tool.status",
-    ToolCallIntent = "pi.gen_ai.tool.call.intent",
-    HandoffFromAgentName = "pi.gen_ai.handoff.from_agent.name",
-    HandoffFromAgentId = "pi.gen_ai.handoff.from_agent.id",
-    HandoffToAgentName = "pi.gen_ai.handoff.to_agent.name",
-    HandoffToAgentId = "pi.gen_ai.handoff.to_agent.id",
-    OneshotKind = "pi.gen_ai.oneshot.kind",
-    GatewayName = "pi.gen_ai.gateway.name",
-    GatewayEndpoint = "pi.gen_ai.gateway.endpoint",
-    GatewayCallId = "pi.gen_ai.gateway.call_id",
-    GatewayRoutedTo = "pi.gen_ai.gateway.routed_to",
+export declare const enum OmpGenAIAttr {
+    AgentStepNumber = "omp.gen_ai.agent.step.number",
+    AgentStepCount = "omp.gen_ai.agent.step.count",
+    RequestReasoningEffort = "omp.gen_ai.request.reasoning.effort",
+    RequestToolChoice = "omp.gen_ai.request.tool.choice",
+    RequestAvailableTools = "omp.gen_ai.request.available_tools",
+    RequestMessages = "omp.gen_ai.request.messages",
+    ResponseText = "omp.gen_ai.response.text",
+    ResponseToolCalls = "omp.gen_ai.response.tool_calls",
+    ResponseUpstreamProvider = "omp.gen_ai.response.upstream_provider",
+    UsageTotalTokens = "omp.gen_ai.usage.total_tokens",
+    UsageServerSideTools = "omp.gen_ai.usage.server_tool_requests",
+    CostEstimatedUsd = "omp.gen_ai.cost.estimated_usd",
+    CostInputUsd = "omp.gen_ai.cost.input_usd",
+    CostOutputUsd = "omp.gen_ai.cost.output_usd",
+    CostUnavailableReason = "omp.gen_ai.cost.unavailable_reason",
+    ToolStatus = "omp.gen_ai.tool.status",
+    ToolCallIntent = "omp.gen_ai.tool.call.intent",
+    HandoffFromAgentName = "omp.gen_ai.handoff.from_agent.name",
+    HandoffFromAgentId = "omp.gen_ai.handoff.from_agent.id",
+    HandoffToAgentName = "omp.gen_ai.handoff.to_agent.name",
+    HandoffToAgentId = "omp.gen_ai.handoff.to_agent.id",
+    OneshotKind = "omp.gen_ai.oneshot.kind",
+    GatewayName = "omp.gen_ai.gateway.name",
+    GatewayEndpoint = "omp.gen_ai.gateway.endpoint",
+    GatewayCallId = "omp.gen_ai.gateway.call_id",
+    GatewayRoutedTo = "omp.gen_ai.gateway.routed_to",
     /** Cloudflare AI Gateway response-cache status (`cf-aig-cache-status`), never prompt-cache. */
-    GatewayResponseCacheStatus = "pi.gen_ai.gateway.response_cache.status"
+    GatewayResponseCacheStatus = "omp.gen_ai.gateway.response_cache.status",
+    /** Caller-level reason a judgment ran (`find`, `ttsr`, `judge_batch`, …). */
+    JudgmentPurpose = "omp.gen_ai.judgment.purpose",
+    /** Questions the caller asked in one judgment request. */
+    JudgmentQuestions = "omp.gen_ai.judgment.questions",
+    /** Questions answered from the local judgment cache instead of the provider. */
+    JudgmentCachedQuestions = "omp.gen_ai.judgment.cached_questions"
 }
 /** GenAI operation names — values for {@link GenAIAttr.OperationName}. */
 export declare const GenAIOperation: {
@@ -114,6 +120,8 @@ export declare const GenAIOperation: {
     readonly ExecuteTool: "execute_tool";
     readonly InvokeAgent: "invoke_agent";
     readonly Handoff: "handoff";
+    /** Typed judgment over a state (System One or a prompted judge model); a project extension value. */
+    readonly Judgment: "judgment";
     readonly GenerateContent: "generate_content";
     readonly TextCompletion: "text_completion";
     readonly CreateAgent: "create_agent";
@@ -121,7 +129,7 @@ export declare const GenAIOperation: {
 };
 export type GenAIOperationName = (typeof GenAIOperation)[keyof typeof GenAIOperation];
 /** Identifies which agent span a callback is reporting on. */
-export type TelemetrySpanKind = "invoke_agent" | "chat" | "execute_tool" | "handoff";
+export type TelemetrySpanKind = "invoke_agent" | "chat" | "execute_tool" | "handoff" | "judgment";
 /**
  * Aggregated usage + cost surface passed to {@link AgentTelemetryConfig.costEstimator}.
  * Mirrors the bucketed shape we already emit as span attributes so the
@@ -144,9 +152,9 @@ export interface CostEstimatorContext {
 }
 /**
  * Cost estimator result.
- *   { usd: number }                — cost is known; emitted as pi.gen_ai.cost.estimated_usd
+ *   { usd: number }                — cost is known; emitted as omp.gen_ai.cost.estimated_usd
  *   { unavailable: string }        — cost is intentionally unknown; emitted as
- *                                    pi.gen_ai.cost.unavailable_reason
+ *                                    omp.gen_ai.cost.unavailable_reason
  *   undefined                      — no opinion; nothing emitted
  */
 export type CostEstimate = {
@@ -177,6 +185,8 @@ export interface CostDelta {
  */
 export interface ChatUsageEvent {
     readonly span: Span;
+    /** Operation that produced the usage: `chat` for model turns, `judgment` for typed judgments. */
+    readonly operation: GenAIOperationName;
     readonly agent: AgentIdentity | undefined;
     readonly conversationId: string | undefined;
     readonly stepNumber: number | undefined;
@@ -195,7 +205,7 @@ export interface ChatUsageEvent {
      *
      * Use this to reconcile gateway-issued ids (e.g. `x-litellm-call-id`) with
      * downstream billing / spend dashboards. Known gateway patterns are also
-     * auto-stamped on the chat span as `pi.gen_ai.gateway.*` attributes.
+     * auto-stamped on the chat span as `omp.gen_ai.gateway.*` attributes.
      */
     readonly headers: Readonly<Record<string, string>> | undefined;
 }
@@ -417,7 +427,7 @@ export interface GatewayHeaderDetection {
 export declare function detectGatewayFromHeaders(headers: Readonly<Record<string, string>> | undefined): GatewayHeaderDetection | undefined;
 /**
  * Bounded Cloudflare AI Gateway response-cache statuses emitted on
- * {@link PiGenAIAttr.GatewayResponseCacheStatus}. Distinct from provider
+ * {@link OmpGenAIAttr.GatewayResponseCacheStatus}. Distinct from provider
  * prompt-cache token counters (`gen_ai.usage.cache_*`).
  *
  * Cloudflare documents `HIT` / `MISS` on `cf-aig-cache-status`; `bypass` covers
@@ -453,6 +463,39 @@ export interface ManualChatTelemetryOptions {
     readonly endSpan?: boolean;
 }
 export declare function recordManualChatTelemetry(telemetry: AgentTelemetry | undefined, options: ManualChatTelemetryOptions): Promise<Span | undefined>;
+/**
+ * One judgment attempt reported to {@link recordJudgmentTelemetry}: a billed
+ * System One request, one prompted-model attempt, or a request answered
+ * entirely from the local judgment cache.
+ */
+export interface JudgmentTelemetryOptions {
+    readonly provider: string;
+    /** Requested model id. */
+    readonly model: string;
+    /** Model that answered, when the provider resolves an alias (`jev-latest` → `jev-1.13.0`). */
+    readonly responseModel?: string;
+    /** Caller-level reason the judgment ran; stamped as {@link OmpGenAIAttr.JudgmentPurpose}. */
+    readonly purpose: string;
+    /** Priced usage of this attempt — the same amount the session ledger bills. */
+    readonly usage: Usage;
+    readonly stopReason: StopReason;
+    readonly errorMessage?: string;
+    /** Epoch ms the attempt started. */
+    readonly startTime: number;
+    /** Questions in the request; omitted for prompted backends, which report per completion attempt. */
+    readonly questions?: number;
+    /** Questions answered from the local cache rather than the provider. */
+    readonly cachedQuestions?: number;
+}
+/**
+ * Emit one `judgment` span per judgment attempt and forward any billed usage
+ * to `onCostDelta` / `onChatUsage` with `operation: "judgment"`. Cost is the
+ * attempt's own priced usage rather than {@link AgentTelemetryConfig.costEstimator},
+ * so spans, metrics, and the session ledger agree on every judgment dollar.
+ * Attempts that billed nothing (cache hits, pre-response failures) keep their
+ * span but fire no usage hooks. No-op when telemetry is disabled.
+ */
+export declare function recordJudgmentTelemetry(telemetry: AgentTelemetry | undefined, options: JudgmentTelemetryOptions): Promise<void>;
 /**
  * Options accepted by {@link instrumentedCompleteSimple}. Mirrors the
  * `streamAssistantResponse` chat-span lifecycle for oneshot LLM calls
@@ -465,7 +508,7 @@ export interface InstrumentedChatSpanOptions {
     /** Step index recorded on the span; defaults to `-1` for non-loop calls. */
     readonly stepNumber?: number;
     /**
-     * Tag stamped onto `pi.gen_ai.oneshot.kind`. Values used by the agent:
+     * Tag stamped onto `omp.gen_ai.oneshot.kind`. Values used by the agent:
      * `compaction_summary`, `compaction_short_summary`, `compaction_turn_prefix`,
      * `handoff`, `branch_summary`, `image_question`. Free-form to allow callers
      * outside this package to add new kinds without bumping the helper.
@@ -535,7 +578,7 @@ export declare function finishExecuteToolSpan(telemetry: AgentTelemetry | undefi
     readonly toolName: string;
 }): void;
 /** Span attribute carrying the terminal {@link ToolStatus}. */
-export declare const EXECUTE_TOOL_STATUS_ATTR = PiGenAIAttr.ToolStatus;
+export declare const EXECUTE_TOOL_STATUS_ATTR = OmpGenAIAttr.ToolStatus;
 /**
  * Record a tool that bypassed the span lifecycle entirely (pre-run
  * interrupt, post-execution tail sweep for calls that never produced a
@@ -570,30 +613,30 @@ export declare function finishInvokeAgentSpan(telemetry: AgentTelemetry | undefi
  * calling this helper.
  */
 export declare function fireOnRunEnd(telemetry: AgentTelemetry, summary: AgentRunSummary, coverage: AgentRunCoverage): void;
-/** Aggregate `pi.gen_ai.agent.*` attributes stamped on the `invoke_agent` span. */
-export declare const enum PiGenAIAggregateAttr {
-    ChatsCount = "pi.gen_ai.agent.chats.count",
-    ChatsTotalLatencyMs = "pi.gen_ai.agent.chats.total_latency_ms",
-    ChatsStopReasonPrefix = "pi.gen_ai.agent.chats.stop_reason.",
-    ToolsCount = "pi.gen_ai.agent.tools.count",
-    ToolsOkCount = "pi.gen_ai.agent.tools.ok.count",
-    ToolsErrorCount = "pi.gen_ai.agent.tools.error.count",
-    ToolsSkippedCount = "pi.gen_ai.agent.tools.skipped.count",
-    ToolsBlockedCount = "pi.gen_ai.agent.tools.blocked.count",
-    ToolsTimeoutCount = "pi.gen_ai.agent.tools.timeout.count",
-    ToolsAbortedCount = "pi.gen_ai.agent.tools.aborted.count",
-    ToolsTotalLatencyMs = "pi.gen_ai.agent.tools.total_latency_ms",
-    ToolsInvoked = "pi.gen_ai.agent.tools.invoked",
-    ToolsAvailable = "pi.gen_ai.agent.tools.available",
-    ToolsUnused = "pi.gen_ai.agent.tools.unused",
-    UsageInputTokensTotal = "pi.gen_ai.agent.usage.input_tokens.total",
-    UsageOutputTokensTotal = "pi.gen_ai.agent.usage.output_tokens.total",
-    UsageCacheReadInputTokensTotal = "pi.gen_ai.agent.usage.cache_read.input_tokens.total",
-    UsageCacheCreationInputTokensTotal = "pi.gen_ai.agent.usage.cache_creation.input_tokens.total",
-    UsageReasoningOutputTokensTotal = "pi.gen_ai.agent.usage.reasoning.output_tokens.total",
-    UsageTotalTokensTotal = "pi.gen_ai.agent.usage.total_tokens.total",
-    CostEstimatedUsdTotal = "pi.gen_ai.agent.cost.estimated_usd.total",
-    ErrorsCount = "pi.gen_ai.agent.errors.count"
+/** Aggregate `omp.gen_ai.agent.*` attributes stamped on the `invoke_agent` span. */
+export declare const enum OmpGenAIAggregateAttr {
+    ChatsCount = "omp.gen_ai.agent.chats.count",
+    ChatsTotalLatencyMs = "omp.gen_ai.agent.chats.total_latency_ms",
+    ChatsStopReasonPrefix = "omp.gen_ai.agent.chats.stop_reason.",
+    ToolsCount = "omp.gen_ai.agent.tools.count",
+    ToolsOkCount = "omp.gen_ai.agent.tools.ok.count",
+    ToolsErrorCount = "omp.gen_ai.agent.tools.error.count",
+    ToolsSkippedCount = "omp.gen_ai.agent.tools.skipped.count",
+    ToolsBlockedCount = "omp.gen_ai.agent.tools.blocked.count",
+    ToolsTimeoutCount = "omp.gen_ai.agent.tools.timeout.count",
+    ToolsAbortedCount = "omp.gen_ai.agent.tools.aborted.count",
+    ToolsTotalLatencyMs = "omp.gen_ai.agent.tools.total_latency_ms",
+    ToolsInvoked = "omp.gen_ai.agent.tools.invoked",
+    ToolsAvailable = "omp.gen_ai.agent.tools.available",
+    ToolsUnused = "omp.gen_ai.agent.tools.unused",
+    UsageInputTokensTotal = "omp.gen_ai.agent.usage.input_tokens.total",
+    UsageOutputTokensTotal = "omp.gen_ai.agent.usage.output_tokens.total",
+    UsageCacheReadInputTokensTotal = "omp.gen_ai.agent.usage.cache_read.input_tokens.total",
+    UsageCacheCreationInputTokensTotal = "omp.gen_ai.agent.usage.cache_creation.input_tokens.total",
+    UsageReasoningOutputTokensTotal = "omp.gen_ai.agent.usage.reasoning.output_tokens.total",
+    UsageTotalTokensTotal = "omp.gen_ai.agent.usage.total_tokens.total",
+    CostEstimatedUsdTotal = "omp.gen_ai.agent.cost.estimated_usd.total",
+    ErrorsCount = "omp.gen_ai.agent.errors.count"
 }
 /**
  * Run `fn` with `span` activated on the OTEL context. Spans created

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-agent-core",
-	"version": "18.3.5",
+	"version": "18.4.1",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"homepage": "https://omp.sh",
 	"author": {
@@ -38,16 +38,16 @@
 		"fmt": "oxfmt --no-error-on-unmatched-pattern 'src/**/*.{ts,tsx}' '{test,bench,examples,scripts}/**/*.ts' '*.ts'"
 	},
 	"dependencies": {
-		"@oh-my-pi/pi-ai": "18.3.5",
-		"@oh-my-pi/pi-catalog": "18.3.5",
-		"@oh-my-pi/pi-natives": "18.3.5",
-		"@oh-my-pi/pi-utils": "18.3.5",
-		"@oh-my-pi/pi-wire": "18.3.5",
-		"@oh-my-pi/snapcompact": "18.3.5",
+		"@oh-my-pi/pi-ai": "18.4.1",
+		"@oh-my-pi/pi-catalog": "18.4.1",
+		"@oh-my-pi/pi-natives": "18.4.1",
+		"@oh-my-pi/pi-utils": "18.4.1",
+		"@oh-my-pi/pi-wire": "18.4.1",
+		"@oh-my-pi/snapcompact": "18.4.1",
 		"@opentelemetry/api": "^1.9.1"
 	},
 	"devDependencies": {
-		"@oh-my-pi/omptype": "18.3.5",
+		"@oh-my-pi/omptype": "18.4.1",
 		"@opentelemetry/context-async-hooks": "^2.9.0",
 		"@opentelemetry/sdk-trace-base": "^2.9.0",
 		"@types/bun": "^1.3.14"

package/src/agent-loop.ts

@@ -64,7 +64,7 @@ import {
 	finishExecuteToolSpan,
 	finishInvokeAgentSpan,
 	fireOnRunEnd,
-	PiGenAIAttr,
+	OmpGenAIAttr,
 	recordSkippedTool,
 	resolveTelemetry,
 	runInActiveSpan,
@@ -2859,6 +2859,13 @@ async function prepareToolCallDispatch(
 		if (toolCall.type !== "toolCall") continue;
 		if ((toolCall as CursorExecResolvedCarrier)[kCursorExecResolved] === true) continue;
 		const tool = resolveToolForCall(context.tools, toolCall, resolveFallbackTool);
+		// A host fallback accepts aliases (`xd://recall`, a mis-separated MCP
+		// name) that providers reject when replayed as a function-call name.
+		// Record the call under the resolved tool's canonical name so history,
+		// persistence, and replay agree; custom-wire calls keep their wire name.
+		if (tool && toolCall.name !== tool.name && toolCall.name !== tool.customWireName) {
+			toolCall.name = tool.name;
+		}
 		const entry: PreparedToolCall = { tool, args: toolCall.arguments as Record<string, unknown> };
 		prepared.set(toolCall.id, entry);
 		let argsForExecution = toolCall.arguments as Record<string, unknown>;
@@ -3316,7 +3323,7 @@ async function executeToolCalls(
 			parent: invokeAgentSpan,
 		});
 		if (toolSpan && toolCall.intent) {
-			toolSpan.setAttribute(PiGenAIAttr.ToolCallIntent, toolCall.intent);
+			toolSpan.setAttribute(OmpGenAIAttr.ToolCallIntent, toolCall.intent);
 		}
 
 		let result: AgentToolResult<any> = { content: [], details: {} };

package/src/compaction/branch-summarization.ts

@@ -86,7 +86,7 @@ export interface GenerateBranchSummaryOptions {
 	convertToLlm?: ConvertToLlm;
 	/**
 	 * Optional telemetry handle. When provided, the branch summary LLM call is
-	 * wrapped in an OTEL chat span tagged with `pi.gen_ai.oneshot.kind = "branch_summary"`.
+	 * wrapped in an OTEL chat span tagged with `omp.gen_ai.oneshot.kind = "branch_summary"`.
 	 */
 	telemetry?: AgentTelemetry;
 	/**

package/src/compaction/compaction-v2-streaming.ts

@@ -597,6 +597,14 @@ function handleCompactionV2Event(
 	if (type === "response.failed" || type === "response.incomplete") {
 		throw new Error(formatCompactionV2Failure(event, type));
 	}
+
+	// A standalone `error` event terminates the stream. Keep its status so a
+	// deterministic 4xx (e.g. context_too_large) is not retried as a dropped stream.
+	if (type === "error") {
+		const message = formatCompactionV2Failure(event, type);
+		const status = numberField(event, "status");
+		throw status === undefined ? new Error(message) : new AIError.ProviderHttpError(message, status);
+	}
 }
 
 function parseCompactionV2Usage(event: Record<string, unknown>): CompactionV2Usage | undefined {
@@ -629,8 +637,9 @@ function formatCompactionV2Failure(event: Record<string, unknown>, type: string)
 		: response && isRecord(response.error)
 			? response.error
 			: undefined;
-	const message = error ? stringField(error, "message") : undefined;
-	const code = error ? (stringField(error, "code") ?? stringField(error, "type")) : undefined;
+	// Responses `error` events carry code/message at the top level.
+	const message = stringField(error ?? event, "message");
+	const code = error ? (stringField(error, "code") ?? stringField(error, "type")) : stringField(event, "code");
 	return `V2 compaction stream ${type}${code ? ` (${code})` : ""}${message ? `: ${message}` : ""}`;
 }
 

package/src/compaction/compaction.ts

@@ -69,6 +69,7 @@ import {
 	defaultConvertToLlm,
 } from "./messages";
 import {
+	assertRemoteCompactionInputFits,
 	buildOpenAiNativeHistory,
 	getPreservedOpenAiRemoteCompactionData,
 	isOpenAiRemoteCompactionApi,
@@ -678,7 +679,7 @@ export interface SummaryOptions {
 	/**
 	 * Optional telemetry handle. When provided, every LLM call emitted during
 	 * compaction is wrapped in an OTEL chat span tagged with
-	 * `pi.gen_ai.oneshot.kind` (`compaction_summary`, `compaction_short_summary`,
+	 * `omp.gen_ai.oneshot.kind` (`compaction_summary`, `compaction_short_summary`,
 	 * or `compaction_turn_prefix`). `undefined` keeps the call paths zero-cost.
 	 */
 	telemetry?: AgentTelemetry;
@@ -1041,7 +1042,7 @@ export interface HandoffOptions {
 	metadata?: Record<string, unknown>;
 	/**
 	 * Optional telemetry handle. When provided, the handoff LLM call is
-	 * wrapped in an OTEL chat span tagged with `pi.gen_ai.oneshot.kind = "handoff"`.
+	 * wrapped in an OTEL chat span tagged with `omp.gen_ai.oneshot.kind = "handoff"`.
 	 */
 	telemetry?: AgentTelemetry;
 	/**
@@ -1748,6 +1749,7 @@ export async function compact(
 						contextWindow: model.contextWindow,
 					});
 				}
+				assertRemoteCompactionInputFits(trimmed, model);
 				const requestOptions = {
 					sessionId: summaryOptions.sessionId,
 					promptCacheKey: summaryOptions.promptCacheKey,

package/src/compaction/openai.ts

@@ -15,7 +15,7 @@
  *   with `{ summary, shortSummary? }`.
  */
 
-import { ProviderHttpError } from "@oh-my-pi/pi-ai/error";
+import { attach, create, Flag, ProviderHttpError } from "@oh-my-pi/pi-ai/error";
 import { getCodexAttestationHeader } from "@oh-my-pi/pi-ai/providers/openai-codex-attestation";
 import { createOpenAICodexCompactionRequestContext } from "@oh-my-pi/pi-ai/providers/openai-codex-compaction";
 import { applyCodexResponsesLiteShape } from "@oh-my-pi/pi-ai/providers/openai-codex/request-transformer";
@@ -122,6 +122,8 @@ export interface TrimRemoteCompactionInputResult {
 	rewrittenOutputs: number;
 	estimatedTokensBefore: number;
 	estimatedTokensAfter: number;
+	/** Whether `input` fits the model window; false means it must not be sent. */
+	fits: boolean;
 }
 
 /** Verdict for one remote-compaction request measured against the model window. */
@@ -199,6 +201,7 @@ export function trimRemoteCompactionInputToContextWindow(
 			rewrittenOutputs: 0,
 			estimatedTokensBefore: before.tokens,
 			estimatedTokensAfter: before.tokens,
+			fits: true,
 		};
 	}
 
@@ -222,6 +225,7 @@ export function trimRemoteCompactionInputToContextWindow(
 			rewrittenOutputs: 0,
 			estimatedTokensBefore: before.tokens,
 			estimatedTokensAfter: before.tokens,
+			fits: false,
 		};
 	}
 
@@ -230,9 +234,29 @@ export function trimRemoteCompactionInputToContextWindow(
 		rewrittenOutputs,
 		estimatedTokensBefore: before.tokens,
 		estimatedTokensAfter: after.tokens,
+		fits: true,
 	};
 }
 
+/**
+ * Refuse a native compaction request whose prepared input cannot fit the model
+ * window, before any network I/O. Re-expanded history behind an unreadable
+ * native boundary can exceed the window even when live context does not.
+ *
+ * @throws Error flagged `ContextOverflow` when `trimmed.fits` is false, so
+ *   compaction callers skip retries and advance to the next method.
+ */
+export function assertRemoteCompactionInputFits(trimmed: TrimRemoteCompactionInputResult, model: Model): void {
+	if (trimmed.fits) return;
+	throw attach(
+		new Error(
+			`Remote compaction input exceeds the context window of ${model.provider}/${model.id}: ` +
+				`estimated ${trimmed.estimatedTokensAfter} tokens > ${model.contextWindow}`,
+		),
+		create(Flag.ContextOverflow),
+	);
+}
+
 /** Race the caller's signal against the request timeout; `timeoutMs <= 0` disables the watchdog. */
 function withRequestTimeout(signal: AbortSignal | undefined, timeoutMs: number): AbortSignal | undefined {
 	if (timeoutMs <= 0) return signal;
@@ -794,6 +818,7 @@ export async function requestOpenAiRemoteCompaction(
 			contextWindow: model.contextWindow,
 		});
 	}
+	assertRemoteCompactionInputFits(trimmed, model);
 	const request: OpenAiRemoteCompactionRequest = {
 		model: requestModel,
 		// Preserve the native transcript. Only oversized trailing tool outputs are

package/src/output-budget.ts

@@ -17,6 +17,17 @@ export const MIN_FITTED_OUTPUT_TOKENS = 1024;
  */
 const PROMPT_ESTIMATE_MARGIN_DIVISOR = 10;
 
+/**
+ * Absolute headway subtracted from the remaining room (unconditionally, on anchored and
+ * fully-local counts alike). Proportional padding covers
+ * tokenizer drift that scales with the prompt, but a host can still count a few tokens more
+ * than any local estimate can see (chat-template framing, reasoning wrappers). Measured
+ * against a 262,144-token host: fitted caps landed +11 to +42 tokens over and were rejected
+ * with 400s. 64 tokens covers the observed drift with margin to spare; the
+ * {@link MIN_FITTED_OUTPUT_TOKENS} floor still applies.
+ */
+export const OUTPUT_FIT_HEADWAY_TOKENS = 64;
+
 /**
  * Output cap for a request, so prompt plus output stays inside the model's
  * context window.
@@ -41,7 +52,8 @@ const PROMPT_ESTIMATE_MARGIN_DIVISOR = 10;
  * OpenRouter-hosted model with no caller cap: the transport omits the catalog
  * default there so each upstream self-caps, and a fitted value would turn into
  * an explicit cap that filters upstreams). Otherwise returns
- * the remaining room (never below {@link MIN_FITTED_OUTPUT_TOKENS}); a
+ * the remaining room minus {@link OUTPUT_FIT_HEADWAY_TOKENS} (never below
+ * {@link MIN_FITTED_OUTPUT_TOKENS}); a
  * prompt that fills the whole window still overflows and is left to the
  * caller's compaction. Near a full window the floor means a turn can stop on
  * `length` instead of failing with a 400.
@@ -67,7 +79,7 @@ export function fitOutputTokensToContextWindow(
 	if (!requested || !contextWindow || contextWindow <= 0) return maxTokens;
 	if (stopsOutputAtContextWindow(model)) return maxTokens;
 
-	const room = contextWindow - countPromptTokens(context, tokenizer);
+	const room = contextWindow - countPromptTokens(context, tokenizer) - OUTPUT_FIT_HEADWAY_TOKENS;
 	if (room >= requested) return maxTokens;
 	return Math.max(MIN_FITTED_OUTPUT_TOKENS, room);
 }

package/src/telemetry.ts

@@ -2,7 +2,7 @@
  * OpenTelemetry instrumentation for the agent loop.
  *
  * Implements the OpenTelemetry GenAI semantic conventions
- * (https://opentelemetry.io/docs/specs/semconv/gen-ai/) plus `pi.gen_ai.*`
+ * (https://opentelemetry.io/docs/specs/semconv/gen-ai/) plus `omp.gen_ai.*`
  * extension attributes for run summaries, dashboard summaries, and cost hints
  * that are useful to downstream observability UIs.
  *
@@ -125,42 +125,48 @@ export const enum OpenAIAttr {
 }
 
 /** Project extension attributes. Kept out of the reserved `gen_ai.*` namespace. */
-export const enum PiGenAIAttr {
-	AgentStepNumber = "pi.gen_ai.agent.step.number",
-	AgentStepCount = "pi.gen_ai.agent.step.count",
-	RequestReasoningEffort = "pi.gen_ai.request.reasoning.effort",
-	RequestToolChoice = "pi.gen_ai.request.tool.choice",
-	RequestAvailableTools = "pi.gen_ai.request.available_tools",
-	RequestMessages = "pi.gen_ai.request.messages",
-	ResponseText = "pi.gen_ai.response.text",
-	ResponseToolCalls = "pi.gen_ai.response.tool_calls",
-	ResponseUpstreamProvider = "pi.gen_ai.response.upstream_provider",
-	UsageTotalTokens = "pi.gen_ai.usage.total_tokens",
-	UsageServerSideTools = "pi.gen_ai.usage.server_tool_requests",
-	CostEstimatedUsd = "pi.gen_ai.cost.estimated_usd",
-	CostInputUsd = "pi.gen_ai.cost.input_usd",
-	CostOutputUsd = "pi.gen_ai.cost.output_usd",
-	CostUnavailableReason = "pi.gen_ai.cost.unavailable_reason",
-	ToolStatus = "pi.gen_ai.tool.status",
-	ToolCallIntent = "pi.gen_ai.tool.call.intent",
-	HandoffFromAgentName = "pi.gen_ai.handoff.from_agent.name",
-	HandoffFromAgentId = "pi.gen_ai.handoff.from_agent.id",
-	HandoffToAgentName = "pi.gen_ai.handoff.to_agent.name",
-	HandoffToAgentId = "pi.gen_ai.handoff.to_agent.id",
+export const enum OmpGenAIAttr {
+	AgentStepNumber = "omp.gen_ai.agent.step.number",
+	AgentStepCount = "omp.gen_ai.agent.step.count",
+	RequestReasoningEffort = "omp.gen_ai.request.reasoning.effort",
+	RequestToolChoice = "omp.gen_ai.request.tool.choice",
+	RequestAvailableTools = "omp.gen_ai.request.available_tools",
+	RequestMessages = "omp.gen_ai.request.messages",
+	ResponseText = "omp.gen_ai.response.text",
+	ResponseToolCalls = "omp.gen_ai.response.tool_calls",
+	ResponseUpstreamProvider = "omp.gen_ai.response.upstream_provider",
+	UsageTotalTokens = "omp.gen_ai.usage.total_tokens",
+	UsageServerSideTools = "omp.gen_ai.usage.server_tool_requests",
+	CostEstimatedUsd = "omp.gen_ai.cost.estimated_usd",
+	CostInputUsd = "omp.gen_ai.cost.input_usd",
+	CostOutputUsd = "omp.gen_ai.cost.output_usd",
+	CostUnavailableReason = "omp.gen_ai.cost.unavailable_reason",
+	ToolStatus = "omp.gen_ai.tool.status",
+	ToolCallIntent = "omp.gen_ai.tool.call.intent",
+	HandoffFromAgentName = "omp.gen_ai.handoff.from_agent.name",
+	HandoffFromAgentId = "omp.gen_ai.handoff.from_agent.id",
+	HandoffToAgentName = "omp.gen_ai.handoff.to_agent.name",
+	HandoffToAgentId = "omp.gen_ai.handoff.to_agent.id",
 	// Marks chat spans emitted outside the agent loop (compaction, handoff, branch
 	// summary, image inspection, …). Lets dashboards split oneshot cost / latency
 	// from main-turn cost without overloading the semconv `gen_ai.operation.name`.
-	OneshotKind = "pi.gen_ai.oneshot.kind",
+	OneshotKind = "omp.gen_ai.oneshot.kind",
 	// Gateway / proxy (LiteLLM, Helicone, Portkey, …) — populated when a known
 	// gateway header pattern is detected on the upstream response. The base
 	// `gen_ai.provider.name` continues to track the *upstream* provider (e.g.
 	// `anthropic`) that the gateway routed to.
-	GatewayName = "pi.gen_ai.gateway.name",
-	GatewayEndpoint = "pi.gen_ai.gateway.endpoint",
-	GatewayCallId = "pi.gen_ai.gateway.call_id",
-	GatewayRoutedTo = "pi.gen_ai.gateway.routed_to",
+	GatewayName = "omp.gen_ai.gateway.name",
+	GatewayEndpoint = "omp.gen_ai.gateway.endpoint",
+	GatewayCallId = "omp.gen_ai.gateway.call_id",
+	GatewayRoutedTo = "omp.gen_ai.gateway.routed_to",
 	/** Cloudflare AI Gateway response-cache status (`cf-aig-cache-status`), never prompt-cache. */
-	GatewayResponseCacheStatus = "pi.gen_ai.gateway.response_cache.status",
+	GatewayResponseCacheStatus = "omp.gen_ai.gateway.response_cache.status",
+	/** Caller-level reason a judgment ran (`find`, `ttsr`, `judge_batch`, …). */
+	JudgmentPurpose = "omp.gen_ai.judgment.purpose",
+	/** Questions the caller asked in one judgment request. */
+	JudgmentQuestions = "omp.gen_ai.judgment.questions",
+	/** Questions answered from the local judgment cache instead of the provider. */
+	JudgmentCachedQuestions = "omp.gen_ai.judgment.cached_questions",
 }
 
 /** GenAI operation names — values for {@link GenAIAttr.OperationName}. */
@@ -169,6 +175,8 @@ export const GenAIOperation = {
 	ExecuteTool: "execute_tool",
 	InvokeAgent: "invoke_agent",
 	Handoff: "handoff",
+	/** Typed judgment over a state (System One or a prompted judge model); a project extension value. */
+	Judgment: "judgment",
 	GenerateContent: "generate_content",
 	TextCompletion: "text_completion",
 	CreateAgent: "create_agent",
@@ -178,7 +186,7 @@ export const GenAIOperation = {
 export type GenAIOperationName = (typeof GenAIOperation)[keyof typeof GenAIOperation];
 
 /** Identifies which agent span a callback is reporting on. */
-export type TelemetrySpanKind = "invoke_agent" | "chat" | "execute_tool" | "handoff";
+export type TelemetrySpanKind = "invoke_agent" | "chat" | "execute_tool" | "handoff" | "judgment";
 
 /**
  * Aggregated usage + cost surface passed to {@link AgentTelemetryConfig.costEstimator}.
@@ -204,9 +212,9 @@ export interface CostEstimatorContext {
 
 /**
  * Cost estimator result.
- *   { usd: number }                — cost is known; emitted as pi.gen_ai.cost.estimated_usd
+ *   { usd: number }                — cost is known; emitted as omp.gen_ai.cost.estimated_usd
  *   { unavailable: string }        — cost is intentionally unknown; emitted as
- *                                    pi.gen_ai.cost.unavailable_reason
+ *                                    omp.gen_ai.cost.unavailable_reason
  *   undefined                      — no opinion; nothing emitted
  */
 export type CostEstimate =
@@ -235,6 +243,8 @@ export interface CostDelta {
  */
 export interface ChatUsageEvent {
 	readonly span: Span;
+	/** Operation that produced the usage: `chat` for model turns, `judgment` for typed judgments. */
+	readonly operation: GenAIOperationName;
 	readonly agent: AgentIdentity | undefined;
 	readonly conversationId: string | undefined;
 	readonly stepNumber: number | undefined;
@@ -253,7 +263,7 @@ export interface ChatUsageEvent {
 	 *
 	 * Use this to reconcile gateway-issued ids (e.g. `x-litellm-call-id`) with
 	 * downstream billing / spend dashboards. Known gateway patterns are also
-	 * auto-stamped on the chat span as `pi.gen_ai.gateway.*` attributes.
+	 * auto-stamped on the chat span as `omp.gen_ai.gateway.*` attributes.
 	 */
 	readonly headers: Readonly<Record<string, string>> | undefined;
 }
@@ -475,6 +485,8 @@ function startSpan(
 		readonly stepNumber?: number;
 		readonly toolCallId?: string;
 		readonly toolName?: string;
+		/** Epoch ms; defaults to now. */
+		readonly startTime?: number;
 	},
 ): Span | undefined {
 	if (!telemetry) return undefined;
@@ -497,7 +509,11 @@ function startSpan(
 	if (options.attributes) Object.assign(attrs, options.attributes);
 
 	const ctx = options.parent ? trace.setSpan(context.active(), options.parent) : context.active();
-	const span = telemetry.tracer.startSpan(name, { kind: options.spanKind, attributes: attrs }, ctx);
+	const span = telemetry.tracer.startSpan(
+		name,
+		{ kind: options.spanKind, attributes: attrs, startTime: options.startTime },
+		ctx,
+	);
 	safeOnSpanStart(telemetry, { ...attrCtx, span });
 	return span;
 }
@@ -548,6 +564,8 @@ function kindToOperation(kind: TelemetrySpanKind): GenAIOperationName | undefine
 			return GenAIOperation.ExecuteTool;
 		case "handoff":
 			return GenAIOperation.Handoff;
+		case "judgment":
+			return GenAIOperation.Judgment;
 	}
 }
 
@@ -683,7 +701,7 @@ export function startInvokeAgentSpan(telemetry: AgentTelemetry | undefined, mode
 /** Stamp the final step count on the `invoke_agent` span. */
 export function applyInvokeAgentFinish(span: Span | undefined, stepCount: number): void {
 	if (!span) return;
-	span.setAttribute(PiGenAIAttr.AgentStepCount, stepCount);
+	span.setAttribute(OmpGenAIAttr.AgentStepCount, stepCount);
 }
 
 /**
@@ -740,7 +758,7 @@ export interface ChatRequestSnapshot {
 
 function buildChatRequestAttributes(stepNumber: number, request: ChatRequestSnapshot, provider: string): Attributes {
 	const attrs: Attributes = {
-		[PiGenAIAttr.AgentStepNumber]: stepNumber,
+		[OmpGenAIAttr.AgentStepNumber]: stepNumber,
 		[GenAIAttr.OutputType]: "text",
 		[GenAIAttr.RequestStream]: true,
 	};
@@ -757,11 +775,11 @@ function buildChatRequestAttributes(stepNumber: number, request: ChatRequestSnap
 	if (request.serviceTier && shouldSendServiceTier(request.serviceTier, provider)) {
 		attrs[OpenAIAttr.RequestServiceTier] = request.serviceTier;
 	}
-	if (request.reasoningEffort) attrs[PiGenAIAttr.RequestReasoningEffort] = request.reasoningEffort;
+	if (request.reasoningEffort) attrs[OmpGenAIAttr.RequestReasoningEffort] = request.reasoningEffort;
 	const toolChoice = serializeToolChoice(request.toolChoice);
-	if (toolChoice) attrs[PiGenAIAttr.RequestToolChoice] = toolChoice;
+	if (toolChoice) attrs[OmpGenAIAttr.RequestToolChoice] = toolChoice;
 	if (request.tools && request.tools.length > 0) {
-		attrs[PiGenAIAttr.RequestAvailableTools] = request.tools.map(tool => tool.name);
+		attrs[OmpGenAIAttr.RequestAvailableTools] = request.tools.map(tool => tool.name);
 	}
 	return attrs;
 }
@@ -779,7 +797,7 @@ function serializeToolChoice(toolChoice: ToolChoice | undefined): string | undef
 
 function applyContentCaptureForRequest(telemetry: AgentTelemetry, span: Span, request: ChatRequestSnapshot): void {
 	const requestMessages = serializeRequestMessagesForTelemetry(telemetry, request);
-	if (requestMessages) span.setAttribute(PiGenAIAttr.RequestMessages, requestMessages);
+	if (requestMessages) span.setAttribute(OmpGenAIAttr.RequestMessages, requestMessages);
 	if (telemetry.contentCapture !== "full") return;
 	const systemInstructions = serializeFullSystemInstructionsForTelemetry(request);
 	if (systemInstructions) span.setAttribute(GenAIAttr.SystemInstructions, systemInstructions);
@@ -789,9 +807,9 @@ function applyContentCaptureForRequest(telemetry: AgentTelemetry, span: Span, re
 
 function applyContentCaptureForResponse(telemetry: AgentTelemetry, span: Span, message: AssistantMessage): void {
 	const responseText = serializeResponseTextForTelemetry(telemetry, message);
-	if (responseText) span.setAttribute(PiGenAIAttr.ResponseText, responseText);
+	if (responseText) span.setAttribute(OmpGenAIAttr.ResponseText, responseText);
 	const responseToolCalls = serializeResponseToolCallsForTelemetry(telemetry, message);
-	if (responseToolCalls) span.setAttribute(PiGenAIAttr.ResponseToolCalls, responseToolCalls);
+	if (responseToolCalls) span.setAttribute(OmpGenAIAttr.ResponseToolCalls, responseToolCalls);
 	if (telemetry.contentCapture === "full") {
 		const outputMessages = serializeFullOutputMessagesForTelemetry(message);
 		if (outputMessages) span.setAttribute(GenAIAttr.OutputMessages, outputMessages);
@@ -1131,6 +1149,7 @@ export async function finishChatSpan(
 	const cost = applyCostEstimate(telemetry, span, message, options.serviceTier, options.stepNumber);
 	if (telemetry) {
 		await emitChatUsage(telemetry, span, {
+			operation: GenAIOperation.Chat,
 			model: message.model,
 			provider: message.provider,
 			serviceTier: options.serviceTier,
@@ -1199,7 +1218,7 @@ function applyChatResponseAttributes(span: Span, message: AssistantMessage): voi
 	span.setAttribute(GenAIAttr.ResponseModel, message.model);
 	if (message.responseId) span.setAttribute(GenAIAttr.ResponseId, message.responseId);
 	if (message.upstreamProvider) {
-		span.setAttribute(PiGenAIAttr.ResponseUpstreamProvider, message.upstreamProvider);
+		span.setAttribute(OmpGenAIAttr.ResponseUpstreamProvider, message.upstreamProvider);
 	}
 	if (message.ttft != null) span.setAttribute(GenAIAttr.ResponseTimeToFirstChunk, message.ttft / 1000);
 	const finishReason = mapStopReason(message.stopReason);
@@ -1215,7 +1234,7 @@ function applyUsageAttributes(span: Span, usage: Usage | undefined): void {
 	span.setAttribute(GenAIAttr.UsageInputTokens, inputTokens);
 	span.setAttribute(GenAIAttr.UsageOutputTokens, outputTokens);
 	const total = usage.totalTokens ?? inputTokens + outputTokens;
-	span.setAttribute(PiGenAIAttr.UsageTotalTokens, total);
+	span.setAttribute(OmpGenAIAttr.UsageTotalTokens, total);
 	if (usage.cacheRead != null) span.setAttribute(GenAIAttr.UsageCacheReadInputTokens, usage.cacheRead);
 	if (usage.cacheWrite != null) span.setAttribute(GenAIAttr.UsageCacheCreationInputTokens, usage.cacheWrite);
 	if (usage.reasoningTokens != null) {
@@ -1223,7 +1242,7 @@ function applyUsageAttributes(span: Span, usage: Usage | undefined): void {
 	}
 	if (usage.server) {
 		const sums = (usage.server.webSearch ?? 0) + (usage.server.webFetch ?? 0);
-		if (sums > 0) span.setAttribute(PiGenAIAttr.UsageServerSideTools, sums);
+		if (sums > 0) span.setAttribute(OmpGenAIAttr.UsageServerSideTools, sums);
 	}
 }
 
@@ -1284,7 +1303,7 @@ export function detectGatewayFromHeaders(
 
 /**
  * Bounded Cloudflare AI Gateway response-cache statuses emitted on
- * {@link PiGenAIAttr.GatewayResponseCacheStatus}. Distinct from provider
+ * {@link OmpGenAIAttr.GatewayResponseCacheStatus}. Distinct from provider
  * prompt-cache token counters (`gen_ai.usage.cache_*`).
  *
  * Cloudflare documents `HIT` / `MISS` on `cf-aig-cache-status`; `bypass` covers
@@ -1323,14 +1342,14 @@ function applyGatewayAttributes(
 ): void {
 	const gateway = detectGatewayFromHeaders(headers);
 	if (gateway) {
-		span.setAttribute(PiGenAIAttr.GatewayName, gateway.name);
-		if (baseUrl) span.setAttribute(PiGenAIAttr.GatewayEndpoint, baseUrl);
-		if (gateway.callId) span.setAttribute(PiGenAIAttr.GatewayCallId, gateway.callId);
-		if (gateway.routedTo) span.setAttribute(PiGenAIAttr.GatewayRoutedTo, gateway.routedTo);
+		span.setAttribute(OmpGenAIAttr.GatewayName, gateway.name);
+		if (baseUrl) span.setAttribute(OmpGenAIAttr.GatewayEndpoint, baseUrl);
+		if (gateway.callId) span.setAttribute(OmpGenAIAttr.GatewayCallId, gateway.callId);
+		if (gateway.routedTo) span.setAttribute(OmpGenAIAttr.GatewayRoutedTo, gateway.routedTo);
 	}
 	const responseCacheStatus = classifyGatewayResponseCacheStatus(headers);
 	if (responseCacheStatus) {
-		span.setAttribute(PiGenAIAttr.GatewayResponseCacheStatus, responseCacheStatus);
+		span.setAttribute(OmpGenAIAttr.GatewayResponseCacheStatus, responseCacheStatus);
 	}
 }
 
@@ -1392,7 +1411,7 @@ function applyCostEstimateForUsage(
 	}
 	if (!result) return EMPTY_COST;
 	if ("unavailable" in result) {
-		span.setAttribute(PiGenAIAttr.CostUnavailableReason, result.unavailable);
+		span.setAttribute(OmpGenAIAttr.CostUnavailableReason, result.unavailable);
 		const cost: AppliedCostEstimate = {
 			costUsd: undefined,
 			inputUsd: undefined,
@@ -1414,9 +1433,9 @@ function applyCostEstimateForUsage(
 		});
 		return cost;
 	}
-	span.setAttribute(PiGenAIAttr.CostEstimatedUsd, result.usd);
-	if (result.inputUsd != null) span.setAttribute(PiGenAIAttr.CostInputUsd, result.inputUsd);
-	if (result.outputUsd != null) span.setAttribute(PiGenAIAttr.CostOutputUsd, result.outputUsd);
+	span.setAttribute(OmpGenAIAttr.CostEstimatedUsd, result.usd);
+	if (result.inputUsd != null) span.setAttribute(OmpGenAIAttr.CostInputUsd, result.inputUsd);
+	if (result.outputUsd != null) span.setAttribute(OmpGenAIAttr.CostOutputUsd, result.outputUsd);
 	const cost: AppliedCostEstimate = {
 		costUsd: result.usd,
 		inputUsd: result.inputUsd,
@@ -1470,6 +1489,7 @@ async function emitChatUsage(
 	telemetry: AgentTelemetry,
 	span: Span,
 	input: {
+		readonly operation: GenAIOperationName;
 		readonly model: string;
 		readonly provider: string | undefined;
 		readonly serviceTier: ServiceTier | undefined;
@@ -1483,6 +1503,7 @@ async function emitChatUsage(
 	if (!hook || !input.usage) return;
 	const event: ChatUsageEvent = {
 		span,
+		operation: input.operation,
 		agent: normalizedTelemetryAgent(telemetry),
 		conversationId: telemetry.conversationId,
 		stepNumber: input.stepNumber,
@@ -1493,7 +1514,9 @@ async function emitChatUsage(
 		cost: costEstimateFromApplied(input.applied),
 		attributes: resolveDynamicAttributes(
 			telemetry,
-			buildTelemetryAttributeContext(telemetry, "chat", { stepNumber: input.stepNumber }),
+			buildTelemetryAttributeContext(telemetry, input.operation === GenAIOperation.Judgment ? "judgment" : "chat", {
+				stepNumber: input.stepNumber,
+			}),
 		),
 		headers: input.headers,
 	};
@@ -1586,7 +1609,7 @@ export async function recordManualChatTelemetry(
 		});
 	if (!span) return undefined;
 	if (options.span && options.attributes) span.setAttributes(options.attributes);
-	if (options.stepNumber != null) span.setAttribute(PiGenAIAttr.AgentStepNumber, options.stepNumber);
+	if (options.stepNumber != null) span.setAttribute(OmpGenAIAttr.AgentStepNumber, options.stepNumber);
 	span.setAttribute(GenAIAttr.ResponseModel, options.responseModel ?? options.model.name);
 	if (options.responseId) span.setAttribute(GenAIAttr.ResponseId, options.responseId);
 	const finishReason = mapStopReason(options.finishReason);
@@ -1602,6 +1625,7 @@ export async function recordManualChatTelemetry(
 			usage: options.usage,
 		});
 		await emitChatUsage(telemetry, span, {
+			operation: GenAIOperation.Chat,
 			model: options.responseModel ?? options.model.id,
 			provider: options.model.provider,
 			serviceTier: options.serviceTier,
@@ -1619,7 +1643,7 @@ export async function recordManualChatTelemetry(
 	}
 	if (options.responseText) {
 		const responseText = stringifyJsonAttribute(summarizeTelemetryTexts([options.responseText]));
-		if (responseText) span.setAttribute(PiGenAIAttr.ResponseText, responseText);
+		if (responseText) span.setAttribute(OmpGenAIAttr.ResponseText, responseText);
 	}
 	if (options.responseToolCalls && options.responseToolCalls.length > 0) {
 		const calls = options.responseToolCalls.map(call => ({
@@ -1628,13 +1652,112 @@ export async function recordManualChatTelemetry(
 			input: summarizeTelemetryValue(call.input),
 		}));
 		const responseToolCalls = stringifyJsonAttribute(limitTelemetryToolCalls(calls));
-		if (responseToolCalls) span.setAttribute(PiGenAIAttr.ResponseToolCalls, responseToolCalls);
+		if (responseToolCalls) span.setAttribute(OmpGenAIAttr.ResponseToolCalls, responseToolCalls);
 	}
 	applyTerminalStatus(span, options.finishReason, undefined);
 	if (options.endSpan ?? options.span === undefined) span.end();
 	return span;
 }
 
+/**
+ * One judgment attempt reported to {@link recordJudgmentTelemetry}: a billed
+ * System One request, one prompted-model attempt, or a request answered
+ * entirely from the local judgment cache.
+ */
+export interface JudgmentTelemetryOptions {
+	readonly provider: string;
+	/** Requested model id. */
+	readonly model: string;
+	/** Model that answered, when the provider resolves an alias (`jev-latest` → `jev-1.13.0`). */
+	readonly responseModel?: string;
+	/** Caller-level reason the judgment ran; stamped as {@link OmpGenAIAttr.JudgmentPurpose}. */
+	readonly purpose: string;
+	/** Priced usage of this attempt — the same amount the session ledger bills. */
+	readonly usage: Usage;
+	readonly stopReason: StopReason;
+	readonly errorMessage?: string;
+	/** Epoch ms the attempt started. */
+	readonly startTime: number;
+	/** Questions in the request; omitted for prompted backends, which report per completion attempt. */
+	readonly questions?: number;
+	/** Questions answered from the local cache rather than the provider. */
+	readonly cachedQuestions?: number;
+}
+
+/**
+ * Emit one `judgment` span per judgment attempt and forward any billed usage
+ * to `onCostDelta` / `onChatUsage` with `operation: "judgment"`. Cost is the
+ * attempt's own priced usage rather than {@link AgentTelemetryConfig.costEstimator},
+ * so spans, metrics, and the session ledger agree on every judgment dollar.
+ * Attempts that billed nothing (cache hits, pre-response failures) keep their
+ * span but fire no usage hooks. No-op when telemetry is disabled.
+ */
+export async function recordJudgmentTelemetry(
+	telemetry: AgentTelemetry | undefined,
+	options: JudgmentTelemetryOptions,
+): Promise<void> {
+	if (!telemetry) return;
+	const attributes: Attributes = {
+		[GenAIAttr.RequestModel]: options.model,
+		[OmpGenAIAttr.JudgmentPurpose]: options.purpose,
+	};
+	const provider = normalizeProviderName(telemetry, options.provider);
+	if (provider) attributes[GenAIAttr.ProviderName] = provider;
+	if (options.questions !== undefined) attributes[OmpGenAIAttr.JudgmentQuestions] = options.questions;
+	if (options.cachedQuestions !== undefined) {
+		attributes[OmpGenAIAttr.JudgmentCachedQuestions] = options.cachedQuestions;
+	}
+	const span = startSpan(telemetry, "judgment", `judgment ${options.model}`, {
+		spanKind: SpanKind.CLIENT,
+		attributes,
+		startTime: options.startTime,
+	});
+	if (!span) return;
+	const model = options.responseModel ?? options.model;
+	span.setAttribute(GenAIAttr.ResponseModel, model);
+	const finishReason = mapStopReason(options.stopReason);
+	if (finishReason) span.setAttribute(GenAIAttr.ResponseFinishReasons, [finishReason]);
+	applyUsageAttributes(span, options.usage);
+	const { cost } = options.usage;
+	span.setAttribute(OmpGenAIAttr.CostEstimatedUsd, cost.total);
+	span.setAttribute(OmpGenAIAttr.CostInputUsd, cost.input);
+	span.setAttribute(OmpGenAIAttr.CostOutputUsd, cost.output);
+	if (options.usage.totalTokens > 0 || cost.total > 0) {
+		const applied: AppliedCostEstimate = {
+			costUsd: cost.total,
+			inputUsd: cost.input,
+			outputUsd: cost.output,
+			costUnavailableReason: undefined,
+		};
+		emitCostDelta(telemetry, {
+			agent: normalizedTelemetryAgent(telemetry),
+			conversationId: telemetry.conversationId,
+			stepNumber: undefined,
+			provider: provider ?? options.provider,
+			model,
+			serviceTier: undefined,
+			usage: buildUsageSnapshot(options.usage),
+			costUsd: applied.costUsd,
+			inputUsd: applied.inputUsd,
+			outputUsd: applied.outputUsd,
+			costUnavailableReason: undefined,
+		});
+		await emitChatUsage(telemetry, span, {
+			operation: GenAIOperation.Judgment,
+			model,
+			provider: options.provider,
+			serviceTier: undefined,
+			stepNumber: undefined,
+			usage: options.usage,
+			applied,
+			headers: undefined,
+		});
+	}
+	applyTerminalStatus(span, options.stopReason, options.errorMessage);
+	safeOnSpanEnd(telemetry, { ...buildTelemetryAttributeContext(telemetry, "judgment", {}), span });
+	span.end();
+}
+
 /**
  * Options accepted by {@link instrumentedCompleteSimple}. Mirrors the
  * `streamAssistantResponse` chat-span lifecycle for oneshot LLM calls
@@ -1647,7 +1770,7 @@ export interface InstrumentedChatSpanOptions {
 	/** Step index recorded on the span; defaults to `-1` for non-loop calls. */
 	readonly stepNumber?: number;
 	/**
-	 * Tag stamped onto `pi.gen_ai.oneshot.kind`. Values used by the agent:
+	 * Tag stamped onto `omp.gen_ai.oneshot.kind`. Values used by the agent:
 	 * `compaction_summary`, `compaction_short_summary`, `compaction_turn_prefix`,
 	 * `handoff`, `branch_summary`, `image_question`. Free-form to allow callers
 	 * outside this package to add new kinds without bumping the helper.
@@ -1726,7 +1849,7 @@ export async function instrumentedCompleteSimple<TApi extends Api>(
 		},
 	});
 	if (chatSpan) {
-		if (oneshotKind) chatSpan.setAttribute(PiGenAIAttr.OneshotKind, oneshotKind);
+		if (oneshotKind) chatSpan.setAttribute(OmpGenAIAttr.OneshotKind, oneshotKind);
 		if (span.attributes) chatSpan.setAttributes(span.attributes);
 	}
 
@@ -1881,7 +2004,7 @@ export function finishExecuteToolSpan(
 }
 
 /** Span attribute carrying the terminal {@link ToolStatus}. */
-export const EXECUTE_TOOL_STATUS_ATTR = PiGenAIAttr.ToolStatus;
+export const EXECUTE_TOOL_STATUS_ATTR = OmpGenAIAttr.ToolStatus;
 
 /**
  * Mapping from non-ok {@link ToolStatus} values to the `error.type` attribute
@@ -1974,66 +2097,66 @@ export function fireOnRunEnd(telemetry: AgentTelemetry, summary: AgentRunSummary
 	}
 }
 
-/** Aggregate `pi.gen_ai.agent.*` attributes stamped on the `invoke_agent` span. */
-export const enum PiGenAIAggregateAttr {
-	ChatsCount = "pi.gen_ai.agent.chats.count",
-	ChatsTotalLatencyMs = "pi.gen_ai.agent.chats.total_latency_ms",
-	ChatsStopReasonPrefix = "pi.gen_ai.agent.chats.stop_reason.",
-	ToolsCount = "pi.gen_ai.agent.tools.count",
-	ToolsOkCount = "pi.gen_ai.agent.tools.ok.count",
-	ToolsErrorCount = "pi.gen_ai.agent.tools.error.count",
-	ToolsSkippedCount = "pi.gen_ai.agent.tools.skipped.count",
-	ToolsBlockedCount = "pi.gen_ai.agent.tools.blocked.count",
-	ToolsTimeoutCount = "pi.gen_ai.agent.tools.timeout.count",
-	ToolsAbortedCount = "pi.gen_ai.agent.tools.aborted.count",
-	ToolsTotalLatencyMs = "pi.gen_ai.agent.tools.total_latency_ms",
-	ToolsInvoked = "pi.gen_ai.agent.tools.invoked",
-	ToolsAvailable = "pi.gen_ai.agent.tools.available",
-	ToolsUnused = "pi.gen_ai.agent.tools.unused",
-	UsageInputTokensTotal = "pi.gen_ai.agent.usage.input_tokens.total",
-	UsageOutputTokensTotal = "pi.gen_ai.agent.usage.output_tokens.total",
-	UsageCacheReadInputTokensTotal = "pi.gen_ai.agent.usage.cache_read.input_tokens.total",
-	UsageCacheCreationInputTokensTotal = "pi.gen_ai.agent.usage.cache_creation.input_tokens.total",
-	UsageReasoningOutputTokensTotal = "pi.gen_ai.agent.usage.reasoning.output_tokens.total",
-	UsageTotalTokensTotal = "pi.gen_ai.agent.usage.total_tokens.total",
-	CostEstimatedUsdTotal = "pi.gen_ai.agent.cost.estimated_usd.total",
-	ErrorsCount = "pi.gen_ai.agent.errors.count",
-}
-
-/** Stamp the aggregate `pi.gen_ai.agent.*` attributes on the given span. */
+/** Aggregate `omp.gen_ai.agent.*` attributes stamped on the `invoke_agent` span. */
+export const enum OmpGenAIAggregateAttr {
+	ChatsCount = "omp.gen_ai.agent.chats.count",
+	ChatsTotalLatencyMs = "omp.gen_ai.agent.chats.total_latency_ms",
+	ChatsStopReasonPrefix = "omp.gen_ai.agent.chats.stop_reason.",
+	ToolsCount = "omp.gen_ai.agent.tools.count",
+	ToolsOkCount = "omp.gen_ai.agent.tools.ok.count",
+	ToolsErrorCount = "omp.gen_ai.agent.tools.error.count",
+	ToolsSkippedCount = "omp.gen_ai.agent.tools.skipped.count",
+	ToolsBlockedCount = "omp.gen_ai.agent.tools.blocked.count",
+	ToolsTimeoutCount = "omp.gen_ai.agent.tools.timeout.count",
+	ToolsAbortedCount = "omp.gen_ai.agent.tools.aborted.count",
+	ToolsTotalLatencyMs = "omp.gen_ai.agent.tools.total_latency_ms",
+	ToolsInvoked = "omp.gen_ai.agent.tools.invoked",
+	ToolsAvailable = "omp.gen_ai.agent.tools.available",
+	ToolsUnused = "omp.gen_ai.agent.tools.unused",
+	UsageInputTokensTotal = "omp.gen_ai.agent.usage.input_tokens.total",
+	UsageOutputTokensTotal = "omp.gen_ai.agent.usage.output_tokens.total",
+	UsageCacheReadInputTokensTotal = "omp.gen_ai.agent.usage.cache_read.input_tokens.total",
+	UsageCacheCreationInputTokensTotal = "omp.gen_ai.agent.usage.cache_creation.input_tokens.total",
+	UsageReasoningOutputTokensTotal = "omp.gen_ai.agent.usage.reasoning.output_tokens.total",
+	UsageTotalTokensTotal = "omp.gen_ai.agent.usage.total_tokens.total",
+	CostEstimatedUsdTotal = "omp.gen_ai.agent.cost.estimated_usd.total",
+	ErrorsCount = "omp.gen_ai.agent.errors.count",
+}
+
+/** Stamp the aggregate `omp.gen_ai.agent.*` attributes on the given span. */
 function applyAggregateAttributes(span: Span, summary: AgentRunSummary, coverage: AgentRunCoverage): void {
-	span.setAttribute(PiGenAIAggregateAttr.ChatsCount, summary.chats.total);
-	span.setAttribute(PiGenAIAggregateAttr.ChatsTotalLatencyMs, summary.chats.totalLatencyMs);
+	span.setAttribute(OmpGenAIAggregateAttr.ChatsCount, summary.chats.total);
+	span.setAttribute(OmpGenAIAggregateAttr.ChatsTotalLatencyMs, summary.chats.totalLatencyMs);
 	for (const [reason, count] of Object.entries(summary.chats.byStopReason)) {
-		span.setAttribute(`${PiGenAIAggregateAttr.ChatsStopReasonPrefix}${reason}.count`, count);
-	}
-	span.setAttribute(PiGenAIAggregateAttr.ToolsCount, summary.tools.total);
-	span.setAttribute(PiGenAIAggregateAttr.ToolsOkCount, summary.tools.ok);
-	span.setAttribute(PiGenAIAggregateAttr.ToolsErrorCount, summary.tools.error);
-	span.setAttribute(PiGenAIAggregateAttr.ToolsSkippedCount, summary.tools.skipped);
-	span.setAttribute(PiGenAIAggregateAttr.ToolsBlockedCount, summary.tools.blocked);
-	span.setAttribute(PiGenAIAggregateAttr.ToolsTimeoutCount, summary.tools.timeout);
-	span.setAttribute(PiGenAIAggregateAttr.ToolsAbortedCount, summary.tools.aborted);
-	span.setAttribute(PiGenAIAggregateAttr.ToolsTotalLatencyMs, summary.tools.totalLatencyMs);
+		span.setAttribute(`${OmpGenAIAggregateAttr.ChatsStopReasonPrefix}${reason}.count`, count);
+	}
+	span.setAttribute(OmpGenAIAggregateAttr.ToolsCount, summary.tools.total);
+	span.setAttribute(OmpGenAIAggregateAttr.ToolsOkCount, summary.tools.ok);
+	span.setAttribute(OmpGenAIAggregateAttr.ToolsErrorCount, summary.tools.error);
+	span.setAttribute(OmpGenAIAggregateAttr.ToolsSkippedCount, summary.tools.skipped);
+	span.setAttribute(OmpGenAIAggregateAttr.ToolsBlockedCount, summary.tools.blocked);
+	span.setAttribute(OmpGenAIAggregateAttr.ToolsTimeoutCount, summary.tools.timeout);
+	span.setAttribute(OmpGenAIAggregateAttr.ToolsAbortedCount, summary.tools.aborted);
+	span.setAttribute(OmpGenAIAggregateAttr.ToolsTotalLatencyMs, summary.tools.totalLatencyMs);
 	if (coverage.toolsInvoked.length > 0) {
-		span.setAttribute(PiGenAIAggregateAttr.ToolsInvoked, [...coverage.toolsInvoked]);
+		span.setAttribute(OmpGenAIAggregateAttr.ToolsInvoked, [...coverage.toolsInvoked]);
 	}
 	if (coverage.toolsAvailable.length > 0) {
-		span.setAttribute(PiGenAIAggregateAttr.ToolsAvailable, [...coverage.toolsAvailable]);
+		span.setAttribute(OmpGenAIAggregateAttr.ToolsAvailable, [...coverage.toolsAvailable]);
 	}
 	if (coverage.toolsUnused.length > 0) {
-		span.setAttribute(PiGenAIAggregateAttr.ToolsUnused, [...coverage.toolsUnused]);
-	}
-	span.setAttribute(PiGenAIAggregateAttr.UsageInputTokensTotal, summary.usage.inputTokens);
-	span.setAttribute(PiGenAIAggregateAttr.UsageOutputTokensTotal, summary.usage.outputTokens);
-	span.setAttribute(PiGenAIAggregateAttr.UsageCacheReadInputTokensTotal, summary.usage.cachedInputTokens);
-	span.setAttribute(PiGenAIAggregateAttr.UsageCacheCreationInputTokensTotal, summary.usage.cacheWriteTokens);
-	span.setAttribute(PiGenAIAggregateAttr.UsageReasoningOutputTokensTotal, summary.usage.reasoningOutputTokens);
-	span.setAttribute(PiGenAIAggregateAttr.UsageTotalTokensTotal, summary.usage.totalTokens);
+		span.setAttribute(OmpGenAIAggregateAttr.ToolsUnused, [...coverage.toolsUnused]);
+	}
+	span.setAttribute(OmpGenAIAggregateAttr.UsageInputTokensTotal, summary.usage.inputTokens);
+	span.setAttribute(OmpGenAIAggregateAttr.UsageOutputTokensTotal, summary.usage.outputTokens);
+	span.setAttribute(OmpGenAIAggregateAttr.UsageCacheReadInputTokensTotal, summary.usage.cachedInputTokens);
+	span.setAttribute(OmpGenAIAggregateAttr.UsageCacheCreationInputTokensTotal, summary.usage.cacheWriteTokens);
+	span.setAttribute(OmpGenAIAggregateAttr.UsageReasoningOutputTokensTotal, summary.usage.reasoningOutputTokens);
+	span.setAttribute(OmpGenAIAggregateAttr.UsageTotalTokensTotal, summary.usage.totalTokens);
 	if (summary.cost.estimatedUsd > 0) {
-		span.setAttribute(PiGenAIAggregateAttr.CostEstimatedUsdTotal, summary.cost.estimatedUsd);
+		span.setAttribute(OmpGenAIAggregateAttr.CostEstimatedUsdTotal, summary.cost.estimatedUsd);
 	}
-	span.setAttribute(PiGenAIAggregateAttr.ErrorsCount, summary.errors.total);
+	span.setAttribute(OmpGenAIAggregateAttr.ErrorsCount, summary.errors.total);
 }
 
 /**
@@ -2068,10 +2191,10 @@ export function recordHandoff(
 	const attrs: Attributes = {};
 	const fromAgent = options.fromAgent ? normalizeAgentIdentity(telemetry, options.fromAgent) : undefined;
 	const toAgent = normalizeAgentIdentity(telemetry, options.toAgent);
-	if (fromAgent?.name) attrs[PiGenAIAttr.HandoffFromAgentName] = fromAgent.name;
-	if (fromAgent?.id) attrs[PiGenAIAttr.HandoffFromAgentId] = fromAgent.id;
-	if (toAgent.name) attrs[PiGenAIAttr.HandoffToAgentName] = toAgent.name;
-	if (toAgent.id) attrs[PiGenAIAttr.HandoffToAgentId] = toAgent.id;
+	if (fromAgent?.name) attrs[OmpGenAIAttr.HandoffFromAgentName] = fromAgent.name;
+	if (fromAgent?.id) attrs[OmpGenAIAttr.HandoffFromAgentId] = fromAgent.id;
+	if (toAgent.name) attrs[OmpGenAIAttr.HandoffToAgentName] = toAgent.name;
+	if (toAgent.id) attrs[OmpGenAIAttr.HandoffToAgentId] = toAgent.id;
 	const name = toAgent.name
 		? fromAgent?.name
 			? `handoff ${fromAgent.name} → ${toAgent.name}`